Information Modelling and Knowledge Bases XVII: Volume 136 Frontiers in Artificial Intelligence and Applications

INFORMATION MODELLING AND KNOWLEDGE BASES XVII Frontiers in Artificial Intelligence and Applications FAIA covers all ...

Author: J. Henno | H. Jaakkola Y. Kiyoki

19 downloads 1415 Views 8MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

INFORMATION MODELLING AND KNOWLEDGE BASES XVII

Frontiers in Artificial Intelligence and Applications FAIA covers all aspects of theoretical and applied artificial intelligence research in the form of monographs, doctoral dissertations, textbooks, handbooks and proceedings volumes. The FAIA series contains several sub-series, including “Information Modelling and Knowledge Bases” and “Knowledge-Based Intelligent Engineering Systems”. It also includes the biannual ECAI, the European Conference on Artificial Intelligence, proceedings volumes, and other ECCAI – the European Coordinating Committee on Artificial Intelligence – sponsored publications. An editorial panel of internationally well-known scholars is appointed to provide a high quality selection. Series Editors: J. Breuker, R. Dieng, N. Guarino, J.N. Kok, J. Liu, R. López de Mántaras, R. Mizoguchi, M. Musen and N. Zhong

Volume 136 Recently published in this series Vol. 135. H. Czap et al. (Eds.), Self-Organization and Autonomic Informatics (I) Vol. 134. M.-F. Moens and P. Spyns (Eds.), Legal Knowledge and Information Systems – JURIX 2005: The Eighteenth Annual Conference Vol. 133. C.-K. Looi et al. (Eds.), Towards Sustainable and Scalable Educational Innovations Informed by the Learning Sciences – Sharing Good Practices of Research, Experimentation and Innovation Vol. 132. K. Nakamatsu and J.M. Abe (Eds.), Advances in Logic Based Intelligent Systems – Selected Papers of LAPTEC 2005 Vol. 131. B. López et al. (Eds.), Artificial Intelligence Research and Development Vol. 130. K. Zieliński and T. Szmuc (Eds.), Software Engineering: Evolution and Emerging Technologies Vol. 129. H. Fujita and M. Mejri (Eds.), New Trends in Software Methodologies, Tools and Techniques – Proceedings of the fourth SoMeT_W05 Vol. 128. J. Zhou et al. (Eds.), Applied Public Key Infrastructure – 4th International Workshop: IWAP 2005 Vol. 127. P. Ritrovato et al. (Eds.), Towards the Learning Grid – Advances in Human Learning Services Vol. 126. J. Cruz, Constraint Reasoning for Differential Models Vol. 125. C.-K. Looi et al. (Eds.), Artificial Intelligence in Education – Supporting Learning through Intelligent and Socially Informed Technology Vol. 124. T. Washio et al. (Eds.), Advances in Mining Graphs, Trees and Sequences Vol. 123. P. Buitelaar et al. (Eds.), Ontology Learning from Text: Methods, Evaluation and Applications

ISSN 0922-6389

Information Modelling and Knowledge Bases XVII

Edited by

Yasushi Kiyoki Keio University, Japan

Jaak Henno Tallinn Technical University, Estonia

Hannu Jaakkola Tampere University of Technology, Finland

and

Hannu Kangassalo University of Tampere, Finland

Amsterdam • Berlin • Oxford • Tokyo • Washington, DC

© 2006 The authors. All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without prior written permission from the publisher. ISBN 1-58603-591-6

Publisher IOS Press Nieuwe Hemweg 6B 1013 BG Amsterdam Netherlands fax: +31 20 687 0019 e-mail: [email protected] Distributor in the UK and Ireland Gazelle Books Falcon House Queen Square Lancaster LA1 1RN United Kingdom fax: +44 1524 63232

Distributor in the USA and Canada IOS Press, Inc. 4502 Rachael Manor Drive Fairfax, VA 22032 USA fax: +1 703 323 3668 e-mail: [email protected]

LEGAL NOTICE The publisher is not responsible for the use which might be made of the following information. PRINTED IN THE NETHERLANDS

Information Modelling and Knowledge Bases XVII Y. Kiyoki et al. (Eds.) IOS Press, 2006 © 2006 The authors. All rights reserved.

v

Preface Information modelling and knowledge bases are becoming very important topics not only in academic communities related to information systems and computer science but also in the business field of information technology. Currently, the structural complexity of information resources, the variety of abstraction levels of information, and the size of databases and knowledge bases are continuously growing. We are facing the complex problems of structuring, sharing, managing, searching and mining data and knowledge from a large amount of complex information resources existing in databases and knowledge bases. New methodologies in many areas of information modelling and knowledge bases are expected to provide sophisticated and reliable solutions to these problems. The aim of this series of Information Modelling and Knowledge Bases is to provide research communities in information modelling and knowledge bases with scientific results and experiences achieved by using innovative methodologies in computer science and other disciplines related to linguistics, philosophy, and psychology. Those interdisciplinary research results include common interests in understanding and solving problems on information modelling and knowledge bases, as well as applying those research results to the practical application areas. The research topics in this series are mainly concentrated on a variety of themes in the important domains: • • • • • • • •

theoretical and philosophical basis of concept modelling and conceptual modelling, conceptual modelling, information modelling and specification, conceptual models in intelligent activity, collections of data, knowledge, and descriptions of concepts, human-computer interaction and modelling, database and knowledge base systems, software engineering and modelling and applications for information modelling and knowledge bases.

It is very significant to recognize, study and share new areas related to information modelling and knowledge bases on which great attention is focused. Therefore, cognitive science, knowledge management, linguistics, philosophy, psychology, logic, and management science are relevant areas, too. This is reflected in the number of research results dealing with multimedia databases, WWW information managements, and temporal-spatial data models. These new directions are pushing the frontier of knowledge and creating novel ways of modelling real worlds. To achieve these aims in this series, the international program committee has selected 16 full papers, 11 short papers and 1 position paper in a rigorous reviewing process from 38 submissions. The selected papers cover many areas of information modelling, concept theories, database semantics, knowledge bases and systems, software engineering, WWW information managements, context-based information access spaces, ontological tech-

vi

nology, image databases, temporal and spatial databases, document data managements, and many more. In Program Committee, there were 32 well-known researchers from the areas of information modelling, concept theories, conceptual modelling, database theory, knowledge bases, information systems, linguistics, philosophy, logic, image processing, temporal and spatial databases, document data managements and other related fields. We are very grateful for their great work in reviewing the papers. We hope that the series of Information Modelling and Knowledge Base will be productive and valuable in the advancement of research and practice of those academic areas. The Editors Yasushi Kiyoki Jaak Henno Hannu Jaakkola Hannu Kangassalo

vii

Program Committee Co-Chairs: Yasushi Kiyoki, Keio University, Japan Hannu Kangassalo, University of Tampere, Finland Members: Alain Pirotte, University of Louvain, Belgium Anneli Heimbürger, Tampere University of Technology, Pori, Finland Arne Sølvberg, Norwegian University of Science and Technology, Norway Benkt Wangler, Skövde University, Sweden Bernhard Thalheim, Kiel University, Germany Björn Nilsson, Astrakan Strategic Development, Sweden Cristine Sernadas, Institute Superior Tecnico, Portugal Eiji Kawaguchi, Keio University, Japan Esteban Zimanyi, Universite Libre de Bruxelles (ULB), Belgium Hannu Jaakkola, Tampere University of Technology, Pori, Finland Isabelle Mirbel-Sanchez, Université de Nice Sophia Antipolis, France Jeffery Xu Yu, Chinese University of Hong Kong, Honk Kong Jörgen Fischer Nilsson, Denmark Technical University, Denmark Koichi Nozaki, Nagasaki University, Japan Manfred A. Jeusfeld, Tilburg University, The Netherlands Marie Duží, Technical University of Ostrava, Czech Republic Martti Juhola, University of Tampere, Finland Michael Schrefl, University of Linz, Austria Olga De Troyer, Vrije Universiteit Brussel, Belgium Pierre-Jean Charrel, Université Toulouse 1, France Seiji Ishikawa, Kyushu Institute of Technology, Japan Takehiro Tokuda, Tokyo Institute of Technology, Japan Tapio Niemi, University of Tampere, Finland Veikko Rantala, University of Tampere, Finland Xing Chen, Kanagawa Institute of Technology, Japan Yoshihide Hosokawa, Nagoya Institute of Technology, Japan Yoshihiro Okada, Kyushu University, Japan Yukihiro Itoh, Shizuoka University, Japan Yutaka Funyu, Iwate Prefectural University, Japan Yuzuru Tanaka, University of Hokkaido, Japan Organizing Committee Jaak Henno, (Organizing Committee Co-Chair), Tallinn University of Technology, Estonia Hannu Jaakkola, (Organizing Committee Co-Chair), Tampere University of Technology, Pori, Finland Hannu Kangassalo, University of Tampere, Finland Eiji Kawaguchi, Keio University, Japan Yasushi Kiyoki, Keio University, Japan Ulla Nevanranta (Publication), Tampere University of Technology, Pori, Finland Yurika Suzuki (Publication), Keio University, Japan

viii

Permanent Steering Committee Hannu Jaakkola, Tampere University of Technology, Pori, Finland Hannu Kangassalo, University of Tampere, Finland Eiji Kawaguchi, Keio University, Japan Setsuo Ohsuga, Japan (Honorary member) Additional Reviewers Hideki Noda, Kyushu Institute of Technology, Japan Jun Sasaki, Iwate Prefectural University, Japan Mina Akaishi, University of Tokyo, Japan Mitsuhisa Taguchi, Tokyo Institute of Technology, Japan Paula Gouveia, Instituto Superior Tecnico, Portugal Per Backlund, University of Skövde, Sweden Satoru Kogure, Shizuoka University, Japan Sven Casteleyn, Vrije Universiteit Brussel, Belgium Tae Yoneda, Iwate Prefectural University, Japan Tetsuya Suzuki, Tokyo Institute of Technology, Japan Wesley Bille, Vrije Universiteit Brussel, Belgium Zhao Peixiang, Chinese University of Hong Kong, Hong Kong

ix

Contents Preface Yasushi Kiyoki, Jaak Henno, Hannu Jaakkola and Hannu Kangassalo Committees Additional Reviewers A Model for Describing and Annotating Documents Nicolas Dessaigne and José Martinez

v vii viii 1

Web Ontology Languages: Theory and Practice – Will They Ever Meet? Marie Duží and Anneli Heimbürger

20

Object Order Based on Concept Analysis and Moebius Inversion Function Petr Gajdoš and Daniela Ďuráková

38

A Semantic Spectrum Analyzer for Realizing Semantic Learning in a Semantic Associative Search Space Yasushi Kiyoki, Xing Chen and Hidehiro Ohashi

50

Theory and Implementation on Automatic Adaptive Metadata Generation for Image Retrieval Hideyasu Sasaki and Yasushi Kiyoki

68

A Meme Media Framework for Introducing a Multimodal User Interface to Existing Web Applications Kimihito Ito

83

Privacy Enhanced Information Systems Sabah S. Al-Fedaghi, Gunar Fiedler and Bernhard Thalheim

94

Description Logics, Data Modelling and Relational Dependencies Tapio Niemi, Marko Niinimäki and Santtu Toivonen

112

Modelling, Searching and Utilising Grid Resources and Services Marko Niinimäki, Santtu Toivonen and Tapio Niemi

127

Conceptual Modelling (and Problem Solving Techniques) for Complex Narrative Events Gian Piero Zarri

138

Role of Models in Knowledge Transfer During OO Software Development Oksana Nikiforova, Marite Kirikova and Wita Wojtkowski

156

Toward a Cooperative Recognition of Human Behaviors and Related Objects Tadahiro Kitahashi, Atsuhiro Kojima, Mirai Higuchi and Kunio Fukunaga

172

A Reliable and Useful Information Distribution System: The “Kuchicomi Network” Jun Sasaki, Tae Yoneda and Yutaka Funyu

180

x

Object-Oriented Model for Fire Spread Prediction Guillermina Pertino, Sabrina Tocho, Arturo Zambrano and Silvia Gordillo

191

Knowledge Entry Maps: Structuring of Method Knowledge in the IT Industry Remko Helms, Sjaak Brinkkemper, Job van Oosterum and Frank de Nijs

205

Building a Conceptual Skeleton for Enterprise Architecture Specifications Veikko Halttunen, Antti Lehtinen and Riku Nykänen

219

Fixed-Point Quantifiers in Higher Order Logics Klaus-Dieter Schewe and José María Turull Torres

237

Discovery of Narrativity on the WWW Based on Perspective Information Access Mina Akaishi, Nicolas Spyratos, Koichi Hori and Yuzuru Tanaka

245

A Computer Science Text Corpus/Search Engine X-Tec and Its Applications Takehiro Tokuda, Yusuke Soyama and Tetsuya Suzuki

253

Deriving Semantic from Images Based on the Edge Information Xing Chen, Tony Delvecchio and Vincenzo Di Lecce

260

Evolving the Database Co-Design Framework by SPICE Hannu Jaakkola, Timo Mäkinen, Bernhard Thalheim and Timo Varkoi

268

Reconstructing Geons in Database Semantics Roland Hausser

280

Query Optimisation as Part of Distribution Design for Complex Value Databases Hui Ma and Klaus-Dieter Schewe

289

Position Paper: Digital Media Archive for Academic Resources Takanari Hayama, Eiji Kawaguchi and Yasushi Kiyoki

297

An Ontology Learning and Reasoning Framework Hele-Mai Haav

302

Strong Relevant Logic as the Universal Basis of Various Applied Logics for Knowledge Representation and Reasoning Jingde Cheng

310

Concept Theory in the Context of Information Systems Jari Palomäki and Momir Radičević

321

Describing Knowledge Architectures Enn Tyugu

329

Author Index

341


1

2

N. Dessaigne and J. Martinez / A Model for Describing and Annotating Documents


3

4



5

6



7

8



9

10



11

12



13

14



15

16



17

18



19


20

Web Ontology Languages: theory and practice – will they ever meet? Marie DUŽÍ* and Anneli HEIMBÜRGER** *VSB-Technical University of Ostrava 17. listopadu 15, 708 33 Ostrava, Czech Republic [email protected] **Tampere University of Technology, Pori, Advanced Multimedia Center P.O. Box 300, FIN – 28101 Pori, Finland [email protected] Abstract. The current state of Semantic Web ontology languages is briefly described, and the ontology languages are characterised from the logical point of view. Generally, these languages are based on the first-order predicate logic enriched with ad hoc higher-order constructs wherever needed. We argue that in the Semantic Web we need a rich language with transparent semantics, in order to build up metadata on the conceptual level of the Semantic Web architecture. A powerful logical tool of Transparent Intensional Logic (TIL) is described, which provides a logico-semantic framework for a fine-grained knowledge representation and conceptual analysis. TIL is based on a rich ontology of entities organised in an infinite ramified hierarchy of types. The conceptual role of TIL in building ontologies is described, and we show that such a system can serve as a unifying logical framework. Concluding we argue that the conceptual and logical level of the Web architecture have an important role, and we should pay a due attention to these levels. Keywords. Ontology languages, Semantic Web, Transparent Intentional Logic (TIL), hierarchy of types, concepts, structured meanings, XML Linking Language (XLink), Resource Description Framework (Schema), RDF, RDFS, Ontology Web Language (OWL), SKIF (Knowledge Interchange Format), Description Logic

1. Introduction The Web was proposed as a tool for representing relationships between named objects, drawing together knowledge from scattered systems into a common framework [2]. The main aim of the Semantic Web initiative is to develop the current Web towards the original proposal. W3C’s Semantic Web Activity develops standards and technologies, which are designed to help machines to understand more information on the Web [14, 29]. The word “semantic” in the context of Semantic Web means “machine-processible” [4]. The main idea is of having the data on the Web defined and linked in such a way that it can be used for more effective information retrieval, knowledge discovery, automation, integration and reuse of information across various applications, organisations and communities. To meet these goals, the World Wide Web Consortium (W3C) has defined a layer model for the Semantic Web (Figure 1), and knowledge representation and ontology languages are being developed. The Semantic Web is vitally dependent on a formal meaning assigned to the constructs of its languages. For Semantic Web languages to work well together their formal meanings must employ a common view (or thesis) of representation [15], otherwise it will not be possible to reconcile documents written in

M. Duží and A. Heimbürger / OWLs: Theory and Practice – Will They Ever Meet?

21

different languages. A common underpinning is especially important for the Semantic Web because it is envisioned to contain several languages, as in Tim Berners-Lee's “layer cake” diagram (Figure 1) first presented at XML 2000 [3]. The diagram depicts a Semantic Web Architecture in which languages of increasing power are layered one on top of the other. Unfortunately, the relationships between adjacent layers are not specified, either with respect to the syntax or semantics. Naturally but unfortunately, the model is being gradually realised in a bottom up way; languages in particular layers come into being in an ad hoc way, without a deep logical insight. Thus the languages often lack an exact semantics and ontological definitions of particular entities; in the syntactic constructs, particular abstract levels are mixed together. Trust level Digital signature, annotation Logical and inference level Rule-based systems, ML, … Ontology level OWL, WordNet, RosettaNet, … Metadata level RDF, RDFS, …

Structure level XML, XLink, XML Schema, … Internet level Unicode, URI, … Figure 1: Semantic Web architecture

The notion ‘ontology’ is understood in many distinct ways: formalisation, conceptual analysis, hierarchical classification, conceptualisation, etc. We are going to show that ‘ontology’ can be understood as a conceptual analysis of a given universe of discourse, i.e., of what (which entities) we talk about, and how, by means of which concepts we capture these entities. Logical forms (constructions) rendering the basic stock of explicit knowledge serve as the base, from which logical consequences can be derived by an inference machine so that to obtain the inferable stock of knowledge [8]. The more fine-grained the analysis is, the more accurate inferences the machine can perform. In an ideal case, the inference machine should be able to derive just the logical consequences of the base: it should not over-infer (derive something that does not follow), nor under-infer (not to derive something that does follow). We show that the current state-of-arts is far from being ideal. Web ontology languages are based on the 1st-order predicate logic (FOL). Though FOL has become stenography of mathematics, it is not expressive enough when used at the natural language area. The obvious disadvantage of the FOL approach is that treating higher-order properties and relations like individuals conceals the ontological structure of the universe, and knowledge representation is not comprehensive. Moreover, when representing knowledge, inference paradoxes are inevitable. Thus for applications where even the full power of FOL is not adequate, it would be natural to extend the framework to a higherorder logic (HOL). A general objection against using HOL logic is its computational intractability. However, HOL formulas are relatively well understood, and reasoning systems for HOLs do already exist, e.g., HOL [11] and Isabelle [25]. Though the Web languages have been enriched by a few constructs exceeding the power of FOL, these additional constructs are usually not well defined and understood. Moreover, particular

22


languages are neither syntactically nor semantically compatible. The W3C efforts at standardization resulted in accepting the Resource Description Framework (RDF) language as the Web recommendation. However, this situation is far from satisfactory. Quoting from Horrocks and Schneider [15]: “The thesis of representation underlying RDF and RDFS is particularly troublesome in this regard, as it has several unusual aspects, both semantic and syntactic. A more-standard thesis of representation would result in the ability to reuse existing results and tools in the Semantic Web.” In the paper we describe a powerful logical tool of transparent intensional logic (TIL), which provides a logico-semantic framework for a fine-grained knowledge representation and conceptual analysis. TIL is based on a rich ontology of entities organised in an infinite ramified hierarchy of types. The language of TIL constructions (hyperintensionally individuated meanings) can serve as an ideal knowledge representation and query language. A common objection against such a rich system, namely that it is too complicated and computationally intractable, is in our opinion rather irrelevant: formal knowledge specification in TIL is semantically transparent and comprehensible, with all the semantically salient features explicitly present. In the Semantic Web we need such a highly expressive language so that to first know what is there, and only afterwards to try deriving the consequences. The fact that for higher-order logics such as TIL there is no semantically complete proof-system is not important. Only when understanding the problem and knowing “what is there”, we can try to solve the problem and derive consequences. Moreover, the TIL framework might at least serve as a both semantic and terminological standard for the development of new languages, as an ideal to which we should aim at. The paper is organised as follows. First, in chapter 2 we recapitulate the current state of Web ontology languages from two viewpoints. In paragraph 2.1 Web ontology languages are described from the inter-relations point of view and in paragraph 2.2 from the logical point of view. We show here that the Web ontology languages are mostly based on the first-order predicate logic approach. In Chapter 3, an expressive system of the Transparent Intensional Logic (TIL) is described. After an informal introduction we provide particular definitions, describe the method of semantic analysis and characterise important conceptual relations. Finally, in Chapter 4, the conceptual and terminological role of TIL in building ontologies is described. 2. Web Ontology Languages The lowest levels of the Semantic Web are based on mark-up languages. By means of schemas and ontology classifications, types of resources and types of inter-relationships between the resources can be specified. At the bottom level, the Unicode and URI layers make sure that international character sets are used and provide means for identifying the resources in the Semantic Web. At the very core level of the Web, there is the XML layer with linking, style and transformation, namespace and schema definitions, which forms the basis for the Semantic Web definitions. On the metadata level we can find RDF and RDF Schema for describing resources with URI addresses and for defining vocabularies that can be referred to by URI addresses, respectively. The ontology level supports the evolution of a shared semantic specification and conceptualization of different application domains. The ontology level is based on the OWL recommended by W3C, which is based on the 1st-order Description logic framework. Based on common logic, the SKIF language accommodates some higher-order constructs. At the logical level, simple inferences based on ontologies can be drawn. As far as we know, the only ontology language supporting inferences at this level is a Semantic Web Rule Language (SWRL) combining OWL and RuleML [16]. At the highest level, the focus is on trust, i.e., how to guarantee the reliability of the data obtained from the Web. Currently, a common research project CoLogNet II (Network of


23

Excellence in Computational Logic II) of fourteen European universities, lead by the Free University of Bozen-Bolzano, Italy, is prepared. The goal of this project can be characterised as the development of a powerful Web-inference machine based on a highly expressive logical semantics. 2.1 Web Ontology Languages from Inter-relating Point of View XML Linking Language (XLink, [32]) is a W3C recommendation which allows elements to be inserted into XML documents in order to create and describe relationships between and inside Web resources. XLink uses XML syntax [30] to create linking and navigation structures. XPointer language [31] is used to identify regions in XML documents. XLink supports simple, unidirectional, unary and untyped links as well as more sophisticated, extended links. Extended links can be bidirectional and n-ary. Arcs are used to describe traversal rules in an n-ary link. XLink also defines a linkbase.xml for storing a set of extended links to separate XML file. Extended links can also have attributes. XLink consists of three attribute categories: semantic, behaviour and traversal attributes. XLink’s attribute space is flexible and can be extended in order to create for example time-sensitive linking structures, which represent context-based inter-relations and low level way of inferring [13]. XLink is at the moment the only formal linking language which is based on an international standard [17]. The Resource Description Framework (RDF) was designed to provide a common way to describe information in a way that it can be read and understood by computer applications [34]. The data model of the RDF includes three basic elements. Resources are anything with an URI address. Properties specify attributes and/or (binary) relations between resources and an object used to describe resources. Statements of the form ‘subject, predicate, object’ associate a resource and a specific value of its property. Both XLink and RDF provide a way of creating relations between resources. However, RDF as a metadata specification is designed for describing the contents of resources and their relations as graphs, while XLink is designed for specifying and traversing hyperlinks. From XLink’s point of view RDF can have two roles: harvesting RDF statements from XLinks [27] and meta-information description of XLink structures. An example of an RDF graph description of an extended link is illustrated in Figure 2. As RDF describes resources with classes, properties and values, by means of RDF Schema (RDFS)an extension of RDFapplication-specific classes and properties can be defined [35]. RDF Schema doesn’t provide actual application-specific classes and properties; instead it provides a way of defining those. By means of RDFS it is possible to create simple vocabularies of a certain application domain, for example in technical documentation. In the Web environment vocabularies allow search engines, agents, authors and users to communicate. In addition to XLink and RDF(S) another important facility to organize Internet resources are XML Topic Maps [24]. With topic maps an associative information structure, which is located outside that information can be created. A topic map usually contains several overlapping levels, which are semantically cross-linked. The core of the XML Topic Maps (XTM) 1.0 specification is formed of topics, which represent the subjects the topic map is about. Topics can be grouped in classes called topic types. A topic type is a category to which one given topic instance belongs. A topic may be linked to one or more information resources that are relevant to the topic in some way. Such resources are called occurrences of the topic. Occurrences are information resources relevant to a topic. Topics can be related through some association expressing given semantics. Just as topic and occurrences can be grouped according to type so can also associations between topics be grouped according to their types.

24


An example of an extended link

xlink:extended

xlink:title xlink:type link1

link2 xlink:locator

xlink:href

xlink:locator xlink:arc xlink:href

L1

L2 xlink:from

xlink:actuate xlink:onLoad

xlink:to xlink:show xlink:new

Figure 2: An example of an RDF graph

Each topic that participates in an association plays a role in that association called the association role. The topic map specification provides a language to represent knowledge, in particular the conceptual knowledge with which one distinguishes information resources semantically. The ontological engineering discipline applied to building topic maps focuses precisely on this aspect. Ontological engineering emphasizes careful conceptual design and construction of topic maps to reflect correctly the semantics of the underlying knowledge. The design and construction of topic maps that provide meaningful expression of knowledge should be based on established principles of ontological engineering. According to the OWL (Web Ontology Language) overview [33], OWL is intended to be used when information contained in documents needs to be processed by applications, as opposed to situations where the contents only need to be presented to humans. OWL can be used to represent the meaning of terms in vocabularies and relationships between those terms. OWL has been designed on the top of XML, XLink, RDF and RDFS in order to provide more facilities for expressing meaning and semantics to represent machine interpretable content on the Web. When thinking in wider sense, XML/XLink, RDF, RDFS and OWL can all be regarded as ontology languages in different levels as far as creating and describing interrelations and navigational structures between and inside XML documents are concerned. XML/XLink is the very basis for defining links and traversal rules between Web resources. However, it also contains a rich and extendable attribute space for describing semantic, behavioural and traversal characteristics of links. RDF provides means for describing metainformation on linking and navigational structures. RDFS defines a framework to describe vocabularies which can be applied to represent linkbases. By means of OWL like languages a vocabulary of terms and some specification of meaning between the terms can be specified. From the linking point of view the Semantic Web is still a network of linked nodes with different granularity levels of inter-relations, and it can be considered from three different perspectives [19]: (a) universal library to be accessed and used by humans in a variety of information use contexts, (b) the backdrop work of computational agents carrying out activities on behalf of humans, and (c) a method for integrating knowledge bases and databases to perform tasks for humans. Current Web ontology languages are, however, far from meeting these points, in particular points (b) and (c), for their expressive power does not enable computational agents to involve an adequate inference machine.


25

2.2 Web Ontology Languages from the Logical Point of View According to Horrocks and Patel-Schneider [15] building ontologies consists in a hierarchical description of important concepts in a domain, along with descriptions of properties of the instances of each concept and relations between them. Current ontological languages correspond roughly in their expressive power to the first-order predicate logic (FOL), with some higher-order ad hoc extensions. None of them (except of languages based on the Description logic framework) makes it possible to express modalities (what is necessary and what is contingent, i.e., which concepts are analytical and which empirical) and higher-order concepts. Concepts of n-ary relations are unreasonably modelled by properties. True, each n-ary relation can be expressed by n unary relations (properties): for instance the fact that G.W. Bush attacked Iraq can be modelled by two facts, namely that G.W. Bush has the property of attacking Iraq, and Iraq has the property of being attacked by G.W. Bush, but such a representation is neither comprehensive nor optimal. The basis of a particular way of providing meaning for metadata is embodied in the model theory for RDF. RDF has unusual aspects that make its use as the foundation of representation in the Semantic Web difficult at best. In particular, RDF has a very limited collection of syntactic constructs, and these are treated in a very uniform manner in the semantics of RDF. The RDF thesis requires that no other syntactic constructs than the RDF triples are to be used and that the uniform semantic treatment of syntactic constructs cannot be changed only augmented [15]. In RDFS we can specify classes and properties of individuals, constraints on properties, and the relation of subsumption (subclass, subproperty). It is not possible, for instance, to specify properties of properties, e.g., that the relation (property) is functional or transitive. Neither it is possible to define classes by means of properties of individuals that belong to the class. A recognition of the limitations of RDFS led to the development of new Web ontology languages such as OIL, DAML-ONT and DAML+OIL [28, 36]. These are used as the basis of a new W3C Web ontology language called the OWL. As a second language for the Semantic Web, OWL has been developed as an extension of RDFS. OWL (like DAML+OIL) uses the same syntax as RDF (and RDFS) to represent ontologies, the two languages are syntactically compatible. However, the semantic layering of the two languages is more problematical. The difficulty stems from the fact that OWL (like DAML+OIL) is largely based on the Description Logic [1], the semantics of which would normally be given by a classical first-order model theory in which individuals are interpreted as elements of some domain (a set), classes are interpreted as subsets of the domain and properties are interpreted as binary relations on the domain. The semantics of RDFS, on the other hand, are given by a non-standard model theory, where individuals, classes and properties are all elements in the domain. Properties are further interpreted as having extensions which are binary relations on the domain, and class extensions are only implicitly defined by the extension of the rdf:type property. Moreover, RDFS supports reflection on its own syntax: interpretation of classes and properties can be extended by statements in the language. Thus language layering is much more complex, because different layers subscribe to these two different approaches. The third group of ontology languages lies somewhere between the FOL framework and RDFS. The group of relatively new languages include SKIF and Common Logic [12]. The SKIF syntax is compatible with functional language LISP, but in principle it is an FOL syntax. These languages have like RDFS a non standard model theory, with predicates being interpreted as individuals, i.e., elements of a domain. Classes are however treated as subsets of the domain, and their redefinition in the language syntax is not allowed. Hence from the logical point of view, the ontological languages can be divided into three groups: the FOL approach, the SKIF approach, and the RDF approach.

26


The FOL approach (DAML+OIL, OWL) is closely connected to the rather expressive Description Logic (DL): Languages of this group talk about individuals that are elements of a domain. The individuals are members of subclasses of the domain, and can be related to other individuals (or data values) by means of properties (n-ary relations are called properties in Web ontologies, for they are decomposed into n properties). The universe of discourse is divided into two disjoint sorts: the object domain of individuals and the data value domain of numbers. Thus the interpretation function assigns elements of the object domain to individual constants, elements of data value domain to value constants, and subclasses of the data domain to data types. Further, object and data predicates (of properties) are distinguished, the former being interpreted as a subset of the Cartesian product of object domain, the latter a subset of the Cartesian product of value domain. DL is rather rich, though being an FOL language. It makes it possible to distinguish intensional knowledge (knowledge on the analytically necessary relations between concepts) and extensional knowledge (of contingent facts). The knowledge base of DL is thus divided into the so-called T-boxes (according to terminology or taxonomy) and A-boxes (according to contingent attributes of objects). Tbox contains verbal definitions, i.e., a new concept is defined composing known concepts. For instance, a woman can be defined: WOMAN = PERSON & SEX-FEMALE, and a mother: MOTHER = WOMAN & child(HASCHILDchild). Thus the fact that, e.g., mother is a woman is analytical (necessary) true. In the T-box there are also specifications of necessary properties of concepts and relations between concepts: the property satisfiability corresponding to a nonempty concept, the relation of subsumption (intensionally contained concepts), equivalence and disjointness (incompatibility). Thus, e.g., that a bachelor is not married is analytically (necessarily) true proposition. On the other hand, the fact that, e.g., Mr. Jones is a bachelor is a contingent unnecessary fact. Such contingent properties (attributes) of objects are recorded in A-boxes. The SKIF approach: SKIF languages are syntactically compatible with LISP, i.e., the FOL syntax is extended with the possibility to mention properties and use variables ranging over properties. For instance, we can specify that John and Peter have a common property: p . p(John) & p(Peter). The property they have in common can be, e.g., that they both love their wives. We can also specify that the property P is true of John, and the P has the property Q: P(John) & Q(P). If P is being honest and Q is being eligible, the sentence can be read as that John is honest, which is eligible. The interpretation structure is a triple , where D is the universe, V is the function that maps predicates, variables and constants to the elements of D, and ext is the function that maps D into sets of n-tuples of elements of D. SKIF does not reduce the arity of predicates. The RDF approach: These languages originally did not have a model theoretic semantics, which led to many discrepancies. The RDF syntax consists of the so-called triples (subject, predicate and object), where only binary predicates are allowed. This causes serious problems concerning compatibility with more expressive languages. RDF(S) has become a Web ontological recommendation defined by W3C, and its usage is world spread. The question is whether it is a good decision. A classical FOL approach would be better, or even its standard extension to HOL would be more suitable for ontologies. Formalisation in HOL is much more natural and comprehensive, the universe of discourse is not a flat set of ‘individuals’, but properties and relations can be naturally talked about as well, which is much more apt for representation of ontologies. Ontologies will play a pivotal role in the Semantic Web by providing a source of shared and precisely defined concepts of entities that can be used in metadata. The degree of formality employed in capturing these concepts can be quite variable, ranging from natural language to logical formalisms, but increased formality and regularity clearly facilitates machine understanding. We argue that conceptual formalisation should be


27

meaning driven, based on natural language. Formal language of FOL is not a natural language; it is a language of non interpreted formulas that enable us to talk about individuals, expressing their properties and relations between individuals, but not to talk about properties of (individuals, properties, functions, relations, properties of concepts, generally of higher-order objects), unless they are simple members of the universe. Thus inference machine based on FOL can under-infer, or paradoxically over-infer. Here is an example: The office of the President of USA is certainly not an individual. It can be hold by individuals, but the holder and the office are two completely distinct things. The office (Church’s individual concept) necessarily has some requisites (like being occupied by at most one individual), which no its occupant has. In FOL, however, we have to treat the office as an individual. The paradoxes like the following arise: The President of USA knows that John Kerry wanted to become the President of USA. George W. Bush is the President of USA. Hence: George W. Bush knows that John Kerry wanted to become George W. Bush. Analysing (as it should be) ‘is’ in the second premise as the identity of individuals, we obtain the obviously non-valid conclusion. True, this shortcoming is usually overcome in FOL approach by introducing a special binary predicate ‘holds’. The second premise would then be translated into: (individual) George W. Bush holds (another individual) the President of USA. But then the inference machine under-infers, because it would not be possible to infer a valid consequence of the above premises, namely that George W. Bush knows that John Kerry wanted to become the President of USA. Another shortcoming of FOL approach is the impossibility to handle (contingently) ‘non-denoting terms’ like the President of Czech Republic in January 2003, or the King of USA. In other words, translating the sentence that the King of USA does not exist into an FOL language, we obtain a paradox of existence (that being an individual, it exists, but actually it does not exist): x (x = k(USA)), but from (x = k(USA)) by existential generalisation we immediately derive that x (x = k(USA)). A theory formalising reasoning of intelligent agents has to be able to talk about the objects of agents’ attitudes, to quantify over them, to express iterated attitudes and/or selfreferential statements, like agent a knows that an agent b knows that (he) believes that P, which in FOL leads to inconsistencies. Such theories have also to enable us to express the distinction between analytical and empirical concepts (what is necessary and what is just contingent), to express empty concepts, to talk about concepts and, last but not least, to express n-ary relations-in-intension between any entities (not only individuals) of our ontology. While this is beyond the expressive power of FOL, many richer logical systems with non-standard operators are proposed: modal, epistemic, intensional, temporal, nonmonotonic, paraconsistent, etc. These logics can be characterized as theories with ‘syntactically driven axiomatization’. They provide ad hoc axioms and rules that define a set of models, each logic partly solving particular kind of problems. Ontology language should be, however, universal, highly expressive, with transparent semantics and meaning driven axiomatisation. We have such a system at hand: the system of transparent intensional logic (TIL). 3. Transparent Intensional Logic (TIL) 3.1 TIL Informal Exposition: Hierarchy of Types TIL is a logic that does not use any non standard operators; in this sense it is classical. However, its expressive power is very high; formalisation of meaning is comprehensive, with transparent semantics, closed to natural language. Notation is an adjusted objectual

28


version of Church’s typed O-calculus, where all the semantically salient features are explicitly present. The two-dimensional ontology of entities we can talk about is in TIL organised in a ramified hierarchy of types. This enables us to logically handle structured meanings as higher-order, hyper-intensional abstract objects, thus avoiding inconsistency problems stemming from the need to mention these objects within the theory itself. Hyperintensionally individuated structured meanings are procedures, structured from the algorithmic point of view. Due to typing, any object of any order can be not only used, but also mentioned within the theory without giving rise to inconsistency. On the ground level of the type-hierarchy, there are entities unstructured from the algorithmic point of view belonging to a type of order 1. Given a so-called epistemic base of atomic types (R-truth values, L-individuals, W-time points, Z-possible worlds), mereological complexity is increased by an induction rule of forming partial functions: where D, E1,…,En are types of order 1, the set of partial mappings from E1 u…u En to D, denoted (D E1…En), is a type of order 1 as well. TIL is an open-ended system. The above epistemic base {R, L, W, Z} was chosen, because it is apt for natural-language analysis, but in the case of mathematics a (partially) distinct base would be appropriate; for instance, the base consisting of natural numbers, of type Q, and truth-values. Derived types would then be defined over {Q, R}. A collection of constructions that construct entities of order 1, denoted by *1, serves as a base for the induction rule: any collection of partial functions (D E1…En) involving *1 in their domain or range is a type of order 2. Constructions belonging to a type *2 that identify entities of order 1 or 2, and partial functions involving such constructions, belong to a type of order 3. And so on, ad infinitum. 3.2 TIL Informal Exposition: Constructions Constructions are structured from the algorithmic point of view; they are procedures, or abstract algorithms, specifying how to arrive at less-structured entities. Since constructions are abstract, extra-linguistic entities, they are reachable only via a verbal definition. The ‘language of constructions’ is a modified version of the typed O-calculus, where Montaguelike O-terms denote, not the functions constructed, but the constructions themselves. The modification is extensive. Church’s O-terms form part of his simple type theory, whereas our O-terms belong to a ramified type theory [26]. Constructions qua procedures operate on input objects (of any type, even higher-order constructions) and yield as output objects of any type. One should not conflate using constructions as constituents of composed constructions and mentioning constructions that enter as input objects into composed constructions, so we have to strictly distinguish between using and mentioning constructions. The latter is, in principle, achieved by using atomic constructions. A construction is atomic if it is a procedure that does not contain as a constituent any other construction but itself. There are two atomic constructions that supply objects (of any type) on which complex constructions operate: variables and trivialisations. Variables are constructions that construct an object dependently on valuation: they vconstruct. Variables can range over any type. If c is a variable ranging over constructions of order 1 (type *1), then c belongs to *2, the type of order 3, and constructs a construction of order 1 belonging to *1: the type of order 2. When X is an object of any type, the trivialisation of X, denoted 0X, constructs X without the mediation of any other construction. 0X is the atomic concept of X: it is the primitive, non-perspectival mode of presentation of X. There are two compound constructions, which consist of other constituents: composition and closure. Composition is the procedure of applying a function f to an argument A, i.e., the instruction to apply f to A to obtain the value (if any) of f at A. Closure


29

is the procedure of constructing a function by abstracting over variables, i.e., the instruction to do so. Finally, higher-order constructions can be used twice over as constituents of composed constructions. This is achieved by a fifth construction called double execution. 3.3 TIL Definitions Definition 1 (Construction) i) Variables x, y, z, … construct objects of the respective types dependently on valuations v; they v-construct. ii) Trivialisation: Where X is an object whatsoever (an extension, an intension or a construction), 0X constructs X. iii) Closure: If x1, x2, …,xn are pairwise distinct variables that v-construct entities of types D1, D2, …, Dn, respectively, and Y is a construction that v-constructs an entity of type E, then [Ox1…xn Y] is a construction called closure, which v-constructs a function of type (E D1…Dn) mapping D1 u…u Dn to E. iv) Composition: If X v-constructs a function f of a type (E D1…Dn), and Y1,…,Yn vconstruct entities A1, …, An of types D1,…,Dn, respectively, then the composition [X Y1 … Yn] v-constructs the value (an entity, if any, of type E) of the (partial) function f on the argument ¢A1, …, An². Otherwise the composition [X Y1 … Yn] does not v-construct anything: it is v-improper. v) Double execution: If X is a construction of order n, n t 2, that v-constructs a construction X’ (of order n–1), then 2X v-constructs the entity v-constructed by X’. Otherwise the double execution 2X is v-improper. vi) Nothing is a construction, unless it so follows from i) through v). Definition 2 (Ramified hierarchy) Let B be a base, i.e. a collection of pair-wise disjoint, non-empty sets. T1 (types of order 1) i) Every member of B is an elementary type of order 1 over B. ii) Let D, E1, ..., Em (m > 0) be types of order 1 over B. Then the collection (D E1 ... Em) of all m-ary (total and partial) mappings from E1 u ... u En into D is a functional type of order 1 over B. iii) Nothing is a type of order 1 over B unless it so follows from T1 i)ii). Cn (constructions of order n over B) i) Let x be a variable ranging over a type of order n. Then x is a construction of order n. ii) Let X be a member of a type of order n. Then 0X, 2X are constructions of order n. iii) Let X, X1,...,Xm (m > 0) be constructions of order n over B. Then [X X1...Xm] is a construction of order n. iv) Let x1,...xm, X (m > 0) be constructions of order n over B. Then [Ox1...xm X] is a construction of order n. v) Nothing is a construction of order n over B unless it so follows from Cn i)v). Tn+1 (types of order n + 1) Let n be the collection of all construction of order n over B. i) n and every type of order n are types of order n + 1. ii) If m > 0, and D, E1,...,Em are types of order n + 1 over B, then (D E1 ... Em) (see T1 ii)) is a type of order n + 1 over B. ii) Nothing is a type of order n + 1 over B unless it so follows from Tn+1 i) ii). Notational conventions An object A of the type D is called an D-object, denoted A/D. That a construction C constructs an D-object will be denoted C o D. We use infix notation

30


without trivialisation for truth-value connectives (conjunction), (disjunction), (implication), for an identity sign = and for binary number relations t, , d. Examples (a) The function +, defined on natural numbers (of type Q), is not a construction. It is a mapping of type (Q QQ), i.e., a set of triples, the first two members of which are natural numbers, while the third member is their sum. The simplest construction of this mapping is 0+. (b) The composition [0+ x 01] v-constructs the successor of any number x. (c) The closure Ox [0+ x 01] constructs the successor function. (d) The composition of this closure with 05, i.e., [Ox [0+ x 01] 05], constructs the number 6. (e) The composition [0: x 00] does not v-construct anything for any valuation of x; it is improper. (f) The closure Ox [0: x 00] is not improper, as it constructs something, even though it is only a degenerate function, viz. one undefined at all its arguments. The constructions 0+, [0+ x 01], Ox [0+ x 01], [Ox [0+ x 01] 05], [0: x 00], Ox [0: x 00], all mentioned above, are members of *1. When IMPROPER is a set of v-improper constructions of order 1, i.e., when IMPROPER is an object of type (R*1), the composition [0IMPROPER 0[0: x 00]] is a member of type *2, and it constructs the truth-value True. The constituent 0[0: x 00] of this composition (a member of type *2) is an atomic proper construction that constructs [0: x 00], a member of *1. It is atomic, because the construction [0: x 00] is not used here as a constituent but only mentioned as an input object. For further details, see [6, 26]. If ARITHMETIC is a set of arithmetic functions, then the composition [0ARITHMETIC 2c] v-constructs True if c v–constructs [Ox [0+ x 01]]. The double execution 2c v-constructs what is v-constructed by [Ox [0+ x 01]], i.e., the arithmetic successor function. Definition 3 ((D-)intension, (D-)extension) (D-)intensions are members of a type (DZ), i.e., functions from possible worlds to the arbitrary type D. (D-)extensions are members of the type D, where D is not equal to (EZ) for any E, i.e., extensions are not functions from possible worlds. Remark Intensions are frequently functions of the type ((DW)Z), i.e., functions from possible worlds to chronologies of the type D (in symbols: DWZ), where a chronology is a function of type (ĮW). Examples of types x being happy is a property of individuals / (RL)WZ. x The president of the Czech Republic is an individual office (‘individual concept’) / LWZ. x That Charles is happy is a proposition / RWZ. x Knowing is an attitude of an individual to a construction, i.e., a relation that is a higherorder intension / (R L n)WZ. 3.4 Semantic Analysis We adhere to the constraint on natural-language analysis dictated by the principle of subject matter: an admissible analysis of an expression E is a construction C such that C uses, as its constituents, constructions of just those objects that E mentions, i.e., only the objects denoted by sub-expressions of E. Any such analysis is the best analysis of E, relative to a set of atomic constructions (see [9], [22]). The principle is central to our general three-step method of logical analysis of language: (i) Type-theoretical analysis Assign types to the objects mentioned, i.e., only those that are denoted by sub-expressions of E, and do not omit any semantically self-contained subexpression of E, i.e., use all of them.


31

(ii) Synthesis Compose constructions of these objects so as to construct the object D denoted by E. (iii) Type checking Use the assigned types for control so as to check whether the various types are compatible and, furthermore, produce the right type of object in the manner prescribed by the analysis. A construction of an intension is of the form OwOt X, w o Z, t o W. If C is a construction of an intension Int, the composition [[C w] t] — the intensional descent of Int to its extension (if any) at w,t — will be abbreviated Cwt. Example of analysis We are going to analyse the sentence, “The President of USA is G.W.Bush”. (i’)

President-of / (LL)WZ—(an empirical function that dependently on the states of affairs assigns an individual to an individual), USA / L — (for the sake of simplicity), the The President of USA / LWZ — an individual office, G.W.Bush / L — (for the sake of simplicity), = / (R LL) — the identity of individuals. The whole sentence denotes a proposition / RWZ.

(ii’)

OwOt [0President-ofwt 0USA] o LWZ (the individual office – role PUSA) [OwOt [0President-ofwt 0USA]]wt o L (the occupant of the office PUSA) [0= [OwOt [0President-ofwt 0USA]]wt 0G.W.Bush] o R OwOt [0= [OwOt [0President-ofwt 0USA]]wt 0G.W.Bush] o RWZ.

(iii’)

OwOt [0=

[OwOt [0President-ofwt 0USA]]wt 0G.W.Bush] (LL) L (R LL) L L R

Abstracting over t: Abstracting over w:

(RW) ((RW)Z), i.e., RWZ.

When being the President of USA, G.W. Bush is identical with the individual that holds the office PUSA. If, however, John Kerry wanted to become the President of USA, he certainly did not want to become G.W. Bush. He simply wanted to hold the office PUSA, i.e., he is related not to the individual, but to the individual office. Now the paradoxical argument is easily solved away: “John Kerry wanted to become the President of USA” OwOt [ Wantwt 0J.Kerry [OwOt [0Becomewt 0J.Kerry OwOt [0President-ofwt 0USA]]]], 0

where Want / (R L RWZ)WZ, Become / (R L LWZ)WZ. The construction OwOt [0President-ofwt 0USA], i.e., a concept of the president of USA, is in the de dicto supposition [6], the undesirable substitution is blocked. Knowing is a relation-in-intension of an agent a to the meaning of the embedded clause, i.e., to the construction of the respective proposition [8]. Thus when the President of USA knows that John Kerry wanted to become the President of USA, he is related to the construction of the proposition, and the former concept of the president of USA is used de re, the latter de dicto. The whole argument is analysed as follows: OwOt [0= [OwOt [0President-ofwt 0USA]]wt 0G.W.Bush] OwOt [0Knowwt [OwOt [0President-ofwt 0USA]]wt 0 [OwOt [0Wantwt 0J.Kerry OwOt [0Becomewt 0J.Kerry OwOt [0President-ofwt 0USA]]]]]

32


Now the former concept of the President of USA is ‘free’ for substitution: we can substitute 0 G.W.Bush for [OwOt [0President-ofwt 0USA]]wt, thus deducing that G.W. Bush knows that John Kerry wanted to become the President of USA, but not that he wanted to become G.W.Bush: OwOt [0Knowwt 0G.W.Bush 0 [OwOt [0Wantwt 0J.Kerry [OwOt 0Becomewt 0J.Kerry OwOt [0President-ofwt 0USA]]]]] The undesirable substitution of 0G.W.Bush for the latter occurrence of the construction OwOt [0President-ofwt 0USA] is blocked. 3.5 Theory of Concepts Category of concepts has been almost neglected in the modern logic (perhaps only Bolzano, Frege and Church studied the notion of a concept). A new impulse to examining concepts came from computer science. Finnish logician Rauli Kauppi [18] axiomatised the classical conception of concept as the entity determined by its extent and content. This theory is based on the (primitive) relation of the intensional containment. Ganter-Wille [10] theory defines formal concept as the couple (extent, content) and makes use of the classical law of inversion between extent and content, which holds for the conjunctive composition of attributes. Due to this law a partial ordering can be defined on the set of formal concepts, which establishes a concept lattice. Actually, ontologies viewed as classifications are based on this framework. All these classical theories make use of the FOL apparatus classifying relations between concepts that are not ontologically defined here. Our conception defines concept as a closed construction [5, 21, 22] in the canonical form. An analogical approach can be also found in [23]. We do not consider only general concepts of properties, but also a concept of a proposition, of an office, of a number, etc. Comparison of the procedural theory of concepts with the classical set-theoretical one can be found in [7]. When building ontologies, we aim at conceptual analysis of entities talked about in the given domain. In TIL, the analysis consists in formalising the meaning of an expression, i.e., in finding the construction of the denoted entity. If a sentence has a complete meaning, the respective closed construction is a complete instruction of how to evaluate the truthconditions in any state of affairs w,t. However, not all the sentences of natural language denote propositions. Sometimes we are not able to evaluate the truth conditions without knowing the (linguistic or situation of utterance) context. In such a case the respective construction is open, contains free variables. For instance, the sentence “he is happy” does not denote a proposition. Its analysis OwOt [0Happywt x] contains a free variable x. Only after x is evaluated by context supplying the respective individual, we obtain a proposition. Thus the sentence does not express a concept of the proposition. On the other hand, sentences with a complete meaning express concepts of a proposition. Constructions (meanings) are assigned to expressions by linguistic convention. Closed constructions are concepts assigned to expression with a complete meaning. Natural language is not, however, perfect. In a vernacular we often confuse concepts with expressions. We say that the concept of a computer had been invented, or that the concept of a whale changed, and so on. As abstract entities, concepts cannot change or being invented. They can only be discovered. We should rather say that the new expressions had been invented to express the respective concept, or that the expression changed its meaning. Moreover, there are homonyms (expressions with more concepts assigned), and synonyms (more expressions expressing the same concept). Anyway, understanding an expression we know the respective concept, we know what to do, which does not, however, mean that we know the result of the procedure.


33

We have to make the notion of a concept still more precise before defining the concept. Constructions are hyperintensionally individuated entities, which is a fine-grained explication of meaning, but from the conceptual point of view it is rather too fine-grained. Some constructions are almost identical, not distinguishable in a natural language, though not strictly identical. We define a relation of quasi-identity on the collection of all constructions, and say that quasi-identical are constructions that are either D-equivalent or K-equivalent. For instance, constructions Ox[x > 00], Oy[y > 00], Oz[z > 00], etc., are Dequivalent. They define the class of positive numbers in a conceptually indistinguishable way. Similarly, conceptually indistinguishable constructions are K-equivalent ones, like 0+, Oxy [0+ x y], where the latter is an K-expansion of the former. Each equivalent class of constructions can be ordered, and we say that the first one is the construction in the canonical form. Concept is then defined as a canonical closed construction, the other quasiidentical constructions point at the concept. Definition 4 (open / closed construction) Let C be a construction. A variable x is R-bound in C, if it is a subconstruction of a construction C’ that is mentioned by trivialisation. A variable y is O-bound in C, if it is a subconstruction of a closure construction of the form Oy and y is not R-bound. A variable is free in C, if it is neither R-bound nor O-bound. A construction without free variables is closed. Examples. Construction [0+ x 01] is open: variable x is free here. Construction Ox [0+ x 01] is closed: variable x is O-bound here. Construction 0[Ox[0+ x 01]] is closed: variable x is Rbound here. Definition 5 (concept) Concept is a closed construction in the canonical form. Concept C1 is contained in the concept C2, if C1 is used as a constituent of C2. Content of a concept C is a set of concepts contained in C. Extent of a concept C is the entity constructed by C. Extent of an empirical concept CE in a state of the world w,t is the value of the intension I constructed by CE in w,t. Examples. The concept of the greatest prime is strictly empty, it does not have any extent: [0Sing Ox ( [0Prime x] y [[0Prime y] [x t y]] )]. (The Sing function – ‘the only x such that’ – returns the only member of a singleon, on other sets (empty or sets of more than one member) it does not return anything.) The content of this concept is: {0Sing, 0Prime, 0, 0, Ox ( [0Prime x] y [[0Prime y] [x t y]] ), 0, 0t, the concept itself}. The empirical concept of the President of the USA identifies the individual office: OwOt [0President-ofwt 0USA] o LWZ. o Its content is the set {OwOt[0President-ofwt 0USA], 0President-of, 0USA}. o Its extent is the office. o Its current extent is G.W. Bush. The concept used to be empirically empty before the year 1789. To make the Web search more effective, important conceptual properties and relations should be followed. Among them, perhaps the most important is the relation called in the Description logic subsuming (or in [18] intensional containment, known also as subconcept-superconcept). Conceptual relations are analytical. In other words, understanding, e.g., the sentences like A cat is a feline, Whales are mammals, No bachelor is married, we do not have to investigate the state of the world (or search in the Web) to evaluate them in any state of the world as being true. This relation can be defined

34


extensionally, i.e., in terms of the extents of concepts. For instance, the property of being a cat has as its requisite the property of being a feline: necessarily in each state of affairs the population of cats is a subset of the population of felines. Or, necessarily, the concept of the President of USA subsumes the concept of the highest representative of USA. We have seen that except of DL, languages based on FOL cannot specify this important relation. Subsuming is in TIL defined as follows: Let C1, C2 be empirical concepts. Then C1 subsumes C2, denoted C1 S C2, iff in all the states of affairs w, t the extent of C1 is contained in the extent of C2. Formally: Definition 6 (subsuming) [0Subsume 0C1 0C2] = wt x [[C1wt x] [C2wt x]], where (Subsume (R n n), C1 o (RĮ)WZ, C2 o (RĮ)WZ, x o Į). Thus for instance, since the concept of bachelor subsumes the concept of being never married ex definitione, once we obtain a piece of information that Mr. X is a bachelor we will not search any more for the information whether X is, or has ever been, married. To follow such necessary relations between concepts, each important term of a domain should be provided with an ontological definition of the entity denoted by the term. The ontological definition is a complex concept composing as constituents primitive concepts of our conceptual system [22]. For instance, the concept woman can be defined using primitive concepts person and female. The concept mother can be further defined using the former, existential quantifier and the concept child: 0 woman = OwOt Ox [[[0sexofwt x] = 0female] [0personwt x]], 0 mother = OwOt Ox [[0womanwt x] y [[0childofwt x] y]] where: x o L, y o L, woman / (RL)WZ, female / (RL)WZ, sexof / ((RL)WZ L)WZ, person / (RL)WZ, childof / ((RL) L)WZ. From the above definitions it follows that necessarily each woman is a female person, each mother is a woman having children, etc. Other analytical relations between concepts that can be defined using Web ontological languages based on the Description logic are equivalence and incompatibility (disjointness in DL): Definition 7 (equivalence, incompatibility): Concepts C1, C2 are equivalent, if they have exactly the same extent (construct the same entity). Concepts C1, C2 are incompatible, if in no w, t the extent of C1 is a part of the extent of C2, and vice versa. Note that concepts C1, C2 are equivalent iff C1 S C2 and C2 S C1, but not if their contents are identical. Examples: Concepts bachelor and a married man are incompatible. Concepts of the proposition that It is not necessary that if the President of USA is a republican then he attacks Iraq and of the proposition that It is possible that the President of USA is a republican and he does not attack Iraq are equivalent. Some procedures can even fail, not producing any output. They are empty concepts. There are several degrees of emptiness. From the strict emptiness, when the respective concept does not identify anything (like the greatest prime) to emptiness, when the respective procedure identifies an empty set. Empirical concepts always identify a nontrivial intension of a type DWZ. They cannot be (strictly) empty, they can be rather empirically empty when the respective identified intension does not have any value or the value is an empty set in a current state of the world w, t. A concept empirically empty in the actual state of affairs is, e.g., the King of France. Not to search for extensions of empty concepts, classes of empty concepts should be defined:


35

Definition 8 (empty concepts): Strictly empty concept does not have an extension (the construction fails). Empty concept has an empty extension. Empirical concept C is strictly empty in w,t, if it does not have an extent in w,t (the intensional descent of I, i.e., Cwt is improper. Empirical concept C is empty in w,t, if its extent in w,t is an empty set (the intensional descent of I, i.e., Cwt, constructs an empty set). 4. Web Ontological Languages and TIL The system of TIL provides a logical framework for building ontologies and defines the features that each ideal knowledge representation ontological language should possess. Moreover, particular “intensional and extensional inter-relations” that should be marked up in and between documents can easily be specified in TIL. Analytical relations between concepts specify “intensional inter-relations”, whereas empirical complex concepts specify “extensional inter-relations”. Due to TIL explicit intensionalisation and temporalisation (parameters w, t), also the time-sensitive links [13] can be easily specified. A common objection against such a rich and highly expressive system is that it is too complicated and computationally intractable. In our opinion this objection is not justified. The TIL specification is fine-grained and semantically transparent; all the semantically salient features are explicitly present, so that a back translation to a natural language is strait-forward. Our notation is basically a classical typed O-calculus with a fixed interpretation in the universe split into an infinite hierarchy of sorts. Thus implementation is not intractable, and not by itself ineffective. The typing generally makes the realisation more effective. For instance, it can be shown that in the typed Prolog the number of necessary resolvents to be generated is rapidly reduced compared to the untyped Prolog. Empirical concepts define problems, i.e., intensions. Their solution consists in evaluating the intension, i.e., finding its value in an actual state of affairs. For instance the concept of the Mayor of Dunedin defines a problem who is currently the Mayor of Dunedin? The fact, that it is Sukhi Turner, is a contingent state of affairs and we have to empirically investigate whether it is so. On the other hand, analytically necessary relations between concepts are important and can make our investigation more effective, because knowing them helps avoiding unnecessary empirical investigations. For instance, if we empirically solve the problem whether Mr. X is a bachelor (for instance by searching in huge amount of Web data), we will not search any more for solving the problem whether Mr. X is married, for we can analytically deduce that these concepts are incompatible. Unfortunately, most Web ontology languages do not make it possible to distinguish analytical vs. empirical, necessary vs. contingent. As for the complexity, it concerns problems not the language. Of course, using a rich language, more complex, or even intractable problems can be specified. Anyway, we have first to understand the problem, i.e., to know what to do, and only afterwards try to solve the problem. Formalising by logical construction makes it possible to understand and to logically handle the problem. Again, a standard objection is the fact that for expressive higher-order languages there are no semantically complete calculi. The question is, whether such a shortcoming of “mathematical beauty” is substantial. In our opinion, it is not. We need to obtain as much information as possible from the Web, and the fact that we cannot mechanically infer all the logical consequences of these facts is not so important. There is however a pragmatic problem. In order that such a system is applicable in practical building of domain ontologies, we have first to create an intelligent thesaurus of the most important concepts of the domain. In this thesaurus each important simple expression is provided with an ontological definition of the denoted entity, i.e., a complex

36


concept of the entity is assigned to the expression. The definition is as fine-grained as the used conceptual system makes it possible. In other words, we have to choose a set of primitive atomic concepts of the domain, from which the composed concepts are created. Moreover, we have to avoid cyclic or circle definitions. 5. Conclusion The nature of the current Semantic Web can be summarised by six principles [20]. Everything can be identified by URI addresses. Resources and links can have types. Partial information is tolerated. There is no need for absolute truth, i.e., everything found from the Web is not true. Evolution is supported. W3C activities aim to standardise no more than is necessary. Conceptual levels and related technologies of the Semantic Web are just the basis. Presenting and utilising meanings as well as drawing inferences are questions at totally different level. This is the practise of the bottom up approach. In this paper we wanted to show that this pragmatic approach has its limitations. We claim that an ontologically deficient conceptual level can not only block some useful inferences, but also prevent our very understanding of Web resources, and the results of information retrieval. Concluding, we can say that an ideal way of building ontologies would be a top-down approach based for instance on the rich conceptual-level ontology of TIL, which provides a unique semantic and terminological framework. If ontological languages are not at least underpinned by a rich logical background, it may take a lot of efforts to reach the goals for which the Web had initially been proposed. Neglecting the conceptual and logical level of the Web architecture may lead to severe problems when integrating particular Web ontology languages and even to impossibility to provide relevant information. In other words, answering the title question, in order that the practice of the Semantic Web meets the theory, we should develop conceptual-level ontology based on a unique and semantically rich logical framework. References [1]

Baader, F., Calvanese D., McGuiness D.L., Nardi D. and Patel-Schneider P.F. 2003: The Description Logic Handbook. Theory, Implementation and Applications. Cambridge: Cambridge University Press. [2] Berners-Lee, T. 1989: Information Management Proposal, (referred 10.1.2005) . [3] Berners-Lee, T. 2000: Semantic Web on XML. XML 2000 Washington DC, (referred 10.1.2005) . [4] Berners-Lee, T., Hendler, J. and Lassila, O. 2001: The Semantic Web. Scientific American, Vol. 284, No. 5, pp. 28 – 37. [5] Duží, M. 2004: Concepts, Language and Ontologies (from the logical point of view). In: Information Modelling and Knowledge Bases XV, Eiji Kawaguchi, Hannu Kangassalo (Eds), Amsterdam: IOS Press, pp 193-209. [6] Duží, M. 2004: Intensional Logic and the Irreducible Contrast between de dicto and de re. ProFil, Vol. 5, No. 1, pp. 1-34, (referred 10.1.2005) . [7] Duží, M., Ćuráková D. and Menšík, M. 2004: Concepts are Structured Meanings. In: Proc. of the 13th European-Japanese Conference on Information Modelling and Knowledge Bases, Y. Kiyoki, H. Kangassalo, E Kawaguchi (Eds). Sweden, pp. 245-264. [8] Duží, M., Jespersen, B. and Müller, J. 2005: Epistemic Closure and Inferable Knowledge. In: The Logica Yearbook 2004, Libor BČhounek and Marta Bílková (eds.), Filosofia Prague, pp. 125-141. [9] Duží, M., Materna, P. 2005: Logical Form. In: Essays on the Foundations of Mathematics and Logic, Giandomenico Sica (ed.), Polimetrica International Scientific Publisher, pp. 115-153. [10] Ganter, B. and Wille, R. 1999: Formal Concept Analysis. Berlin: Springer – Verlag, [11] Gordon, M. J. C. and Melham, T. F. (Eds) 1993: Introduction to HOL: A Theorem Proving Environment for Higher Order Logic. Cambridge: Cambridge University Press.


37

[12] Hayes, P. and Menzel, Ch. 2001: A Semantics for the Knowledge Interchange Format, (referred 10.1.2005) . [13] Heimbürger, A. 2004: Time-sensitive Relationship Management in Technical Manuals. Case: Maintenance Schedules. In: Kiyoki, Y., Kangassalo, H. and Wangler, B. (Eds.). Proceedings of the 14th European-Japanese Conference on Information Modelling and Knowledge Bases (EJC 2004), May 31 – June 4, 2004, Skoevde University, Sweden, pp. 148 – 164. [14] Hendler, J., Berners-Lee, T. and Miller, E. 2000: Integrating Applications on the Semantic Web, Journal of the Institute of Electrical Engineers of Japan, Vol 122, No. 10, pp. 676 – 680. [15] Horrocks, I. and Patel-Schneider, P.F. 2003: Three Theses of Representation in the Semantic Web. WWW2003, May 20-24, Budapest, Hungary, 2003, (referred 10.1.2005) . [16] Horrocks, I., Patel-Schneider, P.F., Boley, H., Tabet, S., Grosof, B. and Dean, M 2004: SWRL: A Semantic Web Rule Language Combiming OWL and RuleML. W3C Member Submission, May 2004, (referred 10.1.2005) . [17] ISO 8879. 1986: Information processing – Text and office systems – Standard generalized markup language (SGML). Geneva: International Organization for Standardization. 155 p. [18] Kauppi, R. 1967: Einführung in die Theorie der Begriffssysteme, Acta Universitatis Tamperensis A/15, Tampere. [19] Koivunen, M. – R. and Miller, E. 2002: W3C Semantic Web Activity. In: Hyvönen, E. (Ed.) Semantic Web Kick-Off in Finland. Visions, Technologies, Research and Applications, November, 2, 2001, Helsinki, Finland. HIIT Publications 2002-01. Pp. 27 – 43. [20] Marshall, C. and Shipman, F. 2003: Which Semantic Web? In: Proceedings of the 14th ACM conference on Hypertext and hypermedia 2003, Nottingham, UK, August 26 - 30, 2003, pp. 57-66. [21] Materna, P. 1998: Concepts and Objects. Acta Philosophica Phennica, Vol.63, Helsinki. [22] Materna, P. 2004: Conceptual Systems. Berlin: Logos Verlag.. [23] Moschovakis, Y.1994: Sense and Denotation as Algorithm and Value. In J. Oikkonen and J. Vaananen, (Eds.) Lecture Notes in Logic, #2 (1994). Berlin: Springer. Pp. 210-249. [24] Park, J. and Hunting, S. (Eds.). 2003: XML Topic Maps. Creating and Using Topic Maps for the Web. Boston: Addison-Wesley. 605 p. [25] Paulson. L. C. 1994: Isabelle: A Generic Theorem Prover. Number 828 in LNCS. Berlin: Springer. [26] Tichý, P.1988: The Foundations of Frege’s Logic. De Gruyer, New Yourk. [27] W3C. 2000: The World Wide Web Consortium: Harvesting RDF Statements from XLinks W3C Note 29 September 2000, (referred 10.1.2005) . [28] W3C 2001: The World Wide Web Consortium. DAML + OIL Reference Description, (referred 10.1.2005) . [29] W3C 2001: The World Wide Web Consortium. Semantic Web Activity, (referred 10.1.2005) . [30] W3C 2001: The World Wide Web Consortium: XML Linking Language (XLink) Version 1.0 W3C Recommendation 27 June 2001, (referred 10.1.2005) . [31] W3C 2003: The World Wide Web Consortium: XPointer Framework W3C Recommendation 25 March 2003, (referred 10.1.2005) . [32] W3C 2004: The World Wide Web Consortium. Extensible Markup Language (XML), 1.0 (Third Edition) W3C Recommendation 4 February 2004, (referred 10.1.2005) . [33] W3C 2004: The World Wide Web Consortium: OWL Web Ontology Language Overview W3C Recommendation 10 February 2004, (referred 10.1.2005) . [34] W3C 2004: The World Wide Web Consortium. RDF Vocabulary Description Language 1.0: RDF Schema W3C Recommendation 10 February 2004, (referred 10.1.2005) . [35] W3C 2004: The World Wide Web Consortium: The Resource Description Framework (RDF), (referred 10.1.2005) . [36] Welcome to the OIL. 2005: Description of OIL, (referred 10.1.2005) . –––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– This research has been supported by the program "Information Society" of the Czech Academy of Sciences, project No. 1ET101940420 "Logic and Artificial Intelligence for multi-agent systems"


38

Object order based on concept analysis and Moebius Inversion Function Petr GAJDOŠ, Daniela ĎURÁKOVÁ Department of Computer Science, VŠB-Technical University of Ostrava tř. 17. listopadu 15, 708 33 Ostrava-Poruba [email protected], [email protected]

Abstract. Ordering objects of interest according to a given criterion often provides a useful piece of information to solve a problem of information processing. A new method of linear order based on Formal Concept Analysis (FCA) and Moebius function is described. The description method uses an example of selecting important key indicators in a specific geographical area. An indicator is defined as a characteristic number (weight) representing in a unique way certain important features of the area in question. The weights are assigned to each indicator in different categories. Using FCA, interdependencies among the indicators are analysed and the indicators are sorted according to their importance and uniqueness in the area description. A new method based on Conjugate Moebius Inversion Function (CMI) was used to evaluate the set of indicators selected as representatives for a pilot area. The output of the method is a sequence of objects / indicators ordered according to their calculated value of importance of every indicator. The method described in the article appears to be an important contribution to the evaluation of regional competitiveness in the 5th Framework Programme RTD project “Iron Curtain”.

Introduction The goal of the project is to provide a method that would organise objects falling under a concept according to the chosen criteria into a hierarchy charactering particular area of interest. The proposed method is based on the Formal Concept Analysis (FCA) and Conjugate Moebius Inversion Function (CMI). Concepts of a given context are partially ordered using FCA in the same way like in [1], and the Moebius function enables us to compute diversity of concepts. This is then the basis that makes it possible to inversely compute weights of particular objects, i.e., the rate of their contributing to the character of the area. The method consists of the following steps: First, collected input data are adjusted into a context data, i.e., the matrix of incidence. The matrix serves as an input into FCA, which computes the set of all concepts of the context.

ˇ P. Gajdoš and D. Duráková / Object Order Based on Concept Analysis

39

To each concept its diversity is computed using CMI, and finally the inverse way is used to compute weights of particular objects. The paper is organised as follows: First, basic theoretical notions, namely formal context, formal concept, diversity, Conjugate Moebius Inverse function are introduced. Then we provide a brief description of our methodology. In the last chapter the proposed method is illustrated by its practical application on real data taken from the 5th FP RTD project “Iron Curtain”.

1. Background This section introduces theoretical background and definitions important for our later evaluation. We define basic notions of formal concept analysis (FCA) and the basic features of the multi-attribute model developed in the theory of diversity (TD).

1.1 Formal Context and Concept FCA has been defined by R. Wille [2] and it can be used for hierarchical order of objects based on object’s features. The basic terms are formal context and formal concept. Definition 1. A formal context C = (G,M,I) consists of two sets G and M and a relation I between G and M. Elements of G are called objects and elements of M are called attributes of the context. In order to express that an object g is in a relation I with an attribute m, we write gIm or ( g , m) ∈ I and read it as “the object g has the attribute m”. The relation I is also called the incidence relation of the context. Definition 2. For a set A ⊂ G of objects we define A↑ = {m ∈ M | gIm for all g ∈ A} (the set of attributes common to the objects in A). Correspondingly, for a set B of attributes we define B ↓ = {g ∈ G | gIm for all m ∈ B} (the set of objects which have all attributes in B). Definition 3. A formal concept of the context (G,M,I) is a pair (A, B) with A ⊆ G, B ⊆ M, A↑=B and B↓=A. We call A the extent and B the intent of the concept (A, B). Power set K(G,M,I) denotes the set of all concepts of context. Definition 4. For concepts (A1, B1) and (A2, B2) ∈ K(G,M,I) we say (A1, B1) is a subconcept of (A2, B2) if A1 ⊆ A2 (which is equivalent to B2 ⊆ B1 ) and we write (A1, B1) ≤ (A2, B2) and say also that (A2, B2) is a superconcept of (A1, B1). In the following table 1 an example of the context describing animals and their properties is adduced:

40


Table 1: An example of the context (i.e., the incidence matrix). M

G dog cat rabbit donkey dove shark

house 1 1 1 1

tail 1 1 1 1 1

hair 1 1 1 1

dump moody swims 1 1 1 1

1 1

Background area (see Table 1) is showing one concept from the set of all concepts over this context. The set of objects consists of animals, G = {dog, cat, rabbit, donkey, dove, shark} and the set of attributes contains some selected features, M = {house, tail, hair, dump, moody, swimming}. Relation I is represented by value 1 in those positions, where the given object has the respective attribute. The relation “≤” is the relation of partial order over the set of all concepts K(G,M,I). The ordered set of concepts forms a lattice, because each two-element set of concepts in K(G,M,I) has a supremum and infimum with regard to “≤” that is defined as follows: inf{(A1, B1), (A2, B2)} = ((A1 ∩ A2), (B1 ∪ B2)↓↑) sup{(A1, B1), (A2, B2)} = ((A1 ∪ A2) ↓↑, (B1 ∩ B2)) The ordered set of concepts is even a complete lattice, where each subset has a supremum and infimum: ↓↑ ⎛ ⎛ ⎞ ⎞ ⎜ ⎟ A , B A , B = t ⎜U t ⎟ ∧( t t ) ⎜I ⎟ t∈T t∈T t∈T ⎝ ⎠ ⎝ ⎠ ↑↓ ⎛⎛ ⎞ ⎞ ( At , Bt ) = ⎜⎜ ⎜ U At ⎟ , I Bt ⎟⎟ ∨ t∈T ⎝ ⎝ t∈T ⎠ t∈T ⎠ The proof and further details can be found in [2]. The hierarchical structure on the set of formal concepts is illustrated by the following lattice created on the set of ‘animal concepts’ of our example:


41

Description: Added attribute in the top-down direction Added object

in the bottom-up direction

C1 ({dog,cat,rabbit,donkey},{tail,hair}) C2 ({dog,cat,rabbit,dove},{house,tail}) supremum ({dog, cat, rabbit, donkey, dove} {tail}) infimum ({dog, cat, rabbit} {tail, hair, house}) Figure 1: The power set K (G, M , I ) of the context of table 1 is described by Hasse diagram. Each concept is characterized by its extent (set of objects) and intent (set of attributes). For example, cat and donkey could have ascribed the notion “moody animals”. We use reduced description of concept lattice to get better readability.

1.2 Diversity, Conjugate Moebius Inverse and basic characteristics Definition 5. Let M be the totality of all features deemed relevant in the specific context, and let I ⊂ G × M be the “incidence” relation that describes the features possessed by objects, i.e. ( g , m) ∈ I whenever object g ∈ G possesses a feature m ∈ M . For each relevant feature m ∈ M , let λm ≥ 0 quantify the value of realization of m. Upon normalization, λm can thus be thought of as the relevant importance, or weight of feature m. The diversity value of a set S is defined as

v( S ) =

∑

λm

(1)

m∈M :( g , m )∈I for some g∈S

The diversity value of a set is given by the total weight of all different features possessed by some objects in S [3]. Note especially that each feature occurs at most once at the sum. In particular, each single object contributes to diversity the value of all those features that are not possessed by any other objects. For any subset A ⊆ G of objects let M A denote the set of those features that are possessed exactly by members of A. Hence every feature in M A is possessed by all elements of A and not possessed by any element of M \ A . We can write v( S ) =

∑ ∑λ

A∩ S ≠ 0 m∈M A

A

(2)


42

Then, for each subset A ⊆ G let λA :=

∑

A∩ S ≠∅

λA denote the total weight of all features in MA

possessed by members of the extension A, with the convention that λA = 0 whenever M A = ∅ . With this notation we write v( S ) =

∑

A: A∩ S ≠∅

λA

(3)

1.2.1 Conjugate Moebius Inverse Theorem 1. For any function v : 2 M → R with v := 0 there exists unique function λ : 2M → R , the Conjugate Moebius Inverse, such that λ∅ = 0 and, for all S, v( S ) =

∑

A: A∩ S ≠ 0

λA

(4)

Furthermore, the Conjugate Moebius Inverse λ is given by the following formula. For all A≠∅,

λA =

∑

(−1)| A|−|S |+1 * v( S C )

(5)

A: A∩ S ≠∅

where S c denotes the complement of S in M. [4] 1.2.2 Basic characteristics The diversity of an object g is the sum of all weights of all features which are related to the object according to the incidence matrix. It conveys information about partial importance of an object but does not clearly display other dependences. Table 2: The incidence matrix and one selected object g1 which owns five attributes G g1

M

m1

m2

m3

m4

m5

m6

x

x

x

x

x

x

x

g2 g3

x

x

g4

x

x

dog =

∑

λm

(6)

m:m∈M and ( g Im)∈I

x x

m7

x

x

x

x

Next characteristic is called the sum of diversities of all objects. Actually, the objects of one concept can “cover” all features.


43

Table 3: The incidence matrix and two selected objects g1 and g3. These objects belong to the concept ({g1, g3}, {m1, m4, m5, m6}). M

G g1

m1

m2

m3

m4

m5

m6

x

x

x

x

x

x x

x

x

x

g2 g3

x

x

g4

x

x

m7

sdoC =

x

x

∑ do

g

(7)

g :g∈C

x

Now we can compute the diversity of this concept mentioned above. Table 4: Selected fields show all possible attributes possessed by objects g1 and g3 of the concept ({g1, g3}, {m1, m4, m5, m6}). G g1

M

m1

m2

x

g2

x

g3

x

x

g4

x

x

m3

m4

m5

m6

x

x

x

x

m7

v( S ) =

x x x

∑

A: A∩ S ≠ 0

x

x

λA

(8)

x

x

The importance of the object g is the final result of our method. The value represents the importance from these aspects: Uniqueness – Is there any other similar object? Range of description – What type of dimension does the object describe? Weight of description – What is the weight of object in each dimension? impog =

sdoC

∑ v ( S )λ do A

g

(9)

C :C ∋ g

Dependently on the character of the solved problem the formula can be composed of some other parameters. In this case the most important indicators, i.e., those ones that have the greatest influence on the further development of the area, have been chosen.

44


2. Description of our method Our method is based (a) on the partial ordering of concepts in the concept lattice and (b) on the inverse calculation of weights of objects using Moebius function. Particular steps are illustrated by figure 2 and briefly described in this chapter. 1. Input data 2. Data transformation

3. Set of concepts creation 4. Basic characteristics computation 5. Sorting objects according to their importance Figure 2: Main steps of our method. First we obtain the input data like a table or matrix of weights for each indicator according to the importance of indicators in four principal competitiveness dimensions. The values are given by experts. The second step - scaling method (see [2] for details) is used to create an input incidence matrix. Every dimension is scaled to a finite number of parts to get the binary values. Then each part determines a new category of an incidence relation. The output of transformation is an incidence matrix that we need as input for the concept calculation. Next the power set of concepts is computed using FCA. We can create the Concept lattice and draw the Hasse diagram, but it is not very important in our method. But it can be useful to show dependences between concepts, if we need it. We use only the list of concepts. After that, we can compute the basic characteristics for each concept according to the formulas (5), (6), (7), (8). Finally, we compute the importance of objects according to the formula (9) and the obtained values provide us the criterion to sort the set of objects.

3. Practical application The data used for a practical application of this new approach originated in the 5th FP RTD project “Iron Curtain” QLRT-2000-01401 [5] (“Innovative models of critical key indicators as planning and decision support for sustainable rural development and integrated cross border regional management in former Iron Curtain areas based on north to south European reference studies.”) The main goal of the project is to find key indicators to evaluate a regional competitive level and territorial capital of selected pilot areas. The four principal


45

dimensions of territorial competitiveness as defined by the LEADER methodology concept are:

• • • •

Economic competitiveness Social competitiveness Environmental competitiveness Positioning in the global context - (relation with the outside world)

From the 6 reference areas (RA) analysed in the “Iron Curtain” project, we have chosen for testing and evaluation of indicators the RA4 located on the border between Norway and Russia, between administrative units of Finnmark County, Sor-Varanger municipality on the Norwegian side and Murmansk Oblast (Region), Pechenga District. It is situated approximately 400 km north of the Arctic Circle and approximately delimited by coordinates 67-69N and 26-29 E or the Norwegian – Russian part of catchment of Pasvik river. Due to its geographic position, historical development, and its dependence on natural resources the area has played a crucial role in history in connection with the development of contacts in the Northern Calotte Area and contact development between Norway and North-west Russia. It plays a unique role in international co-operation in general and in particular in the Barents Euro-Arctic Region.

3.1 Input data from experts For detailed description of the area, 44 indicators were selected and weights were assigned on scale of 1-10 in each dimension of the Leader Concept by the method of expert evaluation.

G

…

…

…

N

42 43 44

TSF UAL URT

7 8 9

6

3

5 3 4

5 4

S Indicator description 4 Average age of residential dwellings/buildings Number of Overnight Stays (per inhabitant) 2 Average Overnight Stays (length of stay..?) 5 border crossing- persons border crossing-goods in tons 7 Children in family day-care centres, 7 kindergartens/daycare 5 Private persons: Extra child support per children. 2 Share of employment in Agricult., forest and fishing 2 Share of employment in Industry …

E 7 8 8 5 7 3 4 5 9 9

…

index 1 2 3 4 5 6 7 8 9 10

…

sign ADB AOI AOO BCP BCT CCC CCD CCE EEA EEI

…

Table 5: Input data – weighted values of indicators of Norway/Russia area. Legend: The Leader competitiveness dimensions: E-Economical, S-Social, N-Environmental, G-Position in the Global Context

1 seaport traffic- freight 6 Average length of unemployment 4 Unemployment rate


46

3.2 Input data transformation This input table is transformed to the incidence matrix (see table 6). Indicators are represented by objects in incidence matrix. Features are represented by scaled dimensions. Table 6: The incidence matrix after scaling. Each dimension is scaled into ten parts. 7

8

… 10 1 … … g10 n1 … … 0 0 … … 0 0 …

10 1 … 10 n10 s1 … s10

0

0

1

0

0

0

0

0

0

0 … 0

5

0

0

0

0

1

0

0

0

0

0 … 0

0 … 0

0 … 0

…

0

…

0

…

0

…

1

…

0

…

0

…

0

…

0

…

0 0

0

…

0 0

3

0

…

4

0 0 … 0 0 … 0

0

0 … 0 0 … 0

0

42 0 43 0

0

0

0

0

0

1

0

0

0

0

0 … 0

0

0

0

0

0

0

1

0

0

0

0 … 0

0 … 0 0 … 0

44 0

0

0

0

0

0

0

0

1

0

0

0 … 0

0 … 0

0 … 0 0 … 0 0 … 0 0

;

…

6

…

5

…

4

…

9 10 1 2 e3 e4 e5 e6 e7 e8 e9 e10 g1 g2 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 3

…

w 1 2 i e1 e2 1 0 0 2 0 0

0 0 … 0 0 … 0

3.3 FCA application: creating a set of concepts In the third step we use FCA and we can compute all concepts over this incidence matrix. The list of selected concepts of power concept set and required values are illustrated in table 7. Table 7: The result of the third step.

C70 C71 C72

{ UAL } { URT } {}

set of attributes

λ

v(S)

sdoC

166

535

{ e7 } { s4 } { e7 s4 } { e8 } { e8 g5 } { s2 } { g3 } { g3 s2 } { e8 g3 s2 }

7 4 0 8 5 2 3 0 0

42 62 11 37 13 51 61 25 13

81 98 11 73 13 93 83 30 13

{ e8 s6 } { e9 s4 } { e1 e2 e3 e4 e5 e6 e7 e8 e9 g2 g3 g4 g5 g6 g9 n2 n3 n4 n5 n6 n8 n9 n10 s1 s2 s3 s4 s5 s6 s7 s8 s9 }

0 0 0

…

0

…

{}

…

…

C2 C3 C4 C5 C6 C7 C8 C9 C10

{ ADB AOI AOO BCP BCT CCC CCD CCE EEA EEI EES ELH ELP EXP FAL FCM FMB FSA FWH GAT GDC GWD HAD HAI HEU HGB HLE HMR HPC HPD HPH HRS LPC MSE MVI NBD NMS NPA NTA RLC RTT TSF UAL URT } { ADB BCT ELH FMB HAD TSF } { ADB GWD HMR HPH MVI NTA URT } { ADB } { AOI AOO GWD HMR UAL } { AOI } { AOO EEA EEI EES ELP GDC NPA } { AOO GAT HAD HEU NPA RLC } { AOO NPA } { AOO } …

C1

set of objects

…

index of concept

14 13 0

14 13 0


47

The set of concepts over the context (see Table 7) represented by scaling matrix. The second column in the table contains a set of objects (extent), the third column contains a set of attributes (intent) appropriate for given concept. The following columns demonstrate values of the weight of a feature (λ), the diversity value (v) and total diversity of objects (sdoC). 3.4 Computation of basic characteristics For the sake of simplicity we demonstrate the computation of basic characteristics on a selected concept, an analogical approach will be applied to other concepts. For instance, we choose the concept called C5 = ({AOI, AOO, GWD, HMR, UAL},{e8}). Applied values are taken from the incidence matrix and the more accurately usage is shown in table 8. (There is only a part of the incidence matrix in this concrete case.) The total diversity of objects is computed based on presented values according to (7). sdo =8 + 5 + 8 + 3 + 2 + 8 + 4 + 8 + 9 + 4 + 8 + 6 = 73 Table 8: The part of the incidence matrix that is necessary to compute sdo value for selected concept C5.

Objects of incidence matrix

8 weight of feature dimension after scaling e8 1 AOI 1 AOO 1 GWD 1 HMR 1 UAL

3 g3 0 1 0 0 0

5 g5 1 0 0 0 0

9 g9 0 0 0 1 0

2 s2 0 1 0 0 0

4 s4 0 0 1 1 0

6 s6 0 0 0 0 1

3.5 Sorting of objects Finally, the importance of object can be computed. Now we have value of importance assigned to each object. The formula (9) is used to obtain the following table 9 where the ordered indicators are presented. Any classical sorting method can be used to obtain the list of objects sorted linearly according to the key value expressed by a real number. Table 9: Weights of all indicators in observed area. order 1 2 3 4 5 6 7 8 9 10

index NTA HMR HPD EEA GDC HEU NPA EEI ELH HLE

impo order index 707,6276 12 HAD 653,2337 13 RTT 554,6846 14 URT 544,3529 15 HPC 501,5607 16 UAL 486,7101 17 NBD 481,8363 18 AOO 476,2059 19 EES 458,6452 20 AOI 446,8805 21 GWD

impo order index 401,3115 23 GAT 380,8132 24 ADB 374,6935 25 BCP 354,2206 26 CCE 325,973 27 CCD 312,0599 28 FCM 305,6665 29 FMB 287,6176 30 CCC 270,1892 31 FAL 265,2764 32 MVI

impo order index 222,5468 34 HRS 218,0484 35 HGB 208,3333 36 RLC 208,3333 37 EXP 207,8784 38 HPH 190,08 39 HAI 177,4138 40 FSA 166,6471 41 NMS 150,4286 42 ELP 143,1532 43 FWH

impo 124,8261 110,1176 109,3224 107,3892 99,64731 94,61667 66,5 64 58,68235 47,68696


48

Final linear order helps participants of Iron Curtain project to select the most important indicators in each observed area. Table 10: Comparison of results shows that the final order is a bit different from simple sum of weights.

ELP 58,7 4 FWH 47,7 2 LPC 30,5 3

2 4 2

...

...

...

...

...

708 653 555 544 502 487 482 476 459 447 ...

...

NTA HMR HPD EEA GDC HEU NPA EEI ELH HLE

...

experts' values sum description of indicator E G N S 4 6 10 4 24 Temperature (weekly average) 8 9 4 21 net migration rate 6 4 6 3 19 Population density 9 5 2 16 Share of employment in Agriculture, forest and fishing 9 4 2 15 GDP per capita 5 3 2 7 17 Share of University graduated 3 3 9 2 17 Share of protected areas in RA 9 4 2 15 Share of employment in Industry 7 4 5 16 Electric energy, production on Pasvik cascade (No/Ru) 4 5 7 16 life expectancy

index impo

6 Electricity prices end-user, national 6 Wood harvest 5 livestock per capita

3.6 Comparison of results of our methods with the opinion of experts The order of indicators does not depend only on the sum of values in each dimension. Our method shows that some more important and relevant relations than those obtained by statistical methods have been discovered. Consulting the results with the experts familiar with the area who were also the authors of weights for individual indicators, it has been proven that this method provides qualitatively new information to the indicator analysis. The above Table 10 illustrates the difference between the results obtained by our method and those obtained by simple sum.

4. Conclusion The demonstrated method of approach has shown that in comparison with subjective evaluation by expert estimates, the Moebius function and FCA offer a well-ordered, quantified hierarchy of indicators with a value indicating the uniqueness or the ability of an indicator to be replaced or substituted by another indicator from the selected set. This hierarchy offers to decision-makers an easy solution to a rather complex and convoluted problem of subjective indicator evaluation and selection of a subset of critical key indicators from the available data. It is obvious that the output depends primarily on the values/weights assigned to each indicator in the four main dimensions of the Leader Project [5] and a possible subjective error of judgement of an expert evaluator may propagate through the FCA approach, but the structure of the output suggest a possibility of further analysis and evaluation of relative


49

position of the indicators, thus highlighting discrepancies and enabling an experienced analyst to determine the subjective validity of the expert evaluation. The parameter impog (the importance of object g) produced by the method provides an aggregated information containing the 3 important descriptors (Uniqueness, Range of Description, Weight of Description), combining the most important features of each indicator and allowing the analyst/evaluator to select the most critical indicators/data fields for monitoring, further analysis and as an input for predictive modelling tools used in advanced decision-support systems. ––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––––– This work has been promoted by the grant project No 1ET101940420. Thanks to Vaclav Snasel for his idea how to use Moebius function to order objects and Marie Duzi for consultations concerning this method. Also,thanks to our colleagues George Lev and Lubor Tvrdy who provided real data from “Iron Curtain” project. It was used to verify our methodology.

5. References 1. 2. 3. 4. 5.

Duží M. and Ďuráková, D. and Menšík M.: Concepts are structured meanings. In Proceedings of the 13th EJC, 2004. Ganter, B. and Wille, R.: Formal Concept Analysis. Springer-Verlag, Berlin, Heidelberg, 1999. Nehring, K.: A Theory of Diversity. Ecometrica 70 (2002) 1155-1198. Nehring, K. and Puppe, C.: Modelling phylogenetic diversity. Resource and Energy Economics (2002). Project web site, http://www.ironcurtainproject.com/


50

A Semantic Spectrum Analyzer for Realizing Semantic Learning in a Semantic Associative Search Space *

**

*

Yasushi KIYOKI , Xing CHEN and Hidehiro OHASHI * Department of Environmental Information Keio University Fujisawa, Kanagawa 252-8520, Japan [email protected], [email protected] ** Department of Information & Computer Sciences Kanagawa Institute of Technology 1030 Simo-Ogino, Atsugi-shi, Kanagawa 243-0292, Japan [email protected] Abstract. In this paper, we present a learning system with a Semantic Spectrum Analyzer to realize appropriate and sharp semantic vector spaces for semantic associative search. In semantic associative search systems, a learning system is essentially required to obtain semantically related and appropriate information from multimedia databases. We propose a new learning algorithm with a Semantic Spectrum Analyzer for the semantic associative search. A Semantic Spectrum Analyzer is essential for adapting retrieval results according to individual variation and for improving accuracy of the retrieval results. This learning algorithm is applied to adjust retrieval results to keywords and retrieval-candidate data. The Semantic Spectrum Analyzer makes it possible to extract semantically related and appropriate information for adjusting the initial positions of semantic vectors to the positions adapting to the individual query requirements.

1. Introduction Database systems are used in various applications and required to support intelligent retrieval and integration for manipulating multimedia databases. A number of databases have spread in wide-area computer network environments, and we have opportunities to obtain significant information from those heterogeneous databases [1,2,6,13]. It is difficult for users to extract appropriate information without knowledge on the contents and structures of those databases. In the research field of semantic associative search systems, the realization of vector spaces is one of the most important issues for resolving semantic relationship between objects and concepts in databases [3,4,5]. Semantic relationship is the essential foundation for extracting the semantically related and appropriate information from multimedia databases [9,10,11,12]. We have proposed a fundamental framework for computing semantic relationship between objects and concepts in multimedia database environments [4,5,6,9]. We have designed a semantic associative search method to extract semantically related and appropriate information concerning user's requests in multimedia database environments. In this paper, we present a learning system with a semantic spectrum analyzer which modifies and adjusts semantic expression vectors to produce correct and appropriate

Y. Kiyoki et al. / A Semantic Spectrum Analyzer for Realizing Semantic Learning

51

retrieval results in semantic associative search. Semantic expression vectors are used as metadata of media data (retrieval-candidate media data) to express attributes, contents and impressions of media data [9]. The learning system is used to create appropriate semantic expression vectors by checking and modifying each element of those vectors. We have designed and implemented learning processes for semantic associative search systems to realize an intelligent retrieval environment for multimedia databases. Our semantic associative search method is advantageous to semantically extract appropriate information from multimedia databases by giving a set of context words as a query [9]. In this method, a set of context words is expressed as a context vector. Then, the context vector is mapped onto the semantic space and correlation computation between the context vector and media data is performed to find semantically correlated media data to the context vector. For improving the quality of semantic associative search, we propose learning processes for semantic associative search to multimedia data. In the initial phase for retrieval, each user specifies a set of context words which is expressed as a query vector. The semantic associative search might not always select appropriate media data from a media database (a set of retrieval candidate data), because media data are not always appropriately expressed as vectors, and the judgment of accuracy is dependent on individual variation among users. Therefore, a learning system is needed for improving accuracy of retrieval results and for adapting retrieval results according to individual variation. Our learning processes are applied to adjust vectors of media data so as to create appropriate retrieval results. Then, learned vectors for media data are stored to be used in subsequent queries. By using those learning processes, we can realize a semantic associative search system environment which provides appropriate and precise retrieval results. We also present an actually implemented learning system with the learning processes for semantic associative search, and also show several experimental results to clarify the feasibility and effectiveness of the system. We have proposed a fundamental method for realizing semantic associative search which extracts semantically related data resources [6,9]. We have also presented a learning mechanism for adjusting retrieval results according to the individual variation and obtaining appropriate retrieval results in our semantic associative search method [7,8]. This learning mechanism can be used for semantic associative search for databases. The main feature of the new learning system in this paper is a Semantic Spectrum Analyzer that is used to adjust vectors of retrieval candidate media data so as to obtain appropriate and precise retrieval results in a semantic space for multimedia databases. The outline of the semantic associative search method is reviewed in Section 2. In Section 3, we introduce a new learning method for realizing appropriate and precise semantic associative search. In Section 4, we present a semantic spectrum analyzer for learning processes and actual system implementation of those processes to clarify the feasibility and effectiveness of the learning processes in an actual multimedia database environment.

2 The Semantic Associative Search Method In this section, the outline of our semantic associative search method based on the Mathematical Model of Meaning is briefly reviewed. This model has been presented in [4,5,9] in detail. In the Mathematical Model of Meaning, an orthogonal semantic space is created for implementing the semantic associative search. Retrieval candidates and queries are mapped onto the semantic space. The semantic associative search is performed by calculating the

52


correlation of the retrieval candidates and the queries on the semantic space. These processing steps are summarized as follows: Step-1. Creation of the semantic space To create the semantic space, a dictionary is selected and utilized. We refer to the basic words that are used to explain all the vocabulary entries in the dictionary as features. For example, in a dictionary, an entry term “small” is explained by the features “great,” “size,” “degree” etc. as shown in Figure 1. small

[a.]

opp. not great in size , degree etc. :

small

-great size degree

ability ... beautiful degree great good size sort ...

small ...

0

...

0

1

-1

0

1

0

woodland ...

0

M Figure 1: The matrix M constructed based on a dictionary

When m terms are given as the vocabulary entries in the dictionary and n features are utilized to explain all these terms, an m by n matrix M is constructed for the space creation. Each term in the matrix M is characterized by the n features. For example, in Figure 1, the term “small” is characterized by the features “great,” “size,” “degree,” etc. When a feature, for example, “degree,” is used for explaining the term “small,” the value of the entity at the row of “small” and the column “degree” is set to “1” as shown in Figure 1. If features are used as the negative meaning, for example, the feature “great” in Figure 1, the columns corresponding to those features are set to the value “-1.” If features are not used to explain terms, the columns corresponding to those features are set to “0.” In Figure 1, the features “ability” and “beautiful” are not used to explain the term “small,” therefore, the characterized value of these two features is “0.” By using this matrix M, the orthogonal space is computed as the semantic space R based on a mathematical method. Step-2. Mapping retrieval candidates onto the semantic space R: In this model, retrieval candidates are characterized by the n features. Then, the characterized retrieval candidates are mapped onto the semantic space through the Fourier expansion. Figure 2 illustrates the retrieval candidates which are mapped on the semantic space R. In the figure, ri, rj and rk represent the axes of the space R and the vectors A, B, C represent three retrieval candidates. Step-3. Semantic associative search: When a sequence of context words which specifies a user’s query is given, a semantic subspace is dynamically selected from the semantic space. The semantic associative search is performed by computing the semantic correlation between the given query context and the retrieval candidates on the selected subspace. Figure 3 illustrates the semantic difference of the retrieval candidates, A, B and C, on two different subspaces according to contexts. In the following, we give details about the mathematical model of meaning.


rj C

53

A

B

ri rk Figure 2: An example of the retrieval candidates on the semantic space R

Figure 3: The semantic difference of the retrieval candidates on the different semantic subspaces

2.1 Creation of the Semantic Space For the semantic space creation, a matrix M is created based on a dictionary. The matrix M is constructed by m vocabulary entities t i (i 1, , m) and n features ( f1 , f 2 , , f n ) . The vocabulary entities are the terms explained in the dictionary. The dictionary can be a special dictionary or a common dictionary, for example, a medical dictionary, the The General Basic English Dictionary [18],” or the “Longman Dictionary of Contemporary English [17].” The features are the words used to explain the terms in the dictionary. The i-th row represents a term ti and the j-th column represents the feature fj. The i,j-th entity, ti,j in the matrix represents whether the term ti is explained by the feature fj or not. If ti is explained by fj, the value of ti,j is set to “1.” If the feature fj is used as the negative meaning to explain the term ti, the value of ti,j is set to “-1.” Otherwise, the value of ti,j is set to “0.” For the m terms and the n features, the matrix M is constructed as an m u n matrix. The matrix M is referred to as the set of metadata for the semantic space creation. The semantic space is created by performing a mathematical method, the Singular Value Decomposition (SVD), on the matrix M.When SVD is performed on the matrix M, and the matrix is decomposed into the product of three other matrices referred to as L, S and R, M = LSR’. Where S is a diagonal matrix that contains singular values, L and R are left and right matrices of the matrix S. R’ is the transposed matrix of R. The diagonal matrix S has the characteristics as the follows: S’=S, SS-1=I, where I is the identity matrix. Both the matrices L and R have orthonormal columns, that is R*R’=I, L*L’=I The semantic space is created based on the following steps. In the first step, we define the matrix D as,

D = LS =

ª d 1,1 « « « d i ,1 « « « d m ,1 ¬

d 1, 2 d i,2 d m,2

d 1, n º » » d i ,n » » » d m , n »¼

,

In the second step, we re-write the matrix R as R >r1 r2 rn @ , where ri 1 d i d n represents the i-th column of the matrix R. Thus, the result of SVD

54


applied on the matrix M, can be written as

DR’ = M=D

ª d 1,1 « « « d i ,1 « « «d m,1 ¬

d 1, 2 d i,2 d m, 2

d 1, n º ª r1 º » « » » «r2 » d i , n » u «r3 » » « » » «» d m, n »¼ ¬«rn ¼»

.

For the i-th row of the matrix M, which represents a vocabulary entity, can be represented as a vector di ª r1 º «r » d i >d i ,1 d i , 2 d i ,n @u « 2 » «» « » ¬rn ¼ . That is, the vocabulary entity di is expressed on the space R, d i d i ,1r1 d i , 2 r2 d i ,n rn . Because RR’=I, R represents an orthogonal space, ªr1 º ª1 0 0º «r » «0 1 0» » R u R' >r1 r2 rn @ u « 2 » « « » «0 0 0» « » « » ¬rn ¼ ¬0 0 1¼ . each column ri ( 1 d i d n ) of the matrix R is an eigenvector representing an axis of the space. We name these eigenvectors “semantic elements.” 2.2 Characterizing Retrieval Candidates In this section, a method for characterizing retrieval candidates is presented. In the method, each retrieval candidate is characterized by the n features. If a retrieval candidate P is explained or defined by t words, o1 , o2 , , ot , we represent the retrieval candidate P as

P oi

ô1 , o2 ,, ot ` . Each word in the retrieval candidate P is characterized by the n features,

u

i ,1

, u i , 2 ,, u t ,n , where, u i ,1 , u i , 2 ,, u t ,n represent the characterized values of the

word oi by the n features. The value of ui , j , ( ui , j ûi ,1 , ui , 2 ,, ut ,n `) is one of the three values, “-1”, “0” or “1”. The absolute value of ui , j , ui , j is an logical value, “0” or “1”.

Therefore, the logical OR operation is performed to characterize the retrieval candidate P, P

t t t § · ¨ s1 u OR u i ,1 , s 2 u OR u i , 2 ,, s n u OR u i ,n ¸ i 1 i 1 i 1 © ¹.

where sk represents a value “+1” or “-1”. The value of sk is calculated by the following formula, k ° ¦ u i ,k °i1 , ° k ® ¦ u i ,k °i1 ° ° 1, ¯

sk

k

¦u

i,k

z0

i 1

k

¦u

i,k

0 .

i 1

The calculation ORti 1 ui,k represents the logical OR operation performed on |u1,k|, |u2,k|, ..., |ut,k|, t

OR u i 1

i ,k

u1,k OR u 2,k OR OR u n ,k

.


The calculation

55

represents that the sign plus “+” or minus “-” is indicated to

t

s k u ORi 1 ui ,k

the calculated result of ORi 1 ui ,k . By defining pk as t

t

sk u OR ui ,k

pk

i 1 , the retrieval candidate P can be represented as P p1 p 2 p n , which is referred to as the “metadata for retrieval candidate.”

2.3 Mapping Retrieval Candidates onto the Semantic Space R

In order to perform the semantic associative search, retrieval candidates must be mapped onto the semantic space. When a retrieval candidate P is mapped on the semantic space R, it is expressed as P d 1r1 d 2 r2 d n rn . That is, the mapping of P on the R is the Fourier expansion of P on the orthogonal space R,

>d1

P

d2

ª r1 º «r » d n @u « 2 » «» « » ¬rn ¼ .

We call di ( 1 d i d n ) as the Fourier coefficient of the retrieval candidate P. We calculate the Fourier coefficients of P under the following formulas: P u >r1

rn @

r2

>d 1

d2

ª r1 º «r » 2 d n @u « » u >r1 «» « » ¬rn ¼

rn @

r2

.

R is the orthogonal space and R*R’ is the identity matrix,

>r1

R u R'

r2

ª r1 º «r » rn @ u « 2 » «» « » ¬rn ¼

ª1 «0 « «0 « ¬0

0 1 0 0

0º 0»» 0» » 1¼ .

Therefore, the Fourier coefficients of the retrieval candidate P are >d 1

d2

dn @ PuR

P u >r1

> p1 > p1

r2

rn @

p2

p n @u >r1

p2

ª r1,1 «r 2 ,1 p n @u « « « ¬rn,1

r2

rn @

r1, 2 r2, 2 rn , 2

r1, n º r2, n »» » » rn, n ¼

. In this model, the norm of each retrieval candidate on the semantic space is normalized into 1, P

d1 d1

d 2 d n u R ' d 2 d n u R '

dˆ

1

dˆ2 dˆn u R '

, where ||V|| represents the norm of the vector V. In the following, we use the normalized vector P to represent the retrieval candidate P.

2.4 Subspace Selection Based on the Semantics of Queries

This process computes correlations between retrieval candidates and a query q. Before the semantic search is performed, a sequence q u1 , u 2 , , u " of " context words are given

56


as a query and a positive real number H s ( 0 H s 1 ) must be given for the subspace selection. Each word in the query q is characterized by the n features ui ui ,1 , ui , 2 ,, ut ,n , which are referred to as the “metadata for context word.” By performing the Fourier expansion on the context word ui ( 1 d i d l ), the word ui is mapped onto the semantic space R and l vectors representing the context words are obtained. The Fourier coefficient of the context word ui on the semantic space can be obtained by using the following calculation: ui u R

ui1

d

ui 2 uin u R

d i , 2 d i ,n . The query on the semantic space is represented as a compound vector q, which is referred to as the query vector, i ,1

l

¦ u u R u R'

q

i

i 1

l § l ¨ ¦ d k ,1 ¦ d k , 2 k 1 ©k 1 q1 q2 qn u R '

l

¦d

k ,n

k 1

· ¸ u R' ¹

. By calculating the absolute maximum value of the Fourier coefficient of the query vector, qmax, we adjust the value of the Fourier coefficient of the query vector between 0 and 1. A vector Cq is defined as the query semantic center based on the adjusted Fourier coefficients of the query vector, § q qn · q2 where, q max MAX q1 , q 2 , , q n ¸ u R' C q ¨¨ 1 q max ¸¹ © q max q max , . A retrieval subspace RS corresponding to the query is selected from the semantic space R by comparing each value of the Fourier coefficient of Cq with the given threshold H, RS

§ ¨ ri qi ri ! H ¨ q max ©

Pi

¦ ^dˆ

· ¸ ¸ ¹

. The selected semantic subspace RS is constructed by the axes ri, rj, rk, ..., on which each Fourier coefficient of the semantic center is greater than the given threshold H . When the subspace RS is selected, the each retrieval candidate vector Pi, Pi ^P1 , P2 , , Pi , ` on the subspace can be represented as q

i, j

rj rj R S

j 1

`.

In order to measure the association between a query and retrieval candidates, we use the Fourier coefficient qj/qmax of the semantic center Cq as the weight to adjust the vectors of retrieval candidates on the subspace. This adjusting emphasizes the semantic elements correlated to the query on the selected subspace. The adjusted vector Pˆ i is represented as Pˆ i

q

qj ˆ ½ d i, j r j r j R S ¾ ¯ max ¿

¦®q j 1

.

Since the norm of the adjusted vector Pˆ i reflects the semantic correlation between the retrieval candidate Pi and the query, we use the norm | Pˆ | as a semantic measure, i

Pˆ i

½ qj ˆ d i, j r j r j R S ¾ ® ¦ j 1 ¯ q max ¿ q

. As the retrieval result, the retrieval candidates are ranked based on the norms of their relative vector Pˆ i . The retrieval candidate with the biggest value of the norm on the subspace is ranked at the top in the retrieval result.


57

3. A New Learning System for Retrieval Candidate Media Data in Semantic Associative Search

In our semantic associative search model, metadata for retrieval candidate data, that are media data to be retrieved, are represented as vectors expressing contents and impressions of those media data. Those vectors are mapped onto a semantic space to compute their correlations to a query that is also represented as a vector. In semantic associative search, it might happen that the media data with the highest relation to a query are not selected. This phenomenon happens when the media data with the mostly related meaning to a query is not located in the correct location in the semantic space. It is caused when media data are not appropriately defined in the original data matrix. In a multimedia database environment, metadata of retrieval candidate data might not be defined as appropriate semantic expression vectors in initial definitions. In this section, we present basic processes of a learning system which modifies and adjusts semantic expression vectors to produce correct and appropriate retrieval results in semantic associative search. The learning system is used to create appropriate semantic expression vectors by checking and modifying each element of vectors in these learning processes. 3.1 An Overview of the Learning Processes

We present a new learning algorithm for adjusting the locations of retrieval candidate data (media data) in a semantic space. This algorithm makes it possible to extract appropriate media data for a user's query without any side effect among media data. In this algorithm, the vector expressing media data which must be highest correlated to a user's query is modified so as to be highest correlated in a semantic space. Figure 4 shows an overview of our learning processes for semantic associative search. multimedia database (Process-1)

……..

metadata

document002

metadata

metadata

……..

metadata

……..

metadata

metadata

metadata

…..

(Process-6)

document N

…..

metadata

…..

metadata

…..

DB design and creation process document001

document meta database

(Process-2) learning in a vector space document vector (original vector)

query

database system

user

modified vector by learning query vector

search results

judgment by a user

(Process-5)

indication of learning information for learning points

(Process-3)

rank 1

document A

…

…

rank i

document D

…

…

rank j

document D

(Process-4) information for learning points

Figure 4: Learning processes for semantic associative search

For example, consider that retrieval candidate data “A” (ex. document “A”), which is expressed as a semantic expression vector located in a semantic space. When a user gives a context word “a” as a query, we assume that the correct answer to be extracted as highest correlated media data is “A” to the context word “a”. The semantic subspace “Sa” is

58


selected by the semantic operator to the context word “a”, and in this subspace, the retrieval candidate data item “A” is not selected as the semantically closest data to the given keyword “a”. In this case, it might happen that the retrieval candidate data with the closest meaning to the context word “a” is not located in the highest correlated location to the context word “a" in the selected semantic subspace. Although the user requires to extract the data “A”, “A” is not currently located at the appropriate position in the semantic space. That is, the retrieval candidate data “A” is not defined appropriately as a vector in the semantic space. Therefore, we need to create a learning system for semantic associative search where retrieval candidate data can be located at its appropriate location in the semantic space. Furthermore, it is important that the learning system includes the concept of adaptability in which each learning process does not affect other retrieval candidate data. In the above example, learning for the media data “A” should not affect the semantic associative search to other media data. To realize this, the locations of other retrieval candidate data items must not be modified to the new locations in the semantic space. If the locations of retrieval candidate data move by the learning of “A”, those locations cannot be correct in subsequent retrievals. As the retrieval candidate data are shared among individual users, their locations must be independent of the learning of “A”. We need a learning system which can avoid affects of learning to other media data and realize appropriate semantic associative search to subsequent queries. We have proposed a learning mechanism for adjusting locations of media data in the semantic space in [7,8]. In this paper, we present new learning processes using a Semantic Spectrum Analyzer to adjust locations of media data with reflecting meanings of their metadata in a semantic space. 3.2 Learning Processes for Vectors Expressing Media Data

As described in Section 2, in the semantic associative search, when a context is given as a query in a form of a set of context words, a semantic subspace corresponding to this context is selected in the semantic space. The highest correlated retrieval candidate data to the query is extracted in the selected semantic subspace. The retrieval candidate data “A” which must be highest correlated to the query “a” is specified by the user. The learning processes are applied when the retrieval candidate data “A” is not extracted, that is, this data is not located in the appropriate location to the query “a” in the subspace. In the computations of correlations between a query and each retrieval candidate data, the learning processes are applied when the retrieval candidate data “A” is not highest correlated to the query. When a semantic subspace is selected according to a given query “a”, the norm of the media data “A” is computed in the semantic subspace selected according to the query “a”. When the “A” is specified as the data which must be highest correlated, that is, the “A” must have the most norm in the semantic subspace, the learning processes are applied to the media data “A”. The query “a” and the target data (retrieval candidate data) “A” are located in the semantic space. Figure 5 shows an overview for the adjustment of a vector in a semantic space. In the learning processes, the retrieval candidate data “A” is modified toward the query “a” so as to make the data “A” is highest correlated to the query “a”. As the constraint in this learning, the change of location of “A” in the semantic space is minimized. This constraint is introduced to minimize the influence of learning in the semantic space. In the selected semantic subspace, the locations (vectors) of “a” and “A” are represented as a’ and A’, respectively.

Y. Kiyoki et al. / A Semantic Spectrum Analyzer for Realizing Semantic Learning q1

59

Semantic Vector Sub-Space

vector R of rank k-1

reversed vector (T2) of 2-nd element of T

NR : norm of vector R

vector S of rank k NS

original vector (T) NT

reversed vector (T1) of 1-st element of T reversed vector (T3) of 3-rd element of T

q2

Figure 5: Adjustment of vectors by a learning process

3.3 Learning Algorithm

The leaning algorithm consists of the following steps. Step-1: When a query (a set of context words), which is represented as a context vector, is given, a semantic subspace is selected according to the query. And then, in the selected semantic subspace, each vector representing retrieval candidate data (media data) is mapped onto the selected subspace. And, norms of those retrieval candidate data, that are representing correlations between each of retrieval candidate data and the query, are computed, and those retrieval candidate data are ranked according to correlations in the descending order. (This means that top ranking of retrieval candidate data are highly correlated to the query, and those are retrieved as the query results.) Step-2: If the target retrieval candidate data “A”, which should be within the top ranking, is not ranked in the appropriate ranking position, the learning process is performed. A table for the learning process is shown to the user. (This table represents the vector of the target retrieval candidate data “A”, and each feature of the vector is shown. The value of a feature is “-1”, “0”, or “1”. This table shows the new correlation values when the user indicates to change (reverse) the value (“0 o 1”, “1 o 0”, “-1 o 0”, “0 o -1”) of each feature, and also shows the new ranking position of “A” if the change (reverse) of the value is actually applied to the feature. That is, this table shows the influence of changing the feature value to the correlation of “A” to the query. Figures 6 and 7 show the structure of tables used for indicating the value change of features in the learning process.) When the change (reverse) of some feature value in the vector of “A” is indicated by the user, a new modified vector for “A” is registered, to be mapped onto the semantic space. Step-3: The norm of the modified vector of “A” is computed, and retrieval candidate data are ranked again according to new correlations in the descending order. The new ranking position is checked by the user. If the new ranking position is not still appropriate to “A”, Steps 2 and 3 are performed repeatedly. If the new ranking position is appropriate to “A”, the current vector of “A” is fixed and registered, and the learning process is terminated.

60

Y. Kiyoki et al. / A Semantic Spectrum Analyzer for Realizing Semantic Learning reversed value patterns to the original vector feature

original vector (T)

f1

f2

f3

T

1

0

1

NTm

RTm

T1

0

0

1

NTm1

RTm1

vector-id

rank

norm

reversed vectors of i-th element for T

T2

1

1

1

NTm2

RTm2

i=1..3

T3

1

0

0

NTm3

RTm3

a vector set of candidates for learning

Figure 6: Vector modification to the original vector

Group A

Group B

fai : features for making norm higher by reverse (0㸢1 or 1㸢0, 0㸢-1 or -1㸢0)

feature

rank

RTmb1

fc1

NTmc1

RTmc1

RTmb2

fc2

NTmc2

RTmc2

…

…

…

…

…

NTmbr

RTmbr

fcs

NTmcs

RTmcs

feature

rank

RTma1

fb1

NTmb1

NTma2

RTma2

fb2

NTmb2

…

…

…

RTmaq

fbr

rank

fa1

NTma1

fa2 …

NTmaq

k=1..s

norm (correlation)

norm (correlation)

feature

fck : features for making norm lower by reverse (0㸢1 or 1㸢0, 0㸢-1 or -1㸢0)

j=1..r

i=1..q

norm (correlation)

faq

Group C

fbj : features of stable norm in reverse

Figure 7: Grouping of vectors after vector modification

4. Semantic Spectrum Analysis and Implementation

In this section, we present semantic spectrum analysis for precise learning, and show an actual implementation of the learning processes to a document database. By these processes, each vector for retrieval candidate data is adjusted as a well-defined vector in the semantic space. 4.1 An Experimental Environment

We have implemented an experimental system for realizing the learning processes in an actual semantic space. In this experiment, we have created a medical semantic space related to the medical field of pulmonary and respiratory organs. To create a data matrix M described in Section 2, we have referred to the Medical dictionary titled ``Stedman's Medical Dictionary by Lippincott Williams and Wilkins [15].” As described in Section 2, the matrix M is constructed by m terms and n features. 539 basic words are used as m terms, and 619 features are used to create the matrix M. That is, 619 features are provided to make the semantic space, and 539 terms are used to represent the words corresponding to the rows in the data matrix M. By using this matrix M, a medical semantic space is created. This space represents the semantic space for computing meanings of context words and media data. Vectors expressing context words and media data mapped onto this space. In this experiment, medical documents are used as media data for retrieval candidates. 160 articles related to


61

the pulmonary and respiratory organs in “Harrison’s Principles of Internal Medicine by McGraw-Hill [16]” are used as retrieval candidate data. Each article are explained as a set of metadata using some of 539 terms, and vectors corresponding to those metadata are integrated as a single vector assigned to the article. Context words given as a query are also described by using some of 539 terms, and a context vector is integrated as a single vector corresponding to a query. In our system, it is possible to use not only 539 terms but also arbitrary various terms, if those terms can be expressed by using 619 features. In this experiment, we use only 539 terms for metadata and query expressions, but it is not limited at all in this system. The procedure for creation of the semantic space is as follows: (1) Each of 539 vocabulary entries corresponds to a row of the matrix M. In the setting of a row of the matrix M, each column corresponding to the explanatory features which appear in each vocabulary entry is set to the value “1”. If features of the vocabulary are used as the negative meaning, the columns corresponding to those features are set to the value “-1”. And, the other columns are set to the value “0”. This process is performed for every vocabulary entry. And then, each column of the matrix is normalized by the 2-norm to create the matrix M. (2) By using this matrix M, a semantic space is created as described in Section 2. This space represents the semantic space for computing meanings of media data and queries. (3) 539 medical terms are mapped onto this semantic space. We have performed several experiments using this image space to clarify the effects of our method. 4.2 Spectrum Analysis in Learning Processes

In this experiment, the context word “genetic” is given as a query, and the retrieval result before learning is shown as a ranking list in Figure 8. As the result of semantic associative search, the document “257CYSTI_DIAGNOSIS” is ranked in ranking 19th. Here, we assume that this document is highly correlated to the context “genetic” and it should be ranked up in the retrieval result. Figure 9 shows very important information in the learning processes. This figure shows the semantic spectrum of the original vector of the document “257CYSTI_DIAGNOSIS” in the subspace which is projected, according to the context word “genetic.” The blue line in each semantic axis represents the semantic spectrum of the context word “genetic,” and the various colored line represents the semantic spectrum of the document “257CYSTI_DIAGNOSIS.” Each color in a various colored line corresponds to one metadata element in a metadata set assigned to this document. That is, the various colored line indicates information on correlations of a metadata element and the context word (query) in detail. By observing this line in detail, we can find which metadata contributes to increase the correlation to the context word (query) or which metadata does not increase the correlation to the context word. Figure 10 shows the interface to the user (the learning indicator) who is responsible for the learning process. This figure represents the vector of the retrieval candidate data “257CYSTI_DIAGNOSIS,” and each feature of the vector is shown. The value of a feature is “-1”, “0”, or “1”. This table shows the new correlation values when the user indicates to change (reverse) the value (“0 o 1”, “1o 0”, “-1 o 0”, “0 o -1”) of a feature, and also shows the new ranking position of “257CYSTI_DIAGNOSIS” if the change (reverse) of the value is actually applied to the feature. That is, this figure shows the influence of changing the feature value to the correlation of “257CYSTI_DIAGNOSIS” to the query.

62


Figure 8: Retrieval result ranking before the learning process

(blue line) : query vector element value (various color line): vector element values of a retrieval candidate document

Figure 9: Semantic spectrum of the original vector of the document (257CYSTI_DIAGNOSIS) before the learning process


63

Figure 10: Interface for the learning process to the document (“257CYSTI_DIAGNOSIS”)

Figure 11 shows that the change (reverse) of the feature value “genetic” in the vector of “257CYSTI_DIAGNOSIS” is indicated by the user, and a new modified vector for “257CYSTI_DIAGNOSIS” is registered, to be mapped onto the semantic space. Figure 12 shows the semantic spectrum (the green lines) of the feature value “genetic” in the subspace. If the feature value of “genetic” is set to “1” in the vector of “257CYSTI_DIAGNOSIS,” the correlation between the “257CYSTI_DIAGNOSIS” and the query increases as shown in this figure.

Figure 11: Learning to the feature “genetic”

(blue line) : query vector element value (various color line): vector element values of “genetic”

Figure 12: Semantic spectrum of “genetic”

64


Figure 13 shows the new semantic spectrum of the document “257CYSTI_DIAGNOSIS” after the learning is applied to the feature “genetic,” that is, after its feature value of the vector of “257CYSTI_DIAGNOSIS” is set to “1” from “0”.

(blue line)

: query vector element value (various color line): vector element values of a retrieval candidate document

Figure 13: Semantic spectrum of the modified vector of “genetic”

the document (257CYSTI_DIAGNOSIS_[75]) by

Furthermore, the learning processes are repeatedly performed. Figure 14 shows that the change (reverse) of the feature value “inherit” in the vector of “257CYSTI_DIAGNOSIS” is indicated by the user, and a new modified vector for “257CYSTI_DIAGNOSIS” is registered, to be mapped onto the semantic space.

Figure 14: Learning to feature “inherit”

Figure 15 shows the semantic spectrum (the grey lines) of the feature value “inherit” in the subspace. If the feature value of “inherit” is set to “1” in the vector of “257CYSTI_DIAGNOSIS,” the correlation between the “257CYSTI_DIAGNOSIS” and the query increases as shown in this figure. Figure 16 shows the new semantic spectrum of the document “257CYSTI_DIAGNOSIS” after the learning is applied to the feature “inherit,” that is, after its feature value of the vector of “257CYSTI_DIAGNOSIS” is set to “1” from “0”. Figure 17 shows the semantic spectrum of the document “257CYSTI_DIAGNOSIS” after the learning is applied to both features “genetic” and “inherit.” In comparing with the spectrum in Figure 9, this spectrum shows that the vector of “257CYSTI_DIAGNOSIS” is highly correlated to the given query after the learning processes, that is, the correlation after


65

learning is higher than that before learning.

(blue line) : query vector element value (various color line): vector element values of “inherit”

Figure 15: Semantic spectrum of “inherit”

(blue line)


Figure 16: Semantic spectrum of the modified vector of the document (257CYSTI_DIAGNOSIS_[511]) by “inherit”

Figure 18 shows the new ranking list in which “257CYSTI_DIAGNOSIS” is highly ranked as the effect of the spectrum analysis in the learning processes. By using this semantic spectrum analysis, we can obtain a very significant learning environment where we can control the correlation of a document to a query very precisely and appropriately. That is, our learning processes provide a new analytical environment for adjusting metadata definitions of retrieval candidate media data to represent their semantics precisely and appropriately in a semantic space. Our learning processes with semantic spectrum analysis can be applied to various vector space models which compute correlations for realizing semantic associative search.

66


(blue line)


Figure 17: Semantic spectrum of the modified vector of the document (257CYSTI_DIAGNOSIS) by “genetic” and “inherit” (the learning process is completed.)

the document of the modified vector

Figure 18: Retrieval result ranking after the learning process

5. Conclusion

In this paper, we have presented a learning system with the semantic spectrum analyzer which modifies and adjusts semantic expression vectors to produce correct and appropriate retrieval results in semantic associative search. This learning system is used to create appropriate semantic expression vectors by checking and modifying each element of those vectors in our learning processes. We have designed and implemented an actual learning system for semantic associative search systems to realize an intelligent retrieval environment for multimedia databases. By using this learning system, we can realize a semantic associative search system environment which provides appropriate and precise retrieval results for multimedia databases. We have also presented an actual learning system with the learning processes for semantic associative search, and also shown several experimental results which clarify the feasibility and effectiveness of the learning system. In the future work, we will extend the learning processes to semantic associative


67

search in multimedia database applications dealing with stories.

References

[1] Batini, C.,Lenzelini, M. and Nabathe, S.B., “À comparative analysis of methodologies for database schema integration,” ACM Comp. Surveys, Vol. 18, pp.323-364, 1986. [2] Bright, M.W., Hurson, A.R., and Pakzad, S.H., “A Taxonomy and Current Issues in Multidatabase System,” IEEE Computer, Vol.25, No.3, pp.50-59, 1992. [3] Deerwester, S., Dumais, S. T., Landauer, T. K., Furnas, G. W. and Harshman, R. A., “Indexing by latent semantic analysis,” Journal of the Society for Information Science, vol.41, no.6, 391-407, 1990. [4] Kitagawa, T. and Kiyoki, Y., “A mathematical model of meaning and its application to multidatabase systems,” Proceedings of 3rd IEEE International Workshop on Research Issues on Data Engineering: Interoperability in Multidatabase Systems, pp.130-135, April 1993. [5] Kiyoki, Y. and Kitagawa, T., “A metadatabase system for supporting semantic interoperability in multidatabases,” Information Modelling and Knowledge Bases (IOS Press), Vol. V, pp.287-298, 1993. [6] Kiyoki, Y., Kitagawa, T. and Hitomi, Y., “A fundamental framework for realizing semantic interoperability in a multidatabase environment,” Journal of Integrated Computer-Aided Engineering, Vol.2, No.1(Special Issue on Multidatabase and Interoperable Systems), pp.3-20, John Wiley ¥& Sons, Jan. 1995. [7] Kiyoki, Y., Kitagawa, T. and Miyahara, T., “A fast algorithm of semantic associative search for databases and knowledge bases,” Information Modelling and Knowledge Bases (IOS Press), Vol. VII, 4.1-4.16, 1995. [8] Kiyoki, Y., Kitagawa, T. and Kurata, K., “An Adaptive Learning Mechanism for Semantic Associative Search in Databases and Knowledge Bases,” Information Modelling and Knowledge Bases (IOS Press), Vol. VIII, May 1996. [9] Kiyoki, Y., Kitagawa, T. and Hayama, T., “A metadatabase system for semantic image search by a mathematical model of meaning,” ACM SIGMOD Record, Vol.23, No. 4, pp.34-41, Dec. 1994. [10] Kolodner, J.L., “Retrieval and organizational strategies in conceptual memory: a computer model,” Lawrence Erlbaum Associates, 1984. [11] Krikelis, A., Weems C.C., “Associative processing and processors,” IEEE Computer, Vol.27, No. 11, pp.12-17, Nov. 1994. [12] Potter J.L., “Associative Computing,” Frontiers of Computer Science Series, Plenumn, 1992. [13] Sheth, A. and Larson, J.A., “Federated database systems for managing distributed, heterogeneous, and autonomous databases,” ACM Computing Surveys, Vol.22, No.3, pp.183-236, 1990. [14] Kiyoki, Y. and Kitagawa, T., “A semantic associative search method for knowledge acquisition,” Information Modelling and Knowledge Bases (IOS Press), Vol. VI, pp.121-130, 1995. [15] “Stedman's Electronic Medical Dictionary VERSION 5.0,” Lippincott Williams and Wilkins, A Wolters Kluwer Company, 2000 [16] “Fifteenth Edition Harrison's Principles of Internal Medicine CD-ROM VERSION 1.0,” McGraw-Hill, 2001 [17] “Longman Dictionary of Contemporary English,” Longman, 1987. [18] Ogden, C.K., “The General Basic English Dictionary,” Evans Brothers Limited, 1940.


68

Theory and Implementation on Automatic Adaptive Metadata Generation for Image Retrieval Hideyasu Sasaki Ritsumeikan University, 1-1-1, Noji-higashi, Kusatsu, Shiga, 525-8577 Japan, [email protected]

Yasushi Kiyoki Keio University, 5322 Endo Fujisawa, Kanagawa, 252-8520 Japan [email protected] Abstract. In this paper, we present the detailed theory and implementation on an automatic adaptive metadata generation system using content analysis of sample images with a variety of experimental results. Instead of costly humancreated metadata, our method ranks sample images by distance computation on their structural similarity to query images, and automatically generates metadata as textual labels that represent geometric structural properties of the most similar sample images to the query images. First, our system screens out improper query images for metadata generation by using CBIR that computes structural similarity between sample images and query images. We have realized automatic selection of proper threshold-values in the screening module. Second, the system generates metadata by selecting sample indexes attached to the sample images that are structurally similar to query images. Third, the system detects improper metadata and re-generates proper metadata by identifying wrongly selected metadata. Our system has improved metadata generation by 23.5% on recall ratio and 37% on fallout ratio rather than just using the results of content analysis even with more practical experimental figures. Our system has its extensibility to the various types of specific object domains with inclusion of computer vision techniques.

1 Introduction Automatic indexing for metadata generation is indispensable to keyword-based image retrieval that is the more precise and speedy than content-based image retrieval (CBIR) in networked multimedia databases. However, the content-based descriptors obtained from images are usable to automatically generate metadata [1]. Content analysis employed in CBIR [2, 3, 4] is to provide a desirable solution to that automatic metadata generation. Metadata generation using content analysis has three issues. First, query images are not always proper for the structural similarity computation by content analysis because images are often distorted or noisy. Second, content-based descriptors are binary indexes that are not proper for keyword-based retrieval that demands textual indexes attached to images. Third, the direct use of the result of content analysis is not satisfactory for metadata generation, because those obtained descriptors are often conflicting in classification of object images. The goal of this paper is to design a system for metadata generation that mitigates those three issues by using content analysis and databases of sample indexes. In Section 2, we formulate our system that consists of three modules. The first module screens out improper query images for metadata generation by using CBIR that computes

H. Sasaki and Y. Kiyoki / Theory and Implementation on Automatic Adaptive Metadata Generation

69

structural similarity between sample images and query images. This module is open to inclusion of additional processes where the screened-out query images are to be improved by image processing techniques available in computer vision. We have realized automatic selection of proper threshold-values in the screening module [5]. We present the detailed theory and implementation on the automatic thresholding mechanism. The second module generates metadata by selecting sample indexes attached to sample images that are structurally similar to query images. The selection of sample indexes as metadata candidates is optimized by thresholding the results of content analysis of sample images and query images [6]. The third module detects improper metadata and re-generates proper metadata by identifying common and distinctive indexes in wrongly selected metadata [7]. In Section 3, we provide mathematical formulation to each process in the modules of our system. In Section 4, we present its cost-efficient implementation using a typical package solution of CBIR with databases. In Section 5, we evaluate the performance of our system using CBIR and databases of query and sample images that are basic geometric figures. We have used more various types of practical figures for the experiments of this research. Our system has generated proper metadata to the images of basic figures at 61% recall ratio and 20% fallout ratio in overall. This result has been bettered off by 23.5% on recall ratio and 37% on fallout ratio than just using content analysis for metadata generation. In Section 6, we discuss the experimental results and conclude this paper with our future work. Our system shows its feasibility with well-selected sample images and indexes that represent typical structural properties of object images in various types of specific domains. Our system has its extensibility to the various types of specific object domains with inclusion of computer vision techniques. Our system is to bridge the contributions in computer vision and database researches, here content analysis and indexing technique, as an automatic solution for precise and speedy image retrieval in networked multimedia databases (Fig. 1). COMPUTER VISION IMAGES

IMAGE PROCESSING

CONTENT ANALYSIS

DATABASE SYSTEM

INDEXES METADATA GENERATION

EMPOWERED CONTENT ANALYSIS E.G. PATTERN RECOGNITION

INDEXING TECHNIQUE

SMART INDEXING FOR KEYWORDBASED RETRIEVAL

Figure 1: Metadata generation bridges computer vision and databases.

2 Methodology for Bridging Content Analysis and Databases Fig. 2 outlines our system that comprises two preprocesses for collection of sample images and sample indexes and distance computation, and three modules for screening, metadata generation and expansive indexing. Our system automatically assigns query images with combinations of textual indexes and numerical indicators as metadata by using content-based descriptors as the results of distance computation on structural similarity between the query

70


IMAGE DATABASE WITH SAMPLE INDEXES

DISTANCE COMPUTATION

EXPANSIVE INDEXING

MODULE

DEFAULT INDEXING

METADATA

INDEXING

MAXIMUM THRESHOLDING

GENERATION

FEATURE EXTRACTION

NORMALIZATION

MODULE

SAMPLE INDEX DATABASE SAMPLE INDEXES { EQUILATERAL; TRIANGLE; ISOSCELES; ANGLE 60 DEGREE }

DB QUERY IMAGES

MINIMUM THRESHOLDING

SCREEN

MEASURE CONVERSION

MODULE

DISTANCE VALUES + SAMPLE INDEX : 0.002 * { TRIANGLE ; ANGLE 60 }

NEW METADATA : 1* { EQUILATERAL; ; ISOSCELES ; TRIANGLE }

Figure 2: Metadata generation architecture.

images and the sample images with sample indexes. The sample indexes represent the structural properties of the sample images stored in databases. Numerical indicators represent the proportion of structural properties of the query images in comparison with the sample images. (1) Preprocess: Collection of Sample Images & Sample Indexes This preprocess is to specify each object domain with a set of sample images for metadata generation. This preprocess as manual preprocessing involves four steps: Select sample images as represent typical types of structural properties that are found in the specified domain, e.g., typical shapes of architectures like water pools and baseball parks in urban surroundings; Select a set of well-defined sample indexes as examples that describe structural properties on selected sample images, e.g., { Triangle, 70◦ ; Isosceles ; . . . } to a top roof of a certain type of a baseball park; Store sample images into an image database and sample indexes that correspond to the stored sample images into another sample index database, and query images requested for metadata generation to the other image database, respectively; and, Combine sample indexes with sample images. Example A sample image of right angle triangle has a set of sample indexes: { Triangle ; Angle 90◦ ; Isosceles }. A sample image of obtuse isosceles triangle has a set of sample indexes: { Obtuse ; Isosceles ; Triangle ; Angle 150◦ }. (2) Preprocess: Distance Computation This preprocess is to compare the query images with the sample images. This automatic preprocessing applies feature extraction and distance computation functions of content analysis to the query images and the sample images in three steps: Store the sample images and the query images into a CBIR system; Fire distance computation functions in the CBIR system; and, Get results of structural distance computation from the CBIR system. In this preprocess, a CBIR system is to be selected from several CBIR systems. That selected system is to provide query images with sets of values that represent structural distance of the query images from the sample images. Those returned values represent certain proportion of structural properties that the query images contain against what the sample images represent in the form


71

Table 1: Sample Data Processing (e.g. Query Image: Right Angle Triangle) SAMPLE IMAGES

SAMPLE INDEXES

DISTANCE VALUES

A/F NORMALIZATION

RIGHT ANGLE TRIANGLE

TRIANGLE ; ANGLE 90◦ ; ISOSCELES TRIANGLE ; OBTUSE ; ANGLE 150◦ ; ISOSCELES RHOMBUS ; QUADRILATERAL ; QUADRANGLE ; POLYGON ELLIPSE TRIANGLE ; ACUTE ; ANGLE 30◦ ; ISOSCELES TRAPEZOID ; QUADRANGLE CIRCLE TRIANGLE ; EQUILATERAL ; ANGLE 60◦ ; ISOSCELES RECTANGLE ; QUADRANGLE ; POLYGON SQUARE ; RECTANGLE ; QUADRANGLE ; POLYGON ; QUADRILATERAL

0.00

0.50

0.00

0.50

0.000160

0.00

0.00256 0.244

0.00 0.00

0.256 0.310 0.877

0.00 0.00 0.00

0.974

0.00

3.603

0.00

OBTUSE ISOSCELES TRIANGLE

RHOMBUS

ELLIPSE ACUTE ISOSCELES TRIANGLE

TRAPEZOID CIRCLE EQUILATERAL TRIANGLE

RECTANGLE

SQUARE

of the sample indexes. Example Table 1 describes that a query image of right angle triangle has the same number of structural distance value with “Right Angle Triangle” and “Obtuse Isosceles Triangle” as 0.00, which means the selected CBIR system evaluates those images as the least distant images to each other. Meanwhile, the larger numbers correspond to the more distant images.

(3) Screen Module This module consists of two processes: measure conversion and minimum thresholding. The measure conversion process is to convert the results of distance computation into more intuitive values that represent structural similarity between the query and sample images. The values on structural distance take 0 to the least distant and most similar images, while the larger numbers correspond to the more distant and less similar images. This process converts the current unintuitive measure on distance computation to another intuitive measure on similarity which evaluates the least distant and most similar images as 1 and the largest distant images as 0. After measure conversion, structural distance values of 0.00 are converted to 1.00 when the query images coincide with certain sample images. The minimum thresholding process is to cut off “noise” information after the measure conversion process. In collection of sample images, prepared images should represent typical classes of structural difference in object domains. Suppose that a query image is evaluated to be equally similar to the more-than-half of sample images in a certain result of measure conversion. This case shows that the query image is not proper for metadata generation because prepared sample images cannot identify any pattern or class of the query image. The minimum thresholding process sets a threshold that is defined as the reciprocal number of the half number of sample images. When the number of sample images is odd, it should be incremented by one. The screen module cuts off noisy query images for further improvement by inclusion of image processing techniques like morphological analysis including thinning, image segmentation of noisy objects, etc.. (4) Metadata Generation Module This module consists of three processes: normalization, maximum thresholding and default

72


indexing. The normalization process is to normalize the results after the screen module as they constitute numerical indicators of proportional structural properties that the sample indexes represent. The maximum thresholding process is to emphasize the values that represent meaningful structural properties of the query images in the form of numerical indicators to the sample indexes. Some values after the normalization still correspond to mutually similar numbers that represent insignificant structural properties of sample images. This process narrows down the range of the values after the normalization process by cutting off the values of lower size than a certain threshold that is the maximum threshold, e.g., 0.95, 0.40, etc., and simply converts those values over the maximum threshold to 1. In addition, when some values are converted to 1, this process evaluates all the other values as “noise” and converts them to 0. The default indexing process is to index query images with combinations of textual indexes and numerical indicators as metadata. This process combines the values after the maximum thresholding process as numerical indicators of proportional structural properties with the sample indexes that represent structural properties, and then constitutes newly generated metadata while those metadata appropriately represent structural properties of the query images. Meanwhile, the default indexing process can generate conflicting indexes that represent several different types of sample images. In that case, a single query image can correspond to different structural types of sample indexes, when several normalized values take the same largest size. The default indexing process simply discards those conflicting indexes from metadata candidates, and does not generate any proper metadata. (5) Expansive Indexing Module This module is to generate new metadata that narrow down the scope of identification of a query image within at most two types of structurally similar images, when that distance computation evaluates a single query image as similar to several different types of sample images. This module involves two steps. The first step extracts “common indexes” within conflicting indexes of all sample images that default indexing inappropriately evaluates structurally “similar” to query images. Common indexes constitute a major class that includes a subclass of the discussed sample images and other subclasses of sample images that include the common indexes but are not evaluated as similar to the query images. The second step extracts “distinctive indexes” from sample indexes in the latter other subclasses of sample images that include the common indexes but are not evaluated as similar to the query images. In the major class, distinctive indexes classify the query images into the former subclass of sample images, and differentiate those query images from the latter other subclasses of sample images. Example Fig. 3 outlines that a query image of obtuse isosceles triangle is evaluated as “similar” to two sample images: right angle triangle and obtuse isosceles triangle. The two sample images have 0.00 as structural distance values, 1.00 as the values after measure conversion, 0.50 as normalized values to the single query image, respectively. First, this process extracts common indexes from the indexes tentatively generated in the default indexing. Those common indexes represent greatest common measure of generated indexes that correspond to the same largest size of normalized values to the single query image of obtuse isosceles triangle as “TRIANGLE” and “ISOSCELES”. Other sample images, which have normalized values lower than the largest normalized values, contain those common indexes. Sample images of equilateral triangle and acute isosceles triangle contain those two common indexes, here, “TRIANGLE” and “ISOSCELES” within their sample indexes. Those two indexes constitute a group of “ISOSCELES TRIANGLE” that represents a major class including right angle

H. Sasaki and Y. Kiyoki / Theory and Implementation on Automatic Adaptive Metadata Generation COLLECTION OF IMAGES & INDEXES

DISTANCE COMPUTATION BY CONTENT ANALYSIS DISTANCE VALUE: 0.000001 ≈ 0.00

QUERY IMAGE 1: EQUILATERAL TRIANGLE

SAMPLE INDEXES { EQUILATERAL; TRIANGLE; ISOSCELES; ANGLE 60 DEGREE } DISTANCE VALUE: 0.03

QUERY IMAGE 2: OBTUSE ISOSCELES TRIANGLE SAMPLE IMAGES WITH SAMPLE TEXTUAL INDEXES

SAMPLE INDEXES {TRIANGLE; ISOSCELES; ANGLE 90 DEGREE } DISTANCE VALUE: 0.03

SAMPLE INDEXES { OBTUSE; TRIANGLE; ISOSCELES; ANGLE 150 DEGREE }

SAMPLE INDEXES { EQUILATERAL; TRIANGLE; ISOSCELES; ANGLE 60 DEGREE }

73

MAIN PROCESSES FOR INDEXING

DEFAULT INDEXING NEW METADATA 1 * { EQUILATERAL; TRIANGLE; ISOSCELES; ANGLE 60 DEGREE }

EXPANSIVE INDEXING

NEW METADATA 1 * {TRIANGLE; ISOSCELES; NOT_(ACUTE) ; NOT_(ANGLE 30 DEGREE) ; NOT _(ANGLE 60 DEGREE) }

DISTANCE VALUE: 0.24 SAMPLE INDEXES { ACUTE; TRIANGLE; ISOSCELES; ANGLE 30 DEGREE }

Figure 3: Exemplary default and expansive indexing.

triangle and obtuse isosceles triangle with equilateral triangle and acute isosceles triangle. Second, this process finds distinctive factors in addition to common indexes, as outlined in Fig. 4. Those two sample images with the lower normalized values also have other sample indexes which are different from the common indexes: “ACUTE”, “EQUILATERAL”, “ ANGLE 30◦ ”, and “ANGLE 60◦ ”. Those different sample indexes work as distinctive factors that constitute a group to represent a subclass within the major class. Distinctive factors are added with negative sign, e.g., “NOT” to constitute distinctive indexes against common indexes: { NOT ANGLE 30◦ ; NOT EQUILATERAL ; NOT ACUTE ; NOT ANGLE 60◦ }. In the expansive indexing, a combination of common indexes and distinctive indexes constitutes a set of new metadata that represents a subclass of right angle and obtuse isosceles triangles, which is different from the other subclass of acute isosceles and equilateral triangles within the major class of isosceles triangles. 3 Mathematical Formulation In this section, we mathematically formulate the processes of our system. (1) Preprocess: Collection of Sample Images & Sample Indexes This preprocess is to collect sample images and sample indexes in the following four steps: collections of query images, sample images and sample indexes, and combinations of sample indexes with sample images. Let define related sets of query and sample images, and sample indexes: Definition 1. (a set of query images) As ı˙ is a continuous natural number, then {I ı˙ } is defined

74


QUERY IMAGE: OBTUSE ISOSCELES TRIANGLE MAJ OR CLASS SAMPLE IMAGES {OBTUSE ; ISOSCELES ; TRIANGLE ; ANGLE 150O }

{ISOSCELES ; TRIANGLE } SUBCLASS SAMPLE IMAGES

{NOT NOT_ NOT EQUILATERAL ; NOT_ NOT ANGLE 60O ; NOT_ NOT ACUTE ; NOT_ NOT ANGLE 30O }

{RIGHT ANGLE ; TRIANGLE ; ISOSCELES ; ANGLE 90O }

{ EQUILATERAL ; TRIANGLE ; ISOSCELES ; ANGLE 60O }

{ ACUTE ; ISOSCELES ; TRIANGLE ; ANGLE 30O }

Figure 4: Exemplary major class and subclass categorization.

as a set of query images: {Iı˙ } := {y|y = ı˙, ı˙ ∈ N }. Definition 2. (a set of sample images) As j˙ is a continuous natural number, then { Iˆj˙} is defined as a set of sample images: {Iˆj˙} := {x|x = j, ˙ j˙ ∈ N }. Definition 3. (a set of sample indexes) As j˙ and k are continuous natural numbers, respecˆ j˙ } is defined as a set of sample indexes: tively, then {W k ˆ j˙ } := {z|z = {W {j˙ × { k}}, j˙ ∈ N , k ∈ N }. k j∈N

Here,

k∈N

represents production operation (same as below).

Let define combinations of sample indexes with sample images: Definition 4. (combinations of sample indexes with sample images) ˆ j˙ × Iˆj˙}}. {Iˆj˙Wˆ } := { {W j˙

j ∈N

(2) Preprocess: Distance Computation Let define a set of structural distance values and a function of distance computation process as follows: Definition 5. (a set of structural distance values) j}}}}. ˙ {Dı˙j˙ } := {d|d = dı˙j˙ ∈ R : {˙ıj˙} = { {˙ı × { ı˙∈N

j∈N ˙

Definition 6. (a function of distance computation between query images and sample images with sample indexes) − {Dı˙j˙ }. φ : { {Iı˙ × { {Iˆj˙Wˆ }}}} → ı˙∈N

j∈N ˙

j˙

Here, → − represents function operation. (3-a) Screen Module: Measure Conversion Process


75

Let define values after the measure conversion by using the first normalization of reciprocal structural distance values, as follows: Definition 7. (values after the measure conversion) ⎧ ⎪ (whenDı˙j˙ = 0) ⎪ ⎪1 ⎪ ⎪ 1 ⎨ Dı˙j˙ Sı˙j˙ := ⎪ (otherwise). ⎪ ⎪ 1 ⎪ ⎪ ⎩ j∈N D ı˙j˙ Here, represents summation operation. Also, subject to the following computation rules: ∞ 1 0 := ∞; := 1; := 1. 0 ∞ 0

Here, the distance computation is converted into an intuitive measure on similarity. (3-b) Screen Module: Minimum Thresholding Process Let define values after the minimum thresholding the results of measure conversion with a certain selected threshold value t: Definition 8. (values after the minimum thresholding) ⎧ 1 ⎨0 (when ∃ Sı˙j˙ < t = [(#{Iˆj˙} + 1)/2] S¯ı˙j˙ := ⎩ Sı˙j˙ (otherwise). Here, [] represents Gauss’ symbol. This thresholding works as automatic noise cut-off mechanism. (4-a) Metadata Generation Module: Normalization Process Let define values after normalizing the results of the screen module: Definition 9. (normalized values) ||S¯ı˙j˙ || :=

S¯ı˙j˙

j∈N

. S¯ı˙j˙

Here, all the values are normalized to represent numerical indicators of proportional structural properties. (4-b) Metadata Generation Module: Maximum Thresholding Process Let define values after the maximum thresholding of the results of normalization with a certain selected threshold value T : Definition 10. (values after the maximum thresholding) 0 (when ∃ ||S¯ı˙j˙ || < T ; 0 < T ≤ 1) ˆ ||Sı˙j˙ || := 1 (otherwise). Here, maximum thresholding works as emphasizes the values that represent meaningful structural properties.

76


(4-c) Metadata Generation Module: Default Indexing Process Let define combinations of the normalized values after the maximum thresholding with sample indexes in the default indexing: Definition 11. (combinations of the values after the maximum thresholding with sample indexes in the default indexing) ˆ j˙ }}}}}. { {||Sˆı˙j˙ ||} × { { {W {Lı˙j˙k } := { k ı˙ ∈ N

j˙ ∈ N

j˙ ∈ N k ∈ N

Let define new metadata generated in the default indexing: Definition 12. (new metadata generated in the default indexing) ˙ such that ||Sˆı˙j˙ || = 1 (when j∈N {||Sˆı˙j˙ ||} = 1) Lı˙j˙k for ∃ j, Mı˙ := (for ∃ ı˙). 0 (otherwise) (5) Expansive Indexing Module Let define common indexes generated in the expansive indexing: Definition 13 (common indexes) ˙ such that ||Sˆı˙j˙ || = 1 (when j∈N {||Sˆı˙j˙ ||} ≥ 2) ∩j˙{Lı˙j˙k } for ∀ j, c Lı˙ := (for ∃ ı˙). 0 (otherwise) Here, ∩ represents cap product operation. Let define distinctive indexes in the expansive indexing: Definition 14. (distinctive indexes) ⎧ ∀ c ⎪ ˙ such that{{Lı˙j˙k } ∩ {Lıc˙ }} = ∅ ⎨{{∪j˙{Lı˙j˙k }} \ {Lı˙ }} for j, d (for ∃ ı˙). Lı˙ := (when{Lıc˙ } = ∅) ⎪ ⎩ 0 (otherwise) Here ∪ represents cup product operation. The bar over a set represents not complementary set but special operation as each element in the set has an index, “NOT”, as negative sign. Let define a set of newly generated metadata in the expansive indexing by combining common indexes and distinctive indexes with metadata generated in the default indexing: Definition 15. (a set of new metadata generated in the expansive indexing) ˆ ı˙ } := {{Mı˙ } ∪ {Lıc˙ } ∪ {Lıd˙ }}. {M

4 System Implementation We have implemented our system based on a typical package solution of CBIR that is QBIC in the DB2. [8, 9]. We have focused just on texture in its implementation. The entire system is implemented in a local host machine (CPU 600MHz; Memory 128 MB). The entire algorithm is programmed in Java and PHP. Fig. 5 outlines the entire data processing. Images and indexes are transferred from sample and query image databases, and a sample index database to QBIC by booting Apache in the local host machine. In DB2, images are transferred via JDBC to QBIC. After cataloging the


CBIR

JDBC

77

TEXTUAL INDEXES AS NEW METADATA

TRIANGLE ; NOT ACUTE ... PHP

CONTENT ANALYSIS INDEXING JAVA

QUERY IMAGE Figure 5: Implementation and data processing.

images stored in QBIC, each path-linkage locates the images, and QBIC defines their data structure. QBIC computes and returns structural distance values between query images and sample images. Those returned values are normalized and combined with sample indexes to generate new metadata by our algorithms programmed in PHP. The results are delivered as outputs in HTML format. 5 Experiments The implemented system has generated proper metadata to the total 4, 420 query images of basic figures at 61% recall ratio and 20% fallout ratio in 9, 821 milliseconds per image in gross average: 8, 762 for indexing and 1, 059 for distance computation. 884 original images were repeatedly used to generate query images for a series of several experiments. This performance shows that our system has improved metadata generation by 23.5% on recall ratio and 37% on fallout ratio rather than just using content analysis. In the experiments, we first evaluate the bottom line performance of the metadata generation system using just content analysis available in CBIR as a benchmark. As experimental objects, we have prepared ten types of basic geometric figures for generating a set of sample images, e.g., triangle, rectangle, ellipse, etc.. However, we have used more various types of practical figures for the experiments of this research. Query images are automatically generated by translating or affine-transporting, scaling by three times or thickening, or adding sparseness to those ten types of basic geometric figures, as displayed in Fig. 6. Geometric figures are proper to evaluate the direct identification of object images for metadata generation in the case of just using CBIR, and simple sample indexes work as objective measures in its bottom line performance evaluation of proper indexing. We then evaluate whether metadata generation is improved by the formulated three modules within reasonable elapsed time.

78


THICKEN SAMPLE IMAGES

TRANSLATED QUERY IMAGES

SPARSE SAMPLE IMAGES

SCALED BY THREE TIMES SAMPLE IMAGES ORIGINAL SAMPLE IMAGES

AFFINE TRANSPORTED SAMPLE IMAGES

AFFINE TRANSPORTED AFFINE TRANSPORTED THICKEN SAMPLE IMAGES SPARSE SAMPLE IMAGES

Figure 6: Query and sample images.

5.1 Performance Evaluation We have measured how many new metadata follow mathematical definitions of query images in five cases: without the screen module (a benchmark as the bottom line performance), applying the screen module, the metadata generation module with the maximum threshold 0.40, the same module with the maximum threshold 0.95 and the expansive indexing module, respectively. The experiments have been processed to query images in the form of bitmap and jpeg formats, respectively. Table 2 describes the results. Query images are provided with their proper metadata when they are indexed with the metadata that follow mathematical definitions of geometric figures represented in the query images. Here, performance is evaluated in recall and fallout ratios. Let define the recall ratio: Definition 16. (recall ratio) ⎧ ⎨0 Recall Ratio := # query images w/ proper metadata ⎩ # all the query images Let define the fallout ratio: Definition 17. (fallout ratio) ⎧ ⎨0 Fallout Ratio := # query images w/ improper metadata ⎩ # query images w/ metadata

(when no metadata generated) (otherwise).

(when no metadata generated) (otherwise).

5.2 Improvement by Screen Module The direct use of the results of content analysis does not allow proper metadata generation even in exploiting databases of sample images. The screen module cuts off “noise” information after the measure conversion process, and improves the fallout ratio of metadata generation results with its minimum thresholding. Fig. 7 describes its typical improved case of scaled images in jpeg format. In overall, the screen module has improved the average fallout ratio by 13.5% from 62% to 48.5%.


79

Table 2: Performance Evaluation ( Improved/Deteriorated) Image Format / Recall & Fallout

b/f Screen (benchmark as bottom line)

a/f Screen

Metadata Generation (Threshold 0.40)

Metadata Generation (Threshold 0.95)

Expansive Indexing

BMP Recall Ratio BMP Fallout Ratio

0.44 0.56

0.44 0.36

0.44 0.36

0.41 0.12

0.61 0.12

JPEG Recall Ratio JPEG Fallout Ratio

0.32

0.32

0.32

0.23

0.61

0.68

0.61

0.60

0.44

0.38

Figure 7: Improvement by screen module.

5.3 Improvement by Maximum Thresholding Even after the screen module, the normalized values do not always properly represent meaningful structural properties of the query images. The maximum thresholding in the metadata generation module emphasizes the only meaningful values. In the applications of the lower threshold 0.40, the average fallout ratio on the images in jpeg format is improved by only 1%. In the applications of the higher threshold 0.95, the average fallout ratios on the images in bitmap and jpeg formats are improved by 24% and 17% compared with those ratios after the screen module. Meanwhile, the recall ratios on the same two types of images are deteriorated by 3% and 9%, respectively. Fig. 8 describes its typical improved case of translated images in bitmap format without any deterioration on its average recall ratio. In overall, the maximum thresholding process has improved the average fallout ratios in the case of threshold 0.40 by 0.5% from 48.5% to 48%, and in the case of threshold 0.95 by 20.5% from 48.5% to 28%, respectively. Meanwhile, the recall ratio of the latter case is deteriorated by 6% from 38% to 32%. 5.4 Improvement by Expansive Indexing In overall, the expansive indexing module has improved the average recall and fallout ratios by 23% from 38% to 61%, and 23% from 48% to 25%, after the maximum thresholding

80


Figure 8: Improvement by maximum thresholding process.

process of the threshold 0.40 instead of 0.95 in the metadata generation module, respectively. Fig. 9 describes its typical improved cases of the translated images and the affine-transported sparse images in bitmap format. 5.5 Elapsed Time of Textual Indexing We have measured elapsed time of default and expansive indexing, respectively. Here, elapsed time is defined as difference between a starter time and its closing time of indexing process in millisecond. This experiment has been processed to translated and affine-transported query images in bitmap and jpeg formats, respectively. Table 3 describes the results. Table 3: Elapsed Time of Indexing Process Format (Default/Expansive Indexing)

Elapsed Time of Indexing to Translated Images (Millisecond)

Elapsed Time of Indexing to Affine-transported Images

Elapsed Time of Indexing to Scaled Images (Millisecond)

BMP (Default) BMP (Expansive)

613 6,441

11,748 5,524

1,235 1,235

JPG (Default) JPG (Expansive)

8,154 8,156

15,040 15,040

9.305 9.305

5.6 Elapsed Time of Distance Computation The proposed method employs distance computation mechanisms used in CBIR to compute similarity between sample and query images. We have measured elapsed time of distance computation of query images in bitmap and jpeg formats, respectively. Elapsed time of distance computation is defined as difference between a starter time and its closing time of distance computation as preprocessing in millisecond. Table 4 describes gross average elapsed time for each distance computation of query images in bitmap and jpeg formats, respectively.


81

Figure 9: Improvement by expansive indexing.

6 Discussion and Conclusion In this paper, we have presented an automatic metadata generation system using content analysis available in CBIR and databases of sample images. We have confirmed the bottom line performance for metadata generation using just content analysis as a benchmark that is shown in the result before the application of the screen module as only 38% recall ratio and 62% fallout ratio in overall. First, our system has screened out improper query images for metadata generation, and improved the average fallout ratio by 13.5% in overall. We have discovered and implemented an automatic solution for the selection of threshold-values in the screening module. Second, our system has its extensibility to bridge database technology with content analysis in the field of metadata generation for keyword-based retrieval, and improved the overall performance by 23% on the recall and fallout ratios, respectively, with more practical types of experimental figures. The experimental results assure that our system is applicable in various types of

82


Table 4: Elapsed Time of Distance Computation Case of Image Computation to Translated Images (BMP) Computation to Affine-Transported Images (BMP) Computation to Scaled Images (BMP) Computation to Translated Images (JPEG) Computation to Affine-Transported Images (JPEG) Computation to Scaled Images (JPEG)

Gross Average (millisecond) 1,066 1,003 4,068 1,072 964 1,036

specific domains with well-selected sample images that represent typical structural properties of images in the object domains. The combination of multimedia database technology for precise and speedy keywordbased image retrieval and empowered content analysis in computer vision is to realize a simple implementation of content-based metadata indexing. Our future work is to integrate modules for improving the screened-out query images by applying techniques available in computer vision. Acknowledgements This research work was supported in part by the Grant by Moritani Scholarship Foundation. References [1] P.J. Moreno, J.M. van Thong, E. Logan, and G. Jones, From Multimedia Retrieval to Knowledge Management, IEEE Computer 35(4) (2002) 58–66. [2] A.W.M. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, Content-based Image Retrieval at the End of the Early Years, IEEE Trans. on Pattern Analysis and Machine Intelligence 22(12) (2000) 1349–1380. [3] Y. Rui, T.S. Huang, and S.F. Chang, Image Retrieval: Current Techniques, Promising Directions and Open Issues, Journal of Visual Communication and Image Representation 10(4) (1999) 39–62. [4] A. Yoshitaka and T. Ichikawa, A Survey on Content-based Retrieval for Multimedia Databases, IEEE Trans. on Knowledge and Data Engineering 11(1) (1999) 81–93. [5] H. Sasaki and Y. Kiyoki, Automatic Adaptive Metadata Generation for Image Retrieval. In: Proc. of the 2005 International Symposium on Applications and the Internet - Workshops (SAINT 2005 Workshops), IEEE Computer Society Press. Trento, Italy, January 31– February 4, 2005, pp. 426–429. [6] H. Sasaki and Y. Kiyoki, A Prototype Implementation of Metadata Generation for Image Retrieval. In: Proc. of the 2004 International Symposium on Applications and the Internet - Workshops (SAINT 2004 Workshops), IEEE Computer Society Press. Tokyo, Japan, March 26–27, 2004, pp.460–466. [7] H. Sasaki and Y. Kiyoki, A Prototype Implementation of Adaptive Metadata Generation to Digital Images, Information Modelling and Knowledge Bases 16 (2005) 134–151. [8] M. Flickner, H. Sawhney, W. Niblack, J. Ashley, Q. Huang, B. Dom, M. Gorkani, J.Hafner, D. Lee, D. Petkovic, D. Steele, and P. Yanker, Query by Image and Video Content: The QBIC System, IEEE Computer 28(9) (1995) 23–32. [9] J. Ashley, M. Flickner, J. Hafner, D. Lee, W. Niblack, and D. Petkovic, The Query by Image Content (QBIC) System, ACM SIGMOD Record 36(7) (1995) 475.


83

A Meme Media Framework for Introducing a Multimodal User Interface to Existing Web Applications Kimihito ITO Meme Media Laboratory, Hokkaido University, Sapporo 060-8628, Japan Abstract. Web applications, which are computer programs ported to the Web, allow us to use various remote services and tools through our Web browsers. There are an enormous number of Web applications on the Web, and they are becoming the basic infrastructure of everyday life. At the same time, multimodal character agents, which interact with human users by both verbal and nonverbal behavior, have recently seen remarkable development. It would be of great benefit if we could easily modify existing Web applications by adding multimodal user interface to them. This paper proposes a framework where IntelligentPad and Multimodal Presentation Markup Language work in collaboration to introduce multimodal character agentsto the front-end of existing Web applications. Example applications include attaching a multimodal user-interface to a news site on the Web. The framework does not require users to write any program code or script.

1

Introduction

For many years, researchers have tried to create new human-computer interfaces to make computers more attractive. Innovations in Web technologies allow end-users to easily use various Web applications through their browsers. A Web application is a program that has an HTML-based front-end for users to utilize some services provided by a remote HTTP server. Many companies and researchers provide Web applications, such as search engines, financial services, real-time news services, scientific analysis tools, and various other kinds of database services. Generally, these services are provided in the form of documents. The use of multimodal character agents is a major topic of interest in the Web development community [1]. However, it is difficult for end-users to introduce multimodal character agents to existing Web applications. For instance, a user may occasionally need character agents to read the latest news published on a website. There is currently no framework that allows end-users to instruct agents to perform a presentation that changes dynamically according to the information on the Web. Character agents [15] converse with human users through verbal and nonverbal behavior. Multimodal character agents have recently seen considerable development, and some of those basic components are built into the Microsoft Windows operating system [13]. Despite their fundamental potential to present information to users, multimodal character agents are not used in many applications.In fact, there are two problems that need to be solved. • The first problem is the lack of a mark-up language to support the authoring of multi-modal presentations in a reasonable amount of time.

84

K. Ito / A Meme Media Framework for Introducing a Multimodal User Interface

Figure 1: One of the major goals of this research is to develop a framework where even end-users can introduce a multimodal user-interface to existing Web applications.

• The second is the lack of a framework to change the user-interface of one piece of software into others. Most application programs are not designed to be used with another user-interface. This paper proposes a new framework where a meme media system IntelligentPad [18] [19] and Multimodal Presentation Markup Language [9] work in tandem to introduce a multimodal user interface into existing Web applications. Recently, Mitsuru Ishizuka’s group at the University of Tokyo designed a markup language called MPML (Multimodal Presentation Markup Language), and succeeded in solving the first problem. IntelligentPad architecture, which was proposed by Yuzuru Tanaka of Hokkaido University in 1989, provides a solution to the second problem. IntelligentPad enables application programs to be functionally linked to other programs. Using IntelligentPad, users may freely and easily change the user-interface of application programs (including Web applications) freely by connecting their data slot to a slot in another application. Aim of this Research This paper aims to develop a framework where even end-users can introduce multimodal user-interface with character agents into existing Web applications (Figure 1). Since there are so many Web applications on the Internet, it is tedious even for professional programmers to define the specialized behavior of characters for each major Web application. In this paper, we focus on a method by which users can instantaneously extend the interface of existing Web applications. A framework where end-users can easily introduce a multimodal user interface with character agents to existing Web applications requires the following capabilities: 1. Easy specification of input and output portions of Web applications to reuse their embedded functions.


85

2. Dynamic and automated generation of character agents’ behaviors according to the contents to be presented. Related Work Web Service technologies such as SOAP (Simple Object Access Protocol) [21] enable us to interoperate services published over the Web. However, they assume that the API (Application Program Interface) library to access such a service is a priori provided by its server side. Bauer and Dengler [4] have introduced a PBD (Programming by Demonstration) method in which even non-savvy users can configure their own Web-based information services to satisfy their individual information needs. They have implemented the method into InfoBeans [3]. By accessing an InfoBox with an ordinary Web browser, users can wrap Web applications. By connecting channels among InfoBeans on the InfoBox, users can also functionally integrate them together. InfoBeans does not allow users to combine wrapped Web applications and local legacy applications. The coordination of character agents and Web applications is beyond its scope. Since the manual scripting of character agent behaviors is very time-consuming, PPP Persona [16] and WebPersona [2] automate presentation design. These systems plan how to present given materials in a temporally coordinated manner. This paper is organized as follows. In the next section, we summarize our previous work for Web application wrapping based on IntelligentPad architecture. We introduce MPML, that is used to define multimodal presentations in Section 3. In Section 4, we propose a new framework where IntelligentPad and MPML work together to introduce a multimodal user interface to existing Web applications. In Section 5, we conclude with our future research plans. 2

Meme Media Architecture and its Application to the Web

2.1 A Meme Media System IntelligentPad Meme media technologies[19] were proposed to enable people to reedit and redistribute intellectual resources, and to accelerate the evolution of intellectual resources accumulated over the Web. The word “Meme Media” was coined by Y.Tanaka and gets its name from the fact that mememedia carry what R. Dawkins called “memes” [7]. IntelligentPad[18] is a meme media system that allows users to combine media objects (called `‘pads”) through view integration. Media objects represented by pads vary widely. They range from multimedia documents to tools such as databases, charts, spreadsheets, and even Web applications. Each pad has slots functioning as data I/O ports. Through ‘drag-and-drop’ and ‘paste’ operations, users can connect one pad to a slot of another pad. This operation simultaneously creates both a composite view and a functional linkage through a slot connection. IntelligentPad represents each component as a pad, which is a sheet of paper on the screen. A pad can be pasted on another pad to define both a physical containment relationship and a functional linkage between them. When a pad P1 is pasted on another pad P2 , the pad P1 becomes a child of P2 , and P2 becomes the parent of P1 . No pad may have more than one parent pad. Pads can be pasted together to define various

86


multimedia documents and application tools. Unless otherwise specified, composite pads are always decomposable and re-editable.

Figure 2: User-Interface and Connection Interface. If a user pastes P1 on P2 with its connection to a slot of P2 , the primary slot of P1 is connected to the slot of P2

In Figure 2, we show an abstract architecture of IntelligentPad. For a more precise architecture of IntelligentPad see the book by Yuzuru Tanaka [19]. Each pad has both a standard user-interface and a standard connection interface. The user-interface of a pad has a card like view on the screen and a standard set of operations like ‘move’, ‘resize’, ‘copy’, ‘paste’, and ‘peel’. Users can easily replicate any pad, paste a pad onto another, and peel a pad off a composite pad. Pads are decomposable persistent objects. You can easily decompose any composite pad by simply peeling off a primitive or composite pad from its parent pad. As its connection interface, each pad provides a list of slots that work as connection jacks of an AV-system component. A pad also a pad provides a single connection to a slot of its parent pad. To set up data linkage between pads, IntelligentPad uses three standard messages, ‘set’, ‘gimme’ and ‘update’. We show an outline of these three messages in Table 1. Table 1: A summary of three standard messages. Message Summary a child sets the specified value to set slotname value its parent’s slot gimme slotname a child requests its parent to return the value of its specified slot update a parent notifies its children that some slot value has been changed

A pad can be embedded in one parent at most. Each pad can have at most one connection to a slot of the parent pad. Connected pads form a tree structure. We do not restrict the maximum depth of the tree. Each pad has one primary slot. When the value of the primary slot of a child is changed, the child sends a ‘set’ message with the new slot value to its parent. Using this value, the parent changes its own slot values. Then, the parent pad notifies all of its children pads of its state change by sending an ‘update’ message. Each child that has received an ‘update’ message sends a ‘gimme’ message to the parent pad, changes its own slot values using the return value of this ‘gimme’ message, and then sends an ‘update’ message to each of its children. Using this mechanism, state changes are propagated from one pad to all the pads connected to it (Figure 3).


87

Figure 3: Three standard messages, ‘set’, ‘gimme’ and ‘update’, between pads

2.2 Wrapping Existing Web Applications with Pads In a previous paper with Y. Tanaka, we proposed a framework that allows end-users to wrap an arbitrary Web application with a pad [11]. Users can visually wrap Web applications with visual components, and visually combine them together to define functional linkages among them. Users can also visually define functional linkages between wrapped Web applications and local tools in pad form such as chart drawing and spreadsheet tools to compose a single integrated tool pad. We define a Web application as an application program that has an HTML-based front-end for the user to utilize some services provided by a remote HTTP server. Here, we summarize the proposed architecture for reusing existing Web applications. Figure 4 shows the architecture. A WebApplicationWrapperPad (‘WAW-PAD’) consits of: 1) a reference to the target Web Application, and 2) references to the document portion to input and to output data to and from a Web application. The URL or NavigationPath [11] is used to realize a reliable reference to the target Web Application. Sometimes a URL is suitable to specify a Web page and sometimes not. Some Web pages may be fetched with a URL directly. In general, however, every page is not accessible with URLs. For example, Web pages that are returned as a result of POST are not specified by URLs. A NavigationPath is a sequence of user

Figure 4: Wrapping Existing Web applications to be Interoperable Components. The framework have been proposed in my previous paper with Y. Tanaka

88


Figure 5: Web application wrapping only through mouse operations

operations. Examples of user operations include opening a page specified by a URL, clicking an anchor, and submitting some data to a form. In other words, a NavigationPath represents how to access the Web page. An HTMLPath is used to realize a stable reference to the document portion to input and to output data to and from a Web application. An HTMLPath expression, which is a specialization of an XPath expression [20], identifies an HTML-node. For instance, we may specify an input form for a Web application with an HTMLPath HTML[1]/BODY[1]/FORM[1]/INPUT[1] The combination of a URL, a NaviagationPath and an HTMLPath allows us to define slots to input and to output data to and from a Web application. Figure 5 shows how users can wrap a Web application using our framework. Here, the user has a WAW-Pad that is a kind of Web browser. With WAW-Pads, we can browse Web pages in the same way as with an ordinary Web browser. Suppose the users access the Google engine page (1). For example, if he inputs the search string “Hypertext 2003” into this input-form, it returns a list of Web pages. Now, he starts to wrap Google. He could directly specify any HTML nodes on this page to work as slots. He defines two slots on this page, the first one for the keyword slot, the second for the first retrieved page (2). This is the wrapper of Google (3). Next, he pastes a TextPad with its connection to the keyword slot, and he pastes a BrowserPad with its connection to the ‘first retrieved page’-slot (4). These paste operations define connections between slots in these pads. Now he has finished creating a new integrated tool by reusing the Google service (5). If he inputs another keyword then he will see the page displaying the retrieved pages. He can change its layout as he likes. He doesn’t need any programming skills to create a wrapper or to reuse the functions embedded in the original page. 3

MPML

MPML (Multimodal Presentation Markup Language) was developed by Mitsuru Ishizuka in order to provide a medium-level description language commonly applicable to various multimodal character systems. MPML provides a simple scripting scheme for users to create a multimodal presentation in XML format [23].


89

Figure 6: User Interface with Connection Interface of MPMLPlayerPad

In a major implementation of MPML, Microsoft Agent [13] is used to give presentations defined in a MPML. The package includes 2D cartoon-style character animation, speech recognition, and a Text-to-Speech (TTS) engine. Character agents in Microsoft Agent are controlled through the use of the scripting language JavaScript. It is still difficult for non-programmers to write program code in JavaScript to control character agents. For further details about MPML, visit its website at the University of Tokyo [10]. 4

Collaboration of IntelligentPad and MPML

4.1 An Architecture allowing Web Applications to Communicate with Users through Multimodal Character Agents As I have shown in Section 2, Web applications can be instantaneously wrapped as IntelligentPad objects. Once a Web application is wrapped as a pad, it can be combined with other pads to compose a new integrated tool easily. For the collaboration of IntelligentPad and MPML, we have developed two pads that allow Web applications to communicate with users through multimodal character agents. These pads are the ‘MPMLPlayerPad’ and the ‘HTML2MPMLPad’. MPMLPlayerPad The MPMLPlayerPad gives a multimodal presentation according to a given script written in MPML. Figure 6 shows its screenshot and an abstract of its connection interface. A character agent is shown with an MPMLPlayerPad on the user’s desktop screen. An MPMLPlayerPad provides a slot named ‘MPML’ as a ’placeholder’ for MPML script. When the MPML script, which is stored in the ‘MPML’ slot, has been altered, the pad gives the presentation according to the updated MPML. MPMLPlayerPad wraps the MPML 2.3a implementation [10], which was developed by Prof. Ishizuka’s group at the University of Tokyo, to control character agents’ behaviors. HTML2MPMLPad Originally, MPML was designed to be written by human hands with or without some support form authoring tools. As we have shown in Section 3, MPML provides a simple scripting scheme for multimodal presentations in the XML format. Due to the simplicity

90


Figure 7: User Interface with the Connection Interface of HTML2MPMLPad

of MPML, it can be also generated by external resources by computer programs such as XSLT. The HTML2MPMLPad converts a text string given in HTML format into MPML format. Figure 6 shows its screenshot and an abstract of its connection interface. The HTML2MPMLPad provides two slots named ‘Text’ and ‘MPML’. When the text string, which is stored in the ‘Text’ slot and given in HTML, has been changed, into a new one, the pad converts the updated HTML text string into MPML. This HTML text string may be a part of the Web page, i.e. a substring corresponding to an element in the DOM tree of the Web page document, or may comprise a whole HTML document. For a given HTML string, the HTML2MPMLPad generates an MPML script by XSL-translation. In the current implementation, we use a conversion to MPML in which two agents read text given in HTML alternately. This is similar to TV news program with two newscasters. The conversion algorithm divides the given HTML text into two parts by detecting the paragraph changes. We use a set of tags to detect paragraph change. This set of tags includes br, p , div ,h1,. . ., h6, and tr. Then it assigns agents to each paragraph to alternate their roles for different paragraphs in turns. Figure 8 shows an example of the conversion from HTML to MPML. The document (a) is a news summary in Yahoo! News, and its HTML source code is shown in (b). MPML code (c) is generated from the HTML (b) by the HTML2MPMLPad. (d) shows the multimodal presentation defined in the MPML(c). In case other types of behavior of agents are required, we must modify the conversion algorithm. Support for users who wish to modify the conversion remains for future work. More sophisticated algorithms may also be applicable that can divide given texts into paragraphs. For instance, the TextTiling algorithm provided by Hearst [8] could be helpful for dividing the given text into several paragraphs that share the same context. Attaching Multimodal Character Agents to Existing Web Applications We now have Web applications, a converter from HTML to MPML, and an MPML player, all in the form of pads in IntelligentPad. It is therefore quite easy even for endusers to combine them together to attach a multimodal user interface to an existing Web application. Figure 9 shows an example application in which two character agents read news posted from external news site. In the sample application shown in Figure 9, WAW-PAD at the bottom has a slot


(a) A document portion in Yahoo News

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

(b) HTML source code of (a)

Yahoo News

91

ORON JGCF VKVNG CIGPVEJCT2GGF[KF2GGF[Z[ CIGPVEJCT)GPKGKF)GPKGZ[ JGCF DQF[ UEGPGKF6GZVECRVKQP6GZV UGS RNC[KF2GGF[CEV9CXG URGCMKF2GGF[ 2TQDG.CPFUQP5CVWTP/QQP 5GPFU2JQVQUURGCM URGCMKF2GGF[ #2(TK,CP2/ '6URGCM URGCMKF)GPKG #'WTQRGCPURCEGRTQDG (TKFC[UGPVDCEMVJGHKTUVFGVCKNGFRKEVWTGUQHVJG HTQ\GPUWTHCEGQH5CVWTP UOQQP6KVCPUJQYKPI UVWPPKPIDNCEMCPFYJKVGKOCIGUQHYJCVCRRGCTGFVQDG JKNN[VGTTCKPTKFFNGFYKVJEJCPPGNUQTTKXGTDGFU ECTXGFD[CNKSWKF(WNN%QXGTCIGURGCM UGS UEGPG DQF[ ORON

(c) Converted MPML of the HTML in (a)

(d) Multi modal presentation according to the MPML in (c)

Figure 8: Conversion from HTML to MPML

named ‘TopNews’. This slot is a priori specified by the user beforehand through mouse operations in the way we have described in Section 2. The ‘Top News’ slot holds the latest important articles from Yahoo! News. This slot is connected to the Text slot in the HTML2MPMLPad. The MPML slot in the HTML2MPMLPad is connected to the MPML slot in the MPMLPlayerPad. These slot connections are all specified by ‘drag & drop’ and ‘paste’ operations. Whenever the top news changes, an update is propagated to the HTML2MPMLPad. The HTML2MPMLPad will then generate a multimodal presentation according to the latest news. Then, the character agents automatically read the latest news. “Peedy the bird” reads the news headline and their dates, followed by “Genie the Genie” reading the description of the news. This framework does not require users to write any programming code either for wrapping an existing Web application or for attaching multimodal character agents to the wrapped Web application. 5

Concluding Remarks

Attachment of multimodal character agents to an existing Web application makes the Web application more attractive. What I have proposed in this paper is a novel framework that enables end-users to introduce a multimodal user-interface to existing Web applications. The framework is based on MPML and IntelligentPad. MPML is a simple but powerful markup language to support the authoring of multimodal presentations. In-

92


Figure 9: A Sample Application of our framework: Two character agents read news posted from an external news site

telligentPad architecture enables Web (or non-Web) application programs to be dynamically linked to other programs. We have newly developed the MPMLPlayerPad and HTML2MPMLPad. Users can easily introduce multimodal character agents to the front-end of existing Web applications through the use of the WAW-Pad, the MPMLPlayerPad and the HTML2MPMLPad. As an application example of our framework, a real-time news-reading agent was constructed as a composite pad. We think a further revision of the conversion algorithm from HTML to MPML will make this framework more appealing and attractive. It might be impressive to control the character’s emotion depending on the contents that the character is reading. For instance, an agent who reads sad news with happy face is not appropriate. Recently many news portal sites use RSS(RDF Site Summary) to provide summaries of their news. Our framework can be applied directly to such RSS feeds for multimodal agents to present the latest news with verbal and non-verbal behaviors. This would be a very practical application of our framework. Acknowledgments I would like to express sincere thanks to Prof. M. Ishizuka and his MPML development team at the University of Tokyo. I had very fruitful discussions with them during my visit to their laboratory. I also wish to express thanks to Prof. Y. Tanaka of Hokkaido University for his kind help and appropriate advice in the course of this research.


93

References [1] André, E. and Rist, T.: From Adaptive Hypertext to Personalized Web Companions, Commun. ACM 45(5), pages 43–46 (2002). [2] André, E., Rist, T. and M¨ uller, J.:, WebPersona: A LifeLike Presentation Agent for the World-Wide Web, In Proc of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent, (1997). [3] Bauer, M. and Dengler, D.: InfoBeans -Configuration of Personalized Information Services, In Proc. of IUI99, pages 153–156, (1999). [4] Bauer, M., Dengler, D., and Paul, G.: Instructible Agents for Web Mining, In Proc. of IUI2000, pages 21–28, (2000). [5] Berners-Lee, T., Cailliau, R., Luotonen, A., Henrik Nielsen, H.F., and Secret, A.: The World-Wide Web. Communications of the ACM, Vol. 37, No.8, pages 76–82, (1994). [6] Davis, H.C., Hall, W., Heath, I., Hill, G., and Wilkins, R. Towards an Integrated Information Environment with Open Hypermedia Systems. In Proceedings of ECHT’92, pages 181–190. (1992). [7] Dawkins, R. The Selfish Gene. 2 edn. Oxford University Press, Oxford (1976). [8] Hearst, M.: TextTiling: Segmenting Text into Multi-Paragraph Subtopic Passages, Computational Linguistics, Vol.23 (1), 33-64, (1997). [9] Ishizuka, M., Tsutsui, T., Saeyor, S. , Dohi, H. , Zong, Y., and Prendinger, H.. MPML: A multimodal presentation markup language with character control functions. In Proc of Agents’2000 Workshop on Achieving Human-like Behavior in Interactive Animated Agents, pages 50–54, (2000). [10] Ishizuka, M.: MPML HomePage, http://www.miv.t.u-tokyo.ac.jp/MPML/ [11] Ito, K., Tanaka, Y.: A Visual Environment for Web Application Composition. In Proc of 14th ACM Conference on Hypertext and Hypermedia, pages 184–193 (2003). [12] Ito, K. CHIP(Collaborating Host-Independent Pads) Homepage. http://ca.meme.hokudai.ac.jp/people/itok/CHIP. [13] Microsoft Corporation: Microsoft Agent, http://www.microsoft.com/msagent/default.asp [14] Microsoft. MSHTML Reference. MSDN Library. [15] Prendinger, H. and Ishizuka, M. (eds.): Life-like Characters – Tools, Affective Functions, and Applications, Springer-Verlag (Cognitive Technologies series),(2003). [16] Rist, T., André, E., and M¨ uller, J.:, Adding Animated Presentation Agents to the Interface. In Proc of Intelligent User Interfaces 1997, pages 79–86 (1997). [17] Rist, T., André, E., Baldes, S.: A flexible platform for building applications with life-like characters, In Proc of Intelligent User Interfaces 2003, pages 158–165 (2003). [18] Tanaka, Y. and Imataki, T.: Intelligentpad: A hypermedia system allowing functional composition of active media objects through direct manipulations. In Proc. of IFIP’89, pages 541–546, (1989). [19] Tanaka, Y. Meme Media and Meme Market Architectures: Knowledge Media for Editing, Distributing, and Managing Intellectual Resources, IEEE Press, John Wiley & Sons,(2003). [20] World Wide Web Consortium: XML Path Language (XPath), http://www.w3.org/TR/xpath [21] World Wide Web Consortium: Simple Object Access Protocol (SOAP) 1.1, http://www.w3.org/TR/SOAP/ [22] World Wide Web Consortium: Document Object Model (DOM) Level 2 HTML, http://www.w3.org/DOM/ [23] World Wide Web Consortium: Extensible Markup Language (XML) , http://www.w3.org/XML [24] World Wide Web Consortium: Synchronized Multimedia Integration Language (SMIL), http://www.w3.org/AudioVideo/


94

Privacy Enhanced Information Systems Sabah S. Al-Fedaghi1 , Gunar Fiedler2 , and Bernhard Thalheim2 1

Computer Engineering Department, Kuwait University, Safat 13060, Kuwait [email protected] 2 Kiel University, Computer Science Institute, 24098 Kiel, Germany fiedler | [email protected]

Abstract. Privacy is becoming a major issue of social, ethical and legal concern on the Internet. The development of information technology and the Internet has major implications for the privacy of individuals. Studies of private or personal information have related it to such concepts as identifiably, secrecy, anonymity, control, etc. Few studies have examined the basic features of this type of information. Also, database models (e.g., ’Hippocratic’ [AKX02] database) for this type of information have been proposed. This paper studies the nature of private information and develops a new conceptual model for databases that contain exclusively private information. The model utilizes the theory of infons to define “private infons”, and develops taxonomy of these private infons based on the notions of proprietary and possession. The proposed model also specifies different privacy rules and principles, derives their enforcement, and develops and tests architecture for this type of databases.

1 Introduction 1.1 Information Privacy The notion of privacy is becoming an important feature in all aspects of life of modern society. “[ P]rivacy will be to the information economy of the next century what consumer protection and environmental concerns have been to the industrial society of the 20th century.” [Gle96] One of the most important factors that contributed to this increasing significance of the concept of privacy has been the appearance of global computer networks. They have allowed information to be moved around the world and made national boundaries irrelevant to most types of communication. This has raised concerns about privacy in the on-line environment and has affected different aspects of dealing with private information in the areas of security, commerce, government, etc. We distinguish today between media privacy, territorial privacy, communication privacy, bodily privacy, and information privacy. The first four forms of privacy are now well supported by laws, constitutional rights and other legal frameworks. Although a number of laws protecting information privacy have been introduced in most developed countries, information privacy is challenged and not well protected due to the technological development, e.g. the commercial success, misbehavior, tools for the “glass box customer” and WWW misuse. Already [Wes67] coined principles of privacy that must be preserved by any system: openness and transparency, individual participation, collection limitation, data quality, use limitation, reasonable security, and accountability.

S.S. Al-Fedaghi et al. / Privacy Enhanced Information Systems

95

1.2 Database Privacy Protecting sensitive and classified data stored in database management systems requires measures beyond controlling access to files managed by operating systems. An example in this context is the problem of inference, which is the deduction of unauthorized information from the observation of authorized information - where unauthorized information is deduced from the legitimate responses. There is a lot of research work in the area of applying privacy protection in the context of databases. In the medical field, a great deal of interest is in anonymizing textual information. Sweeney’s pioneering work [Swe96] is based on removing the personally identifying information from the text so that the integrity of the information remains intact even though the identity of the persons remains confidential. The problem in this case is identification detection problem where identifiers of persons are detected and anonymized. Sweeney’s recognition methodology aims at detecting information that can personally identify any person. One important issue is the definition of “personal information.” Is it the whole text, the paragraph, the sentence, the phrase or only the word that denotes the identity? For tabular data, several techniques are proposed at the relation level. For example, the kanonymization technique [Sam01] assumes that a relational table with a prime key that refers to a person is “personal information”. Its main concern is anonymizing entries in the table in order to block any attempt to reach “identifiablity” that stems from these entries. Systems that use such techniques aim at protecting individual identifiable information and simultaneously maintaining the entity relationship in original data. Still, the definition in these works of “personal information” is not clear. Implicitly, it is understood that the privacy aspect comes from associating the attribute name with the identifying key of the relation. Statistical Databases aims at providing statistical information without compromising sensitive information about individuals. However, statistical databases goal is to prevent disclosure of private information and is limited to implied information. Agrawal et al. [AKX02] is first in proposing ‘Hippocratic’ database systems that put privacy as central concern. They articulate what it means for a database system to responsibly manage private information under its control. Nevertheless, their proposed model can be considered as an extension of the relational database model. It lacks modeling features that put privacy as a central concern of the conceptual schema. Generally, in database design the conceptual schema encompasses the enterprise’s view that encompasses all types of information. We concentrate on building conceptual schema exclusively for private information. We envision a system where not only enterprises have ‘private information’ databases but every individual has his/her own ‘private information’ database that includes his/her own private information and other’s private information possessed by him/her. Our approach is comprehensive in terms of analyzing the overlapping between different conceptual schemas. Suppose that we have one enterprise and 2000 individuals. Our model is concerned with developing a framework of 2001 conceptual schemas and their sub-schemas. In the e-environment the corresponding (private) databases interact with each other utilizing the knowledge in their schemas. Private information agents in all databases work on behalf of their owners to facilitate private information transactions. For example, the agent of individual x accesses the database of the enterprise to update its owner’s private information, then the agent record in the individual’s database that such information is consented to the enterprise’s database. Or, the agent of the enterprise communicates with the individual’s database to verify that the private information it has about the individual, is consented. Such an approach can be applied also to current proposed ontologies of the real world (e.g., RDF) through integrating descriptions of private information into the model.

96


1.3 Our Contribution Our approach aims at developing a systematic methodology that proceeds from a working definition of “private information” to a complete set of conceptual schemas that includes privacy-based decomposition and constraints. We achieve this goal through: Framework: This paper introduces a formalism to specify informational privacy assertions based on a theoretical foundation. We introduce a new definition of private information and categories of this type of information to entangle issues of compound private information (John and Mary use contraceptives is ‘private’ information of both John and Mary). We also introduce a classification of relationships of “private information pieces” to its possessors and proprietors. The proprietor here refers to the identified person in the assertion. Model: Our theoretical framework is applied to the restrictions and conditions of possession. Thus, the private information in possession of z can be classified according to whether this possession is according to: consent/no-consent of the proprietor, legal/illegal possession of private information, and awareness/no-awareness of the proprietor of this possession. 1.4 Related Work The comprehensiveness of our approach extends from the system’s level (e.g., conceptual schema design) to the specification level that includes functionalities of subsystems such as constrains and policies. At the system level, J. Biskup introduced in his recent script [Bis05] a general approach for development of security preserving systems. According to his model participants in the information society include nearly every individual, group, institution, association or private company. Informational activities among participants must be in some way based on trust. Orthogonally we distinguished between known and unknown services provided[Gib05]. An individual is seen as an actor being involved in a large variety of different social roles including determining which personal information he/she is willing to share. Since there are potentially conflicting interests, the design of a the design of a security/privacy system takes into consideration the specific needs and wishes of communities of participants. We visualize the general architecture of this system as shown in Figure 1. Participants are threatened by certain environments and explicitly forbid services in other environments. Services are provided by a system in a selected environment. Participants may trust service provider(s) and require a service. We may add other restrictions such as acquiring only trusted services. Notice that our main concern here is private information transactions; while Biskup’s model deals with all aspects of security. The correctness properties may be specified by three path constraints in the security preserving system: • A security system is reliable correct if all the participants’ requests can be handled by services in which the participant has trusts. • The system is a confine system if no illicit services pop up with services used. • The system is a trusted system if whenever a service pops up with the requested services, it is whitelisted. Privacy system is not a subsystem of the security system and ‘private information trust’ does not exactly correspond to the general notion of trust. Some security measures may compromise privacy protection as in network monitoring. Nevertheless, the security tools used by

97


Participant 6

TrustsIn

kk

?

Malware Service

Known Requires

ThreatenBy

Checked

Y 6

j

-

Service

j

?

ProvidedBy 66

Y

Forbidden

-

System

Prosecuted Service

+

Environment

Disclosure Legislation

Whitelisted

BoundWith

Figure 1: The structure of a security system model.

a security system are very important for any privacy system since they can provide mechanisms to facilitate the privacy aspects of a system. Generally, data security involves protecting the information itself against various risks, such as the risks of being accessed or modified by unauthorized persons. ‘Data security’ is not an aspect of ‘privacy protection’ unless the involved data represents ‘private information’. Consequently, in our model participants in the ‘private information society’ include every individual and enterprise. This ‘society’ could be a hospital and its patients or it could be the whole society with all individuals and enterprises participating in private information exchanges. In the context of such ontologies as RDF, our conceptualization can be thought of as a ‘mode’ of RDF that deals with privacy-related information. At the technicalities level, there is a great deal of work in the area of protecting private information. In authentication systems (e.g., Microsoft Passport) the private information is generally limited to identification/contact information. Some published works propose agents that are trusted by a user to store and make available private information such as user profiles, other published works propose vocabularies for composing policies to allow or deny access to the private information [Lee02]. The Platform for Privacy Preferences (P3P) has provided tools and services that improve trust and offer an automated way for users to gain more control over the use of their private information. APPEL (A P3P Preference Exchange Language) can be used by the user agent to make automated and semi-automated decisions regarding the acceptance of privacy policies from P3P enabled Web sites. P3P enhancements also allow negotiating for a different privacy policy that can be used in any subsequent release of private information. Recent works in this field address semantic issues for privacy management. It is pointed out that a standard method of exchanging privacy policies (a privacy ontology) is needed for the Semantic Web [KHM02, TDT04]. P3P Project extensions allow data and conditions to be expressed in an RDF privacy policy [W3C02]. There are also several projects that exploit the development of privacy protection agents using semantic web technology and the W3C’s Platform for Privacy Preference protocol [Gre02]. These works lack an encompassing model for private information databases. This situation is similar to if we had relational constraints, operations, policies, etc. without having the underling relational database model. A basic step in developing a private information model is building ontology to identify general types of private information. Our proposed ontology

98


provides a deeper level of meaning of sub-categorization of private information. For example, these general types of private information are used in analyzing the relationship between privacy and ethics [AF05b] and in refining the notion of anonymity in the context of private information [AF05a]. 1.5 Problems Found In A Privacy Project Private information includes sensitive information that many people consider it deserving the strongest protection. We give here one case related to health information. The collection and processing of medical information have witnessed great changes since the early 90’s. The setting for information began to shift quickly from manual forms to electronic medium. This has raised concerns about protecting the privacy of medical records. The German government issued in 2002 the health card project as a competitive project between the German states. The Schleswig-Holstein health card project is considered the most advanced [LSH]. All relevant patient data are stored on the SIM1 of a health identity card. The data must be protected by one pin and, additionally, might be centrally stored. Data to be stored on the card consist of personal health condition, all data associated with medical treatment, physician visits, and specific disease and disorder information. Moreover, access keys are provided to read the full medical record prepared by doctors. The German health card project is currently based on a simple access model that consists of: Assigning views to each role such as the pharmacist view, and Securing and encrypting information based on pin codes managed by patients. The storage model used in the Schleswig-Holstein health card is based on: Central storage of all relevant information that is accessible by everybody who has the access portfolio, and Local partial storage of any recent information on the personal health card of each patient. The patient information is accessible by those who are permitted to access it either by using the health card of the patient directly or by a pin number provided by the patient and pointer to the central storage. These models are not appropriately designed to handle private information. They even do not have a clear definition of ‘health information’. ‘Health information’ (e.g., the payment method) is not necessary private information if it does not refer to an identifiable individual. They are not ‘private information models’ but general database models designed for any type of data such as inventory data, maintenance data, budget data etc. This generality may cause difficulties in controlling the flow of private information. For example, once information has been retrieved, the patient has no control on, or even knowledge of, its distribution. In our case we envision cooperating among private information databases with their software agents that check each other and report abnormal situations. A person can examine his/her private information in an enterprise just as he/she can check his/her account in a bank. 1 SIM=Subscriber Identity Module, a chip with a processor and some memory. SIMs are typically used for user identification, encryption and data storage.

99


2 Infons Owned by Agents 2.1 Infons of Agents Infons have been introduced by [Dev91] to represent possible facts and are extensively used in situation theory [She96]. The theory of infons can be understood as a generalized theory of notions introduced in [Kau67]. They are less rigid than concepts considered in [GW98]. An example of an infon might be the predicate Is Village In(Weißig,Saxony) or the dual (negated) predicate ¬ Is Village In(Dresden,Saxony). Infons can be composed of infons, e.g., the infon Possesses(A100352 , ¬ Is Village In(Dresden,Saxony)) which says that the agent A100352 possesses the infon ¬ Is Village In(Dresden,Saxony). An infon is a discrete item of information and may be parametric. The parameters are objects, and the so-called anchors assign these objects such as agents to parameters. An action results in changing many infons in the world, and the change in an infon is essentially dependent on the situation where the action occurs. Primary infon sets I are (temporary, epistemic) subsets of the set of all infons with the restriction that all dual infons (same infons with different polarity) also belong to I and that all instantiations of parameters of an infon I are in I. More formally the second condition says that the infon A(f ) is obtained by replacing each parameter x˙ ∈ (para(A) ∩ dom(f )) by the object f (x) ˙ in A for A ∈ I and an anchor f . Infons not belonging to a primary infon set are called secondary. An infon constraint relation is a subset of the Cartesian product of the set of all infons and the set of all secondary infons. We may use [Kau67] to develop a general logic of infons. Infons can be ordered by the generalization/specialization order A B specifying that A is at least as specific as B. Similarly, is the opposite relation. The relation specifies that two infons are similar. We can now derive operations: ∃

Homogeneity of two infons defines the relation A B := ∃X(A X ∧ B X), i.e. two infons can be homogenized if a more specific infon X exists. ∃

Inhomogeneity of two infons is the opposite, i.e., A B := ¬∃X(A X ∧ B X). ∃

Compatibility of two infons requires the existence of a more general infon, i.e., A B := ∃X(X A ∧ X B). ∃

Incompatibility of two infons is given through A B := ¬∃X(X A ∧ X B). Derived predicates are ∃

∃

divergency := ∧ ∃ ∃ isolation := ∧

∃

∃

potential homogenizability := ∧ ∧ ∧ , and ∃

∃

heterogeneity := ∧ . We may directly derive a number of general laws and other operators: ∃

• A B ↔ ∃x∀y(x y → (A y ∧ B y)) ∃

• A B ↔ ∃x∀y(x y → (A y ∧ B y)) • Product: C = A B := ∀y(C y ↔ (A y ∧ B y)))

100


• Sum: C = A B := ∀y(C y ↔ (A y ∧ B y)) • The negation specifies the negation of all properties of the infon but not the infon itself: ∃

B = A¯ := ∀x(x B ↔ x A). • The difference between two infons is given through all those properties that are contained in the first infons but form an infon that is incompatible with the second infon: ∃

C = A B := ∀x(C x ↔ (A x ↔ B x)) This general logic and general computation theory can be used to infer new infons based on the set of primary infons. Agents A from a general agent set Z are divided into two disjoint sets: individuals V and non-individuals N. An individual represents a natural person while a non-individual represents an artifical person like a company, a government agency, etc. We may, furthermore, model that an infon belongs to an individual or agent by using the general predicates. The relationship between infon and the sets V and N is categorized as follows: • Possesses is a relation between members in Z and primary infons. Possesses (z ∈ Z, f ) means that infon f is possessed by z. For example, an individual may possess private information of another individual or, a company may have in its database, private information of someone. • Knows is a relation between individuals in Z and primary infons. Knows(v ∈ V, f ) means that v has knowledge that f is in possession of z ∈ Z. The predicate can be combined with itself, for instance, we can state Knows(A100352 ,Possesses(A210981 , ¬ Is Village In(Dresden,Saxony))) that is, A100352 Knows that A210981 possesses the misinformation Is Village In(Dresden,Saxony). Also, Knows(A180753 , Knows(A100352 ,Possesses(A210981 , ¬ Is Village In(Dresden,Saxony)))) .

• Belongs is a relation between entities in Z and primary infons. Belongs(v ∈ V, f ) means that v is a referent in f . For example, if f is John is an alcoholic, then f “belongs” to John. The infon f is a private information assertion that identifies John, hence, Belongs(John , f). Infons are similar to memes discussed in [Bla99]. Memes are the units of cultural evolution and selection. They can be folded and be used for derivations. We will use situation theory for development of a theory of information flow. This paper focuses on infons that represent private information. 2.2

Private Infons

We propose an Information Privacy Model (IPM) that focuses on linguistic forms of private information taking infons as its basic components. A ‘model’ here refers to abstractions representing the knowledge, organization, expressions, and constraints acquired from the world around us. The realization of this model means introducing a set of rules and data structures for which the rules are true. The model in this case, is formed from certain graphs, called schemata, that serve as patterns or frames for the model. The schema is ‘conceptual’ in the sense that, it represents a ‘mental’ model that is built from concepts and conceptual relations. ’Language’ is the main vehicle that describes entities, actions and states in the domain


101

of knowledge, which in our case is informational privacy. Linguistic elements will be our starting point in identifying the meaning of ‘private’ information. Infons are categorized according to the number of their referents as follows: Zero infon is an infon that has no referent that signifies a single individual (person) A ∈ V. For example, Spare part number 54321 is in slot a-2, is a zero infon because it does not refer to a person. The zero infon is not private information because it refers to nonindividual object in N. Some zero infons may be embedded into private information. For example, This dog with collar belongs to John embeds the zero infon The dog has a collar, and the non-zero assertion John has a dog. Atomic infon is an infon that has a single referent that signifies a single individual A ∈ V. Compound infon is an infon that has more than one referent that signifies individuals in V. Let A be an infon. If A is true then A is said to be information. Consequently, there are zero information, atomic information, and compound information according to the number of the referents. Atomic (compound) information becomes private if it refers to identifiable individual(s). Private information is also related to who possesses it. A single piece of atomic private information may have many possessors; where its proprietor (i.e., the referent in the private information assertion) may or may not be among them. Atomic infons can be possessed by any agent in Z. Individuals can have private information of other individuals. Companies and government agencies can possess a great deal of private information about individuals. “Ownership” of atomic private information is limited in IPM to the subject of the assertion, i.e., the individual to whom the atomic private information refers. If A is atomic private information of A ∈ V; i.e., A is its subject, then A is proprietary private information of A and A is its proprietor. Every piece of private information has its proprietor. A proprietor of private information may or may not be its possessor and vice versa. Individuals can be proprietors or possessors of private information; however, non-individuals can only be possessors of private information. The notion of proprietorship here is different from the legal concept of ownership. The ‘legal owning’ of a thing is equated with the exclusive possession of this thing with the right to transfer this ownership of the thing to others. “Proprietorship” of private information is nontransferable in the absolute sense. Others may possess or (legally) own it but they are never its proprietors (i.e., it cannot become their proprietary data). Two or more individuals “have” the same piece of compound assertion when each individual has at least one atomic assertion embedded into the compound assertion. For example, John, Jim and Alice hate each other is a compound assertion of John, Jim, and Alice. It embeds the atomic assertions Someone hates someone, where someone can be John, Jim, or Alice. These compound assertions can be represented by the three atomic assertions and the meta-assertion that states that they are components of a single compound assertion. In logic, we have the predicates hates(John, y), hates(Jim, x), and hates(Alice, z) that are mapped to the predicate hates(x, y, z) with objects {John, Jim, Alice}. It is not possible that two individuals have identical atomic private information because simply they have different identities. Atomic private information is the “source” of privacy. Compound private information is “private” because it embeds atomic private information. Also, the concept of proprietorship is applied to compound private information, which represents ‘shared proprietorship’ but not necessarily shared possession or ‘knowing’.

102


2.3 Reducing Compound Infons to Atomic Infons A piece of compound private information is not a collection of atomic private information. Compound private infons are fundamental structures obtainable from relations and individuals. The infon InLove(A1 , A2 ) does not have the same ’collectivity’ or compositionality as the infons In(A1 , London) and In(A2 , London). The latter is pseudo (not real) compound private information. It is a collection of the atomic private information. The structure of compound private information is important if we are interested, not only in breaking up the compound assertion into its embedded atomic assertions, but also in reconstructing the original compound assertion. For example, John, Jim and Alice hate each other can be reconstructed uniquely from its atomic assertions, while the pseudo compound assertion John and Jim hates Alice is not. Notice that the later embeds the atomic assertion x and someone hates someone. A compound infon is reducible to a set of atomic infons but it has more than that. The ‘compound’ characterization refers to compound identities. If the atomic private information InLove(A1 , x) is composed in the compound private information InLove(A1 , A2 ) then the possessor of the remaining part, InLove(y, A2 ), has no private information of A1. If the infon John and Alice are in love is partially anonymized and released as Someone and Alice are in love, then no private information of John is released. That is the reason that we have concentrated on atomic private information. Atomic infons are “pure” private information while compound information is not proprietary information of a single individual. It is shared privacy, thus its control is shared among its proprietors. Nevertheless, “pure” compound information (the part that connects the resultant atomic infons) is an essential part of privacy-based databases. Suppose that we have the compound private infon A(A1 , A2 , ..., An ) where A1 , A2 , ..., An refer to different identifiable individuals (e.g., hates(John, Jim, Alice)). Privacy-reducibility of A(A1 , A2 , ..., An ) refers to producing n atomic private infons and the zero infon: Infon 1 Infon 2 ... Infon n . For example, John, Jim and Alice hate each other embeds the atomic assertions x hates someone, where x can be John, Jim, or Alice and the meta-assertion that states that they are components of a single compound assertion. The process of breaking up the compound infon to its atomic infons is a straightforward procedure. It can be described as follows: For each Aj , j) in A(A1 , A2 , ..., An ) by the nonproduce its atomic private infon by replacing all Ai (i = identifiable description of Ai . Additionally, we introduce the zero-information meta-infon described previously. Notice that the zero-information meta-infon is a zero infon because it refers to assertions not individuals. Thus, we have isolated the private information of each individual and at the same time have preserved the “possibility” of recovering any compound private information. In the restrictive structure of databases, reconstructing the original compound private information requires access to the n + 1 databases (n atomic infons plus one zero infon). For example, consider the relational schema MARRIED(HUSBAND, WIFE), WIFE-INFORM(NAME, AGE), and HUSBAND-INFORM(NAME, AGE). The first table contains zero privacy information that links tuples in the other two ’private information’ relations. So the original compound infon John whose age is 40 is married to Alice whose age is 30 can be represented by the atomic infons: (1) (John whose age is 40) is married to (someone whose age is 30) and (2) (Someone whose age is 40) is married to (Alice whose age is 30). MARRIED(HUSBAND, WIFE) includes exclusively pointers and releasing it does not compromise the privacy of any husband or wife in the database. For example, if John is the 4th tuple in HUSBAND-INFORM


103

and Alice is the 9th tuple in WIFE-INFORM, the tuple (4, 9) can be one way of implementing the zero privacy information that links John and Alice. The methodology of syntactical construction of the original assertion is not of central concern here. Whether the resultant set of atomic assertions is semantically equivalent to the original one is an interesting problem. The reduction process of the original compound private infon is utilized here to identify “privacy centers” and use these ‘centers’ in different privacy related applications. 2.4 Possession and Property of Infons Every individual has his/her stock of private information. Some of this information is his/her own and some is in his/her possession. In this section we study different types of atomic assertions associated with a single individual in order to identify their relationships to the proprietor. We divide the set into two disjoint sets of atomic assertions Individual.Proprietary and Individual.Possession as shown in Figure 2. Each set is defined as follows: Infons of individuals

Proprietary infons

Not know (the individual)

Possession (infons of others)

Known by individual

Not known by other individuals

Known by other individuals

Known Public to Shared with others through consent others by contracts

Figure 2: The taxonomy of infons of an individual.

1. Possession: This is the set of pieces of atomic private information of the others that are in the possession of an individual. For example if individual A knows that individual A has cancer then this information is in the set Individual.Possession.A of A . 2. Proprietary: This is the set of atomic private assertions (atomic infons) of the individual him/herself. It has two categories: (a) Not Know (NKnow): This is the set of atomic private information, which the proprietor does not know about it, e.g. results of medical analysis. This set is included in the tree for completeness sake. We often assume that if the object Individual (as a data structure) represents an entity of type individual then this set should be empty. However Individual can be part of another schema, in this case NKnow may not be empty. (b) Known: This is the set of atomic private information assertions (infons) that the proprietor does know about it. It includes two types of atomic private information assertions as follows. (b1) Not Known by other individuals: This is the set of atomic private information assertions that are known only by their proprietor and not known by others.

104


(b2) Known by other individuals: This is the set of private information assertions that the individual knows that it is known by others (in possession of others). It is divided into many sets of pieces of atomic information that are in possession of subsets of others: (b21) Infons that are shared with others by contracts. The contracting schema depends on the individuals. (b22) Infons that are shared with others: (1) with the consent of the individual or, (2) against his/her will. An example of the latter type is private information (e.g., names) of victims obtained from a publicly released police report. (b23) Infons that are shared with everybody and are thus public. Contracts are specified through the “who-how-whom-what-on what basis” frame of contracting. Contracts may be built using parametric secondary infons. The basis for contracts of information sharing is usually another contract. We may envision that contracts use cases of information sharing in the style of UK jurisprudence. The ‘who-whom’ specification might use a specification of collaboration between individuals or non-individuals. The style of information sharing is based on a specification of workflows which might include message passing activities whenever an infon is used by an agent who is not its proprietor.

2.5 States of Possession of Infons We have developed a general theory of infons, agents, possessions and proprietorship. Let us sharpen the requirements for a privacy enhanced information system. The general state of possession of infons is displayed in Figure 3. Consent of possession 6

State of possession

- Legality of possession

Awareness of possession Figure 3: States of possession of infons monitored by the privacy enhanced system.

We distinguish three perspectives of possession: Legality of possession that ranges in the hierarchy of private information from proprietary through Known by contract to Known illegal. Awareness of possession that can be recorded for each individual. Consent of possession that may be based on contracts, on consent or be none. We will concern with non-individual entities or agents (companies, government agency, etc.) that possess private information about individuals. Each non-individual agent is ‘an enterprise’. By our definition of private information an ’enterprise’ cannot be a proprietor of private information. We will use paths within the schemas to identify sets of atomic infons. For


105

instance, Az .Possession.Ax produces all private infons of Ax in possession of Az . The states of possession of private information by enterprises may be restricted as follows: Awareness of the proprietor: This constraint indicates that the individual should be aware of any private infons in the possession of the enterprise. This fact can be embedded into the conceptual schemas structures. The consent of the proprietor: This constraint indicates that the individual should give his/ her consent of any private information in the possession of an enterprise. The consent implies awareness, however, awareness may not indicate consent. For example, private information about an ex-patient may be in the possession of the hospital without his/her consent. Greedy deletion of information: This constraint indicates that any information, which is stored without the consent or without awareness of the proprietor must be deleted from the system. Restrictions on visibility of information impact: This constraint concern implied information. An enterprise agent is, in general, able to draw an inference of what certain information holds if the agent knows the bearing of other information to the information the agent seeks. For instance, a conclusion could be drawn by a health insurance company if a patient requests medicine for sexually transmitted diseases. A typical situation of possession is displayed in Figure 4.

AZ Possession

AV Consent t

...

...

...

No Consent Know

Not Know

Figure 4: A conceptual structure of private information with the notions of consent and awareness. AV denotes a possessor and t denotes an infon.

Additionally, this model allows representing other states of possession, e.g. illegal possession. Consider the case of a police detective, say AZ , who has authority to watch a suspect, say AV , secretly. The collected private information of the suspect AV would be stored in AZ .Possession.AV .NConsent.NKnow where NConsent denotes the set of non-consented infons and NKnow denotes its subset of infons that the proprietor does not know that the detective possesses them. Now suppose that the detective forced the suspect to surrender his ID card. All information in the ID card will be stored in AZ .Possession.AV .NConsent.Know because the suspect, in this case, knows this fact.

106


We notice here that this discussion assumes that we are modeling the state of private information in the universe under consideration. Each entity in this universe has its own private information corpse of data classified according to our model. For example, if the universe contains two individuals and one enterprise, then the “universal” conceptual schema has three sub-schemas. One of the individuals may not keep track of his/her private information state. Still, he/she has in his/her possession private information of others, has his/her proprietary information which includes information he/she knows that it is in possession of the enterprise, information in possession of the other individual illegally, etc. 2.6 The Model of Possession Management Possession management can be supported by an information system that encapsulates the information obtained from other agents. We may envision distributed privacy enhanced information systems and systems with a centralized storage engine. Since there is no conceptual difference between the two approaches we concentrate in this paper on the centralized engine. The structure of the infons management schema2 is displayed in Figure 5. Individual

Privacy classification schema

Signifies zero: card = (0,0) atomic: card = (1,1) compound: card = (2,n)

6 ?

6

^

Agent = Non-individual · ∪ Individual

possesses

6

Infon

6

applies

]

true -

Known by others

Known possession

Information

Consent ?

Kind of sharing

derivable

through

-

Contract

6

Trivial information

Figure 5: The infon management schema for privacy support.

The database restricts the infons known by others to those which the individual knows. An infon belongs (signifies in the figure) to one individual (atomic infon), to a number of individuals (compound infon), or none (zero infon). True infons are information (i.e., the information type is a subtype of the infon type). Derivable information is considered to be trivial depending on the derivation system. Individuals and non-individuals may possess infons which belong to an individual forming thus an association between the signifies relationship types and either non-individuals or individuals. We may assume that each owned infon 2 The development of the database schema is based on the higher-order entity-relationship model [Tha00a] that supports co-design of structuring and functionality of information systems by a high-level specification language.


107

is also possessed by its proprietor (it may however not be the case in applications such as the health card since some infons (e.g., having cancer) may be withheld from individuals). The possession may be known (recorded through the property Known class) or unknown. Known possessed infons may be known by other individuals or non-individuals depending on whether the sharing of information is consented or contracted sharing. Additionally infons may be classified within a privacy classification schema. 3 Management of Infons of Individuals The infons management is based on the separation of matters of concern that include the following: Data are managed by the database systems of individuals and non-individuals. The database systems of the individuals are equipped with an export facility for data that might be provided to other agents. The export is based on predefined private information schemas or views. The result of an export enactment is small units of content that includes data together with meta-data describing the data and its possible usage. Infons are small units of content that are classified according to their truth and the applicable classification schema. Compound private information extends the infons of an agent. Agents have their own profiles [Tha00b] (consisting of private data - if applicable -, access restrictions, privacy and security profiles) and their own portfolios [Tha01] (consisting of permitted tasks, forbidden services, access rights, applicable contracts and consent). Deliverables are generated by the privacy enhanced information system upon requests by agents. 3.1 The Model of Infons Possession Management The playout of units of content is based on a view management system. This system binds the infons to individuals depending on the kind of possession, proprietorship, and the strategy for orchestration in the system. The general structure is shown in Figure 6. The strategy for privacy orchestration is based on a workflow specification. Each workflow step uses views that are defined in terms of infons. These views compile compound private information that consists of two spheres of private information: 1. Atomic private information embedded into compound private information C. So in general: C(AV1 , AV2 , ..., AVn ) subset of atom(AV1 ) ∧ subset of atom(AV2 ) ∧ ... ∧ subset of atom(AVn ). For example, in the view of a husband we can find the atomic private information that he is married, date of marriage, marriage certificate number, etc. but the identity of the wife is excluded.

2. Zero-information that links two or more atomic information in atomic databases. Thus, in general an individual has: (A) Proprietary database: This database includes proprietary private information as described previously. (B) Possession database: This database includes others’ private information in possession of the individual as described previously.

108


Zero privacy infon

-

Infon

y

Atomic privacy infon

6

stage / epistemic truth

derived information j Shared privacy infon

rules / Kind of privacy/ applicable for Supported principles- Strategy possession privacy usage proprietorship orchestration proprietorship / identification / truth ?

Individual

Figure 6: Main data structures in views in the infons possession management.

(C) Compound database: This zero-information database facilitates linking shared private information as described previously. Furthermore, a non-individual has: (A) Possession database: This database includes others’ private information in possession of the non-individual as described previously. (B) Compound database: This zero-information database facilitates linking shared private information as described previously. 3.2 System Support The architecture of the envisioned privacy enhanced information system is displayed in Figure 7. Although any database management system could be used as a basis for an privacy enhanced information system we choose the content management system CoreMedia SCT (Smart Content Technology) [Cor04] for the implementation of our experimental prototype. CoreMedia SCT supports generation, maintenance and controlled delivery of complex content structures. The system components of CoreMedia SCT like the storage engine or the delivery server are designed to support scaleable, highly performant, highly available, and secure information systems for a distributed collaboration of users. CoreMedia SCT repositories are used for storing the content. Due to the possibilities of the SCT document model and programming interface, any encryption technology can be seamlessly integrated. Additionally, we propose an infons possession tracking system. The content managed by the CoreMedia SCT might be the suite of encrypted infons, the workflow content that manages agents, atomic information, the signification, the possession, the known possession, or the sharing of infons. CoreMedias playout facilities deliver this content to individuals and non-individuals depending on the profiles and portfolios of agents. The preprocessor controls the legitimacy of requests issued by agents. It adds the portfolio and profile of the issuer to the request. Depending on the strategy for privacy orchestration,

109


Privacy Enhanced Information System Content Management System

SCT Server

SCT Server

Encrypted ... Encrypted DBS DBS

Infon Tracker Infon Possession Tracking DBS

Pre- processor ? - Filter

?

request

Agent -

-

deliverable

Tracker

Figure 7: The general architecture of the proposed privacy enhanced information system.

the extended request is issued to the content management system. The filter is based on the view depicted in Figure 7. The delivery to an agent is recorded in the content management system. 3.3 Recording Possession of Infons Within a health enhanced card project we separate six private information schemas: patient, hospital, doctor, pharmacist, insurance and information desk schemas. These schemas are limited to private information and embed in their structure information about proprietary of, possession of, and knowledge of private information. The actual complete database may exist in one system (e.g., hospital database) or may be scattered on different computers (e.g., patient’s laptop, doctor’s laptop) or in-between situation. All parties participate in this comprehensive picture of the patient’s private information. It is a portion of a global picture of any individual’s private information. When a person introduces new data that requires the possession of his/her private information, the corresponding infons, views and usage data are generated. The patient’s computer and the hospital’s system can exchange information and check each other according to their schemas. This multi-view of private information that is designed according to possession, knowledge, etc. is not limited to enterprises but -in a comprehensive database- applies also to individuals. Thus, the system, upon receiving a request to add new private information AV .t to its database, would activate certain procedures to ensure that AV is aware that AV .t is now in the possession of AZ . On his/her part, the individual AV can be permitted to access AZ .Possession.AV to look at all private information about him/her in possession of AZ . So the relationship between AV and AZ is reciprocal with regard to possession of private information. Various scenarios can be incorporated in this relationship with regard to updating the private data, correcting it, etc. In ideal situation AZ .Possession.AV should always be equal to AV .Known.AZ . Thus, the individual and the enterprise have identical copy of the private information of the individual that is in possession of the enterprise. This can also be applied for sub-schemas of the enterprise that contain portions of this private information. Several routines can be performed automatically on both sides, the proprietor and possessor, and appropriate reports of discrepancies can be produced. The enterprise schema of AZ may have “hidden data” in AZ .Possession.NKnow that includes private information that is legally hidden from the proprietor of that information.

110


For example, in a hospital DB system, the patient may be aware of AZ .Possession.Know.AV but is not aware of AZ .Possession.NKnow.AV . 4 Conclusion This paper introduces a formalism to specify information privacy infons based on a theoretical foundation. Current privacy research lacks such formalism. The new formalism can benefit in two areas. First in introducing precise definitions of informational privacy notions. It also, can be used as a base to develop a formal and informal specification language. Informal specification language can be used as a vehicle in specifying different privacy constrains and rules. Further work can develop a full formal language that can be used in privacy systems. The paper also discusses the health enhanced card project that is currently finalized [Fie04] in a project of the Kiel information systems engineering group. The experience we gained has shown that further development on the basis of security meta-profiles, on patient information facilities, infons tracking functions, and encryption mechanisms is necessary. We have shown that in a protected environment these requirements can be satisfied. We are in doubt that a completely satisfying solution can be provided in open environments although most systems that can be envisioned or are currently used, will be open systems. Our current investigation are centered around development of reliable correct systems similar to [WLT05] by providing a framework that traces and protects users against ‘bad’ services displayed in Figure 1. We preferred closed privacy supporting systems. Although closed privacy supporting systems is preferable, an open privacy enhanced system may be developed using enforcement techniques [Tha00a] or information wallets [AF05b]. The management of infons has to be extended in order to cope with changes from zero to private infons and vice versa. The logics of known infons, possessed infons and not known infons can be extended to trusted, whitelisted, known, and forbidden services using modal epistemic intuitionistic logics. References [AF05a]

S. S. Al-Fedaghi. A systematic approach to anonymity. In 3rd Int. Workshop on Security in Information Systems WOSIS-2005, Miami, May, 2005.

[AF05b]

S.S. Al-Fedaghi. The ’right to be let alone’ and private information. In Proc. 7th Int. Conf. Enterprise Information Systems, Miami, 2005.

[AKX02] R. Srikant Agrawal, J. Kiernan, and Y. Xu. Hippocratic databases. In Proc. 28th VLDB, pages 143–154, 2002. [Bis05]

J. Biskup. Information systems security. Lecture dortmund.de/issi/teaching/lectures/skript-sicherheit/, 2005.

script

http://ls6-www.cs.uni-

[Bla99]

S. Blackmore. The Meme Machine. Oxford University Press, Oxford, 1999.

[Cor04]

CoreMedia. Homepage of the CoreMedia AG. http://www.coremedia.com, 2004.

[Dev91]

K. Devlin. Logic and Information. Cambridge University Press, 1991.

[Fie04]

G. Fiedler. Description of the health card development practicum. System development practicum collocated with the program “Intelligent Information systems”, http://www.is.informatik.unikiel.de/∼fiedler/teaching/ws2005/iis/Praktikum.pdf, Oct. 2004.

[Gib05]

S. Gibson. Spyware was inevitable. CACM, 48(8):37–40, August 2005.

[Gle96]

J. Gleick. Behind closed doors; big brother is us. New York Times, Sept. 29 1996.


111

[Gre02]

D. Green. When the web starts thinking for itself. http://www2.vnunet.com/Ebusinessadvisor/1137710, 2002. 20-12-2002.

[GW98]

B. Ganter and R. Wille. Formal concept analysis - Mathematical foundations. Springer, Berlin, 1998.

[Kau67]

R. Kauppi. Einführung in die Theorie der Begriffssysteme. Acta Universitatis Tamperensis, Ser. A, Vol. 15, Tampereen yliopisto, Tampere, 1967.

[KHM02] A. Kim, L. J. Hoffman, and C. D. Martin. Building privacy into the semantic web: An ontology needed now. In Semantic Web Workshop, Hawaii, USA, 2002. http://semanticweb2002.aifb.unikarlsruhe.de/proceedings/. [Lee02]

R. Lee. Personal data protection in the semantic web. ME Thesis, MIT, 2002.

[LSH]

Landesregierung Schleswig Holstein. Project information Gesundheitskarte Schleswig-Holstein. http://www.gesundheitskarte-sh.de.

[Sam01]

P. Samarati. Protecting respondents’ identities in microdata release. IEEE Transactions on Knowledge and Data Engineering, 13(6), November/December 2001.

[She96]

J. Seligman and D. Westerst˚ahl (eds). Logic, Language and Computation. CSLI Lecture Notes, Stanford, CSLI Publications, 1996.

[Swe96]

L. Sweeney. Replacing personally-identifying information in medical records, the scrub system. Journal of the American Medical Informatics Assoc., pages 333–337, 1996.

[TDT04] A. Tumer, A. Dogac, and I. H. Toroslul. A semantic-based user privacy protection framework for web services. http://www.srdc.metu.edu.tr/webpage/publications/2004/TumerDogacToroslu.pdf, 2004. [Tha00a] B. Thalheim. Entity-relationship modeling – Foundations of database technology. Springer, Berlin, 2000. See also http://www.is.informatik.uni-kiel.de/∼thalheim/HERM.htm. [Tha00b] B. Thalheim. The person, organization, product, production, ordering, delivery, invoice, accounting, budgeting and human resources pattern in database design. Technical Report Preprint I-07-2000, Brandenburg University of Technology at Cottbus, Institute of Computer Science, 2000. See also: http://www.is.informatik.uni-kiel.de/∼thalheim/slides.htm. [Tha01]

B. Thalheim. Abstraction layers in database structuring: The star, snowflake and hierarchical structuring. Technical Report Preprint I-13-2001, Brandenburg University of Technology at Cottbus, Institute of Computer Science, 2001. See also: http://www.is.informatik.unikiel.de/∼thalheim/slides.htm.

[W3C02] W3C. W3C Note 25, An RDF schema for P3P. http://www.w3.org/TR/2002/NOTE-p3p-rdfschema20020125/, January 2002. [Wes67]

A.F. Westin. Privacy and Freedom. Atheneum, New York, 1967.

[WLT05] M. Warkentin, X. Luo, and G.F. Templeton. A framework for spyware assessment. CACM, 48(8):79– 84, August 2005.

Acknowledgement. The authors thank the reviewers for their comments and Joachim Biskup for his advices that led to improvement of the paper.


112

Description Logics, Data Modelling and Relational Dependencies Tapio Niemi1 , Marko Niinimäki2 and Santtu Toivonen3 Institute of Physics at CERN, CH-1211 Geneva, Switzerland 2 NorduGrid at CERN, CH-1211 Geneva, Switzerland 3 VTT Information Technology, P.O.Box 1203, FIN-02044 VTT, Finland Email: [email protected], [email protected], santtu.toivonen@vtt.fi 1 Helsinki

Abstract. Description Logics are a well-known formalism used in artificial intelligence. We present an approach where a Description Logic is used as an intuitive modelling language. Unlike many other modelling methods, this provides a sound semantic basis for modelling. We indicate that the database implementation is effective and unambiguous, since there is a direct mapping between our formalism and dependencies in the relational model. We also give an example applying the method to the Semantic Web.

1 Introduction According to Nebel in [29], Description Logics (DL) describe the domain of interest in terms of concepts, which represent classes of individuals, and roles, which are binary relations used to specify properties or attributes of individuals as well as links with other individuals. Description Logics has been applied to artificial intelligence and information/database modelling. A novel application of DL is the Semantic Web [23, 3]. The language used for this representation is often called a concept language (as in [14]). The expressions of such language describe concepts and roles. The language together with a computer program that stores these expressions and derives information based on it, is called a Terminological Knowledge Representation System. The following informal example (adapted from [14]) describes a typical domain of interest both verbally and visually. In Figure 1, the ellipses represent concepts and arrows between them represent “terminology”: for example the link between Mother and Parent says that mothers are parents; this is sometimes called an IS-A relationship. Moreover, as Nardi and Brachman state in [28] “the concept of Parent has a property that is usually called a role, expressed by a link from the concept to a node for the role labeled hasChild. The role has what is called a value restriction, denoted by the label v/r, which expresses a limitation on the range of types of objects that can fill that role. In addition, the node has a number restriction expressed as (1,NIL), where the first number is a lower bound on the number of children and the second element is the upper bound, and NIL denotes infinity.”

T. Niemi et al. / Description Logics, Data Modelling and Relational Dependencies

113

v/r hasChild

Person

Female

(1.nil)

Woman

Parent

Mother

Mother is a subtype of Parent. Mother is a subtype of Female. Woman is a subtype of Female. Woman is a subtype of Person. Parent is a subtype of Person. Parent has 1 or more children that are Persons.

Figure 1: Person example

Figure 1 is a “typical” Description Logic example, where IS-A relationships are dominant. For instance, we can say that Parent IS-A Person that has at least one child. In figure 2 we present another example where relationships other than IS-A dominate. The “has” relationships in the figure will be given explicit semantics later. In this paper we discuss how Description Logics can be used as a database modelling language. A review of Description Logics can be found in Section 2. In Section 3, we present an unambiguous mapping between Description Logic constructs (concept - concept relationships and concept - role relationships) and database constructs (based on dependencies in relations). This of course makes database semantics obvious for people who know about Description Logics. To benefit database designers without background in Description Logics, a simple graphical presentation is introduced in Section 4. The Section after that discusses the Semantic Web as an application area for description logics, and is followed by conclusions and an outlook on the future work. Related Work “Ontologies”, i.e. formal conceptualizations, and their presentation by Description Logics have been described in many papers since [21]. In [8], Borgida and Brachman present a method of abstracting concepts from the Universe of Discourse (UoD) and then expressing the structure the UoD using Description Logics. However, they do not present a graphical formalism related to the methodology, nor a relation between description logics and databases. These are discussed by Borgida, Lenzerini and Rosati in [9], where they present a mapping between Description Logics and the Entity Relationship Model and use Description Logics as a database query language. In [11], Calvanese et al. present mappings of Description Logics with semantic and object oriented data models. However, as far as we know, the approach of combining Description Logics with relational model dependencies, and providing a graphical formalism for this kind of design has not been used before. Different approaches to data modelling in general are numerous. Data models in general are discussed by Elmasri and Navathe in [17]. The Entity-Relationship formalism (The ER Model, or simply ER), presented by Chen in [12] is an influential semantic data model that popularised many notions (entities, attributes, relationships,..) that are still commonly used. For instance, the popular “Unified Modelling Language”, UML, employs some ER conven-

114


company

customer person

has−product has−customer

has−employee product

has−employee

has−component

employee component has−salary salary

has−component−number component−number

Employee is a person. Company is related to employee-product. Employee-product is related to employee. Employee-product is related to product. Employee-product is related to salary. Product is related to component. Component is related to component number.

Figure 2: Company example

tions; for details, see e.g. [25]. Several extensions to ER have been suggested; e.g. Extended Entity Relationship formalism (EER), discussed in [4], where subclass-superclass relationships can be expressed. Though famous, ER is not the only semantic data model. Other merited ones include IFO (see [1]), whose semantics is based on database updates and is thus very well defined; SDM (see [22]) for its relative “semantical relativism”, where it is not very important if something is a relationship or an attribute1; NIAM (see [33]), and more recently its successor Object Role Modelling (ORM) [31]. Most of these modelling languages employ a method by which the model can be translated into a database schema (see e.g. [17], [1]). In that respect, our approach is not a novel one. However, Duzi in [16] discusses the criteria for conceptual modelling languages and explicitly mentions expressibility, clarity, semantic stability, semantic relevance, validation mechanisms, abstraction mechanisms and formal foundation. We think that by combining Description Logics and modelling, these criteria can be matched. 2 Description Logics and Concept Languages In the 1960’s, so-called semantic networks were introduced as knowledge representation formalisms to express “word meanings” (see [32]). Though initially successful, these formalisms were semantically vague. Historically, Description Logics were devised to overcome their semantic ambiguities. Their main features include (i) a language by which the user can state concept descriptions, thus forming a terminology; (ii) a “classifier”, a method implemented in a computer program by which subconcept - superconcept relationships can be determined, thus allowing the terminology to form a (concept) subsumption hierarchy (see [2]). This 1 Hammer

and McLeod in [22] use the term “relative viewpoint”.


115

normally takes the form of a lattice with a top concept indicated by and bottom concept indicated by . From the point of view of predicate logics, it is not too difficult to device a “language” by which one can express the example in Figure 1. We can simply see concepts (like Person) as unary predicate symbols and roles (like hasChild) as binary predicate symbols. Using standard predicate logic conventions, we can write: Person:

Person x

Parent is a subtype of Person:

x Parent x

Person x

Parent has a child:

x Parent x

y hasChild x y

All children of any parent are persons:

x Parent x

y hasChild x y

Parent is a person who has a child who is a person:

x Parent x

Person x

Person y

y hasChild x y

Person y

However, it has been customary to introduce a “shorter” syntax for Description Logics, by omitting variables. We do this by examining the expressions used in the example in Figure 1. Atomic concepts:

PERSON, FEMALE, WOMAN, PARENT. To these we add (represents the top of subsumption hierarchy) and (represents the bottom of the hierarchy).

Atomic roles:

hasChild.

Value restrictions:

hasChild PERSON (all children are persons).

Existential quantification:

hasChild (states the role “hasChild”, normally used in an intersection, see below).

Intersection:

PERSON

Negation:

for example

hasChild

(a person who has a child).

FEMALE.

We omit the formal syntax here, but state that atomic concept or constructs builds using atomic concepts, atomic roles, value restrictions existential quantification, intersection and negation are called concepts. Terminological axioms make statements about how concepts are related to each other. They take the form of inclusions (e.g. WOMAN PERSON) or equalities (e.g. PARENT PERSON hasChild , meaning “Parent is a person who has a child”). The terminological axioms that can be seen in Figure 1 can be presented as follows: MOTHER

PARENT

MOTHER

FEMALE

WOMAN

FEMALE

WOMAN

PERSON

PARENT

PERSON

hasChild

hasChild.PERSON.

116


FEMALE

PERSON WOMAN

PARENT

MOTHER

Figure 3: Concept lattice

We define the semantics of the language like in predicate logics, as follows: Interpretation I consists of a non-empty set ΔI and an interpretation function which assigns to every atomic concept A a set AI ΔI and to every atomic role R a binary relation RI ΔI ΔI . The semantics of value restrictions, existential quantifications, intersections and negations are “natural” for the point of view of logics; similarly I ΔI and I 0/ (for details, see e.g. [10]). We say that an interpretation I satisfies C D if C I DI . I satisfies C D if C I DI . If T is a set of terminological axioms and I satisfies each of them, we say that I is a model of T . Subsumption can be formally defined as follows: Let T be a set of terminological axioms. A concept C is subsumed by a concept D with respect to T if C I DI for every model I of T . Subsumption is important since the following properties can be reduced to it (see [2]): C is unsatisfiable C is subsumed by . C and D are equivalent C is subsumed by D and D is subsumed by C. C and D are disjoint C D is subsumed by . Typically the classifier of a description logic system functions in a way that the user enters terminological axioms and the classifier builds the subsumption hierarchy, normally a lattice. The lattice (apparently) formed of the concepts of Figure 1 is shown in Figure 3. The expressive power of the language affects the performance of the classifier. This has been the source of extensive studies e.g. in [15]. However, for the type of language presented in this paper, the subsumation algorithm is known to be decidable. Intuitively, a set of terminological axioms is called a DL schema. However, a proper definition of a DL schema is based on our modelling language in Section 4. 3 Database Dependencies and DL-modelling We apply a conceptual modelling method based on dependencies of the relational database theory [30]. According to Fagin “dependencies” are first-order statements on database relations [18]. In the dependency based modelling method, universal relation with null values is used as the abstract model for the application area. The relation contains all the concepts as its attributes. Values of attributes can be thought of as identifiers of concepts, however, we do not assume the existence of real object identifiers. Although the universal relation is not a realistic choice for the database schema, it gives a clear theoretical base by simplifying definitions of conceptual relations and enables derivation of interference rules. The idea is that one row represents only one binary relationship between instances if the relationship is functional or so-callad free binary relationship. Attributes not


117

participating in the relationship have null values in the row. A null value means that the value does not exist. A relationship is valid only if all attributes participating in the relationship have real values. Therefore, we do not have problems with null values in dependencies. In addition, we assume that unknown-value nulls do not exist in the application area. In the case of inclusion dependencies, this is naturally slightly different, since an inclusion dependency is a relationship among domains of attributes. Therefore, tuples representing an inclusion dependency have only one non-null attribute value. As mentioned earlier, dependencies are statements about relations. A dependency is valid in a relation if the statement is true in this relation. On the other hand, dependencies can be seen as relationships among attributes. In this work we only focus on binary dependencies, i.e. dependencies whose left and right hand side contain only one attribute. This is a natural assumption, since both the IS-A relationship and the role in Description Logics are binary relations. For conceptual modelling binary relations are enough, since possible higher degree dependencies can be represented by binary relationships by creating a new concept for the relationship. The definitions of our dependency relationships are based on the relational model, thus they come from the extensional level. However, we consider them as intensional information on the conceptual level. Their semantics denote that if there is a database instance, then the instance has to obey the given dependencies. The idea is that if there is some relationships (dependency) on the instance level between the instances of some concepts, then there is a dependency between these concepts, i.e attributes on the conceptual level. As mentioned earlier, a Description Logic contains two types of dependency relations: ISA and the role. The IS-A dependency corresponds to the relational inclusion dependency. The role relation of DL is divided into different dependencies according to the cardinalies of the concepts: the key dependency, the functional dependency, and the free binary dependency. In this way, the relationships of the Description Logic have a one-to-one mapping to the dependencies of the relational model: the one-to-many role relationship corresponds to the functional dependency, the one-to-one role relationship to the key dependency, i.e. symmetric functional dependency, and the many-to-many role relationship to the free binary dependency. Next, we give definitions to the functional dependency ( ), the key dependency (KD), the free binary dependency (FBD), and the IS-A (inclusion) dependency on the basis of the dependencies in the relational model [Cod72a]. These dependencies are a basis of mapping Description Logics schemata to relational databases. In our formalism the set of all concepts is denoted by C. The concepts are the same as in the corresponding Description Logic. The ’world’ is considered to be a universal relation r with attributes C. Definition 1 (Binary Functional Dependency). Let A B C. A relation r over C satisfies the functional dependency A B, if for any tuples t t r, where t A null t B null, and t A t A , then t B t B . Lemma 1. Binary functional dependency is reflexive and transitive. Proof. Follows from the properties of the functional dependency in the relational model. A key is a concept that can identify the instances of another concept. In the relational model the key is defined through the concept of the super key: Let R U be a relation schema. t S , where A set of attributes S U is a superkey if a relation r over R cannot have t S t and t are distinct tuples in r. A superkey is a key if it is minimal, i.e. no attribute can be

118


removed without violating the superkey property. If S is a superkey of a relation r over R U , then S U , and, trivially, U S. A functional dependency is called a key dependency if its left hand side is a superkey. The definition above could be used for the key dependency here. However, it is more practical to define the key dependency in a different way: it is simply a symmetric functional dependency. Thus, the key dependency, denoted by KD, is defined as follows: Definition 2 (Binary key dependency). Let A B satisfies A KD B if r satisfies A B and B A.

C and r be a relation over C. Now, r

Lemma 2. Binary key dependency is reflexive, transitive and symmetric. Proof. The proof follows from the properties of the functional dependency and the definition of the binary key dependency. The correspondence of IS-A relationship of the Description Logic is straightforward to the inclusion dependency of the relational model. Definition 3 (Binary inclusion dependency). Let A B C and r is a relation over C. Now, there is an inclusion dependency A B if for each tuple t r there exists a tuple t r such that t A t B . The inclusion dependency can also be defined using the projection operator of the relational algebra: There is an inclusion dependency between A and B, if π A r πB r . Lemma 3. The inclusion dependency is reflexive, transitive, and antisymmetric. Proof. The subset relation is reflexive, transitive and antisymmetric. The free binary relationship denotes a connection between concepts, which cannot be expressed by functional or IS-A dependencies. Definition 4 (Free binary dependency). Let A B C, and r is a relation over C. There is a free binary dependency between A and B (A FBD B) if there exists at least one tuple t r, such that both t A and t B have non-null values. It follows from the definition that the functional dependency is a special case of free binary dependency. A free binary dependency (FBD) between A and B corresponds to the following Description Logic expression about concepts A and B: A

b B.

Thus, it is possible that role b of A has a null value, but there is at least one B that stands in relation “b” with A. Theorem 1. The equivalence of free binary relationship of Description Logics and the free binary dependency of the relational model. Let C be the set of concepts and A B

C. If and only if A

B

, then A FBD B.

119


Proof. Let IC be the set of all possible instances of the concepts in C, i.e. I C = D C dom D , and, in the same way, I A =dom(A) and I B =dom(B). Further, let r be a relation over C. Clearly, all values of the attributes of r belong to I C . We first assume A B . This denotes that for some tuple t r such that t A null, there exists t[B] null. That is, according to the definition of the free binary dependency A FBD B. Next, we assume A FBD B. Now, there exists a tuple t r such that t A null and t[B] null. For that we get A B . A functional dependency between A and B corresponds to the following Description Logic expression about concepts A and B (“functional relationship of Description Logics”): A

bB

bB

This means that A implies B, for instance, a city (A) implies a zip code (B), or an that an employee exists implies that there is a salary for him/her. Theorem 2. The equivalence of the functional relationship of Description Logics and the functional dependency of the relational model.

A

Let C be the set of concepts, A B b B b B, then A B.

C and b a role (a relation in C

C). If and only if

Proof. Let IC be the set of all possible instances of the concepts in C, i.e. I C = D C dom D , and, in the same way, I A =dom(A) and I B =dom( B). Further let r be a relation over C. Clearly, all values of the attributes of r belong to I C . We first assume A bB b B. Now, for all tuples t inr such that t A neg null, there exists t B neq null, and if t[A]=t’[A], then t[B]=t’[B]. Thus, A B according to the definition of the functional dependency. Next, we assume A B. According to the definition of the functional dependency, we get that if t[A]=t’[A], then t[B]=t’[B]. Therefore, A b B b B. An inclusion dependency of the relational model naturally correponds with the IS-A relation in Descrption Logics, i.e. A B. For instance, EMPLOY EE PERSON. Theorem 3. The equivalence of the subsumption of Description Logics ( sion dependency of the relational model ( RM ). Let C be the set of concepts and A B

C. If and only if A

DL

DL )

B, then A

and the inclu-

RM

B.

Proof. Let IC be the set of all possible instances of the concepts in C, i.e. I C = D C dom D , and, in the same way, I A =dom(A) and I B =dom( B). Further let r be a relation over C. Clearly, all values of the attributes of r belong to I C . We first assume that A DL B. Now, AI BI according to the definition of the subsumption. Let t be an arbitrary tuple in r. Now, t A AI . Since AI BI , then t A BI . Thus, A RN B. Next, we assume that A RN B. Now, AI BI and, further, A DL B according to the definition of the subsumption.

120


Now we have shown that it is possible to map each relationship in a DL schema into a dependency in a relational database. This makes it possible to present the information content of the DL schema using the relational model. In the following subsection we will show how to construct a relational instance to correspond a DL schema. The approach is based on Armstrong relations. 3.1 Relational Instance of a DL Schema The equivalence of the relationships in DL and the dependencies in relational model enables us to present a DL schema as a relation. The concepts of the DL schema will be attributes in the relation schema and the relationships among concepts are represented by tuples in the relation. This relational instance is actually the Armstrong relation [18]. An Armstrong relation obeys a given set of dependencies but nothing else. There exist algorithms for computing Armstrong relations for a set of functional dependencies, for example in [27, 6]. However, the computation time and the size of the minimum Armstrong relation for a set of functional dependencies can be of exponential size in number of attributes [6]. This obviously seldom makes generating Armstrong relations for DL schemata meaningful. In case of only functional dependencies the situation is easier. Next we give an algorithm (Algorithm 1) that constructs an Armstrong relation of IS-A relations of a DL schema. The relation instance contains minimal amount of instances of concepts but, for clarity, each tuple contains only one instance in our example. Thus, the amount of tuples is not minimal. Input: A DL schema with C is the set of concepts, D the IS-A relation among C. Output: A relation r over R(U) representing the IS-A relationships of the DL schema as inclusion dependencies of relational model. 1: 2: 3: 4: 5: 6: 7: 8: 9:

Set U := C, i:=0. Let r be a relation over R(U). for all A C do Add a new tuple t to r such that t[A]=Ii and t[U-A]=null. for all (A,B) D do Add a new tuple t to r such that t[B]=Ii and t[U-B]=null. end for i:=i+1. end for Algorithm 1: Constructing relation from DL schema

The algorithm constructs an instance for each concept. The instances are denoted by Ix . At the first phase a tuple is constructed for the instance at hand. In the tuple only the attribute corresponding the present concept has a non-null value. After that an additional tuple is created for each IS-A relationship in which the concept participate. This is repeated for each concept. (NB: set D represent the IS-A relation among concepts and the relation is transitive by definition, so we do not need to compute the closure first.) We use Figure 1 as an example. There, we can conclude the following relationships among the concepts.


121

IS-A relationships: MOTHER WOMAN MOTHER PARENT PARENT PERSON WOMAN PERSON WOMAN FEMALE Role-based relationships: parent - person Parent-person relationship is represented by a free binary dependency between parents and children, since it is neither an inclusion dependency nor functional dependency. In the following relational table each row has only one value. In principle, the table could have less rows, for example rows 16 and 17 could be combined but this would also mean that mother and parent have another relationship than the IS-A relationship represented by the inclusion dependency. Rows 1-5 indicate all entities, rows 6-9 persons, rows 10-12 females, 13-14 women, 15-16 parents, and 17 mothers. As we can notice, there is and inclusion dependency corresponding each IS-A relationship.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

all i1 i2 i3 i4 i5

person

female

woman

parent

mother

i1 i2 i3 i4 i1 i3 i5 i1 i3 i1 i2 i1

In Figure 1 in the first page, we have one role-based relationship in addition to IS-A relationships. Also this relationship can be represented by a relational instance. This is illustrated on the following table, in which rows 18-19 represent the parent-child relation. For instance, row 18 determines that person ’i3’ is a child of a parent ’i1’. all 18 19

person i3 i4

female

woman

parent i1 i1

mother

122


4 DL-modelling, Graphical Presentation Here, we propose simple textual and graphical conventions to present our DL modelling language. We call any set of expressions that conforms with the textual conventions a DL schema, and any diagram that conforms with the graphical conventions a DL diagram. The building blocks of a schema are concepts and binary relations between concepts. In the graphical convention the concepts are designated by an oval with a name of the concept; and the relations are designated by lines or arrows. The relations correspond to the IS-A and “has” relations in Figure 2, on one hand, and to IS-A, free binary relations and functional relation of Section 3 on the other hand. As Figures 4 - 7 indicate, we use a simple line to indicate a free binary relationship; an open-ended arrow for a functional relationship; a two head open-ended arrow for a key relationship; and an arrow for an IS-A relationship. A free binary relationship: free binary relation(product,customer) product has−customer customer

Figure 4: A free binary relationship

A functional relationship: functional relation(employee,salary) employee has−salary salary

Figure 5: A functional relationship

A key relationship: key relation(employee,employee ID) employee has−ID employee ID

Figure 6: A key relationship


123

An IS-A relationship: is a relation(employee,person) person

employee

Figure 7: An IS-A relationship

5 Applying Description Logics: the Case of Semantic Web An important application area for description logics is the Semantic Web. The Semantic Web is an extension to the current web, in which the content is given machine-undersandable meaning [7]. Perhaps the most essential building blocks of the Semantic Web are ontologies, i.e., formal conceptualizations of some domains [20]. The ontologies are intended for knowledge sharing in an agreed-upon manner in order to enable this machine-understandability. Web Ontology Language (OWL) is intended for expressing ontologies in the Semantic Web [5]. Being a descendant of DAML+OIL [23] and especially OIL [19], OWL has strong influence from the description logics community. Some examples considered in the previous sections could be defined in OWL and serialized in XML as follows: 1

The first class definition defines the concept of Mother as being a subclass (earlier in this paper referred to as subtype) of Parent. Next, a property hasChild is defined by specifying, that it can occur in the domain of Parent and receive values from the range of Person. Finally, the class of Parent is defined first by saying that it is a subclass of Person. In addition, a restriction is set to the hasChild property, when it appears in the domain of Parent. The restriction declares that a parent must have at least one child by using the owl:minCardinality property with the value 1.

124


Semantic Web as an open system provides an excellent test platform for the applicability of description logics. Anyone can create their own ontologies and put them in the Semantic Web just as they put their home pages on the current web. To really make use of the description logics based ontologies, however, reasoning engines capable of handling them must be implemented. FaCT and Racer are examples of such reasoners [24, 3]. 6 Summary and Future Work We have presented a mapping between Description Logics and dependencies in the relational model. This enables us to use a Description Logic as a database modelling language. The result of the modelling activity, i.e. a schema, is therefore semantically well established (by the semantics of Description Logics) and can be immediately implemented in relational database design and/or by a relational database system: The relational presentation enables us to map a DL schema straightforwardly into a relational database schema. This can be simply applying well-known normalisation algorithms for relational databases. Other possible applications are Semantic Web and integration of heterogenous data sources. For designing the DL schema we have presented a language that has a simple syntax and illustrative graphical notation. An interesting application of mapping Description Logics to relational model is a possibility to apply inference rules of relational dependencies. In case of only unary inclusion and functional dependencies as in our model, this problem is much easier than in the case of more complex dependencies [26, 13]. In this way it is possible to study using solid relational theory how conceptual level dependencies interact, i.e. what new dependencies can be derived from the given ones. Acknowledgements The work was partly supported by the Academy of Finland, project 203554. References [1] S. Abiteboul. and R. Hull. IFO: A formal semantic database model. ACM Transactions on Database Systems, 12(4), 1987. [2] F. Baader and W. Nutt. Basic description logics. In F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. Patel-Schneider, editors, The Description Logic Handbook. Cambridge University Press, 2003. [3] Franz Baader, Ian Horrocks, and Ulrike Sattler. Description logics as ontology languages for the semantic web. In Dieter Hutter and Werner Stephan, editors, Festschrift in honor of J o¨ rg Siekmann, Lecture Notes in Artificial Intelligence. Springer, 2003. [4] C. Batini, S. Ceri, and S. Navathe. Conceptual Data Base Design: An Entity-Relationship Approach. Benjamin/Cummings, 1992. [5] S. Bechhofer et al. OWL Web Ontology Language Reference. World Wide Web Consortium, February 2004. W3C Recommendation, available at: http://www.w3.org/TR/owl-ref/. [6] C. Beeri, M. Dowd, R. Fagin, and R. Statman. On the structure of armstrong relations for functional dependencies. J. ACM, 31(1):30–46, 1984. [7] Tim Berners-Lee, James Hendler, and Ora Lassila. The semantic web. Scientifi c American, May 2001.


125

[8] A. Borgida and R. J. Brachmann. Conceptual modeling with description logics. In F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. Patel-Schneider, editors, The Description Logic Handbook. Cambridge University Press, 2003. [9] A. Borgida, M. Lenzerini, and R. Rosati. Description logics for data bases. In F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. Patel-Schneider, editors, The Description Logic Handbook. Cambridge University Press, 2003. [10] A. Borgida and P. F. Patel-Scheider. A semantics and complete algorithms for subsumption in the classic description logic. Journal of Artifi cial Intelligence Research, 1, 1994. [11] D. Calvanese, M. Lenzerini, and D. Nardi. Description logics for conceptual data modelling. In J. Chomicki and G. Saake, editors, Logics for Databases and Information Systems. Kluwer Academic Publisher, 1998. [12] P. Chen. The entity-relationship model - towards a unified view of data. ACM Transactions on Database Systems, 1(1), 1976. [13] S. S. Cosmadakis, P. C. Kanellakis, and M. Y. Vardi. Polynomial-time implication problems for unary inclusion dependencies. J. ACM, 37(1):15–46, 1990. [14] F. M. Donini, M. Lenzerini, and D. Nardi. The complexity of concept languges. In Proceedings of the 2nd Int. Conf on Principles of Knowledge Representation and Reasoning, Cambridge, MA, 1991. [15] F.M. Donini, M. Lenzerini, D. Nardi, B. Hollunder, W. Nutt, and A. M. Spaccamela. The complexity of existential quantification in concept languages. Artifi cial Intelligence, 53(2/3), 1992. [16] M. Duzi. Logical foundations of conceptual modelling. A manuscript, 2000. [17] R. Elmasri and S. Navathe. Fundamentals of Database Systems. Benjamin/Cummings, 2nd edition, 1994. [18] R. Fagin. Horn clauses and database dependencies. J. ACM, 29(4):952–985, 1982. [19] Dieter Fensel, Ian Horrocks, Frank van Harmelen, Stefan Decker, Michael Erdmann, and Michel C. A. Klein. Oil in a nutshell. In Proceedings of the 12th European Workshop on Knowledge Acquisition, Modeling and Management, pages 1–16. Springer-Verlag, 2000. [20] Thomas Gruber. A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2):199–220, 1993. [21] N. Guarino. Understanding, building and using ontologies. International Journal of Human Computer Studies, 46(2/3), 1997. [22] M. Hammer and D. McLeod. Database description with SDM: A semantic database model. ACM Transactions on Database Systems, 6(3), 1981. [23] I. Horrocks. DAML+OIL: a description logic for the semantic web. Bull. of the IEEE Computer Society Technical Committee on Data Engineering, 25(1):4–9, March 2002. [24] Ian R. Horrocks. Using an expressive description logic: FaCT or fiction? In Anthony G. Cohn, Lenhart Schubert, and Stuart C. Shapiro, editors, KR’98: Principles of Knowledge Representation and Reasoning, pages 636–645. Morgan Kaufmann, San Francisco, California, 1998. [25] I. Jacobson, G. Booch, and J. Rumbauch. The Unifi ed Software Development Process. Addison-Wesley, 1999. [26] P. C. Kanellakis, S. S. Cosmadakis, and M. Y. Vardi. Unary inclusion dependencies have polynomial time inference problems. In Proceedings of the fi fteenth annual ACM symposium on Theory of computing, pages 264–277. ACM Press, 1983. [27] H. Mannila and K. Raiha. Test data for relational queries. In Proceedings of the fi fth ACM SIGACTSIGMOD symposium on Principles of database systems, pages 217–223. ACM Press, 1986. [28] D. Nardi and R. J. Brachmann. An introduction to description logics. In F. Baader, D. Calvanese, D. McGuinness, D. Nardi, and P. Patel-Schneider, editors, The Description Logic Handbook. Cambridge University Press, 2003.

126


[29] B. Nebel. Terminological reasoning is inherently intractable. Artifi cial Intelligence, 1990. [30] T. Niemi, J. Nummenmaa, and P. Thanisch. Applying dependency theory to conceptual modelling. In Topics in Conceptual Analysis and Modeling, pages 271–290. Czech Academy of Sciences’ Publishing House Filosofia, 2000. [31] G. Nijssen and T. Halpin. Conceptual Schema and Relational Database Design: a fact oriented approach. Prentice-Hall, 1989. [32] D. Mac Randal. Semantic networks. In G.A. Ringland and D.A. Duce, editors, Approaches to Knowledge Representation. Research Studies Press, 1988. [33] G. Verheijn and J. van Bekkum. Niam: An information analysis method. In T.W. Olle, H.G. Sol, and A. Verrijn-Stuart, editors, Information Systems Design Methodologies: A Comparative Review (CRIS ’82 Proceedings). Elsevier Science Publishers, 1982.


127

Modelling, searching and utilising Grid resources and services Marko Niinimäki 1, Santtu Toivonen2, Tapio Niemi3 1 NorduGrid, University of Oslo, Norway 2 VTT Information Technology, Finland 3 Helsinki Institute of Physics, CERN, Geneva, Switzerland [email protected], santtu.toivonen@vtt.fi, [email protected]

Abstract. As Grid technologies become more mature, tools and methods are needed to enable interoperability and common understanding of the components in different Grid implementations. In this paper, we survey the existing efforts (GLUE, OGSA), and unify them with emerging WWW technologies. We concentrate on modelling the structure of a Grid computing element (CE) and propose a design of a Grid resource broker so that it could make better use of the information about the CEs when user requests some Grid resource or service. We expand our view to consider Grid services and data management, too.

1 Introduction As Foster et al. state in [1], “Grid systems and applications aim to integrate, virtualize, and manage resources and services within distributed, heterogeneous, dynamic ”virtual organizations”.” In order to do so, these systems and applications need information, in some standardized or at least agreed-upon form of description, about the resources and services. Several technical frameworks have been proposed for this purpose, for instance Monitoring and Discovery Service MDS [2] used by Globus Grid software, or Abstract Job Objects used by Unicore software (see [3]). A bit less on the technical level and explicitly addressing the resource description, there are Grid Laboratory Uniform Environment GLUE schemata (see [4]). In this paper, we survey these frameworks from the perspective of modelling; in other words, we study what features of the resources and services are considered relevant, and how they are represented. based on this, we present a practical resource broker design. Related work relevant in this context includes [5], where integrating GLUE schemata with Unicore’s Abstract Job Objects is discussed and The Universal Description, Discovery and Integration (UDDI) protocol, that enables data source and service interoperability in the World Wide Web (see [6]). Resource Description Framework (RDF) [7] and Web Ontology Language (OWL) [8] are World Wide Web Consortium recommendations for the Semantic Web and previously applied to Grid context for instance in [9]. In [10], the authors describe an integration of OWL and UDDI via OWL-S, which is an OWL-ontology for describing Web services [11]. Moreover, on the field of Semantic Grid some research related to discovering Grid resources is done, e.g. ([12, 13, 14, 15]). Tangmunarunkit et al. in [14] introduce an ontology based resource matchmaker for the Grid. The idea is that the use of Semantic

128

M. Niinimäki et al. / Modelling, Searching and Utilising Grid Resources and Services

Web technologies relaxes the demand of exact syntax matching but “semantical” matching is enough. This is very practical in decentralised Grid environments. The rest of the paper is organized as follows. In Section 2, we describe our domain of application, namely, what is Grid and what are Grid resources and services. In Section 3 we present GLUE’s description method, and proceed to RDF and OWL that we consider to be suitable tools for giving Grid descriptions and providing capabilities for reasoning about them. In Section 4, we describe our design of a Grid search system (Resource Broker). Section 5 contains a summary and discussion. 2 Background 2.1 Grid In a recent paper [16], Bote-Lorenzo et al. define Grid as a “large-scale geographically distributed hardware and software infrastructure composed of heterogeneous networked resources owned and shared by multiple administrative organizations which are coordinated to provide transparent, dependable, pervasive and consistent computing support to a wide range of applications.” In practical terms, Grid can be seen as an approach to Internet based distributed computing that promotes standards to combine resources (like computing power or storage capacity) and services (like specific information sources X and Y, or an application to compute new information based on X and Y) wrapped in a authentication and authorization method that verifies the users’ identity and authorizes them to use the resources and services, based directly on that identity, or indirectly by the user’s membership in some organization or group (see e.g. [17]). In order to implement this approach, several Grid projects and consortiums have been launched, among them the Globus Alliance1 , The Legion Project of the University of Virginia2 , The Condor Project of the University of Wisconsin-Madison 3, the Unicore project of Central Institute of Applied Mathematics, Forschungszentrum Julich in Germany 4 , Nordic NorduGrid5 , the transatlantic collaboration DataTAG6 , the European Union funded DataGrid project7 and its follow-up EGEE (Enabling Grids for E-science in Europe)8 . These, in turn, either promote the use of Grid technologies, establish Grid platforms, or develop Grid software. Grid software packages, so-called middle wares include Globus, Legion, Condor, Unicore, the EDG suite, and NorduGrid’s Advanced Resource Connector ARC. Unfortunately, different middle wares rarely communicate with each other. For instance, if organization A uses Globus software to provide Grid services, until recently organization B that uses ARC could not have queried or used those services. To enable interoperability, standards are needed. The Global Grid Forum (GGF, http://www.gridforum.org) promotes standards; the first large set of recommendations being prepared by GGF is called Open Grid Services Architecture (see [1]). 1 http://www.globus.org 2 http://legion.virginia.edu 3 http://www.cs.wisc.edu/condor 4 http://www.fz-juelich.de/unicore 5 http://www.nordugrid.org 6

http://datatag.web.cern.ch

7 http://www.eu-datagrid.org 8

http://www.eu-egee.org


129

Figure 1: An XRSL description of a job

2.2 Grid resources and services Most, if not all, Grid middle wares enable the user to execute a job in a remote computer. The job is a computer program ”executable”, or a description that expresses what to execute; and what are the input and output and if there are some specific resource requirements (see Figure 1).9 Following this stereotypical example, the job description of Figure 1 could be stored in a file called test.xrsl, and submitted by the user to a specific computer, say, computer1.example.org by using the middle ware’s job submission command. Assuming the user is authenticated and authorized by computer1.example.org, the job would run, and its results could be later retrieved by the user. However, Grid users are probably not interested in, or would not even know the name of the computer where their job is to be executed; it is quite sufficient for them to have it executed somewhere and recover the results quickly. For this purpose, a system working on behalf of the user needs information about the requirements of the job (e.g. memory, operating system and disk space) and the resources available in different computers available to the user. Usually this implies, too, that the grid service that allocates jobs to computers needs to know something about the user; not all computers are available to all users. In the context of this paper we consider Grid resources simply as anything a Grid user might be interested in. Grid services are standardized or at least published ways of accessing these resources in the Grid community in question. Thus, disk space is a resource, but a Grid method of putting files or records on a disk, or recovering them, is a service. In the following sections, we concentrate on Grid resource descriptions, but it should be noticed that 9 XRSL stands for eXtended Resource Specifi cation Language, see [18] for details. It should be noticed that different Grid implementations have incompatible specifi cation languages for jobs; however, JSDL (see [19]) may emerge as a standard.

130


for services, our approach can be combined with Web Service Definition Language (WSDL, see http://www.w3.org/TR/wsdl), that is already utilized in Grid. 3 Resource description In [4] DataTAG’s GLUE project’s goals are described as follows. “The Glue-Schema activity aims to define a common conceptual data model to be used for grid resources monitoring and discovery.” GLUE 1.1 divides Grid into computing (Computing Element, CE), storage (Storage Element, SE) and network. Slightly surprisingly, there is no separate concept of a “principal” (a user, a group of users or a virtual organization). Figure 2 shows GLUE’s CE description in UML class diagram format. World Wide Web Web Consortium’s Web Ontology Language (OWL) was promoted to a recommendation in February 2004 [8]. It is based on Resource Description Framework (RDF) [7, 20, 21], which provides a set of conventions for representing information in the World Wide Web. The most basic method of representation is stating that something (a subject) has a ”property” (a predicate) whose value is something (an object, see [20]); for instance document S has a property called ”creation date” whose value is ”16th August 1999”. With RDF, one can represent aggregates (like someone having an address that consists of predicates such as ”street address”, ”zip code”, and ”country”), sets (for example a document having many editors), as well as types and class hierarchies (someone having a predicate ”type” with an apparent value of ”human being”, whereas ”human being” can be represented as a subclass of ”mammal”). OWL is an extension of RDF. It enables more expressive descriptions for resources. Unlike RDF as such, OWL is capable for expressing various restrictions on properties attached to classes. For example, one can declare minimum, maximum, or exact cardinalities for the properties with regard to specific classes. In addition, one can define unions of classes, disjointness, concepts for ontology versioning, etc. [8] The syntax of RDF is typically defined in RDF/XML, which is a specific eXtensible Markup Language, XML [22]. XML as such can be described as a set of syntactic standards for ”markup”, i.e. expressing something by means of elements, subelements and attributes. A specific XML language defines what element names are valid for this particular language, and where they can appear (for instance some attribute names like ”person’s address” can only be applied to a ”person” element). The following example, adapted from [20], illustrates the syntax of RDF/XML. 1. 2. 4. 5. August 16, 1999 6. 7. There, lines 2 and 3 indicate that the XML name spaces to be used in the description are ”rdf” and ”exterms”. In the description, lines 4 and 5, we can see that a name space is attached in front of an XML element name (Description) or an attribute (creation-date).


Figure 2: GLUE CE description (to be continued on the next page)

131

132



133

Since a namespace actually states what we see as relevant elements and attributes in the domain of application, how they related and how they should be named, a namespace would serve well the purpose of describing Grid services and resources. Such namespace, expressed in a suitable language, would provide a unifying background, an ontology, for our Grid resource descriptions. Moreover, relying on the element and attribute names declared in the name space, we can effectively search through Grid RDF instances (e.g. in a set of Grid computing element RDF’s, select those that indicate the operating system to be Linux). XML documents are typically queried using Xpath [23], but a specific query languages RQL and RDQL has been created for RDF searching (see [24, 25]). We use OWL for describing the concepts found in Figure 2. We do not go into details of the ontology in this paper, but merely present some excerpts in Figure 3. There, one can see how the element names (for instance “OperatingSystem”) correspond to class names of Figure 2. The full ontology (the source of the excerpts) is available in http://wiki.hip.fi/xml/ontology/glue.xml. 4 Validating Grid resource descriptions and finding information in them In the previous section we “modelled” Grid and expressed the model in a sufficiently formal manner, in an OWL schema. A Grid service providers will be able to express (“publish”) their services by creating an instance of the schema. An excerpt of an instance satisfying the OperatingSystem part of the schema in Figure 3 in shown in Figure 4. In order to enable real and consistent provision and use of Grid services, instance validation is essential. To our knowledge, however, there are currently no tools available for explicitly validating instances against OWL ontologies. However, Jena project ([26]) has published software frameworks, using which such validator could be implemented. Let us assume that the providers of Grid services publish their Grid CE information as in Figure 4, and that a system (actually a service) known as a resource broker has access to these instances. Based on the technologies presented above, we can design the resource broker so that it answers the user’s question in Section 2.2: “Where can I run the job?” 10 As we indicated in Section 3, RDF and consequently OWL supports class hierarchies. As a trivial example of utilising such hierarchies, we consider the following case: it is sufficient for the user that his job is executed in a computer with Unix operating system. However, there are several Unix implementations and the name of the operating system (like “Linux”, “SunOS”, “MacOsX”) does not necessarily reveal the fact that the operating system belongs to Unix family. The fragment of a schema in Figure 5 expresses this for Linux.11 RDF descriptions can be queried using, e.g., RQL [24] that is an SQL based query language and can operate with concepts of RDF, i.e. classes, properties, etc. The syntax of RQL remains SQL with its SELECT, FROM, WHERE clauses. The example query in Figure 6 returns all Unix computers. In practice, the resource broker would function as shown in Figure 7. There, (1) the information service knows about the CE’s on the basic of the OWL instances that they publish. 10 In

current Grid systems, fi nding a “match” like that is called candidate set generation that a part of OGSA selection service (see [1]). Naturally, it relies on an information service like MDS. 11 For simplicity, we consider these IS-A statements a part of our ontology in http://wiki.hip.fi/xml/ontology/glue.xml.

134


.. 1 1 1

Figure 3: Excerpts of GLUE CE presented in OWL

"3.0" "Linux" "Debian"

Figure 4: A part of an instance


135

Figure 5: Stating that Linux is Unix

SELECT ?x WHERE (?x ), (?x ?y) AND (?y eq "Unix") USING glue for , rdf for ; Figure 6: A query retrieving Unix computers

The user has a job description that defines that he wants to run his job in a Unix computer (2). The resource broker combines the resource information with the information that Linux, SunOS and MacOsX are Unix operating systems (3) as in the query of Figure 6, and thus creates a set of candidates that excludes the computers running Windows. The set (or the best candidate, e.g. the one with least load) is sent to job scheduler (4) that executes the job. 5 Summary and conclusions In this paper, we have presented a framework, based on existing technologies for modelling, searching and utilising Grid resources. As an example, we use a case where a user wants to run a job in some Unix computer without the need for him to know technical details like computer or cluster name, or names of different Unix operating system names. To achieve this, we have used GLUE’s description of a computing element (CE) and rewritten it as an OWL schema. Instances conforming the schema can be made available and they contain information about actual CE’s. These instances can be accessed by a resource broker in order to find a match for the user’s requirements. Since most Grid middle ware packages contain a tool for job submission, our design can be used for unifying the job submission interface. However, for future development, we see much potential in describing Grid data sources, too. Acknowledgements The authors wish to thank Sergio Andreozzi of the GLUE team for his permission to use Fig. 2. This work was partly supported by the Academy of Finland, project 203554.

136


Figure 7: Resource broker design

References [1] Foster, I., Kesselman, C., Nick, J.M., Tuecke, S.: The physiology of the Grid. (Draft 1.0 as of July 2004, available at http://www.globus.org/research/papers/ogsa.pdf) [2] Czajkowski, K., Fitzgerald, S., Foster, I., Kesselman, C.: Grid information services for distributed resource sharing. In: Proceedings of the Tenth IEEE International Symposium on High-Performance Distributed Computing (HPDC-10), IEEE Press (2001) [3] Erwin, D., ed.: Unicore Plus Final Report - Uniform Interface to Computing Resources. Unicore Forum, Forschungszentrum Julich (2003) [4] Andreozzi, S.: GLUE - DataTag Project, work package 4: Interoperability between Grid domains. (Available at http://server11.infn.it/datatag/wp4) [5] Brooke, J., Fellows, D., Garwood, K., Goble, C.: Semantic matching of Grid resource descriptions. In: Proceedigs of the 2nd European Across-Grids Conference (AxGrids 2004), 28-30 January 2004, Cyprus. (2004) [6] UDDI.org: UDDI techical white paper. Technical report, UDDI (2000) Available at http://www.uddi.org/pubs/Iru UDDI Technical White Paper.pdf. [7] W3C: Resource Description Framework (RDF), The World Wide Web Consortium (2004) Available at http://www.w3.org/RDF/. [8] Bechhofer, S., et al.: OWL Web Ontology Language Reference, World Wide Web Consortium (2004) W3C Recommendation, available at: http://www.w3.org/TR/owl-ref/. [9] Pouchard, L., Cinquni, L., Strand, G.: The earth system Grid discovery and semantic web technologies. In: Semantic Web Technologies for Searching and Retrieving Scientifi c Data, ISWCII. (2003) [10] Paolucci, M., Kawamura, T., Payne, T., Sycara, K.: Importing the semantic web in UDDI. In Bussler, C., et al., eds.: Web Services, E-Business, and the Semantic Web, CAiSE 2002 International Workshop, WES 2002, Toronto, Canada, May 27-28, 2002, Revised Papers. Volume 2512 of Lecture Notes in Computer Science., Springer (2002) 225–236 [11] Martin, D., et al.: Bringing semantics to web services: The OWL-S approach. In: Proceedings of the First International Workshop on Semantic Web Services and Web Process Composition (SWSWPC 2004), San Diego, CA (2004)


137

[12] Harth, A., Decker, S., He, Y., Tangmunarunkit, H., Kesselman, C.: A semantic matchmaker service on the grid. In: Proceedings of the 13th international World Wide Web conference on Alternate track papers & posters, ACM Press (2004) 326–327 [13] Ludwig, S., van Santen, P.: A grid service discovery matchmaker based on ontology description. In: EuroWeb 2002 Conference. (2002) [14] Tangmunarunkit, H., Stefan Decker, C.K.: Ontology-based resource matching in the grid the grid meets the semantic web. In: 1st Workshop on Semantics in Peer-to-Peer and Grid Computing. (2003) [15] Brooke, J., Garwood, K., Goble, C.: Interoperability of grid resource descriptions: A semantic approach. In: The First GGF Semantic Grid Workshop. (2003) Available on: http://www.semanticgrid.org/GGF/ggf9/john/. [16] Bote-Lorenzo, M.L., Dimitriadis, Y.A., Gomez-Sanchez, E.: Grid characteristics and uses: a Grid defi nition. (To be published by Springer) [17] Niinimaki, M., White, J., Cerff, W.D., Hahkala, J., Niemi, T., Pitkanen, M.: Using virtual organizations membership system with EDG’s Grid security and database access. In: Proceedings of GLOBE’04, August 30 - September 3, 2004 in Zaragoza, Spain. (2004) [18] Pukacki, J., Wolniewich, M.: XRLS language. Technical report, PROGRESS Project (2003) Available at http://progress.psnc.pl/English/architektura/xrsl.html. [19] The JSDL Working Group: JSDL. Available at https://forge.gridforum.org/projects/jsdl-wg (2005) [20] W3C: RDF primer. Technical report, The http://www.w3.org/TR/rdf-primer.

World

Wide Web

Consortium (2004) Available at

[21] Hendler, J., Berners-Lee, T., Miller, E.: Integrating applications on the semantic web. Journal of the Institute of Electrical Engineers of Japan 122 (2002) [22] Bray, T., J.Paoli, Sperberg-McQueen, C., Maler, E., Yergeau, F., Cowan, J.: Extensible markup language (XML) 1.1 W3C recommendation 04 february 2004. (Available at http://www.w3.org/TR/2004/RECxml11-20040204/) [23] Clark, J., DeRose, S.: Xml path language (xpath) version 1.0 w3c recommendation. (Available at http://www.w3.org/TR/xpath) [24] Karvounarakis, G., Alexaki, S., Scholl, M., Christopides, V., Plexousakis, D.: RQL: A declarative query language for RDF. In: Proceedings of the World Wide Web conference 2002, May 7-11, 2002, Honolulu (WWW2002), IEEE Press (2002) [25] Seaborne, A.: RDQL - a query language for RDF. Technical report, W3C Member Submission (2004) Available at http://www.w3c.org/Submission/RDQL/. [26] Jena: Jena - a semantic web framework for java, Jena http://jena.sourceforge.net.

Project (2004) Available at


138

Conceptual Modelling (and Problem Solving Techniques) for Complex Narrative Events Gian Piero ZARRI LaLICC, University Paris4-Sorbonne Maison de la Recherche – 28, rue Serpente 75006 Paris, France

[email protected], [email protected]

Abstract. An application of NKRL (Narrative Knowledge Representation Language) techniques on (declassified) ‘terrorism’ documents has been carried out in the context of the IST Parmenides project. To allow the broadest possible exploitation of the ‘semantic content’ of the original documents, this application has required implementing the integration between the two main inferencing modes of NKRL, ‘hypotheses’ and ‘transformations’. The paper describes the conceptual problems encountered and the solutions adopted.

Introduction An in-depth experiment concerning the use of the inferencing capabilities of NKRL on a corpus of news supplied by the Greek Ministry of Defence (MoD) and related to terrorism in Philippines between 1993 and 2000 has been carried out in the context of the EC-supported Parmenides project (IST 2001-39023). MoD was one of the partners of the Parmenides consortium; NKRL is the “Narrative Knowledge Representation Language”, see [1-3]. To get the best from NKRL, the experiment has required the integration of the two main modalities of inferencing of NKRL, ‘hypotheses’ and ‘transformations’. ‘Transformation rules’ try to automatically replace a retrieval query that failed with one or more different queries that are not strictly ‘equivalent’ but only ‘semantically close’ to the original one. ‘Hypothesis rules’ allow building up causal-like explications of given events according to pre-defined reasoning schemata. Integrating the two inferencing modes corresponds to make use of ‘transformations’ to modify the reasoning steps to be executed within a ‘hypothesis’ context. In this paper, we recall firstly, in Section 2, the main principles underpinning NKRL and, in particular, the functioning of its inference engine(s). We will then illustrate, in Section 3, the theoretical problems that the integration implies – e.g., that of finding a correspondence between the hypothesis and transformation variables in an ‘integrated’ context – and the solutions adopted. Section 4 will illustrate briefly how these solutions have been practically implemented; Section 5 will make some comparison with related work. Section 6, a short Conclusion, will end the paper. Many of the examples used will refer to the (declassified) MoD corpus; the author makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy or completeness of such information.

G.P. Zarri / Conceptual Modelling for Complex Narrative Events

139

2. NKRL and the NKRL inference techniques 2.1 General information about NKRL NKRL innovates with respect to the current ontological paradigms, both the ‘traditional’ one – see, e.g., [4] – and that inspired by the Semantic Web research – [5, 6] – by associating with the usual ontologies of concepts an ‘ontology of events’, i.e., a new sort of hierarchical organization where the nodes correspond to n-ary structures called ‘templates’. Instead of using the usual object (class, concept) – attribute – value organization, templates are generated from the association of quadruples connecting together the symbolic name of the template, a predicate, and the arguments of the predicate introduced by named relations, the roles. The different quadruples associated within a given template have in common the ‘name’ and ‘predicate’ components. If we denote then with Li the generic symbolic label identifying a specific template, with Pj the predicate, with Rk a generic role and with ak the corresponding argument, the NKRL core data structure for templates has the following general format: (Li (Pj (R1 a1) (R2 a2) … (Rn an)))

(1) .

Predicates pertain to the set {BEHAVE, EXIST, EXPERIENCE, MOVE, OWN, PRODUCE, RECEIVE}, and roles to the set {SUBJ(ect), OBJ(ect), SOURCE, BEN(e)F(iciary), MODAL(ity), TOPIC, CONTEXT}. An argument of the predicate can consist of a simple ‘concept’ (according to the traditional, ‘ontological’ meaning of this word) or of a structured association (‘expansion’) of several concepts. Templates represent formally generic classes of elementary events like “move a physical object”, “be present in a place”, “produce a service”, “send/receive a message”, “build up an Internet site”, etc. They are included in an inheritance hierarchy, HTemp(lates), see the ‘father’ and ‘position’ fields associated with the template in Table 1a – and, for more details and a full description of HTemp, [7]. HTemp implements then the new ‘ontology of events’. When a particular event pertaining to one of these general classes must be represented, the corresponding template is ‘instantiated’ to produce what, in the NKRL's jargon, is called a ‘predicative occurrence’. To represent a simple narrative like: “On November 20, 1999, in an unspecified village, an armed group of people has kidnapped Robustiniano Hablo”, we must then select firstly in the HTemp hierarchy the template corresponding to “execution of violent actions”, see again Table 1a. This template is a specialization (‘father’ code) of the particular PRODUCE template corresponding to “perform some task or activity”. In a template, the arguments of the predicate (the ak terms in (1)) are represented by variables with associated constraints – which are expressed as concepts or combinations of concepts, i.e., using the terms of the NKRL standard ‘ontology of concepts’, called HClass, ‘hierarchy of classes’, in NKRL. A full description of HClass is also given in [7]. The constituents (as SOURCE in Table 1a) included in square brackets are optional. When deriving a predicative occurrence (an instance of a template) like mod3.c5 in Table 1b, the role fillers in this occurrence must conform to the constraints of the father-template. For example, in this occurrence, ROBUSTINIANO_HABLO (the ‘BEN(e)F(iciary)’ of the action of kidnapping) and INDIVIDUAL_PERSON_20 (the unknown ‘SUBJECT’, actor, initiator etc. of this action) are both ‘individuals’, instances of the HClass concept individual_person: this last is, in turn, a specialization of human_being_or_social_body, see, in Table 1a, the constraint on the

140


variables var1 and var6. kidnapping_ is a concept, specialization of violence_, see var3, etc. Throughout this paper, we will use small letters to represent a concept_, capital letters to represent an INDIVIDUAL_. The ‘attributive operator’, SPECIF(ication), is one of the four operators that make up the AECS sub-language, used for the construction of ‘structured arguments’ (‘expansions’); apart from SPECIF(ication) = S, AECS includes also the disjunctive operator, ALTERN(ative) = A, the distributive operator, ENUM(eration) = E, and the collective operator, COORD(ination) = C. The interweaving of the four operators within an expansion is controlled by the so-called ‘precedence rule’, see [1, 2]. The SPECIF lists, with syntax (SPECIF ei p1 … pn), are used to represent the properties or attributes which can be asserted about the first element ei, concept or individual, of the list – e.g., in the SUBJ filler of mod3.c5, Table 1b, the attributes weapon_wearing and (SPECIF cardinality_ several_) are both associated with INDIVIDUAL_PERSON_20. Table 1. Building up and querying predicative occurrences.

a) name: Produce:Violence father: Produce:PerformTask/Activity position: 6.35 NL description: ‘Execution of Violent Actions on the Filler of the BEN(e)F(iciary) Role’ PRODUCE

var1 var3 var4 var6 var8

SUBJ var1: [(var2)] OBJ var3 [SOURCE var4: [(var5)]] BENF var6: [(var7)] [MODAL var8] [TOPIC var9] var10] [CONTEXT {[modulators], abs}

= = = = =

, , , , , var9 = var10 = , , var2, var5, var7 =

b) mod3.c5)

PRODUCE

SUBJ

(SPECIF INDIVIDUAL_PERSON_20 weapon_wearing (SPECIF cardinality_ several_)): (VILLAGE_1) OBJ kidnapping_ BENF ROBUSTINIANO_HABLO CONTEXT #mod3.c6 date-1: 20/11/1999 date-2:

On November 20, 1999, in an unspecified village (VILLAGE_1), an armed group of people has kidnapped Robustiniano Hablo.

c) PRODUCE SUBJ : human_being : OBJ : violence_ BENF : human_being : {} date1 : 1/1/1999 date2 : 31/12/2001

There is any information in the system concerning violence activities?


141

To deal with those ‘connectivity phenomena’ that arise when several elementary events are connected through causality, goal, indirect speech, co-ordination and subordination etc. links, the basic NKRL knowledge representation tools have been complemented by more complex mechanisms that make use of second order structures created through reification of the conceptual labels of the predicative occurrences, see [1, 2] for further details. A simple example concerns the filler of the CONTEXT role in the occurrence mod3.c5 of Table 1b: in this case (‘completive construction’), the ‘context’ of the kidnapping is supplied by a whole predicative occurrence, mod3.c6, telling us that the kidnapping happened when Robustiniano Hablo was on his way home with his father. More complex examples of second order constructions are represented by the ‘binding occurrences’, i.e., lists of (reified) symbolic labels (ci) of predicative occurrences; the lists are differentiated making use of specific binding operators like GOAL and CAUSE.

2.2 ‘Search patterns’ and low-level inference procedures The basic building block for all the NKRL inference procedures is the FUM, Filtering Unification Module, see also [3]. It takes as input specific NKRL data structures called ‘search patterns’. Search patterns can be considered as the formal counterparts of natural language queries; their aim is that of supplying the general framework of information to be searched for, by filtering or unification, within an NKRL knowledge base. Basically, they can be assimilated to specialized/partially instantiated templates, where the ‘explicit variables’ that characterize the templates (vari, see Table 1a) have been replaced by concepts/individuals compatible with the constraints imposed on these variables in the original templates. In a search pattern, a concept acts as an ‘implicit variable’. It can unify the individuals representing its instances, and all the subsumed concepts in HClass along with their own instances, where the unified individuals and concepts appear in the NKRL occurrences that constitute the answer to the search pattern. A simple example of search pattern, translating the query: “There is any information in the system about violence events occurred during the year 1999?” is reproduced in Table 1c: it supplies, using FUM, the occurrence mod3.c5 (Table 1b) as one of the possible answers. The two timestamps, date1 and date2 associated with the pattern constitute the ‘search interval’ used to limit the search for unification to the slice of time that it is considered appropriate to explore, see [2]. Note that the search pattern of Table 1c – as, by the way, the answer mod3.c5 of Table 1b – derives from the template Produce:Violence (6.35) of Table 1a where all the explicit variables vari have been replaced by HClass concepts (implicit variables) corresponding to some of the original constraints.

2.3 ‘Hypothesis’ rules To introduce the main concepts concerning the functioning of the NKRL inference engine in a ‘hypothesis’ context, we will make use of a ‘standard’, very simple example. Let us then suppose we have directly retrieved, thanks to FUM and an appropriate search pattern, the occurrence conc2.c34, see Table 2a, which corresponds to the information: “Pharmacopeia, an USA biotechnology company, has received 64,000,000 USA dollars from the German company Schering in connection with a R&D activity”.

142


We will suppose, moreover, that this occurrence is not explicitly related with other occurrences in the base by second order elements, see Section 2.1 above. Under these conditions, we can activate the InferenceEngine module of NKRL, asking it to try to link up automatically the information found by FUM with other information present in the base. If this is possible, this last information will represent, in a way, a sort of ‘causal explanation’ of the information originally retrieved – i.e., in our example, an ‘explanation’ of the money paid to Pharmacopeia by Schering. A possible hypothesis rule that could fit our case is hypothesis h1 reproduced in Table 2b. Table 2. An example of hypothesis rule.

a) conc2.c34) RECEIVE

SUBJ OBJ SOURCE TOPIC date1 : date2 :

(SPECIF PHARMACOPEIA_ (SPECIF biotechnology_company USA_)) (SPECIF money_ usa_dollar (SPECIF amount_ 64,000,000)) (SPECIF SCHERING_ (SPECIF pharmaceutical_company GERMANY_)) r_and_d_activity

b) HYPOTHESIS h1

premise: RECEIVE SUBJ OBJ SOURCE

var1 money_ var2

var1 = company_ var2 = human_being, company_

A company has received some money from another company or a physical person. first condition schema (cond1): PRODUCE

SUBJ OBJ BENF TOPIC

(COORD var1 var2) var3 (COORD var1 var2) (SPECIF process_ var4)

var3 = mutual_commitment, business_agreement var4 = artefact_

The two parties mentioned in the premise have concluded a general sort of business-oriented agreement about the creation of a new product. second condition schema (cond2): PRODUCE

SUBJ OBJ MODAL CONTEXT

var1 var4 var5 var3

var5 = industrial_process, technological_process

The company that received the money has actually created the product mentioned in the first condition schema.

From an algorithmic point of view, InferenceEngine works according to a standard backward chaining approach with chronological backtracking, see, e.g., [8]. The differences with respect to other examples of use of this approach (Mycin, PROLOG …) are mainly linked with the complexity of the NKRL data structures that implies, after a


143

deadlock, the execution of difficult operations of restoration of the program environment to return to the previous choice point. Four ‘environment variables’ are used: x

x x

x

VALAFF (valeurs affectables in French), holds the values provisionally affected to the variables vari of the three schemata of Table 2 (premise, cond1 and cond2) that

implement the reasoning steps of the hypothesis: these values can be deleted after a backtracking operation; DESVAR holds the final values associated with the variables vari when the successful processing of one of the reasoning schemata has been completed; RESTRICT holds all the constraints (HClass terms) associated with the variables vari of the different reasoning schemata: these constraints will be used to build up systematically all the search patterns that can be derived from these schemata, see below; OCCUR holds the list of the symbolic names of all the occurrences retrieved by the search patterns derived from the reasoning schemata: the values bound to vari that have been retrieved in these occurrences are used to build up the VALAFF lists.

The first set of operations corresponds to the execution of the EXEPREM submodule, and consists in trying to unify, using FUM, the premise of the hypothesis, see Table 2b, and the event (the payment in our case, see conc2.c34 in Table 2a) to be ‘explained’ – more exactly, in trying to unify (using FUM) the event and the different search patterns derived from the premise by systematically substituting to the variables var1 and var2, see Table 1b, the associated constraints. As already stated, search patterns processed by FUM can only include implicit variables (concepts). This first step allows then i) to verify that the hypothesis tested is, in principle, suitable to ‘explain’ the particular event at hand, and ii) to obtain from the external environment (the event, i.e., conc2.c34) some values for the premise variables var1, var2. In our case, the premise variable var1 can only be substituted by the constraint company_; on the contrary, two substitutions, var2 = human_being and var2 = company_ are possible for the variable var2. A first search pattern will be then built up by substituting human_being for var2 (a value human_being is provisionally associated with var2 in VALAFF), i.e., a first unification with the event to explain will be tried by using a search pattern corresponding to a payment done by an individual person instead of a company. This unification obviously fails. The engine then ‘backtracks’ making use of a second sub-module, REEXEC: REEXEC is the sub-module to be systematically used during the execution of a hypothesis rule i) to backtrack when a deadlock occurs, and ii) to reconstruct, making use of the environment variables, the data structures (environment) proper to the previous choice point. The association var2 = human_being is removed and, using the constraint values stored in RESTRICT, the engine builds up a new pattern making use now of the value var2 = company_, that will unify the value SCHERING_ in conc2.c34. The engine can then continue the processing of the hypothesis h1; the two values var1 = PHARMACOPEIA_ and var2 = SCHERING_ will then be stored in DESVAR and passed to the first condition schema (cond1), see Table 2b. The search patterns derived from the first condition schema cond1 – by taking into account the values already bound in DESVAR to var1 and var2 and by replacing systematically, as usual, all the other variables with the associated constraints – will be tested by a third sub-module, EXECOND; this last is called whenever there exist conditions favourable for advancing in the hypothesis, in other words, for being able to process a new condition schema. EXEPREM and EXECOND perform then

144


both the forward traversal of the choice tree, with REEXEC being systematically called whenever the conditions for a backtracking exist; the difference between EXEPREM and EXECOND derive mainly from the fact that, in an EXECOND context, the unification of the search patterns derived from the condition schemata is tested against the general base of predicative occurrences to (try to) find possible unifications with these occurrences while, in an EXEPREM context, the unification concerns only the search patterns derived from the premise and the (unique) starting occurrence. As usual, many deadlocks are generated in the course of the EXECOND operations. Some are due, as in the premise case, to the chronological utilization of the constraints. For example, when trying to make use of a pattern derived from cond1 where the variable var3 has been substituted by its first constraint mutual_commitment, see Table 2b, a failure will be generated and REEXEC will be invoked again. The occurrences we must retrieve in the knowledge base about the relationships between Pharmacopeia and Schering concern, in fact, different sorts of commercial agreement between Pharmacopeia and Schering – e.g., r_and_d_agreement and sale_agreement, see below, both specific terms in HClass of business_agreement (the second constraint on var3) – and not a private arrangement like mutual_commitment. We will, eventually, find in the base an instantiation of cond1 corresponding to an event of the form: “Pharmacopeia and Schering have signed two agreements concerning the production by Pharmacopeia of a new compound, COMPOUND_1”. The values associated with the variables var3 (r_and_d_agreement and sale_agreement) and var4 (COMPOUND_1) in cond1 will then be used to create the search patterns derived from cond2. It will then be possible to retrieve an occurrence corresponding to the information: “In the framework of an R&D agreement, Pharmacopeia has actually produced the new compound”, see, e.g., [3] for more details. The global information retrieved through the execution of the hypothesis can then supply a sort of ‘plausible explanation’ of the Schering’s payment: Pharmacopiea and Schering have concluded some agreements for the production of a given compound, and this compound has been effectively produced by Pharmacopeia.

2.4 ‘Transformation’ rules The ‘transformation rules’ are used to obtain a plausible answer from a repository of predicative occurrences also in the absence of the explicitly requested information, by searching semantic affinities between what is requested and what is present in the repository. The principle employed consists in using these rules to automatically ‘transform’ the original query (i.e., the original search pattern) into one or more different queries (search patterns) that are not strictly ‘equivalent’ but only ‘semantically close’ to the original one. For example, let us suppose that, in a ‘Philippine terrorism’ context, we are asking: “Search for the existence of some links between ObL (a well-known ‘terrorist’) and Abubakar Abdurajak Janjalani, the leader of the Abu Sayyaf' group”. In the absence of a direct answer, the corresponding search pattern can be transformed into: “Search for the attestation of the transfer of economic/financial items between the two”, which could lead to retrieve this information: “During 1998/1999, Abubakar Abdurajak Janjalani has received an undetermined amount of money from ObL through an intermediate agent”. From a formal point of view, transformation rules are made up of a left-hand side, the ‘antecedent’ – i.e. the formulation, in NKRL format (search pattern), of the ‘query’ to be transformed – and one or more right-hand sides, the ‘consequent(s)’ – the NKRL


145

representation(s) of one or more queries that must be substituted for the given one. A transformation rule can, therefore, be expressed as: A (antecedent, left-hand side) B (consequent(s), right-hand side). The ‘transformation arrow’, ‘’, has a double meaning: x

operationally speaking, the arrow indicates the direction of the transformation: the left-hand side A (the original search pattern) is removed and replaced by the righthand side B (one or more new search patterns); the standard logical meaning of the arrow is that the information obtained through B implies the information we should have obtained from A.

x

More formal details can be found, e.g., in [3]. A representation of the above transformation is given in Table 3. Note that the left-hand side (antecedent) of this transformation corresponds to a partial instantiation of the template Behave:FavourableConcreteMutual that is routinely used to represent into NKRL format a (positive) mutual behaviour among two or more entities. With respect now to the implementation details, the InferenceEngine version to be used for transformations is quite identical to that used for executing the hypothesis rules. The sub-module ANTEXEC (execution of the antecedent) corresponds, in fact, to the EXEPREM sub-module; CONSEXEC (execution of the consequent(s)) corresponds to EXECOND. REEXEC is the same in the two versions. Table 3. A simple example of ‘transformation’ rule.

t1)

BEHAVE

var1 var2 var3 var4

= = = =

SUBJ OBJ MODAL

(COORD1 var1 var2) (COORD1 var1 var2) var3

RECEIVE

SUBJ OBJ SOURCE

var2 var4 var1

human_being_or_social_body human_being_or_social_body business_agreement, mutual_relationship economic/financial_entity

To verify the existence of a relationship or of a business agreement between two (or more) people, try to see if one of these people has received a ‘financial entity’ (e.g., money) from the other.

3. Integrating the two inferencing modes of NKRL We will now introduce the general framework of the problem concerning the possibility of making use of ‘transformations’ in a ‘hypothesis’ context. 3.1 Strategies for the execution of transformations in a hypothesis context Executing transformations in a hypothesis context means taking a search pattern built up by EXECOND during the execution of a hypothesis rule and ‘transforming’ it by executing the following operations: x

after having created a list of all the transformation rules stored in the system whose left hand side (antecedent) is generally congruent with the general format of the

146

x x


current search pattern, find in this list a transformation rule, if any, whose antecedent can actually unify the original pattern; in this case, execute the transformation by substituting to the original search pattern the new one formed using the right hand side (consequent) of the transformation; if the new search pattern is successful (i.e., if it can unify with the occurrences of the knowledge base) store in VALAFF (in the original hypothesis environment) new values for the variables vari of the condition schema under execution.

Several strategies can be adopted with respect to i) the ‘number’ of transformations to execute and ii) the ‘circumstances’ of their execution, see [9] for the details. In the present version of the NKRL software, the user is allowed to select a particular transforming strategy, see Figure 1. After having answered ‘Yes’ to the question about the execution of internal transformations, on the left, she/he is asked whether she/he wants to carry out ‘positive’ or ‘negative’ transformations. Conventionally, executing ‘negative’ transformations means that the transformation process is activated only after the failure of a search pattern derived directly from a hypothesis condition schema (i.e., a search pattern able to find an unification in the knowledge base cannot be transformed); executing ‘positive’ transformations means that all the search patterns derived directly from a condition schema are transformed, independently from the fact they have been successful or not. The last query on the right of asks for the ‘depth’ of the transformation, i.e., if a pattern that is the result of an internal transformation can be transformed in turn, and how many ‘transformation steps’ are allowed.

Figure 1: Parameters for internal transformations.

3.2 The correspondence among variables The main ‘theoretical’ problem concerning the integration hypotheses/transformations concerns the possibility of finding a correspondence between i) the variables, vari, originally present in the condition schema from which the search pattern to transform has been derived and ii) those, totally disjoint, varj, that appear in the transformation rules to be used. A correspondence among vari and varj must, in general, be found. We must now operate a distinction between ‘global’ and ‘local’ variables. In the context of the ‘integration’ problem, a ‘local’ variable is a variable that is used in only one condition schema, condi, without appearing explicitly in all the subsequent condition schemata condi+1 … condn. A global variable, on the contrary, is a variable that, after having been introduced in condi, is utilized again in at least one of the subsequent n-i condition schemata. Note that the variables introduced by the premise are not interested by this distinction, given that the search patterns derived from the premise are never


147

submitted to the transformation operations: if the premise patterns cannot unify the starting occurrence, this means that the hypothesis is not suitable for explaining this occurrence. To give an example, variables var3 and var4 introduced in the cond1 scheme of Table 2 are ‘global’, because they are re-utilized in cond2; var5 in cond2 is local, given that cond2 is the last condition scheme of the hypothesis. We can now say that: x

x

If the new variables introduced by the condition schema condi at the origin of the search pattern to be transformed are all local – i.e., none of them appear in the subsequent n-i condition schemata – this means that no new values for vari are needed in order to continue with the processing of the hypothesis. In this case, the successful transformation of the search pattern is only useful i) to guarantee, in an indirect way, that condi is congruent with the data of the knowledge base and, more importantly, ii) to introduce in the hypothesis' answer some new results (occurrences) retrieved by the transformed pattern(s). If some of the new variables introduced by condi are global – i.e., they appear in at least one of the subsequent n-i condition schemata – associating new values with these variables through the transformation process will increase the possibility of constructing new search pattern making use of those subsequent schemata, enlarging considerably the search space and augmenting then the probability of obtaining new interesting results. In this case, a correspondence between vari (the variables of the condition schema) and varj (the variables used in the transformation) must be found.

3.3 Executing the internal transformations We will analyse now some concrete examples. In Table 4b, we reproduce part of a typical ‘terrorism’ hypothesis used in the Parmenides project, namely, the hypothesis that allows, among other things, to explain the kidnapping of Robustiniano Hablo (Table 4a and Table 1b above) in terms of: x

x x x

Robustiniano Hablo has been kidnapped by members of a separatist movement or terrorist organization – e.g., the Abu Sayyaf Group, a separatist group in the Southern Philippines; this separatist movement or terrorist organization carries out a particular form of kidnapping, kidnapping for ransom; this form of kidnapping implies that the kidnapped person is wealthy; Robustiniano Hablo is not, may be, wealthy himself, but he is the son of a rich Philippine businessperson.

Note that, to perform the last reasoning step, it is necessary to use a ‘two steps’ transformation (a transformation having two related consequent schemata) saying that, in a ransom kidnapping context, i) someone who is a very strict relative (son, spouse…) of a wealthy person, ii) can be considered as being wealthy himself. See [9] for more details. If we examine now cond2 in Table 4b, we can see that the unique new variable introduced by this condition schema is var4 (var3 has been introduced in cond1); var4 is now ‘local’, in the sense that it will no more used in the subsequent condition schemata. We can now suppose to have, among the transformations of the system, the transformation rule reproduced in Table 5b. Using the version of InferenceEngine for

148


transformations, the antecedent part of this rule will unify the search pattern, see Table 5a, derived from cond2 in Table 4b, where var3 has been replaced by ABU_SAYYAF_GROUP – we can suppose that this value has been obtained during the processing of cond1 – and var4 by the constraint ransom_kidnapping. The execution of the transformation will produce the new search pattern reproduced in the lower part of Table 6. If this pattern can unify some information in the knowledge base of NKRL occurrences, the success of the transformation will give rise to ‘variants’ of the ‘normal’ results: in our case, the generic (‘positive’, see the previous Section) result concerning the fact that the Abu Sayyaf Group carries out ransom kidnapping will be reinforced by the a more specific information telling us that the Abu Sayyaf Group has received some money for freeing from captivity a given person. In this case, no new values will be introduced in the VALAFF and DESVAR environment variables of the h2 hypothesis, that will then continue unchanged with the processing of cond3. Table 4. The ‘kidnapping’ hypothesis rule.

a) mod3.c5)

PRODUCE

SUBJ OBJ BENF CONTEXT date-1: date-2:

(SPECIF INDIVIDUAL_PERSON_20 weapon_wearing (SPECIF cardinality_ several_)): (VILLAGE_1) kidnapping_ ROBUSTINIANO_HABLO #mod3.c6 20/11/1999

b) HYPOTHESIS h2

premise: PRODUCE SUBJ var1 OBJ kidnapping_ BENF var2

var1 = human_being_or_social_body var2 = individual_person

A human being has been kidnapped. first condition schema (cond1): BEHAVE SUBJ var1 MODAL member_of TOPIC var3

var1 = human_being_or_social_body var3 = separatist_mouvement, terrorist_organization

The kidnappers are member of a separatist movement or of a terrorist organization. second condition schema (cond2): PRODUCE SUBJ OBJ

var3 var4

var3 = separatist_mouvement, terrorist_organization var4 = ransom_kidnapping

This organization performs ransom kidnapping. …

If we consider now the cond1 condition schema of hypothesis h2, see Table 4b, we can see that the (unique) new variable introduced in this schema is var3: this one is now ‘global’ given that it will be used in cond2 and the subsequent condition schemata. The transformations operating on the search patterns derived from cond1 will then be able, in


149

principle, to produce ‘new’ values for var3, to be stored in VALAFF/DESVAR and to be used as they had been obtained through the ‘normal’ hypothesis operations – ‘new’ means here different (and possibly unexpected) values with respect to those that could be obtained via the usual procedures. At the difference then of the previous case (‘local’ variables) – where the final result of the execution of successful transformations in a hypothesis context was only that of producing ‘locally’ (i.e., for the condition schema actually transformed) some ‘variants’ of the standard results – the success of the ‘internal’ transformations could (see below) now lead also to the addition of totally new branches to the choice tree, likely in principle to produce new results for each of the condition schemata included between the transformation point and the end of the hypothesis. Table 5. A transformation rule for cond2 of h2.

a) search pattern derived from cond2 of Table 4b: PRODUCE SUBJ : ABU_SAYYAF_GROUP : OBJ : ransom_kidnapping : {} date1 : date2 :

The Abu Sayyaf group performs ransom kidnapping.

b) ‘ransom kidnapping’ transformation:

t2) PRODUCE

var1 var2 var3 var4

= = = =

SUBJ OBJ

var1 var2

RECEIVE

SUBJ OBJ SOURCE TOPIC

var1 money_ var3 (SPECIF captivity_freeing var4)

separatist_mouvement, terrorist_organization ransom_kidnapping human_being human_being

To verify if a given organization performs ransom kidnapping, try to see this organization has received some money for freeing from captivity one or more human being(s).

Table 6. The original search pattern derived from cond2 and the transformed one (using t2). PRODUCE SUBJ : ABU_SAYYAF_GROUP : OBJ : ransom_kidnapping : {} date1 : date2 : RECEIVE SUBJ : ABU_SAYYAF_GROUP : OBJ : money_ : SOURCE : human_being : TOPIC : (SPECIF captivity_freeing human_being) {} date1 : date2 :

A problem of ‘variable correspondence’ occurs now. Let us suppose we will use the rule of Table 7, i.e., the membership in an ‘organization’ can be verified by checking, among other things, if the ‘members’ receive some form of permanent or occasional ‘salary’ from the organization.

150


Table 7. A transformation rule for cond1 of h2. ‘permanent or occasional salary’ transformation: t3) BEHAVE

var1 = var2 = var3 =

SUBJ MODAL TOPIC

var1 member_of var2

RECEIVE

SUBJ OBJ SOURCE

var1 var3 var2

human_being political_group/party irregular_payment, salary_

To verify if a person is a member of a given organization, try to see if, among other things, this person receives a permanent or occasional salary from this organization.

In this case, after having entered the specific transformation environment and having activated the sub-module ANTEXEC (see Section 2.4 above) to execute the unification between the search pattern derived from cond1 of h2 (see also Table 8a) and the antecedent of t3 in Table 7, the current situation is globally represented in Table 8. Table 8. Environment variables after the unification of the pattern derived from cond1 and t3.

(a)

BEHAVE SUBJ : INDIVIDUAL_PERSON_20 : MODAL : member_of : TOPIC : separatist_movement {} date1 : date2 : VALAFF (hypo, h2) var1(hypo) = INDIVIDUAL_PERSON_20 var2(hypo) = ROBUSTINIANO_HABLO var3(hypo) = separatist_movement

(b)

BEHAVE SUBJ : human_being : MODAL : member_of : TOPIC : political_group/party {} date1 : date2 : VALAFF (transfo, t3) var1(transfo) = INDIVIDUAL_PERSON_20 var2(transfo) = separatist_movement

(c)

RECEIVE SUBJ : INDIVIDUAL_PERSON_20 : OBJ : irregular_payment : SOURCE : separatist_movement : {} date1 : date2 :

The top search pattern (a) is then the pattern derived from cond1: returning, in fact, to Table 4 above, we can see that – because of the unification, in the original hypothesis environment and using EXEPREM, of the premise of h2 with the starting occurrence mod3.c5 – the variables var1(hypo) and var2(hypo) introduced by the premise have taken, respectively, the values INDIVIDUAL_PERSON_20 and ROBUSTINIANO_HABLO, as reflected by the state of the environment variable VALAFF(hypo) in Table 8. The value separatist_movement associated with var3(hypo) in VALAFF(hypo), see again Table 8, derives from the operations performed by EXEPREM (in a hypothesis environment) to


151

build up the search pattern (a) from cond1. The search pattern (b) has been built up from ANTEXEC (in a transformation environment) from the antecedent part of the transformation t3 of Table 7; after the unification of (a) and (b) – this unification, executed by FUM, is ‘oriented’, in the sense that (a) has a ‘predicative occurrence’ role – the variables var1(transfo) and var2(transfo) of t3 have, respectively, the values INDIVIDUAL_PERSON_20 and separatist_movement, see the status of VALAFF for the transformation environment in Table 8. From the comparison of the values affected to the variables in the two versions of VALAFF, for hypothesis (VALAFF hypo) and transformation (VALAFF transfo) it is now evident that there is a correspondence between var3(hypo) and var2(transfo). This fact will be then registered into a CORRESP(ondence) table, see Table 9: as already stated above, only var3(hypo) is of interest for the further processing of the hypothesis. The processing of the transformation in its proper environment will continue, and the CONSEXEC sub-module will then build up the search pattern (c) of Table 8 from the consequent of transformation t3 of Table 7. Let us suppose now that (c) is able to unify (at least) a predicative occurrence in the NKRL knowledge base. In this case, a value will be bound to var2(transfo); according to the information stored in CORRESP, see Table 9, this value will also be bound to var3(hypo) and inserted into the VALAFF/DESVAR variables of the original hypothesis (h2) environment. Supposing now, see Section 3.1 above, that the strategy chosen for the execution of the ‘internal’ transformations is a ‘positive’ one – this implies a transformation of all the search pattern built up by EXECOND from a condition schema independently from the fact that these pattern have been successful or not – four (main) different cases must be envisaged. x

x

x

In the first one, we suppose that var3(hypo), before the execution of the transformation, was already linked with, e.g., ABU_SAYYAF_GROUP – i.e., during the ‘normal’ execution of the hypothesis h2, a search pattern directly derived from cond1 has been able to retrieve the information that INDIVIDUAL_PERSON_20 (representing collectively the group that has realized the kidnapping) was part of Abu Sayyaf. We suppose now that the transformed pattern (c) of Table 8 retrieves the information telling that INDIVIDUAL_PERSON_20 receive some form of occasional salary from Abu Sayyaf: var2(transfo) is also bound to ABU_SAYYAF_GROUP, and passing this value to var3(hypo) will add nothing from the point of view of an ‘augmented’ development of the hypothesis. As in the previous case of ‘local’ variables, the only real benefit linked with the transformation will be a confirmation of the links between INDIVIDUAL_PERSON_20 and Abu Sayyaf through the discovery that the kidnappers receive also some money from this group. If we suppose now that, in the ‘normal’ execution of the hypothesis h2, all the search patterns directly derived from the condition schema cond1 failed, the possibility of obtaining the value ABU_SAYYAF_GROUP for var3(hypo) via the transformation and the passage through var2(transfo) allows continuing with the processing of the hypothesis h2 otherwise destined to irremediably fail. We can suppose now that pattern (c) of Table 8 is able to find an unification within the knowledge base telling us that INDIVIDUAL_PERSON_20 receives some form of occasional salary from another group, e.g., from the Moro Islamic Liberation Front (another separatist group) – MORO_ISLAMIC_LIBERATION_FRONT will be then, in this case, the final value bound to var2(transfo) after the unification with the occurrence. If, as in the previous case, var3(hypo) was ‘empty’ before the execution of

152

x


the transformation, activating the process of internal transformation will allow, as in the previous case, to continue with the processing of h2 otherwise destined to fail. Eventually, we can now suppose that var2(transfo) takes the value MORO_ISLAMIC_LIBERATION_FRONT while var3(hypo) is already bound to ABU_SAYYAF_GROUP via the ‘normal’ hypothesis processing. This means, in practice, that the group of kidnappers is linked in some way to both the Abu Sayyaf and Moro Islamic Liberation Front groups. According to the correspondence between var2(transfo) and var3(hypo), see CORRESP, MORO_ISLAMIC_LIBERATION_FRONT must also be bound to var3(hypo) and stored, accordingly, into the VALAFF/DESVAR variables of the original hypothesis (h2) environment. This now value will be used in the further processing of the hypothesis in parallel with the original one (ABU_SAYYAF_GROUP), leading then (possibly) to a totally new and particularly interesting set of results. Table 9. The CORRESP table for the example. original condition schema: variable name

value

transformation consequent: variable name

var3

separatist_movement

var2

4. Some simple remarks about the software solutions Integrating the two versions, for hypotheses and transformations, of InferenceEngine, corresponds to solve a complex ‘coroutine’ problem, where the main difficulty is generated, as usual in NKRL, by the existence of complex data structures to be managed, stored and reloaded. The integration can be implemented according to two approaches. The first one is the classical ‘coroutine’ solution, where the ‘transformation’ version of the engine has the same priority of the ‘hypothesis’ version, and then it starts its execution, as the last one, from the main method of InferenceEngine. The problem of this solution concerns the fact that it should imply the complete execution of the transformation version (the ‘internal’ version) until a result has been obtained, requiring then an efficient way – Java's threads could be used in this context – to return to the hypothesis version (the ‘external’ one). On the other hand, this solution should also imply the fact of having a separate display (text output) for the results of the ‘internal’ execution, with some difficulties then in co-ordinating them with the displaying of the ‘external’ (hypothesis) results. The second approach, more manageable and simpler to implement, consists in just integrating the InferenceEngine Java object corresponding to the ‘internal’ (transformation) version in the execution of the ‘external’ (hypothesis) version. This allows the external hypothesis version of InferenceEngine – which works now as the ‘main’ program – to run the internal version until it has a result and to get back this result simply as a function execution return. This approach allows to implement a ‘transparent’ running of the internal module, and implies also the advantage of having the results of the ‘internal’ execution naturally integrated with those displayed by the interface of the main (hypothesis) version. We have then chosen the second approach. Giving that, i) during the functioning of InferenceEngine, REEXEC is the only sub-module invoked in order to reconstruct the environment proper to a previous choice point and to allow then EXECOND to build up a


153

new search pattern, and that ii) executing transformation operations – in our case, within a hypothesis context – amounts exactly to build up new search patterns, we can conclude that only REEXEC must be modified to implement the ‘integration’ operations. In practice, during the execution of an internal transformation, REEXEC will run the ‘internal’ (transformation) Java object until a result has been found and returned: the internal object is then stopped, and it will wait for a new REEXEC call, producing other results if these are possible for the current transformation. For each call to REEXEC, there will be then one internal module pending and waiting for further results – each main program execution level will potentially have an internal module object reference pointing to such object. From a Java programming point of view, InferenceEngine includes now three different objects: Hypothesis, Transformation and InternalTransformation. The first two are practically unchanged. The InternalTransformation object is modelled on the Transformation one: as already stated, it will run trying to find a result and then, if successful, it will return this result. If this is not possible, it will throw a NoMoreResultsException(). REEXEC now first executes the InternalTransformation code to get the next result and, if this is not possible, catches the exception and continues its execution trying to build a new model for this level. See again [9] for more information and a detailed example. 5. Comparison with similar work Comparison of what expounded in the previous Sections with work accomplished in a Semantic Web framework – Semantic Web (SW) seems to be today the winning paradigm in a Knowledge Representation and Reasoning context – is not very easy because of a fundamental, ‘epistemological’ difference between the NKRL and the SW approaches. Semantic Web languages like RDF and OWL are, in fact, inherently ‘binary’, in the sense that, for these languages, a property can only be a binary relationship, linking two individuals or an individual and a value. The (very scarce) proposals intended to extend the SW languages in order to deal with n-ary relationships appear, often, as ‘odd’ and ‘artificial’ to some extent. For example, a very recent working paper from the W3C Semantic Web Best Practices and Deployment Working Group (SWBPD WG) about defining n-ary relations for the SW languages, see [10], proposes some extensions to the binary paradigm to allow the correct representation of very specific ‘narratives’ like: “Christine has breast tumor with high probability” or “Steve has temperature, which is high, but failing”. The solutions proposed, not really convincing, range from the introduction of fictitious ‘individuals’ to represent the n-ary relations to the rediscovery of semantic networks solutions of the seventies. The SW languages seem then difficult to use outside the set up and management of static ‘ontologies’ of concepts and individuals. Unfortunately, this corresponds to leave aside a big amount of important, ‘economically relevant’ information, which is buried into general ‘narrative’ information resources, including most of the corporate knowledge documents, the news stories, the normative and legal texts, the medical records, many intelligence messages, etc., as well as a huge fraction of the information stored on the Web. In these ‘narrative documents’; the main part of the information content consists in the dynamic description of ‘events’ that relate the real or intended behaviour of some ‘actors’ (characters, personages, etc.) – the term ‘event’ is taken here in its more general meaning, covering also strictly related notions like fact, action, state, situation etc. These actors try to attain a specific result,

154


experience particular situations, manipulate some (concrete or abstract) materials, send or receive messages, buy, sell, deliver etc. To sum up the, SW languages are ‘binary’ and NKRL ‘n-ary’. It is well evident, then, that also the tools (rules) used to manage the data structures must be deeply different in the two approaches. Note that rules have not been included in the standard descriptions of SW languages like RDF and OWL, and the whole SW rule domain seems to be still in a very early state of development. Languages like RuleML [11], TRIPLE [12], and SWRL [13] – all based, roughly, on extensions of the inferential properties of Prolog/Datalog to deal with OWL-like data structures – appear to be, for the time being, more working proposals than implemented languages. We can also note that an advantage of the existing, Owl-compatible reasoning tools like RACER [14] is that of being characterized by the use of sound and complete inferencing algorithms supported by the description logics theory. Unfortunately, the reduced expressiveness of their main reasoning component, the automatic classification mechanism, linked with the reduced expressiveness of the underpinning data structures, do not allow them to come close to the power of the unification-based, reasoning procedures in the NKRL style. For an approach to ‘inferencing’ similar to that described in this paper, it is then better to look into the ‘semantic network’ domain – in the widest meaning of these words – given that systems dealing with n-ary sorts of data structures are relatively frequent in this field. Conceptual Graphs (CGs) for example, see [15], are based on a powerful graph-based representation scheme that can be used to represent n-ary relationships between complex objects in a system. A conceptual graph is a finite, connected, bipartite graph that makes use of two kinds of nodes, i.e., ‘concepts’ and ‘conceptual relations’; every arc of a graph must link a conceptual relation to a concept. ‘Bipartite’ means that every arc of a conceptual graph associates necessarily one concept with one conceptual relation: it is not possible to have arcs that link concepts with concepts or relations with relations. Very interesting work concerning indexing and querying of knowledge bases of CGs making use of unification techniques can be found, e.g., in [16, 17]; Corbett [18] adds to the standard CGs unification procedures the possibility of dealing with constraints, leading then to unification algorithms very similar to those used in FUM. There is, however, a fundamental difference between the NKRL- and CGs-based approaches to the set up and managing of n-ary structures. Even if CGs have the possibility of defining general conceptual structures, similar to the NKRL templates, for describing narrative-like phenomena, an exhaustive and authoritative list of these structures under the form of ‘canonical graphs’ does not exist, and its construction seems have never been planned. The practical consequence of this state of affairs is then the need, whenever a concrete application of CGs theory must be implemented, of defining anew a specific list of canonical graphs for this particular application. On the contrary, a fundamental (and apparently unique) characteristic of NKRL is given by the fact that its catalogue of ‘basic templates’ (coinciding with HTemp and including 150 templates, very easy to extend and customize) is in practice part and parcel of the definition of the language – as already stated, all the different sorts of inferential rules used in NKRL are then obtained via the partial instantiation of the templates of the catalogue. This approach is very important for practical applications, and it implies, in particular, that: i) a systembuilder does not have to create himself the structural and inferential knowledge needed to describe and exploit the events proper to a (sufficiently) large class of narratives (in the most general meaning of this word); ii) it becomes easier to secure the reproduction or the sharing of previous results.


155

6. Conclusion In this paper, we have first introduced the two main inferencing modes in NKRL, ‘hypotheses’ and ‘transformations’. We have then discussed the conceptual problems linked with their integration, and illustrated concretely the solutions adopted. The paper ends up with some comparisons, from an inferencing and knowledge representation point of view, between the NKRL approach and work done in a Semantic Web and Conceptual Graph context. References [1]

[2]

[3]

[4]

[5]

[6] [7] [8] [9] [10]

Zarri, G.P. (1997) “NKRL, a Knowledge Representation Tool for Encoding the ‘Meaning’ of Complex Narrative Texts”, Natural Language Engineering – Special Issue on Knowledge Representation for Natural Language Processing in Implemented Systems, 3: 231-253. Zarri, G.P. (1998) “Representation of Temporal Knowledge in Events: The Formalism, and Its Potential for Legal Narratives”, Information & Communications Technology Law – Special Issue on Models of Time, Action, and Situations, 7: 213-241. Zarri, G.P. (2002) “Conceptual Modelling and Knowledge Management for Narrative Multimedia Documents”, in Information Modelling and Knowledge Bases XIII, Kangassalo, H., Welzer, T., Jakkola, H., Rozman, I, and Kawaguchi, E., eds. Amsterdam: IOS Press. Noy, F.N., Fergerson, R.W., and Musen, M.A. (2000) “The Knowledge Model of Protégé-2000: Combining Interoperability and Flexibility”, in Knowledge Acquisition, Modeling, and Management – Proceedings of EKAW’2000. Berlin: Springer-Verlag. Bechhofer, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D.L., Patel-Schneider, P.F., and Stein, L.A., eds. (2004) OWL Web Ontology Language Reference – W3C Recommendation 10 February 2004. W3C (http://www.w3.org/TR/owl-ref/). Horridge, M. (2004) A Practical Guide to Building OWL Ontologies with the Protégé-OWL Plugin (Edition 1.0). Manchester: The University of Manchester. Zarri, G.P. (2004) NKRL Manual, Part II (Parmenides Technical Report). Paris: University of Paris IV/Sorbonne. Clocksin, W.F., and Mellish, C.S. (1981) Programming in PROLOG. Berlin: Springer-Verlag. Zarri, G.P., and Bernard, L. (2004) Using NKRL Inference Techniques To Deal With MoD 'Terrorism' Information (Parmenides Technical Report). Paris: University of Paris IV/Sorbonne. Noy, N., and Rector, A., eds. (2005) Defining N-ary Relations on the Semantic Web – W3C

Working Draft 24 May 2005. W3C (http://smi-web.stanford.edu/people/noy/nAryRelations/ n-aryRelations-2nd-WD.html). [11]

[12]

Boley, H., Tabet, S., and Wagner, G. (2001) “Design Rationale of RuleML: A Markup Language for Semantic Web Rules”, in Proceedings of SWWS'01, The First Semantic Web Working Symposium. Stanford: Stanford University. Sintek, M., and Decker, S. (2002) “TRIPLE – A Query, Inference, and Transformation Language for the Semantic Web”, in Proceedings of the First International Semantic Web Conference – ISWC 2002 (LNCS 2342). Heidelberg: Springer-Verlag.

[13]

Horrocks, I., Patel-Schneider, P.F., Bechhofer, S., and Tsarkov, D. (2005) “OWL rules: A Proposal and Prototype Implementation”, Web Semantics: Science, Services and Agents on the World Wide Web 3: 23-40.

[14]

Haarslev, V., and Möller, R. (2003) “Racer: A Core Inference Engine for the Semantic Web”, in Proceedings of the 2nd International Workshop on Evaluation of Ontology Tools (EON2003), Sanibel Island (Florida, USA), October 20th, 2003. Sowa, J.F. (1999) Knowledge Representation: Logical, Philosophical, and Computational Foundations. Pacific Grove (CA): Brooks Cole Publishing Co. Ellis, G. (1995) “Compiling Conceptual Graph”, IEEE Transactions on Knowledge and Data Engineering 7: 68-81. Willems, M. (1995) “Projection and Unification for Conceptual Graphs”, in Proceedings of the Third International Conference on Conceptual Structures. Heidelberg: Springer-Verlag.

[15] [16] [17]

[18] Corbett, D. (2003) Reasoning and Unification over Conceptual Graphs. New York: Kluwer Academic/Plenum Publishers.

156


Role of Models in Knowledge Transfer during OO Software Development Oksana NIKIFOROVA1 Riga Technical University, Latvia, [email protected] Marite KIRIKOVA Riga Technical University, Latvia, [email protected] Wita WOJTKOWSKI Boise State University, USA, [email protected] Abstract. Formal and semi-formal transfer of knowledge in the conventional object-oriented software development is seriously impaired. This occurs because it is impossible to assure both completeness and consistency of the initial body of knowledge from which the knowledge transfer may commence. In this paper we propose that better utilization of the formal transfer of knowledge requires the focus on the functional aspects of the problem domain to a much higher degree than it is currently customary. We propose that the two hemisphere model based approach, where the problem domain knowledge is expressed in terms of the business process model and the conceptual model, offers effective as well as efficient knowledge transfer mechanism, if sufficiently complete and consistent, explicit and structured, problem domain knowledge is available.

Introduction If software developers use explicit models at all, they use them mostly as simple sketches of the design ideas, often discarding them once they have written the code. This mode of use is sufficient for the traditional, code-centric development. For more advanced approaches, such as the Model Driven Approach (MDA) recently proposed by the Object Management Group [1], the role of explicit models is considerably more important [2]. MDA is a novel approach that separates business logic from the underlying technology platform. In essence, MDA moves development focus from the programming language code to models, specifically, models expressed in the Unified Modeling Language (UML) and its profiles [3]. The stated objective of the MDA is to increase productivity and reduce time-to-market. This is accomplished by facilitating software development through the use of concepts closer to the problem domain at hand, rather then those offered by the programming languages [2]. Using MDA, developers create platform independent application domain level model that may be transformed into a platform specific model targeted at a specific technology such as Corba, Enterprise JavaBeans, or .NET. [1], [4]. Model transformation bridges the gap between the application domain level model and implementation [2]. With MDA, the platform independent application domain level model becomes the primary 1

The research has been supported by European Social Fund under the contract No. 2004/0002/VPD1/ESF/PIAA/04/NP/3.2.3.1/0002/0007

O. Nikiforova et al. / Role of Models in Knowledge Transfer During OO Software Development

157

artifact in the development of software. Productivity of the model-driven approach comes about because of the automated transformation of the platform independent application domain model into the platform specific model, from which the software code may be generated. If this transformation is not automated, the productivity decreases [4]. In the traditional object-oriented software development, during the implementation, the transfer of knowledge from the application domain model to the implementation components is made possible because of the tacit knowledge of the developers. In MDA, when the transition to the implementation level is fully automated, however, code modification and fine-tuning based on the tacit knowledge are not possible. The concept of the “tacit” refers to the knowledge in human heads, while the concept of the “explicit” concerns the knowledge represented in a documented, thus formal form. We posit therefore, that when the transition process is fully automated, the application domain knowledge must be of high quality and encompass all knowledge needed for further software development. By high quality we understand that knowledge adequately reflects the problem domain needs, is consistent, and has agreed upon level of completeness with respect to the problem domain. By all knowledge needed for further development we understand those domain details, which are relevant in software development, but usually are not considered at the problem domain level (e.g., detailed data structures). Assuming that the fully automated knowledge transfer from the application domain level to the implementation is possible, this paper addresses the following research questions: o Do conventional object-oriented approaches are suitable for the construction of high quality application domain level models? o Can high quality application domain level models be developed through formal knowledge transfer from the problem domain models to the application domain models? In order to answer first question, we analyze four conventional object-oriented approaches with respect to the type of knowledge and the knowledge transfer. To obtain the answer to the second question we analyze the two-hemisphere model based approach to the software development [5], [6] that directly deals with this issue. In this paper our focus is on the knowledge transfer. The paper is organized as follows. First we discuss knowledge transfer that takes place during the analysis stage of the object-oriented software development. Next we present the results of our appraisal of four conventional object-oriented software development approaches from the point of view of the knowledge transfer. We follow our discussion with a brief on the two-hemisphere model based approach to the software development, and examine it with respect to the knowledge transfer. We conclude the paper with the suppositions concerning formal knowledge transfer. 1. Peculiarities of the Knowledge Transfer in Software Development Knowledge transfer is the process that changes the contents and/or the form of representation of knowledge. Knowledge transfer also includes such operations as knowledge integration and knowledge generation. When considering the level of abstraction, we distinguish between the problem domain knowledge, application domain knowledge, and implementation level knowledge (Figure 1). When considering knowledge representation, we differentiate between the tacit and explicit knowledge. Explicit knowledge can be represented in an unstructured, semi-structured, and structured forms. A

158


model is a form of knowledge representation that can be tacit as well as explicit; a diagram, in turn, is a model, which is represented as a structured explicit knowledge. In this paper we will be concerned with knowledge represented in the UML that is declared as a standard [3] used in representation of the software model. UML defines a notation for a set of diagrams to demonstrate the different aspects of the application and implementation domains. Figure 1 shows the minimal set of the UML diagrams for the object-oriented software development.

Problem domain

Use-Case Model

Application domain

Class Diagram

Interaction diagram

Implementation domain Component diagram

Figure 1 A minimal set of UML diagrams for the object-oriented software development

Specifically, components diagram reflects the design of the software to be developed. Class diagram, at the application domain level, shows an overall structure of the software system and encapsulates the responsibility of each class (its attributes, methods and associations with other classes). Interaction diagrams show the sequence of message passing between objects during system execution. The interaction diagrams are created for each use-case of a system that reflects the way the system serves its users. The modeling efforts in the object-oriented software development usually start with the identification of use-cases for the software to be developed. A use-case reflects interactions between the system to be built and the actor (an outside object in a particular role) that has a particular purpose of using the system. Use case description is a semistructured explicit knowledge, because it has a structured frame, which is filled with unstructured text. On the other hand, the use case diagram and the sequence diagram, which represent a part of knowledge included in the use-case, are structured explicit knowledge.


159

When knowledge at a particular level of abstraction is processed it can acquire different form and/or content, i.e., it is transferred from one form of representation into another. In the object-oriented software development we can distinguish between the following types of knowledge transfer: o inside the problem domain o from the problem domain into the application domain o inside the application domain o from the application domain into the implementation In Figure 2 an example of knowledge transfer from a problem domain to application domain is shown by dotted line arrows, knowledge transfer inside the application domain and knowledge transfer from application domain to implementation are shown by the continuous line arrows.

ca ses as so c ia tio sc ns rip ts fo ru se c

us e

ts j ec ob

ac to rs

as es

Problem domain

objects as classes Application domain

Use case model

use cases actors associations scripts for use cases

actors as actors

steps of scripts as messages

parameters as attributes Interaction diagrams

actors objects messages

messages as operations collaboration as associations

well-d

Implementation domain

efin

at ses are ed clas

Class diagram

classes attributes operations associations

ts s ponen encie to com pend tached as de s n tio ia c o ass

Component diagram

components dependencies

Figure 2 Examples of knowledge transfer

As noted in the Introduction, if we assume that the fully automated knowledge transfer from the application domain level to the implementation is possible, we need to consider the preconditions for the use of the formal (that is subject to the full automation) knowledge transfer, from the application domain level to implementation. First we note that when the knowledge is to be transferred from the application domain level to the implementation level automatically, tacit human knowledge does not enter this process. This leads to the second observation: successful software implementation depends on the application domain model, i.e. structured explicit knowledge. Here we note that these observations signal an important issue for the object-oriented software development: how to define reliable classes that could be translated into the programming language [2]. The automatic knowledge transfer from the application domain into implementation can lead to high-quality software code only if the application domain knowledge reflects real needs of the problem domain. Research in the area of requirements engineering has shown that consistency [7] and completeness [8] are two significant features of the

160


knowledge required to describe the real needs of a problem domain. In this context it is necessary to observe that (1) completeness of knowledge is always relative, and (2) to acknowledge that the knowledge is not complete if it is inconsistent and that it may be incomplete even it is consistent [7]. Moreover,knowledge, at the application domain level, can be syntactically consistent and semantically consistent. It is syntactically consistent if all syntax requirements of the structured forms of knowledge (diagrams) are met. Knowledge is semantically consistent if different forms of knowledge utilized for an application domain level do not contradict each other. In order to meet the problem domain needs for software, the consistency between application domain level knowledge and problem domain knowledge is critical. The use of formal methods for knowledge transfer enhances software development speed and efficiency [2]. However it does not guarantee the effectiveness of the software development. The reason is rather simple: tacit knowledge is always richer than the explicit knowledge extracted from it. Rich tacit knowledge can be very useful in object-oriented software development, when the knowledge transfer is mainly informal and semi-formal knowledge transfer from problem domain to implementation takes place. This has been unequivocally demonstrated by the success of the agile software development approaches [9]. Consequently raising the formality level of the knowledge transfer implies higher requirements for completeness and consistency. A consistency of knowledge is required not only inside the application domain level, but also between the problem domain level and application domain level. The later type of consistency can be proved only if the problem domain knowledge is explicit, structured, complete and consistent. If such knowledge exists at the problem domain level then formal knowledge transfer from problem domain knowledge into application domain level is possible. Therefore, from the point of view of utilized knowledge transfers, applicability of particular object-oriented software development methods depends on availability and the possibility of obtaining explicit, structured, complete and consistent knowledge at the application domain and/or problem domain levels. 2. On the Possibility of Knowledge Transfer in the Conventional Object-Oriented Approaches We now analyze four conventional object-oriented approaches with respect to the knowledge transfer from the problem domain into the application domain model. We distinguish between the following three modes of knowledge transfer: formal, semi-formal, and informal. Formal transfer of knowledge may be done automatically by software development tools that follow the given transfer algorithm. Formal (and automatic) transfer is possible only if the source knowledge is explicit and structured. Automatic transfer may increase the speed of the software development considerably [10]. The semi-formal transfer of knowledge may also increase this speed. The semi-formal transfer of knowledge is executed by the human expert who uses specific transfer principles and is supported by a software tool that performs part of the transfer activities according to the transfer rules that are built into the tool. Thus, the semi-formal transfer is possible if at least part of source knowledge is presented as explicit knowledge. A minimum requirement for the semi-formal transfer is a possibility to track the link between knowledge element in the source knowledge and corresponding knowledge sub-structure in the target knowledge representation. Informal knowledge transfer fully relies upon the human intellect and is fraught with the problem of subjectivity.


161

In this paper our analysis of the knowledge transfer for different software development approaches centres around the following concerns: o Knowledge type(s) in the problem and application domains o Means for proving the completeness of knowledge o Means for proving the consistency of knowledge o Mode(s) of the knowledge transfer inside the problem or the application domain o Mode(s) of the knowledge transfer from the problem domain into the application domain o Means for the check of the knowledge consistency between the problem domain level and the application domain level Different software development approaches are illustrated by the simplified metamodels (Figure 3 to Figure 6). The results of the analysis are summarized in Tables 1-4. We posit that these results offer an interesting insight into the nature of the knowledge transfer. They also aid the assessment of the applicability of the given approach when using the formal knowledge transfer from the application domain level into the implementation level. 2.1. Informal Approach In the informal approach [11] the software development process is based on a problem statement and the conversations with the customers and the domain experts. The metamodel of the approach is shown in Figure 3.

Problem domain Problem statement

Conversation with customer and domain experts

Use-Case Model

Application domain

Interaction diagram

Class Diagram

Figure 3 Informal approach

Knowledge at problem domain level consists of the tacit knowledge of the customers, the domain experts and the developers, and the unstructured explicit knowledge reflected in the problem statement. Application domain knowledge consists of the semistructured use-cases and the structured use-case diagram, sequence diagrams, and class diagram. The knowledge transfer from the problem domain into the application domain is informal, unless the links between the pieces of the problem statement and the corresponding application domain knowledge substructures are formally recorded. Thus we note that the informal approach assumes that the knowledge externalized by the

162


customers and the experts is recorded only in the application domain level use-case model (Table 1). Table 1. Features of the Knowledge and the Knowledge Transfer: Informal Approach Problem domain Knowledge type(s)

Unstructured explicit Tacit

Means for proving the completeness of knowledge

Informal

Means for proving the consistency of knowledge

Informal

Mode(s) of the knowledge transfer inside the domain Knowledge transfer from the problem domain into the application domain Means for the consistency check between the problem domain and the application domain

Application domain Semi-structured explicit Structured explicit Tacit Informal Semi-formal (syntax only) Informal Semi-formal Informal Semi-formal

Informal Informal Semi-formal (if utilized) Informal

2.2. Problem Domain Model and the Application Domain Model Based Approach In the problem domain model and the application domain model based approach [12] the problem domain level objects/classes model supports the knowledge transfer (Figure 4).

Conversation with user Problem domain Application boundary

Conversation with domain experts

Problem domain

Domain model

Application

Use-Case Model

Application model Application domain

Class Diagram Interaction diagram

Figure 4 The problem domain model and the application domain model based approach

As indicated in Figure 4, the problem domain knowledge acquisition starts with an identification of an application boundary. Conversations with the users and the domain experts constitute the main source of knowledge for the problem domain analysis and the application boundary identification. Further, two prong analysis is used for the knowledge acquisition, i.e., first, a problem domain model (by capturing the domain knowledge) is built and then an application domain model (by examining the use-cases of the particular


163

application) is developed. In this approach there is no formal strategy on how to identify the application boundary. Nevertheless, several semi-formal strategies for the domain class identification exist [13]. Similar to the approach discussed above is a method of software development [13] that is based on a detailed knowledge about problem domain and application domain. Detailed knowledge is obtained by exploiting different sources of knowledge including patterns. The strategy of knowledge acquisition and transfer utilized in this method also uses objects as its key concepts in problem domain modeling, and usecases - as a modeling tool for application domain analysis. The problem domain model and the application domain model based approach to the software development is much richer in the explicit knowledge, when compared to the informal approach (see Tables 1 and 2). Table 2. Features of the Knowledge and the Knowledge Transfer: The Problem Domain Model and the Application Domain Model Based Approach

Knowledge type(s) Means for proving the completeness of knowledge Means for proving the consistency of knowledge Mode(s) of the knowledge transfer inside the domain Knowledge transfer from problem domain into the application domain Means for the consistency check between the problem domain and the application domain

Problem domain Structured explicit Unstructured explicit Tacit Informal Informal Informal Semi-formal (if utilized) Informal Semi-formal (if utilized)


Informal Semi-formal

Two types of class models and the use-case model are utilized in the problem and application domain analysis, i.e., the declarative explicit knowledge, in terms of object/class model is provided for both, the problem domain and the application domain analysis, while the explicit procedural knowledge (the use-case model) is given only for the application domain. Use of the explicit knowledge allows for the semi-formal checks of consistency and completeness with respect to the initial knowledge to be transformed into the implementation model. On the other hand, here no formal means exist to prove the consistency of this knowledge. 2.3. Requirements Specification Based Approach Several approaches [14], [15], [16], [17] base identification of the use-cases on an initial (not final) requirements specification document that customarily is referred to as user requirements. The general meta-model of the requirements specification based approach is shown in Figure 5.

164


Problem domain

Requirement specification document

Use-Case Model Conceptual model Application domain

Interaction diagram

Class Diagram

Figure 5 Requirement specification based approach

The requirements specification based approach assumes that the conceptual model has to be built as a basis for class identification. The identification of concepts, in turn, is part of an investigation of the problem domain [15], [16]. Conceptual model creation is based on the use-cases identified for the software system and depends on the use-case descriptions and other documents, such as the user requirements specification, in which the relevant concepts may be found. The approach based on the requirements specification document uses a semi-formal source of knowledge for the construction of the initial body of knowledge to be transformed during the software development process. The approach, on one hand, provides opportunities for the semi-formal consistency and completeness checking between the problem domain and the application domain, but on the other hand, heavily depends on the quality of knowledge in the requirements specification provided in the form of the specification document. On the basis of the requirements specification, explicit procedural and declarative knowledge is provided in the form of the use-cases and the conceptual model (Table 3.) Table 3. Features of Knowledge and Knowledge Transfer: Requirements Specification Based Approach Problem domain Knowledge type(s)

Semi-structured explicit


Informal

Means for proving the consistency of knowledge

Mode(s) of the knowledge transfer inside the domain Knowledge transfer from the problem domain into the application domain Means for the consistency check between the problem domain and the application domain

Informal

N/a Informal Semi-formal Informal Semi-formal (if utilized)



165

Use-cases reflect the application domain knowledge. Conceptual model comprises both the problem domain and the application domain knowledge elements. The semi-formal methods of the knowledge transfer, from the conceptual model into the class diagram can thus be applied. However, the representation of the knowledge in a form of the use-cases does not support the formal methods for the consistency and the completeness checks of the knowledge. 2.4. Rational Unified Process based approach In the Rational Unified Process based approach the business modeling, in terms of the business use-cases and the business object models, is used for analysis of complex and multidimensional application environments [18], [19]. A meta-model of the approach is shown in Figure 6.

Problem domain

Business Object Model

Business Use-Case Model

Application domain

Use-Case Model Interaction diagram

Class Diagram

Figure 6 Rational Unified Process based approach

Business level, i.e., the problem domain models, may be translated into the systems level models in the following way [18]: o Business workers become actors in the systems to be developed. This is possible because the business object model contains the business workers as one type of its objects o Behaviors described for the business workers are things to be automated; they assist in finding the software systems use-cases and define needed functionality o Business entities may become things maintained by the help of the software system; they may be used to find entity classes in the application domain model Rational Unified Process based approach clearly shows dependencies between the problem domain models and the application domain models. It is important to note however that the knowledge transfer strategy utilized in this approach does not give explicit answers to such questions as: “Should all business systems workers become actors in the use-case diagrams?”, “Do all potential system’s classes (concepts) are identified in the Business Object Model?” Thus it becomes obvious that only the semi-formal knowledge transfer is here possible. Nevertheless, the representation of the domain knowledge in the form of the use-case diagram is a step forward toward the ‘explicitness’ of the problem domain knowledge. The approach does not enable the use of the formal means for checking the consistency and the completeness of the knowledge represented in the diagrams. Specific details concerning the features of the knowledge and the knowledge transfer for the Rational Unified Process based approach are presented in Table 4.

166


Table 4. Features of Knowledge and Knowledge Transfer: Approach based on Rational Unified Process Problem domain Knowledge type(s)

Semi-structured explicit Structured explicit


Informal Semi-formal

Means for proving the consistency of knowledge Mode(s) of the knowledge transfer inside the domain Knowledge transfer from the problem domain into the application domain Means for the consistency check between the problem domain and the application domain

Informal Semi-formal Informal


Informal Semi-formal Informal Semi-formal

3. Suitability of the Application Domain Model to the Formal Knowledge Transfer to Implementation As previously stated, for the application domain knowledge to be formally transferable into implementation we require these conditions: o Knowledge should be explicit o Knowledge should be structured o Knowledge should be relatively complete o Knowledge should be consistent Analysis presented in Tables 1 trough 4 demonstrates that the object-oriented software development approaches we scrutinized can meet first two requirements. However, structured explicit knowledge (reflected by diagrams) is not the only type of knowledge to be relayed upon at the application domain level. In all approaches to software development it is backed up by tacit and/or semi-structured explicit knowledge. When knowledge transfer is informal or semi-formal, both tacit and explicit knowledge contribute to the overall completeness of knowledge. If formal knowledge transfer takes place, only completeness of structured explicit knowledge matters. Customarily, informal completeness and consistency of knowledge is based on the consensus and the agreements between the stakeholders involved in software development. Formal completeness and consistency of knowledge is achieved only if formal means exist to prove it. Completeness of knowledge at the application domain level depends on the completeness of knowledge at the problem domain level. Therefore we can claim that the complete application level knowledge can be achieved by the successful transfer of the complete problem domain level knowledge into the application domain knowledge. All four approaches to software development discussed above may produce informally complete application domain knowledge, if complete problem domain knowledge is available. However none of these approaches can guarantee or prove the success of the transfer. The most vulnerable, in this context, is requirements specification based approach, if informal or semi-formal knowledge transfer is not enhanced by the developers’ tacit knowledge at the problem domain level. Only the Rational Unified Process based approach provides the semi-formal means for the knowledge completeness checking at the problem domain level. None of the approaches is suitable for achieving formal completeness of the


167

application domain level knowledge. Formal completeness of the application domain knowledge could be proved only if all problem domain knowledge to be transferred was represented as a structured externalized knowledge, and its consistency and completeness were proved. Complete knowledge is always consistent [7]. Therefore consistency of the knowledge at the application level and the problem domain level is a precondition of the formal knowledge completeness. We wish to point out that none of the approaches we analyzed utilizes the formal consistency check of the knowledge. Therefore we can conclude that conventional object-oriented approaches cannot provide formally complete and consistent application domain knowledge. If consistency and completeness of the knowledge cannot be proved formally it means that the software development approach is suitable only for comparatively small and simple problem domains where it is reasonable to depend on the informal and semi-formal means of the check for completeness and consistency. Next we discuss the two-hemisphere model based approach that utilizes both, the semi-formal and the formal knowledge transfer. We posit that this approach represents a step toward formally complete and consistent application domain knowledge. 4. Knowledge transfer in Two-Hemisphere Model Based Approach All approaches reviewed so far allow only for the semi-formal knowledge transfer. Formal transfer is not feasible because of a quite considerable impact of a tacit dimension of the problem domain knowledge, from which the transfer could initiate. If the completeness and the consistency of the explicit problem domain knowledge is not proved, then semi-formal transfer may give better results than the formal one, because it allows for the improvement of the initial knowledge during the transfer. On the other hand, from the point of view of speed of the software development, formal knowledge transfer is more desirable than a semi-formal one [2]. That is why we propose to commence and base the object-oriented software development process on the two-hemisphere problem domain model [5], [6] where one “hemisphere” reflects functional (procedural) aspects of the business and the software system, and another “hemisphere” reflects the corresponding conceptual structures (Figure 7).

Problem domain Model of system functioning

Application domain

Conceptual Model

Use-Case Model Interaction diagram

Figure 7 Two-hemisphere model based approach

Class Diagram

168


The process of the model construction may start from the conceptual model as well as from the functional one. The co-existence and inter-relatedness of these models enables the utilization of the formal and the semi-formal knowledge transfer, as well as the utilization of the knowledge completeness and consistency checks [5]. Initial version of the two-hemisphere model based approach was proposed by Nikiforova [20], where the general framework for the software development has been discussed and its application has been demonstrated in practice. The approach supports gradual model transfer from the problem domain models into the program components. Functional model is reflected in the form of a business process model. Although the notation of the business process model may be optional, it must reflect the following components of the business processes [8]: external entities (external processes), subprocesses (the number of levels of decomposition is not restricted), performers, trigger conditions, information flows, and information (data) stores. Real-world classes relevant to the problem domain and their relationships are presented in the concepts model (or system conceptual model). Conceptual model shows the things that exist in the problem domain and their relations to other things. The notational conventions of the business process diagram allow identification of the concepts through the analysis of all the data stores in the functional model [20]. Data stores from the functional model can be transformed into the concepts of the conceptual model. Both functional and conceptual models serve as an initial source for further knowledge transfer [20]. Formal transfer from the functional (business process) model into the use-cases is possible, if processes to be performed by software system are identified in the business process model [21]. Processes to be performed by software system become use-cases in the use-case model, performers of related processes become actors in the usecase model, and scenarios for realization of the use-cases may be defined by the decompositions of the business processes (sub-processes) corresponding to the use-cases [20]. Interaction diagram is developed for each use-case and is based on its realization scenario (or sequence of sub-processes). Appropriate interacting objects are extracted from the conceptual model. The class diagram is based on the conceptual model and is formed according to the information in the interaction diagram. Here, the class diagram is already a structure of a software application and contains only those classes, whose objects interact during the use-case realization [20]. Features of the knowledge at the problem domain and the application domain levels as well as features of the knowledge transfer are shown in Table 5. Table 5. Features of the Knowledge and the Knowledge Transfer: Two-Hemisphere Model Based Approach

Knowledge type(s)

Means for proving the completeness of knowledge Means for proving the consistency of knowledge Mode(s) of the knowledge transfer inside the domain Knowledge transfer from the problem domain into the application domain Means for the consistency check between the problem domain and the application domain

Problem domain Structured explicit Informal Semi-formal Formal Formal (syntax only) Semi-formal Informal Semi-formal Semi-formal Formal

Formal

Application domain Structured explicit

Semi-formal Formal Formal (syntax only) Semi-formal Informal Semi-formal


169

Possibilities of the semi-formal and formal knowledge transfer in the twohemisphere approach at its current stage of the development are shown in Figure 8. Problem domain

information flows as concepts

Business process model

Conceptual model

concepts associations attributes

units of information stores as attributs hierarchy of information stores as associations

con

attributs

cep ts a s

associacions

obj ect

s

sub-processes as scripts for use cases

performance as associations

processes as use cases

performers as actors

processes performers sub-processes information flows information stores

objects as classes steps of scripts as messages

parameters as attributes

Use case model

use cases actors associations scripts for use cases

Class diagram actors as actors

Interaction diagrams

actors objects messages

messages as operations collaboration as associations

classes attributes operations associations

Figure 8 Possibilities of the formal knowledge transfer in the two hemisphere model based approach

Semantic completeness of the functional model depends on the richness of the knowledge behind the model and cannot be formally proved. However, existence of the conceptual model gives some semi-formal means for checking the semantic completeness and consistency of the business process model. Informal semantic completeness of the business process model can be achieved [8], if the wide spectrum of the business aspects of the enterprise (to which the process belongs) are taken into consideration (e.g., mission, system of values, goals and constraints, general functions, physical functions, technical infrastructure, etc.) Syntactic completeness of a business process model can be proved formally (or automatically) by checking the presence of all components mentioned above, for each business process (sub-process) reflected in the model. This possibility enables the formal knowledge transfer during the object-oriented software development process. It also contributes to the semi-formal semantic consistency checking, because syntactically incomplete model cannot be semantically complete. As seen from Table 5 not all types of knowledge transfer are formal. Nevertheless the approach gives an opportunity to develop a relatively complete and consistent application domain level platform independent model, if the business process model is explicit, structured, and semantically and syntactically complete. Completeness and consistency of the application level knowledge then is proved by formal consistency check between the problem domain and the application domain levels. We note that the comprehensive discussion regarding the methodology and tool support for the two hemisphere model based approach is outside the scope of this paper.

170


Conclusion One of the most desirable attributes of the software development process is high speed of the development. Agile methods have offered best results for rapid software development [9]. However the applicability of agile approaches is restricted by their ability to handle only relatively simple problem domains [22]. Another approach that aims at increasing the productivity of the software development process is the model-driven software development [1], [2]. The level of complexity of the problem domain does not restrict the applicability of this approach [1]. Success of both approaches may be attributed to the speed of knowledge transfer during software development process. One of the main differences between these approaches is the type of knowledge behind the software code. Agile approaches significantly rely upon the tacit knowledge, while the model driven development mainly utilizes models (diagrams) in the form of explicit structured knowledge. Effectiveness and efficiency of agile methods is based on the transfer of knowledge performed by humans. On the other hand productivity of the model-driven approach depends on the automated transformation of the platform independent application domain model into the platform specific model, from which the software code may be generated. Nevertheless, the possibility of automatic (or formal) knowledge transfer from the application level to the implementation does not warrant that the software code will be useful, and therefore the gains of the automatic transfer may be insignificant. The level of warranty could be elevated if the completeness and the consistency of the application domain model were proved. Completeness of the application domain model is rooted in the problem domain knowledge, therefore completeness and consistency of the problem domain knowledge is a precondition for a high quality application domain models. It is possible to prove the completeness and the consistency of the problem domain knowledge only if this knowledge is externalized and structured. Once it is externalized and structured, the possibility of the formal (automatic) and the semi-formal knowledge transfer, from the problem domain into application domain, occurs and thus offers additional opportunity to accelerate the software development process. The analysis of several conventional object-oriented software development approaches we undertook shows that the use of the formal and the semi-formal transfer of knowledge in the development process are hindered by the impossibility to achieve and assure completeness and consistency of the initial body of knowledge. Therefore we conclude that these conventional approaches are not suitable for the formal knowledge transfer from the application domain into implementation stage for complex problem domains where informal consistency and the completeness checks of the application domain knowledge are insufficient. We propose that the two hemisphere model based approach is as a step forward in the utilization of the formal knowledge transfer. The approach, by expressing the problem domain knowledge in terms of the business process model, puts the emphasis on the functional aspects of the problem domain to a higher extent than it is common for the conventional object-oriented software development approaches. Through the analysis of the two-hemisphere model based approach we show that the formal knowledge transfer from the problem domain into implementation is possible. Moreover, it may offer effective means for software development, when sufficiently complete and consistent, explicit and structured, problem domain knowledge is available.


171

References [1] MDA Guide Version 1.0.1 from http://www.omg.org/docs/omg/03-06-01.pdf [2] Sendall Sh. & Kozaczynski W. Model Transformation: The Heart and Soul of Model-Driven Development, IEEE Software, Sept. / Oct., pp. 42-45, 2003 [3] Object Management Group, OMG Unified Modelling Language Specification, Version 1.5., from www.omg.org/technology/documents/formal/uml.htm [4] Weis T, Ulbrich A, & Geihs K. Model Metamorphosis, IEEE Software, Sept. / Oct., pp. 46-51, (2003) [5] Nikiforova O. & Kirikova M. Enabling Problem Domain Knowledge Transfer during Object-oriented Software Development, Constructing the Infrastructure for the Knowledge Economy:Methods and Tools, Theory and Practice, H. Linger, J. Fisher, W. Wojtkowski, W.G. Wojtkowski, J. Zupancic, K. Vigo and J. Arnolds (Eds.), Kluwer Academic/Plenum Publishers, New York, pp. 481-493, 2004 [6] Nikiforova O. & Kirikova M. Two-Hemisphere Model Driven Approach: Engineering Based Software Development, Proceedings of the 16th International Conference Advanced Information Systems Engineering CAiSE’2004, A. Persson and J. Stirna (Eds.), LNCS 3084, Springer-Verlag Berlin Heidelberg, 2004, pp. 219-233., [7] Kirikova M. Consistency of Information in Requirements Engineering, Information Modelling and Knowledge bases IX, P-J. Charrel, H.Jaakola, Hkangassalo, E Kawaguchi (Eds.), IOS Press, Amsterdam, Berlin, Tokyo,Washigton, DC, pp. 192 – 205., 1998 [8] Kirikova M. Towards Completeness of Business Models, Information Modelling and Knowledge Bases X, H. Jaakola et al. (Eds.), IOS Press, pp. 42-54., 1999 [9] Ambler Sc. W. (2003) Agile Model Driven Development is Good Enough, IEEE Software, September/October, pp.71-73. [10] Avison D. & Fitzgerald G. Information Systems Development: Methodologies, Techniques and Tools, McGraw-Hill Education, 2003. [11] Quatrany T. Visual Modeling with Rational Rose 2000 and UML (2nd ed.) Addison-Wesley, 2000 [12] Rumbaugh J. “Getting started: Using Use-cases to Capture Requirements”, Journal of Object-oriented Programming, Sept., pp 8-11, 1994 [13] Mathiassen L., Munk-Madsen A., Nielsen P. A. & Stage J. Object-oriented Analysis & Design, Marko Publishing House, 2000 [14] Larman Cr. Applying UML and Patterns: An Introduction to Object-oriented Analysis and Design, Prentice Hall PTR, 1998 [15] Fawler M. Analysis Patterns: Reusable Software Models, Addison-Wesley, 1996 [16] Martin J. & Odell J. Object-oriented Methods: A Foundation, Prentice Hall, 1995 [17] Rumbaugh J. Models Through the Development Process, Journal of Object-oriented Programming, May 1997, Vol. 10, No 2, pp. 5-8, 14 [18] Leffingwell D. & Widrig D. Managing Software Requirements: A Unified approach, Addison-Wesley, 2000 [19] Kruchten Ph. The Rational Unified Process: An Introduction”, Second Edition, Addison-Wesley, 2002. [20] Nikiforova O. General Framework for Object-Oriented Software Development Process, Scientific Proceedings of Riga Technical University, Series – Computer Science. Applied Computer Systems, 13 vol., pp.132-144, 2002 [21] Kirikova M. Business Modelling and Use Cases in Requirements Engineering, Information Modelling and Knowledge Bases XII, H.Jaakkola, H.Kangassalo E. Kawaguchi (eds.), IOS Press, Ohmsha, Amsterdam, Berlin, Oxford, Tokyo, Washington, DC, pp. 410-420., 2001 [22] Turk D, France R., & Rumpe B. Limitations of Agile Software Process, from http://www.agilealliance.com/articles/articles/LimitationsofAgile.pdf (2003)

172


Toward a Cooperative Recognition of Human Behaviors and Related Objects Tadahiro Kitahashi 1, Atsuhiro Kojima 2, Mirai Higuchi 2 and Kunio Fukunaga 2 1࡮Kwansei Gakuin University, Gakuen, Sanda Hyogo 669࡮337, Japan 2࡮Osaka Prefecture University, Gakuen-cho, Sakai Osaka 599࡮531, Japan Abstract: Recently we proposed a new scheme of aiming at scene recognition by understanding human activities, objects and environment from video images including these constituents. The most significant difference of this approach from conventional ones is the proposal of a method for cooperative understanding of humans’ actions and related objects through analyzing motion of humans’ face and hands. It is based on the idea that usefulness and functions of an object a person is going to deal with could be inferred from analyzing the person s movements because of that human motions, especially trajectories of face and hands, and the relative position and movement of the objects are closely related to the usage and the states of the objects. In this paper, we will describe the above mentioned scheme focusing on the structure of database and its usage in inference for recognizing human movements and objects. e-mail: [email protected], fax:+81-79565-9077, [email protected], [email protected],

Introduction The recent trend in the field of computer vision could be recognition of human face and gestures as well as environment recognition including human movements. Recently, Higuchi, Kojima and Fukunaga propose a method to recognize objects and their arrangement in a scene by recognizing human movements using extracting trajectories of parts of a human body, mainly face and hands together with depth information from a conventional stereoscopic method. Extending the concept, we intend to recognize usage of objects by means of inferring indirectly from observation of human movements interacting with object without shape models that are commonly used so far [2] in object recognition. Consequently unlike the conventional recognition, objects are identified as rather objective of behaviors than categorical labels. A formulation of interactions between human and objects was proposed by Siskind et al as a description of the scene dynamics [3, 4]. They present a qualitative perception based on kinetic analysis for simple scenes; for example, describing that a block is put on the table or picked up by human hand. Indeed, such a basic work of human-object interaction is actually essential in this field, but their bottom-up approach may not be applicable to the analysis of human and object movement in video images of a daily life scene. From the view point of modeling objects by functions, Kitahashi et al. proposed a functional model discussing the relationship between the structure of an object and the structure and the material of its user[5]. Applying the concept to characterizing dynamic objects such as tools, they presented a method for inferring the adaptability of an object as a tool such as a wrench and a bottle-opener from the shape and usages. For most of objects, however, it is not easy to infer the function directly from its shape. We will show the function, or more precisely the usage of objects can be inferred indirectly from watching how people actually handle the objects. Moore et al. proposed a method to recognize objects using model-based object recognition in accordance with tracking human actions [6]. The interesting point of their method is that it employs two complemented models, that is, shape models and action-based ones at the same time.

T. Kitahashi et al. / Toward a Cooperative Recognition of Human Behaviors and Related Objects

173

However, they do not positively utilize the concepts of intrinsic function of an object. In contrast, we more focus on intrinsic functions of an object. For example, a table can be defined as a thing vertically supportable other objects with a flat and rigid surface superior to visual features of it. Thus, it can be identified in an image sequence by recognizing a space where a person put on an object brought with. We believe that humans recognize actual functions of an object by mean of touching it directly or at least observing that a person is using it in some scene. As mentioned above, the approach of cooperative recognition of human movements and related objects in a human action provides a novel and timely scheme for circumstance recognition just challenged by many researchers in the Computer Vision field. Also, cooperative recognition of human movements and related objects may be an active recognition scheme in contrast with the conventional static scheme of object recognition merely based on the apparent shape. As well, the object recognition through human movements allows us to classify the object according its usage in the action, and also in some cases, for example a case of eating a rice-ball, to infer the material of the object by detecting a kind of shape deformation of the object during the human movement. At the same time, the object classification inversely makes it possible to assign the movement to a type of the human actions. This scheme results in labeling the recognized object by not an individual conceptual label but a functional label such as edible goods, portable ones and so on.

1. A Model of Objects and Human Movements Kojima et al.[7] have shown that the concept hierarchies of human actions and objects are useful for identifying actions and related objects. A hierarchical model of objects based on the frame representation is shown in Figure 1(a). Each node has slots indicating the structures and function of the object such as IS-A, HAS-A-PART, FUNCTION and ATTRIBUTE. This hierarchy starts from the top node : UNKNOWN.

FREE SPACE

SPACE

Fig.1(a) Conceptual Hierarchy of Objects

174


Fig.1(b) Conceptual Hierarchy of Actions This node branches into more than two nodes representing objects(occupied space) and vacant space. This can be verified by analyzing human activities. For example, a region where a person’s body passes through can be considered to be unoccupied vacant space. On the other hand, objects consist of VSO (Vertically Supporting Object) and HSO (Horizontally Supporting Object). The former represents the objects that can vertically support other objects like a desk and a shelf, and the latter is the ones that can support other ones horizontally like a white board and a wall. These categorical labeled objects are descendant nodes of VSO and HSO. It is noteworthy that VSO, HSO, FREE_SPACE and so forth are unusual as the labels of object models, since they represent rather functions or usages of objects than the appearance of them. However, they are more familiar with human actions than other attributes like shape and color of an object. Functions are distinguishable from other appearance based features at the point that they have no visual form in the real world, and can solely be realized in the course of a human activity. In this sense, the functional attributes have a close relationship to human actions. On the other hand, a hierarchical model of human actions shown in Figure 1(b) consists of nodes with slots labeled as IS-A, OBJECT, GOAL, PRECOND and EFFECT. These slot names are exact to describe the action and its relationship to the environment. It is remarkable that OBJECT and GOAL slots have functional attributes as their values. As a consequence, the two kinds of items in human actions hierarchy and objects one could be linked by the labels shared between the nodes of the hierarchical object model and OBJECT or GOAL slot of a node in the hierarchical action model. As you can see, the novel description provides an essential mechanism of intelligence for cooperative recognition of human movement and objects, that The expression is shared with the two kinds of hierarchical databases. For example,

2. Cooperative Recognition of Human Behaviors and Objects Therefore, referring to the characteristics of the process of typical human movements interacting with objects as mentioned above, we acquire a procedure for recognition of human movements and objects related to them. Namely, the procedure to recognize human movements could usually start with capturing a move of a head and hands from an image sequence. Actually, the head movement represents the transfer of the whole body, while that of the hands does


175

various kinds of manipulations.

2.1 Recognition of the Head and Hands and Tracking their Motion. Going into the details of human movements, when a person wants to do something somewhere, she/he usually walks close to the target, moves his/her arms so as to reach hands and fingers to the object and then accomplishes his/her goal, in many cases, by manipulating the object. Accordingly, the recognition process should also trace the same sequence of human actions as begins with discerning a horizontal move of the human head, since the move of a head represents that of whole body. When detecting a break of the head translation, which means stop of walking, we should turn the visual attention to hands and some objective related to them, and to tracking their motion. Extracting skin-color regions, we distinguish hand’s candidates from them by comparing the size of the extracted plausible regions. After discerning the candidate regions of hands, we have to track their move to the target position for object manipulation. At the point where the hand stops, we may again switch the target of detection to the objects for hand-manipulation and finally pay attentions to detailed movement of hands and fingers for their manipulations of the object. The feature described in the last paraphrase is verified by a couple of examples such as Fig. 2 and Fig.3 Especially, the time-varying velocity graph shown in Fig.6 depicts the features mentioned in the above description. This is the first step of solving the problem. As will be described in below, more important point of this paper is that if the sequence of the hand actions is successfully extracted, it just means a simple movement of the hand, without observing the cooperation of the body parts and change the size, location of the object. Then, what can happen by the scheme?

2.2 Change in Objects according to Object Manipulation by Hands Actually when we apply an intended manipulation to an object, small or big change will often occur on some factors of the object such as its position and its shape. For example, as shown in Fig.2 and 3, an object disappears at a point around the mouth, on the moving process of a hand and an object. This fact allows us to infer that the object must be a kind of foods based on the knowledge about daily life. Generally speaking, it is obvious in the situation like this example that the change in the object surely results from humans’ action of manipulating objects with fingers and hands. Few computer vision people have ignored the cooperation between human act and the object so far.

Fig. 2 Extracted regions of the face and the hand and the trace of the hand movement

176


ᴾᴾᴾᴾᴾᴾᴾᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ ᴾ Fig.3 ᴾ Disappear of an Object around Mouth Taking into account the relationship, we could create another kind of object recognition scheme than the conventional ones in favor of detecting human movement. It is an evidence of the assertion that the proposed scheme uses none of shape models that are inevitable in the conventional pattern recognition. Moreover, in addition to object recognition, the recognition in turn makes it possible to interpret the human movement as a meaningful act such as eating. It is not an easy problem to understand the meaning of a human gesture solely from processing the image sequence with the technique of usually employed in current image understanding.

2.3 Cooperative Recognition of Human movement and Object Thus, detecting disappearance of an object in a process of a go-and-back movement of a hand reveals that the object is an edible thing, and also this object recognition reflectively makes it clear that the simple go-and-back movement itself means an eating procedure. In this manner, we can understand these kinds of video images by recognizing human actions and objects related to them in a scene at the same time. The novel approach of a scene understanding method could be called as a dynamic scene understanding in contrast with a static one. One more new aspect of our method is that it brings the concept of usage in the recognition of an object. It is a significantly different point of this method from the traditional object recognition.

2.4 Inference Mechanism in the Databases of Cooperative Recognition ᴾ ᴾ There could be a couple of methods to implementᴾ the above mechanism. We employ a scheme of referring items with sharing slot values in the hierarchical database. As for the example of

ACTION

STAND UP

SQUAT DOWN

CO-MOVE with OBJECT IS-A: OBJECT:

ACTION PORTABLE OBJECT

RELEASE OBJECT

INDIRECTLY (not in HAND) DIRECTLY (in HAND) IS-A:

CO-MOVE with OBJECT

IS-A: PRECOND: EFFECT: INSTR:

CO-MOVE with OBJECT EXISTING INSTR. CHANGE LOCATION WITH HAND A KIND OF TOOL

EATING

Fig. 4 Focused Parts of Action Hierarchy

IS-A: INDIRECTLY OBJECT: EDIBLE OBJECT GOAL: MOUTH


177

OBJECT PORTABLE OBJECT

RIGID OBJECT

IS-A: OBJECT FUNCTION: CO-MOVE with HAND

IS-A: OBJECT FUNCTION: SUPPORTABLE

EDIBLE

SHEET

HSO

VSO

IS-A: PORTABLE OBJECT FUNCTION: DIVIDABLE

IS-A: PORTABLE OBJECT FUNCTION: ATTRIBUTE: THIN, LIGHT

IS-A: RIGID OBJECT FUNCTION: HORIZONTALLY SUPPORTABLE

IS-A: RIGID OBJECT FUNCTION: HORIZONTALLY SUPPORTABLE

Fig.5

Focused Parts of Object Hierarchy

eating procedure, based on the results of image processing, that is, detecting a hand motion together with an object region, the reference begins with the node INDIRECTLY(a child of CO-MOVE with HAND) in the action hierarchy shown in Fig.4. As the node includes CO-MOVE with HAND, the reference moves to the node PORTABLE of the object hierarchy depicted in Fig.5, the slot value of which is inherited from the parent CO-MOVE. Recognizing disappearance of the object region at the mouth from an image sequence, the referring point at PORTABLE descends to one of its child node EDIBLE. Then, the reference point shifts to the node EATING that shares the slot value with EDIBLE. In this manner, the system deduces that the human movement and the related object are rating process and an edible thing, respectively.

3. Detecting Human Movement from Trajectories of Human Head and Hands Returning to the situation mentioned at the beginning, humans usually act as, so to say, coarse to fine, that is, at first they walk and approach roughly close to the target point, then control the arm for a hand to reach the precise location for interacting with the object and finally manipulate or deal with it with a hand and fingers. In this sense, a human movement usually consists of sequential and hierarchical movements.

Average Velocity Time 㪌㪇

Move hand to a fork

Move hand to a cake Touch a cake

㸣㪋㪇

䵛

Pick up a fork

䊐䊧䌼

㪊㪇

䊛

Stick a cake Move hand upward Put a cake into mouth Move hand downward Put a fork on a table Move hand onto kneel

ᢙ

㪉㪇

㪈㪇

㪇㪇

㪈㪇

㪉㪇

㪊㪇

㪋㪇

䋨 ↹ ⚛ ᢙ 䋩 Motion㩷㩷㸢 Velocity ⒖ofേ 〒a㔌Hand

Fig. 6

Change of hand movement along eating process of a cake with a folk

㪌㪇

178


3.1 Characteristics of Human Movements Assuming the structure of a movement, we can design a simple strategy of image processing to recognize which part of human’s body should be paid attention at every moment. The rule is to distinguish an active phase and a still one in the move of each part of the body. It will be easily discriminated by drawing the chart of time sequence of variation of the parts location. The average value is a useful threshold to separate the two kinds of the phases. Fig. 2 and Fig.6 show an example of the rule. A hand of a man in Fig.2 moves in the manner of time variation of its velocity shown as the chart in Fig.6. The vertical axis of the chart indicates the flow of time, and the horizontal one does the amount of displacement of a hand in a short term, every five frames in this experiment. The average value of them is indicated with a red line. The periods in left hand side of the red line correspond to the still phase of hand movement. The rest does to the active one. In the active periods, humans mainly transfer themselves or the parts of their body in this case a hand. On the other hand, in the still phase they focus their attention on some tasks than transfer. This is a key notion of an algorithm described in follows.

3.2 The Detection of the Task Switching As discussed in the previous section, a procedure to recognize the interaction of humans and object based on humans’ movements could be summarized on the basis of hierarchical and sequential structure of humans’ movement. 1. Discriminate the active periods and the still ones by differentiating the movement data of the head and fingers. 2. During the still period of the motion of the body parts, for example the body, in parent level, we should pay attention to the movement, for example the hand, in lower levels. These processing are dealt only with subtraction of a current frame from a previous one.

4. Conclusion Image understanding techniques have recently improved to the extent of recognizing real time recognizing of the environment containing walking people. However, they have been recognizing with each of them and also other things separately, as treating humans merely as objects of dynamically changing form. On the other hand, in this paper, focusing on the analysis of cooperative movement of human body-parts and the manipulated objects, we have made it clear that the notion leads us to a novel scheme for computer vision and to understanding of usage of objects and meaning of cooperative human actions with objects. The method is strongly based on the knowledge of the scene of interest. This may open a way to connecting image processing with Knowledge Base and some of AI technology. However, the method described in 3.3 is not available in real time processing. It is a great disadvantage of it. We are developing a method to deal with it, and going to present it in near future.

References [1] M. Higuchi, S. Aoki, A. Kojima, and K. Fukunaga, ˈScene Recognition Based on Relationship Between Human Actions and Objects ˉ , Proc. of 17th International Conference on Pattern Recognition, Vol. 2, pp.73-78, Cambridge, United Kingdom, Aug. 2004. [2] A.R. Pope, ˈModel-Based Object Recognition: A Survey of Recent Researchˉ, Technical Report 94-04, University of British Columbia Dept. of Computer Science, Jan. 1994.


179

[3] R. Mann, A. Jepson, and J.M. Siskind, ˈComputational Perception of Scene Dynamicsˉ, Computer Vision and Image Understanding, Vol. 65, No. 2, 1997, pp.113-128. [4] A.P. Fern, R.L. Givan and J. Siskind, ˈSpecific-to-General Learning for Temporal Events with Application to Learning Event De.nitions from Videoˉ, Journal of Artificial Intelligence Research, Vol.17, December 2002, pp.379-449. [5] T. Kitahashi, M. Ashida, S. Dan and H. Ogawa,ˈFunctional Model and Shape Modelˉ, Technical Report AI-TR-92-3, ISIR, Osaka University, 1992, pp.1-10. [6] D. Moore, I. Essa, and M. Hayes, ˈ Exploiting Human Actions and Object Context for Recognition Tasksˉ, In Proc. of IEEE International Conference on Computer Vision 1999, Corfu, Greece, March 1999. [7] A. Kojima, T. Tamura and K. Fukunaga, ˈNatural Language Description of Human Activities from Video Images Based on Concept Hierarchy of Actionsˉ, International Journal of Computer Vision, Vol. 50, No.2, pp.171-184, 2002. [8] A. Kojima, T. Tamura, and K. Fukunaga, ˈTextual Description of Human Activities by Tracking Head and Hand Motions ˉ , In Proc of 16th International Conference of Pattern Recognition, Vol. 2, pp.1073-1077, Quebec City, Canada, Aug. 2002.


180

A Reliable and Useful Information Distribution System: the “Kuchicomi Network” Jun SASAKI, Tae YONEDA, and Yutaka FUNYU Iwate Prefectural University, Faculty of Software and Information Science Sugo 152-52, Takizawa-mura, Iwate-ken, Japan Abstract. As the Internet and mobile Internet have been expanding widely, we can now obtain many kinds of information easily and in large quantities. However, it is still difficult, for example, for tourists to get reliable and useful information about locations that they have never visited. In this paper we propose a reliable and useful information distribution system, called the "Kuchicomi Network," to provide “kuchicomi,” or word-of-mouth, information that is supplied by local informants through mobile phones or PCs connected to the Internet. We introduce the concept of, and business model for, a “Kuchicomi Network” and explain the system structure. We then describe features of the system and outline the development of a prototype system.

1. Introduction As recent telecommunication technologies such as the Internet and mobile Internet have rapidly advanced, large quantities of many kinds of information can now be easily obtained over the Internet. However, it is still sometimes difficult for users to get highly reliable and useful information. Tourists, in particular, often seek information about good hotels or restaurants located in the destinations they plan to visit for the first time. Though a lot of information about these locations can often be found on various Web sites or in guide books, such information is often part of commercial advertisements, and therefore the tourist may not be able to obtain satisfactory information from such sources. To address this issue, we propose a reliable and useful information distribution system called the “Kuchicomi Network.” The word “kuchicomi” in Japanese refers to information obtained by word-of-mouth [1]. There have been studies conducted on word-of-mouth communication, but they all focus on the propagation mechanisms of word-of-mouth information [2]. In this paper, we focus not on the propagation mechanisms but the reliability of word-of-mouth communication. Therefore, we prefer to use the Japanese word “kuchikomi” in place of “word-of-mouth” to represent our interpretation of the concept. A recent report published in Japan stated that 82% of the people surveyed believe more in information obtained through kuchicomi than through general advertisements [3] [4]. We have attempted to apply this characteristic of reliability found in kuchicomi information to an actual information system connected to the Internet. Further, we plan to develop the Kuchicomi Network, a kind of a Web site that will initially provide reliable, useful, and up-to-date information to tourists in Japan. The Kuchicomi Network we are developing currently has a database of kuchicomi information that consists of reliable and useful information collected by informants known as

J. Sasaki et al. / A Reliable and Useful Information Distribution System

181

Local Supporters (LS). LS live in each local city and have been authorized as official informants by a third-party certification company. In this system, LS input high-quality local information, such as information about hotels, restaurants, shops, or places of interests, based on their own experiences or other objective evaluations. Tourists then, through a mobile phone or PC, access the system and select their desired destination. They then determine which of the available LS are the most similar to themselves, and from the information provided by the LS they can get reliable and useful information easily before or during their period of travel. In this paper, we propose the concept of, and business model for, the Kuchicomi Network and introduce the system structure. We then describe the features of the system and outline the development of a prototype system.

2. The Concept of and Business Model for the "Kuchicomi Network" 2.1 Quality of Information Before we describe the concept of the system, let us clarify our view on the quality of information. We feel that the quality of information depends on both its reliability and usability for users. Over the past decade, the Internet has grown to such an extent that our society has been called the “Information Society” because we can get huge quantities of information through the Internet. Moreover, mobile systems have begun taking on more important roles as this information society continues to develop [5] [6]. We can now get just about any type of information, anywhere and at anytime. This has been referred to as a “Ubiquitous Society” [7] [8]. However, at the same time it is still difficult to get really reliable and useful information easily. The goal of this research is to develop a new mechanism which provides high quality information that users can obtain easily. As the development of such a mechanism designed for general Internet users would take a very long time to complete, we decided to begin our research with the development of a more concrete system for tourists. Generally speaking, when we wish to travel to a city that we have never visited and try to get detailed information about the city, including good restaurants and convenient hotels, we often refer to guide books written by travel agents. We may also try to get such information through the Internet, or we may ask our friends. Such sources provide us with general information about the city. In addition, if we include the use of mail magazine services provided by travel agencies, we can get also get a lot of travel information through e-mail. However, almost all of the restaurants and hotels described in such travel books have some kind of business relationship with the publishers because they pay advertising charges in order to be featured in their books. In fact, most travel information that we can obtain through the Internet is also provided by World Wide Web (WWW) servers run mainly by travel agencies. On the other hand, we also have the option of taking our friends’ advice about the city we wish to visit, but we cannot rely entirely on such information because it is limited to within the scope of their experience. Further, in the case of mail magazine services, they also consist largely of commercial advertisements for travel agencies. Thus, it is difficult to get objective information from such sources about, for example, restaurants with a good reputation and high hospitality ratings that serve delicious meals at a reasonable cost. This is because such restaurants do not need to place advertisements in travel books due to the fact that their business is already very successful. Moreover, the homepages made by such restaurants are often too small and hard to find for the average tourist to utilize as travel information.

182


As a result, we often end up travelling only on low quality information in spite of living in an advanced information society. In order to solve this problem, a mechanism to obtain high quality information is strongly needed.

2.2 The Concept of the “Kuchicomi Network” How can we get high quality information? Our proposal is that the “Kuchicomi Network” for travellers using mobile phones can solve the problem. The Kuchicomi Network is a destination-information supplying system that supplies a traveller’s mobile phone with information on a particular destination (city) through the Internet. The features of the system are as follows. (1) Information Sources Reliable and useful information about the destination city is obtained from the system. In the system, Local Supporters (LS), or local informants, are able to supply reliable, useful, and up-to-date information (kuchicomi information) because they are actually living in the destination city and have experienced excellent and reasonable services at particular hotels and restaurants, etc. In order to meet user needs more satisfactorily, the kuchicomi information is classified according to sex, age, etc., and such information is updated daily.The LS thus become a kind of “real” agent of the user, as opposed to a virtual agent such as a mobile agent [9]. LS are registered to LS database in the system after a strict examination by managers of the “Kuchicomi Network” business. LS provide “kuchicomi” information with self evaluation value on the recommendation level, such as weak recommend (+1), usual recommend (+2) or strong recommend (+3). Then a user gets the information and has an actual experience on the information. After that, the user evaluate the information and send an evaluation value such as bad (-1), normal (0), good (+1), very good (+2) or excellent (+3) to the manager. The manager totally evaluates the LS quality by both the self evaluation of LS and user evaluation. If a LS provides no information for long time or is evaluated as very low level, the LS may be removed from the LS database. As a result, we think that the reliability and usefulness of information are kept in high level. But we do not implement the evaluation function to the prototype system. (2) The Management System of the Kuchicomi Network The Kuchicomi Network is managed by managers, advisors, and operators as well as by LS. Managers are responsible for the management of the whole system and any business aspects of the network. They can control the access rights of other users as they relate to the system functions. They can also manage the information flow and add new services. Advisors have the role of checking all kuchicomi information. If LS input is found to be wrong, false, too old, or redundant, advisors can then delete or update the information and inform the changes to the relevant LS. Advisors also evaluate the quality of kuchicomi information, and when specific kuchicomi information is referred to frequently and evaluated to be good by many users, the information will be judged by advisors to be of high quality. The LS who supply high quality kuchicomi information may then get a prize or some type of award. We are now studying possible mechanisms that can be used to evaluate kuchicomi information automatically. The role of advisors is important to keep the quality of kuchicomi information and the incentives for LS at a high level. When kuchicomi information supplied by LS is not in the form of electronic media or is in the wrong format, operators will then input this information into the system or change its format so that it can be used. Further, operators also have the responsibility of adding access


183

maps or pictures to the original information if necessary. Thus, as a result of the work of the network operators, most kuchicomi information will be easy to understand and well organized for the network users. (3) System Structure Our Kuchicomi Network was developed at our university by a new development method known as XP (Extreme Programming). We were thereby able to develop the system in a very short time, and a flexible system structure could be realized. We also provide an easy management environment for the Kuchicomi Network for the managers, advisors, operators, and LS in each local area. As the framework of the system is clearly understandable and has individual functions, new services can be added easily to the system by each local manager. (4) Accounting Method In order to realize such a high-quality-information supplying system, a reasonable and efficient accounting method is required. If, for example, the current charging methods used by a particular mobile phone service provider are adopted in the Kuchicomi Network directly, the service will then be limited to that particular provider. If we wish for the service to be expandable to multiple providers, then the implementation becomes very difficult because of the different charging methods employed by each provider. To solve this problem, we investigated a new accounting method for the Kuchicomi Network. After a detailed survey of available methods, an accounting method that incorporates a prepaid card using QR code was adopted for use in the system. We developed the accounting system uniquely and included a certification function for identification (ID) with access keys and management functions of the access number and effective usage period.

2.3 Business Model Figure 1 shows the business model of the Kuchicomi Network. First, a user has to buy a prepaid card using QR code, which has a unique access key to the system. The user can then access the system through his/her mobile phone by using the access key for a specific number of times or period of time, such as 10 times or 1 month, respectively. The fee from the user’s payment will be available for distribution among the system resources, such as hardware, software maintenance, and payment for the staff, which includes managers, advisors, operators, and LS. To better understand and the implementation process of these functions, we referred to a study on e-business support functions [10].

Figure 1: Business model of the Kuchicomi Network.

184


3. System Structure As shown in Figure 2, the Kuchicomi Network system has a certification system, a database of Local Supporters (LS), and database of kuchicomi information. The certification system includes the management function of user’s access rights, and it counts access times or effective periods for each user. Profiles such as sex and age of the LS are stored in the LS database. Kuchicomi information supplied by LS is accumulated in the kuchicomi information database. Kuchicomi information is further classified into fields, such as hotels, restaurants, or sightseeing spots, etc. The system and databases each exist individually but work together relationally.

Kuchicomi Network Object

Certification

Location

User

DB of LS

LS profile The like ges

Sex

Getting

LS

Kuchicomi

Object

Field Field B Field C

Input Kuchicomi

Classified fields

DB of Kuchicomi

Figure 2: System architecture.

A user can access the destination-information supplying site by his/her mobile phone through the Kuchicomi Network after undergoing a certification process of access rights by the certification system. The user selects the desired location, interest fields, and LS who have similar profiles to the user. Then he/she will receive reliable and useful information.

4. Development of a Prototype System 4.1 Outline of a Prototype System Figure 3 shows a prototype system structure of the proposed Kuchicomi Network. There are managers and operators for system management and LS for supplying kuchicomi information. Approximately 200 LS will be recruited to supply about 800 pieces of kuchikomi information.


185

The LS will actively strive to provide local kuchicomi information in order to activate the local economy. This can be started initially in the Tohoku area of Japan. Figure 4 shows the server structure of the prototype system. The server is constructed in a Linux-based environment and adopts a new framework, which is used to construct Web applications efficiently. We developed the framework for the Kuchicomi Network by referring to XOOPS technology [11]. The certification system and the databases of LS and kuchiomi information are developed individually from the application program of the Kuchicomi Network. We have been developing the Kuchicomi Network using new development methods of UP (Unified Process) [12] and XP (eXtreme Programming) [13]. They are known as “Agile Methods” and are normally used when developing flexible and small-sized Web application software [14][15]. Recently, however, they have become popular for use in developing larger sized software packages as well [16]. A pre-prototype system of this prototype system of the Kuchicomi Network was developed from May to August of 2004 by five students at Iwate Prefectural University under the UP method. The full prototype system was then developed from November to December of 2004 by the five students at Iwate Prefectural University along with a business partner (the Ginga Tsushin Corporation) under the XP method. The XP method offers comparatively faster development times when compared to the UP method because no-document communication is possible and onsite-customer methods are available. These differences in experimental development methods along with a new documentation method for XP are reported and proposed in another paper [17]. Kuchicomi Network Managers Access Server

User Get Kuchicomi

DB Evaluation

LS

Input Operators

PC, FAX or Paper Sheet Figure 3: Structure of the prototype system.

Application Program Framework Certification System

PHP 4.1.2

My SQL 3.23.39

Apache 1.3.31

Red Hat Linux Figure 4: Server structure of the prototype system.

186


4.2 Usability In this section, we explain how to use the Kuchicomi Network from the viewpoint of a user. First a user must purchase a prepaid card in order to access the Kuchicomi Network. The card in the prototype system (explained in Figure 5) includes instructions on usage and a QR code with a unique access key. A user will then be allowed to access the network 10 times after inputting the QR code through a mobile phone.

Figure 5: Prepaid card in the prototype system

When users access the destination information through a mobile phone, they select an area, prefecture, and city where they wish to visit by pushing the corresponding keys. Figure 6 shows a case in which a user wishes to visit Morioka City of Iwate Prefecture in the Tohoku area. Next, the image shown in Figure 6 changes to the image shown in Figure 7, in which some new fields appear, such as hotels, restaurants, and shops, etc. Users can select the desired field according to their interests. After the desired field has been selected, LS who have kuchicomi information regarding the selected field in that location appear, as shown in Figure 8. Users can also confirm the sex, age, and other information (i.e., personal taste data) of each LS. Users can also select a LS whose personality matches their own. The mobile phone then displays the kuchicomi information, as shown in Figure 9. The kuchicomi information consists of both abstract and detailed information. In Figure 9, for example, the user selects “Hotel information” supplied by “Jun SASAKI” and gets detailed information about “Hotel Metro” through e-mail. The reason for using e-mail here is because access to the Kuchicomi Network a subsequent time should not be necessary in order to confirm such detailed information. The Kuchicomi Network is feasible and workable with current hardware and software techniques since packet communication services for mobile phones, such as “i-mode” by the NTT Docomo Corporation, are popular in Japan. In other words, the system can be realized by utilizing the current Japanese infrastructure.


Figure 6: Selecting an area, prefecture, and city.

-Hotels -Restaurants -Shops -Sightseeing -

Figure 7: Selecting needs.

Profile

-Y. FUNYU (M,66) -J. SASAKI(M,48) -T. YONEDA(F,30) -

J.SASAKI(M,48) -Morioka -090-878??? -Drinking -Guitar -

Figure 8: Selecting LS.

187

188


Hotels -Hotel -Hotel -Grand -

By J. SASAKI(M,48) Metro: Near Morioka St.. RICH: River side.. Hotel: On the Mt. Iwa..

Details

Hotel Metro -Good Service -Very comfortable -*** --

Get

Get the information by e-mail if needed.

Figure 9: Getting Kuchicomi.

5. Affects and Future Issues of the Kuchicomi Network We believe that a fully realized Kuchicomi Network would have the following remarkable effects. First, a user would be able to get reliable and useful information because kuchicomi information is composed of the direct opinions of experienced users, and this characteristic differentiates it from commercial advertisements. Second, as many users will be able to get more local information, including that of less well-known places, through the Internet and then actually go to visit those places, local economies will be activated. Further, the Kuchicomi Network will be available for local companies or communities to perform marketing research on their products or services based on the network’s subjective evaluations. The Kuchicomi Network is now being further developed in order to be even more useful and expandable. Some research issues that have yet to be solved are as follows. (1) How to maintain the high reliability of kuchicomi information. In its current state, human advisors check the information in the network. We believe some type of automatic information filtering tool for kuchicomi information will be necessary in the future. We expect that more studies on recommendation systems will soon become available to construct such a system [18] [19]. (2) How to maintain incentives for informants (Local Supporters) to input kuchicomi information. There is currently no function in the network for measure the value of the information. We are studying whether the giving of advantage points to LS who supply often-accessed information would be effective as an incentive. (3) Construction of a prototype system. We must first complete the development of the prototype system. A field test is currently scheduled for the Tohoku area of Japan from April 2005. (4) Evaluation method on the system. We must decide what the evaluation items on the system will be and how to evaluate them. Determining the effectiveness of system will take several months. (5) Push-type system. In our proposed system, a user must actively get the kuchicomi information. Thus it is a kind of pull-type system. A push-type system would be more convenient for some users because it would allow for the automatic distribution of kuchicomi information through mobile phones [19]. We have an idea for a new type of


189

push-style system, in which kuchicomi information would be provided automatically according to the user’s location and time of travel.

6. Conclusion In this paper, we first proposed the concept of, and business model for, the “Kuchicomi Network” and explained its system structure. However, some additional developments are still required in the design of the system in order for it to be a business success. Secondly, we described the features of a prototype system and outlined its development. In addition, some remaining research issues were discussed. We would like to complete the development of the prototype system and carry out a field test. After a field test has been completed, we wish to revise the system in order to create a successful business and expand it throughout the world.

Acknowledgements We thank Mr. Takahiro Tabata, Mr. Ryuhachiro Suzuki, and Mr. Yoshinobu Sasaki of the Ginga Tsushin Corporation along with Mr. Hiroaki Shimomukai, Mr. Toru Nagasama, Mr. Satoshi Nakada, and Mr. Hidaka Yano of Iwate Prefectural University for their work in developing the Kuchicomi Network and their useful advice related to this research. References [1] Hiroto Endo and Masato Noto, “Information Recommendation System by Work-of-Mouth Communication Model”,Technical Report of IEICE, SITE2003-3(2003-5). [2] Yusuke Ito, Hirosi Yoshida and Masayuki Numao, “A Communication Ware of Mouth”, IPSJ Report, 124-2, p.9-15(2001.5.17), [3] Kaeko HINO, “Kuchi-community Marketing 2”, Asahi Shinbun shya, 2003, (ISBN4-02-257859-9) [4] HER STORY; http://www.herstory.co.jp [5] Hideaki Yumiba and Masami Yabusaki, “Mobile Service History and Future”, IEICE Trans., Commun., Vol.E85-B, No.10, pp.1878-1886, October 2002. [6] Subhsis Saha Mark Jamtgaard and John Villasenor, “Bringing the Wireless Internet to Mobile Devices”, IEEE Computer, p.54-58, June 2001. [7] Tomoko Itao, Tetsuya Nakamura, Masato Natsuo and Tomonori Aoyama, “Context-Aware Construction of Ubiquitous Services”, IEICE Trans., Commun., Vol.E84-B, No.12, pp.3181-3197, December 2001. [8] William G. Grisword, Patricia Shanahan, Steven W. Brown, Robert Boyer, Matt Ratto, R. Benjamin Shapiro and Tan Minh Truong, “Active Campus: Experiments in Community-Oriented Ubiquitous Computing”, IEEE Computer, p.73-81, October 2004. [9] Im Y. Jung and Heon Y. Yeom, “An Efficient and Reliable Message Passing by Mobile Agent”, 15th Proceedings of IEEE ICOIN, pp.900-905(2001). [10]Hiroshi Tokunaga, “A Study on e-business Support Functions”, IEICE Trans., Vol.J85-B, No.7, pp.1078-1085, July 2002. [11] Mamoru Takai, “Construction Guide of Community Site by XOOPS”, Printed in Japan, Gijutu Hyoron-Shya, 2004,(ISBN4-7741-1925-3C3055) [12] Craig Larman, “Practical UML”, Pason Education, ISBN: 4894713861 (2003/10) [13] Kent Beck, “Extreme Programming Explained”, Addison-Wesley Pub co.,1st edition (October 5, 1999) . [14] Barry Boehm, “Get Ready for Agile Methods, with Care”, IEEE Computer, p.64-69, January 2002. [15]Victor Skowronski, “Do Agile Methods Marginalize Problem Solvers?”, IEEE Computer, p.118-120, October 2004. [16] Mikael Lindvall, Dirk Muthing, Aldo Dagnino, Christina Wallin, Michael Stupperich, David Kiefer, John May and Tuomo Kahkonen, “Agile Software Development in Large Organizations”, IEEE Computer, p.26-34, December 2004.

190


[17] Hiroaki Shimomukai, Hidaka Yano, Jun Sasaki, Tae Yoneda and Yutaka Funyu, “Agile development method for Web Apprication”, IPSJ conference in Japan, March 2005 [18] Emi Yano, Yusuke Kitano, Emi Sueyoshi, Isao Shinohara, Pinyapong Sineenard and Toshikazu Kato, “Developing of Recommendation System with Anonymous Kansei Model”, IPSJ Trans.,Vol.44, No.SIG 8 (TOD 18), pp.46-54, June 2003. [19] Takako Hashimoto, Takashi Katooka and Atsushi Iizawa, “Development of Push Type Contents Distribution System of Reflect Viewer’s Preferences”, Tehnical Report of IEICE., pp.157-162, DE 2003-58(2003-07).


191

Object-Oriented Model for Fire Spread Prediction Guillermina PERTINO, Sabrina TOCHO, Arturo ZAMBRANO, Silvia GORDILLO LIFIA – Facultad de Informática, Universidad Nacional de La Plata (1900) La Plata, Buenos Aires, Argentina [gpertino,stocho,arturo,gordillo]@sol.info.unlp.edu.ar Abstract. Fire is an extremely complex phenomenon and therefore fire spread prediction is not trivial. Spread prediction in grassland fires differs from the prediction in forest fires as the factors influencing each one are not the same. Moreover, input values for the algorithms differ depending on the factors they consider to have influence on fire behavior, even though the algorithms might be applicable to surfaces with similar characteristics. Selecting the most suitable algorithm in each particular case is not a simple task. In this paper, we describe an easily extensible object oriented model for fire spread prediction in, potentially, any surface. This is of great importance since new mathematical algorithms can be added in order to predict fire spread in new surfaces. Additionally, we present an approach to derive unavailable data, required by the algorithms, from other available data. Finally, we provide a way to manage the selection of the “most suitable algorithm” from a collection of applicable algorithms in a particular type of surface, depending on the needs of the user (accuracy, execution time, etc.).

1.

Introduction

When dealing with fire spread prediction, a critical element of the analysis is the interaction among three factors: forest fire, local climate and the characteristics of the surface. This is extremely complex [1] since weather conditions influence fuels inflammability and determine the risk of ignition of fire. “…The rate of spread of a fire is affected by a variety of issues including wind speed, moisture content of the fuel, fuel particle size, vegetation height, fuel bulk density, percentage of dead fuels and topography…” [2]. Predicting the behavior of a grassland fire differs from the prediction of a forest fire since factors affecting each one are not exactly the same. There are many common issues in both grassland and forest fires; however, there are others that may not be significant for a particular prediction. For example, species of grass may not be significant when it comes to predicting grassland fire behavior whereas species of trees might be an important factor in accurate prediction of forest fires behavior. In spite of the complexity of predicting fire behavior, there are many mathematical algorithms which take into account different factors in order to make estimates. The results obtained by these mathematical algorithms are usually tested in experimental fires in order to determine the degree of accuracy. So, different algorithms vary in their accuracy and in the factors considered to have a significant influence on fire behavior. Current GISs (Geographic Information Systems) try to benefit from the use of Object Orientation, but there is no research work in the area of Object Oriented modeling for fire prediction applications. In this paper, we present an object-oriented model for forward ROS (Rate of Spread) prediction; the advantage of our approach is the possibility

192

G. Pertino et al. / Object-Oriented Model for Fire Spread Prediction

to manipulate a variable set of algorithms in the application to study a wildfire and determine which algorithm is the best under particular conditions. The use of many algorithms, together with the possibility to be adopted according to the circumstances, gives rise to a more detailed and accurate analysis. Section 2 presents a brief description and a comparison of three different forward ROS prediction algorithms which were taken as a base to study fire spread prevention. In Section 3 we present the available ROS applications and explain the need for a ROS prediction application which can be applied to any domain, offering the possibility to select one out of many algorithms to make the prediction analysis. In Section 4 we show the suggested object-oriented model for ROS prediction in detail. In Section 5, we suggest some extensions that can be made to the model, such as estimating the fire perimeter from the forward rate of spread. Finally, the conclusions are stated in Section 6. 2.

Forward Rate of Spread Prediction Algorithms

Factors influencing the fire behavior are many, and they all have an impact on fire in different intensities. In order to develop algorithms to study fire spread, some of these factors are considered to have an insignificant influence on fire and thus, they are not taken into consideration in the prediction. Based on more than a hundred experimental fires, Cheney, Gould and Catchpole [3] determined, in the particular case of grasslands, that “…wind speed had most effect on fire spread…”. In the case of forest fires, wind speed and slope, have the most effect. As it is said in [4], “…slope steepness affects rate of spread… by preheating unburned fuels ahead of the fire. On a slope, the flames are tilted towards uphill fuels (even in the absence of wind), thus enhancing preheating through radiation by reducing the distance between the flame and unburned fuels… Therefore, fires burn more rapidly and intensely upslope than either down slope or on the level…”. The first algorithm discussed in this section is one presented by Cheney et al. which can be used to predict fire forward spread in grasslands. The following is an algorithm by McArthur, applicable to areas with the same characteristics. Finally, we present an algorithm for fire forward spread prediction in forests, also introduced by McArthur. All three algorithms are planned to predict the forward ROS of a fire which grows without the presence of any barrier, i.e. the existence of a river, road or anything that could restrict the natural spread of fire, is not taken into account by the algorithms in question. 2.1.

Fire Spread in Grasslands (Cheney et al.)

In [5], an algorithm to predict fire forward ROS in grasslands is proposed. Fire spread is estimated “…from wind speed at 10 m, dead fuel moisture and degree of grass curing in three defined pasture types…”. In their extensive study about the effect of fuel characteristics on fire spread in grasslands, grass was divided into three defined types. It was determined that “…in continuous pastures, neither fuel load nor grass species … had a significant effect on the rate of spread…” [5]. As a consequence, for a fire prediction model, the influence of fuel can be analyzed according to natural undistributed pastures, cut or grazed pastures and eaten-out or discontinuous ones. For their algorithm, Cheney et al. suggested conceiving it in the simplified form of a function. The function has the following signature:


193

RsCh = ƒ(i, U10, Mf, C) where, RsCh = ROS by Cheney et al. i = pasture type U10 = wind speed at 10 m Mf = moisture content of dead grass C = curing stage of the grass

It should be considered that this algorithm does not take into account the land topography. As a consequence, it is only applicable to plain geographic regions or with a low inclination. 2.2.

Fire Spread in Grassland Fires (McArthur)

The McArthur Grassland Fire Danger Meter Mk5 [6], estimates, among other things, the fire spread in grasslands based on grass curing, temperature, relative humidity, wind speed at 10 meters and fuel weight. Pasture dryness is estimated depending on the number of green areas. The pastures that are 100% dry are those which lack green areas. The wind speed is measured at 10 meters of height and it refers to the average speed in 10 minutes. In a similar way to the one proposed by Cheney et al., the fire spread of the Mk5 can be seen as a function, with the following signature: RMk5G = ƒ(C, T, H, U10, Pf) where, RMk5G = ROS of the Mk5 for grassland fires C = grass curing (0 – 100%) T = temperature (0 – 45°C) H = relative humidity (0 – 100%) U10 = wind speed at 10 meters (0 – 70 km/hr) Pf = fuel weight (0 – 25 tonnes/ha)

When the wind is gusty, the fire behavior is erratic, particularly with thin fuels, which respond to wind changes quickly. The spread rates are lower than the ones indicated in discontinuous pastures. It also varies depending on the topography as the Mk5G is only applicable to gentle slopes [6]. 2.3.

Fire Spread in Forest Fires (McArthur)

The McArthur Forest Fire Danger Meter Mk5 [7], is similar to the McArthur Grassland Fire Danger Meter Mk5 since it estimates, among other things, fire forward ROS. “…The meter is designed for general fire danger forecasting purposes and is based on the expected behavior of fires burning for an extended period in high eucalypt forest carrying a fuel quantity of 12.5 tones per hectare and traveling over level to undulating topography…” [7]. The ROS is predicted from temperature, relative humidity, wind speed, fuel load, drought factor and ground slope. The drought factor is a metric of the availability of fuel, influenced by recent precipitations and the time of the year.

194


Following the function scheme of the previous algorithms, the ROS prediction of forest fires Mk5 can be seen as: RsMk5F = ƒ(T, H, U10, Pf, S, G) where, RsMk5F = ROS of the Mk5 for forest fires T = temperature (0 – 45 °C) H = relative humidity (0 – 100%) U10 = wind speed (0 – 70 km/hr) Pf = fuel load (tonnes/ha) S = drought factor (0 to 10) G = ground slope (°)

The atmospheric instability is not considered to affect fire behaviour. The ROS indicated by the meter is generally for unstable conditions, which is the reason why the estimation could be lower under conditions of stability. 2.4.

Comparison Between the Algorithms

With the three algorithms described above, fire ROS can be predicted in grasslands with a low inclination of slope, and forests. Neither the RsCh nor the RsMk5G makes predictions regarding topography. Due to the unlikelihood of grasslands in surfaces with slope variations, it is not necessary to consider this case. Table 1 shows which factors influence fire in the three algorithms previously presented. Table1. Factors influencing fire behavior

Influence factor Wind speed Recent rain Fuel distribution Fuel weight Moisture content of dead grass Grass curing Temperature Relative humidity Topography

RsCh

RsMk5G

RsMk5F

According to [6], the ROS estimated by the RsMk5G in dry discontinuous pastures is higher than the real ROS. Also, fire behavior is erratic when it is windy, particularly with thin fuel which responds quickly to wind changes. The predictions obtained with the RsCh in grazed pastures at moderate wind speeds (25 – 45 km/hr), seem to be higher when compared to wildfires [5]. The RsMk5F considers climatic unstable conditions when predicting the ROS, and the estimation could be lower than the real rate in stable conditions [7]. 3.

The Need for an Application to Predict Fire ROS

Nowadays, there are many systems that estimate fire forward ROS. Most of them were made to predict fire ROS in particular areas, for example grasslands or forests, but not


195

both. This means that in order to predict the forward ROS of different kinds of fire, many predicting applications are required. In Canada, FBP (Fire Behavior Prediction) Extension is used to predict fire behavior. This software requires two static inputs, elevation and fuel type. As it is stated in [8], “…the fuel types recognized by the system are those specified by the Canadian Forest Fire Behavior Prediction System…”. Users can also input other parameters, such as wind and slope, not necessarily obligatory, to make the prediction more accurate. In other applications, such as Farsite, the number of inputs is larger. Sometimes people do not have all the required information available, which is why the prediction can only be made with the help of other software that estimates the missing information. Both, FBP Extension and Farsite, require all their essential inputs, so that, it is not possible to perform the prediction when some input is missing even when some of them can be derived from the others. It would be useful to have software capable of predicting, as accurately as possible, using the information available. Automatic input derivation from existing data would also be useful. As we explain later, this can be achieved deriving some inputs from others. This allows us to have a broader range of algorithms applicable to the same areas, regardless their inputs. Applications for predicting fire behavior are normally used in the context of governmental agencies working in the area of natural resources. Wide area countries with many topographies and ecosystems may need to deal with fire in several contexts, such as those presented here (grasslands and forests). That is why we propose an architecture that allows the integration of algorithms for, potentially, any context. 4.

Object Oriented Modeling

Object orientation focuses on modeling objects as humans perceive them in reality, by combining their structure and behavior. GISs serve as repositories of observations humans make about spatially related objects and their properties. The amount of spatial data related to a GIS is usually large. As stated in [9]: “…Object-oriented programming … is most useful for application areas like GISs because it naturally supports the treatment of complex objects… Compared with conventional data models, an object oriented design is more flexible and better suited to describe complex data structures…”. In order to design an application to predict the fire forward ROS, we propose to use object orientation because of all the benefits obtained. The object oriented paradigm allows the definition and manipulation of complex entities (concrete and abstract, discrete and continuous). Its semantics features, such as composition and inheritance facilitate the specification of entities representing pertinent factors influencing fire and the prediction analysis. Moreover, by using object orientation it is possible to model algorithms as polymorphic objects, allowing the exchange of them at runtime, making the whole system more flexible and adaptable. 4.1.

Fire Behavior and Influence Factors

Wildfire and factors that affect its behavior are continuous phenomena since they spread along a continuous space. These kinds of phenomena are usually modeled using the so-called continuous fields. Storing and using continuous data in a computer is not a trivial task, since ideally it requires measuring the phenomenon in all positions belonging to the spatial domain, which

196


are, in nature, infinite. To cope with this problem, a discretization process is realized by using a sample to measure some representative positions, while values for the remainder points are estimated through interpolation. In [10] an object oriented architecture that allows the manipulation of continuous fields was presented. According to this work, continuous fields were divided into two categories: vectorial and scalar continuous fields. We will use the UML (Unified Modeling Language) specification [11] as the graphical modeling language. The value of a scalar field in a particular position is a scalar magnitude, whereas the value on each position in a vectorial field is a vector. For example, in a vectorial field representing wind, the value in a determined position would be the magnitude and direction of the wind. A continuous field’s domain is the geographical zone in which the phenomenon is exerted; which means that the domain limits the zone in which interpolation of samples can be done. We will use these concepts to model fire prediction. Figure 1 shows a simple representation of the continuous field’s architecture.

Figure 1. Continuous field’s architecture

Following the model defined in [10], the ContinuousField class has a Representation associated. This is achieved using the Bridge structural pattern, which allows to “…decouple an abstraction from its implementation so that the two can vary independently…” [12]. As we said before, we have to organize the information in a way that best resembles reality. Nowadays, the most used data spatial models for continuous fields are: regular point grid, irregular point grid and triangulated irregular networks (TIN), among others. The ContinuousField class also has an EstimationMethod associated, represented in a separated hierarchy and implemented using the Strategy pattern, which allows to “…define a family of algorithms, encapsulate each one, and make them interchangeable. Strategy lets the algorithm vary independently from clients that use it...” [12]. The EstimationMethod represents an algorithm that allows value interpolation basing the calculation on near sample points.

197


If we consider the case of temperature, measures are going to be in a range of values, for example 0 to 45 °C, so temperature would be represented using scalar continuous fields. In Figure 2, we propose a way to represent fire and influence factors by using continuous fields. In order to simplify the diagram, the package ContinuousField represents the Continuous Fields architecture. Fire is represented as a vectorial continuous field because the value in a determined position would be the temperature and the direction in which the fire is spreading in that position. In fact, each specified fire in the application will be an instance of the VectorialCF class.

Figure 2. Fire and factors that influence fire

We have to make a distinction between what we represent as the Influence class and InfluenceType class. An instance of InfluenceType represents just the concept of a factor influencing fire. It does not represent data in a particular place, nor does it represent particular measured values. Dealing with influence types allows the addition, deletion or even the study of conceptual factors more than values. In the InfluenceType class we can see the instantiation of the Type Object structural pattern which allows to “…decouple instances from their classes so that those classes can be implemented as instances of a class. Type Object allows new "classes" to be created dynamically at runtime, lets a system provide its own type-checking rules, and can lead to simpler, smaller systems…” [13]. Instead, an instance of Influence is a set of values for a particular region that are influencing fire behavior. The intention of Figure 3 is to make this idea clear. vectorialField : VectorialCF

fire : VectorialCF

scalarField3 : ScalarCF

wind : Influence

algROS : RsCh

curingStage : Influence

wind : InfluenceType

pasture : InfluenceType

pasture : Influence

grassMoisture : Influence



curingStage : InfluenceType

grassMoisture : InfluenceType

Figure 3. Instance diagram for a prediction using the RsCh

198


As we explained before, the RsCh requires four inputs to make the prediction. Then, a given instance of RsCh knows four influences as collaborators. Each influence has its type (InfluenceType) and holds influence data. In the diagram there is an instance of the class RsCh. We explain this class in the next section and in order to understand the diagram, we just need to think of it as a class that given a fire and a set of influences, predicts the forward ROS of the fire. 4.2.

Rate-of-Spread Algorithms

In the previous section, we described the representation of fire and the factors that influence its behavior. We now need some way to model how factors influence fire, in order to estimate the forward ROS for a particular fire. There are many mathematical algorithms to predict fire ROS. We previously mentioned three, but many more exist and new algorithms can be developed. Despite the difference between predicting the ROS in grasslands and forest fires, it is necessary to model different ROS predicting algorithms. Moreover, multiple algorithms can be applied to a particular situation. This is the case of the RsCh and the RsMk5G, which use different influence factors to make the prediction. The results obtained by different algorithms also differ from each other in how accurate they are. Figure 4 shows a way to represent the ROS algorithms. We have an abstract class ROSAlgorithm which defines the common interface to the different algorithms. Each one (RsCh, RsMk5G and RsMk5F) is defined as a concrete subclass of this class and implements the interface of its abstract super class.

Figure 4. Rate of spread algorithms

The influence factors considered by each algorithm are different. This is the reason why each algorithm “knows” what types of influence it particularly needs in order to make the estimation. The ROS algorithm hierarchy shows another instantiation of the Strategy behavioral pattern. In this way, each algorithm works with its own set of influence types, independently from the others, and each algorithm may declare factors as needed or not, also in an independent way. The Mediator is the one in charge of giving to the algorithm all the information it needs to make a prediction, which is the fire and the influences. We will explain this class in detail further on.


4.3.

199

Primitive and Derived Influences

In many cases, the information related to all the influence factors for the area in which a fire started is not available or we do not have the information in the format the algorithm expects, but we still might want to predict its forward ROS. Since some influences can be derived from others, this could be possible even in these conditions. To cope with this problem, we introduce the concept of derived influence. A derived influence is calculated from one or more existing influences. There are two types of influence derivation; derivation based on one influence which could be seen as data adaptation, and derivation based on more than one influence which is far more complex than the other. For instance, we can take the algorithm before proposed, the RsCh. It could happen that for a specific territory, the information related to the pasture is not divided into the three defined types required by this algorithm, so, the estimation could not be made using it. In this case, the pasture information could be derived from the distribution and the weight of the grass. Once obtained this division, the estimation can be made using the derived influence. Note that this could produce an increment in the error of the final result. In Figure 5, we propose a way to represent different kinds of influences. For this, we modify the structure of the Influence class from Figure 2, in order to specify primitive and derived influences.

Figure 5. Different influences

The representation is obtained using the Composite structural pattern which allows to “…compose objects into tree structures to represent partwhole hierarchies. Composite lets clients treat individual objects and compositions of objects uniformly…” [12]. The Influence class declares the common interface of the objects in the composition. Primitive represents an atomic influence, because it can be defined independently of other influences. Derived represents an influence whose values are calculated based on other influences values. 4.4.

Algorithm Selection Criteria

Many algorithms are applicable to the same area, which is why it is necessary to have some way to select the best algorithm for a particular prediction from a group of

200


algorithms. The term best implies some kind of subjectivity, in some opportunities the best algorithm could be the most accurate, whereas in other cases, as when looking for the quickest estimation, the best algorithm could be that whose calculation time is smallest. The calculation time would be determined by the magnitude order of the algorithm. We could also include other selection criteria. Figure 6 shows how we can represent different selection criteria.

Figure 6. Different selection criteria

In this case, we also use the Strategy behavioral pattern. The class SelectionCriterion defines the common interface to all the selection criteria. SmallestMarginOfError and SmallestExecutionTime implement different algorithms that select the best ROS algorithm from a set using the interface of SelectionCriterion. The smallest margin of error criterion selects, from a set of algorithms, the one with a smaller margin of error in the result. The smallest execution time selects, from a set of algorithms, the one with smallest magnitude order and therefore, the one with better response time. Adding new subclasses makes it possible to deal with new desired selection criteria. 4.5.

Mediator

We now introduce a class responsible of coordinating all the classes directly involved in the forward ROS prediction. This class defines the cooperative behavior of the following classes: Influence, SelectionCriterion and ROSAlgorithm. As this class encapsulates cooperative behavior, the Mediator pattern is applied. Mediator allows to “…define an object that encapsulates how a set of objects interact. Mediator promotes loose coupling by keeping objects from referring to each other explicitly, and it lets you vary their interaction independently…” [12]. Figure 7 shows the Mediator pattern reified in our model. Given a fire who’s forward ROS is intended to be estimated and the desired selection criterion, the Mediator calculates the ROS using the most appropriate algorithm from a group of available algorithms, for that particular situation. The method calculateROS(selC: selectionCriterion, f: fire): int is the one in charge of making this calculation. The mediator’s selection of the best algorithm can be divided into three steps. First, all the unsuitable algorithms for that particular fire are discarded from the whole set of available algorithms. For example, if we are dealing with a grassland fire, the Mk5F algorithm would be unsuitable and therefore eliminated. In the following step, the algorithms which do not have information available on that domain of all the factors it needs in order to make the estimation are eliminated.


201

Figure 7. Mediator

The information of the influence factors can be either primitive or derived. In this case, all derived influences are instantiated. Algorithms that surpass these steps are applicable to the current situation and all the information required by them, is available. The last step consists of selecting the best algorithm depending on the selection criteria wanted. This would be achieved making a ranking of algorithms and keeping the first algorithm in the ranking. Once it has selected the best algorithm, the mediator can estimate the ROS. The selected algorithm will be in charge of dealing with this task. Figure 8 shows an extension of the instance diagram presented in figure 3 for a prediction using the RsCh.

Figure 8. Instance diagram extension for a prediction using the RsCh

The main difference between both diagrams is the inclusion of the Mediator in the extension. The extension diagram shows which instances would be known by the mediator, i.e. the fire, the predicting algorithm, and the influences needed, when predicting the forward ROS using the RsCh algorithm. The RsCh knows all the influence types it needs to make the estimation and each influence its influence type.

202

4.6.


Forward Rate of Spread Object Oriented Model

If we put together all the classes before mentioned, we would obtain the forward ROS prediction object-oriented model showed in Figure 9.

Figure 9. Object oriented model for fire forward ROS prediction

The addition of a new rate of spread algorithm or selection criterion could be easily achieved. In the case of a ROS algorithm, it is necessary to define a concrete subclass of the abstract ROSAlgorithm class, and redefine the ros(influences: Collection): int message. Incorporating a new selection criterion is quite similar because it is only necessary to define a concrete subclass of the abstract SelectionCriterion class, and redefine the message bestAlgorithm(algs: Collection): ROSAlgorithm. 5.

Fire Perimeter Estimation and Feedback

Up to this moment, we can only estimate the forward ROS of fire, which can be useful to determine what to do and how to react in order to extinct it as the speed in which it is growing can be predicted. However, this is quite limited as it would be extremely useful to know at what speed the whole fire is propagating, including not only the front but also the tail and the sides. It is possible to predict the ROS of the tail and sides of the fire making a projection of the forward ROS. Even though this projection facilitates the comprehension of fire behavior, it would be more useful to be able to predict the fire expansion perimeter after a determined time interval.


203

We must consider that this type of fire perimeter prediction is not as accurate as the one obtained with a fire perimeter expansion algorithm, as could be the method proposed in [14], which is based on the Huygen’s principle. Although with the projection, the estimation of the fire perimeter is made in a simplified way. The fire expansion perimeter is obtained from the wind direction and the projections of the forward ROS of the fire. The wind direction determines the propagating direction of the projections. Figure 10 shows the fire, altogether with the direction and speed of propagation of the different points of the fire.

Figure 10. Fire perimeter estimation

When a wildfire starts, the geographical region were the fire evolves, changes. This is because fire influences its surrounding environment. For example, fire influences fuel curing and weight. Imagine a continuous grassland. When fire starts and after a period of time, the fuels curing and weight in the zone change depending on the direction and rate of spread of the fire. Once the fire is ignited, fuel on the grassland can no longer be considered as continuous grasslands because fire has changed its curing and weight. For a later rate of spread prediction in that same grassland, these changes should be taken into account otherwise the prediction would not be accurate. Not all the factors that influence fire behaviour are affected by fire, as it is the case of topography. In the model proposed this feedback of the factors that influence fire behaviour is not contemplated but it could be easily achieved knowing how fire influences these factors and making the necessary modifications to them after a fire. 6.

Conclusions and Future Work

Fire spread prediction is extremely complex and existing applications are oriented to one type or a reduced group of surfaces, for example forests with a gentle slope. Therefore they can be hardly used in other contexts, thus reducing system's applicability. The object-oriented model we proposed in this paper has the following properties. First, it is possible to specify and manipulate multiple algorithms and selection criteria as first class objects; this allows users to choose which criterion best suits their needs. Second, since algorithms and criteria are specified as objects it is possible to add new ones, allowing both, to predict fire behavior in surfaces not previously foreseen and to establish new criteria to decide which algorithm is the best one. Third, alternative calculus based on derived influences is allowed; this is useful when some kind of influence is not present for a given algorithm, thus giving robustness to the system. Fourth, the object model was built from widely used designs patterns, which results in model flexibility and extensibility. Finally, as a consequence of the nature of object-orientation, we obtain a model that is closer to the real domain, and easier to understand, even by no-computer scientists. In order to complete this project, and take full profit from it, an implementation is needed. We believe that all the needed information is available to accomplish this.

204

7.


Acknowledgements

We would like to thank Pertino Cecila, Balzamo Martín, Carzalo Hernán and Simoncini Germán who read early versions of this paper and provided valuable feedback on its content. 8.

References

[1] Bradley, M., Schomer, C. L., Sumikawa, D. A., Walker H. and Younker, L. W. “The National Wildfire Prediction Program: A Key Piece of the Wildfire Solution”. The Joint Fire Science Conference and Workshop. 1999. [2] McCormick, B. “Bushfires: Is Fuel Reduction Burning the Answer?”. Department of the Parliamentary Library, 2002. [3] Cheney, N. P., Gould, J. S. and Catchpole, W. R. “The Influence of Fuel, Weather and Fire Shape Variables on Fire Spread in Grasslands”. International Journal of Wildland Fire 3(1): 31-44, 1993. [4] Pearce, G. “Applications of rate of spread information to fire suppression planning”. Fire Technology Transfer Note, Number 4, October 1994. [5] Cheney, N. P., Gould, J. S. and Catchpole, W. R. “Prediction of Fire Spread in Grasslands”. International Journal of Wildland Fire 8(1): 1-13, 1998. [6] McArthur Grassland Mk5 Meter. www.esb.act.gov.au/firebreak/grassmk5.html [7] McArthur Forest Mk5 Meter. www.esb.act.gov.au/firebreak/forest-5.html [8] Englefield P., Lee B. and Suddaby R. Spatial Fire Management System. http://gis.esri.com/library/userconf/proc00/professional/papers/PAP489/p489.htm [9] Egenhofer M. J. and Frank A. U. “Object-Oriented Modeling for GIS”. URISA Journal 4 (2): 3-19, 1992. [10] Zambrano A. F. “Manipulacion de Campos Continuos y su aplicacion a los Sistemas de Informacion Geografica”. 2001. [11] Object Management Group. “OMG Unified Modeling Language Specification – Version 1.5”. March 2003. [12] Gamma, E., Helm, R., Johnson, R. and Vlissides, J. Design Patterns: Elements of Reusable ObjectOriented Software. Addison-Wesley, 1995. [13] Bobby, W., Johnson, R. Type Object in Pattern Languages of Program Design 3. Addison-Wesley Software Pattern Series. 1997. pp. 47-65. [14] Knight, I. and Coleman J. “A Fire Perimeter Expansion Algorithm Based on Huygens´ Wavelet Propagation”. International Journal of Wildland Fire 3(2): 73-84, 1993.


205

Knowledge Entry Maps: structuring of method knowledge in the IT industry Remko Helms, Sjaak Brinkkemper Institute of Information and Computing Sciences University Utrecht, The Netherlands Job van Oosterum, Frank de Nijs Centric IJsselstein, The Netherlands Abstract. For professionals in the IT industry it is important to have easy access to knowledge about the day-to-day business processes. Such knowledge, which is called method knowledge, can be made easy accessible through a knowledge repository system. One factor that determines the accessibility of method knowledge in such a repository is the way in which this knowledge is structured and made accessible. This paper presents a meta-modeling technique for modeling knowledge structures. The technique is called Knowledge Entry Map and supports the design of knowledge repository systems based on topic maps. To validate the technique it has been applied in two case studies in the professional IT industry, a software developer and a service provider. The case studies demonstrate the applicability of the meta-modeling technique for capturing the structure of method knowledge in IT organizations. Moreover, the case studies provided the idea of a generic method knowledge structure for project-oriented and product-oriented IT organizations.

1. Knowledge in the professional IT industry Knowledge intensive companies in the IT industry such as software developers and service providers depend on their employees for the proper execution of their business processes. To this end, companies use methods to standardize their processes. A method is an integrated collection of procedures, techniques, product descriptions, and tools for effective, efficient and consistent support of business processes [1], [2]. Knowledge about a company’s methods, i.e. method knowledge, is seen as a valuable asset that needs to be shared among employees [3]. Knowledge repository systems can be used for this purpose [4], [5], [6]. They represent a category of knowledge systems that focus on sharing explicit knowledge between members of an organization [6], [7]. Types of knowledge repository systems that are widely discussed in literature are Lessons-Learned Systems and Expertise Locators. Typically, such a system consists of a repository in which all the knowledge resources are stored, e.g. electronic documents and video files, and a retrieval mechanism to search for the knowledge in the repository. The companies involved in this research both used a knowledge repository system to share their method knowledge. Retrieving knowledge from these systems was based on a keyword and/or taxonomy mechanism. However, employees in both companies experienced problems in finding the right knowledge resource using these retrieval mechanisms. This was partly caused by the fact that the amount of knowledge in the repository gradually expanded over time. Hence, searches resulted in dozens of hits. To

206

R. Helms et al. / Knowledge Entry Maps: Structuring of Method Knowledge in the IT Industry

solve this problem it was decided to use another way of retrieval that is based on a different way of structuring knowledge: topic maps [8], [9], [10], which emerged from the field of knowledge engineering However, the design of a topic map is not arbitrary because it can easily contain hundreds of topics. This makes the design of a topic map cumbersome and might even negatively influence the quality of the topic map. Therefore, a meta-modeling technique is proposed in this paper: Knowledge Entry Map, which captures the structure of a topic map on a meta-level [19]. The structure of this paper is as follows. In section 2 the research method is explained, followed by the description of the Knowledge Entry Map technique in section 3. The metamodeling technique has been evaluated in two cases and the results are described in section 4. In section 5 the results of the case studies are discussed. Section 6 describes related work and in section 7 the conclusions are presented. 2. Research approach The main goal of the research is to develop a technique for meta-modeling of knowledge structures that is based on topic maps. For this purpose we use the Design Research approach, a widely accepted method in the field of Information Systems research. The philosophy behind Design Research is that new scientific knowledge can be generated by means of constructing an artifact [15], [16], [17]. In this case the artifact is the metamodeling technique called Knowledge Entry Map. The steps in developing the technique according a Design Research approach are described below. Problem awareness: Awareness was raised by participating in the development of a knowledge structure for a software developer. When the idea of knowledge maps was adopted at the company, we encountered the problems as discussed in section 1. Suggestion and Development: Literature research on data and knowledge modeling learned that especially meta-models can be helpful in this respect (section 3 and section 6). The advantage of such a meta-model is that it provides an overview of the structure instead of all the details. Therefore, a meta-modeling technique is developed, which is called Knowledge Entry Map (section 3). Using this technique should result in a better design of topic maps and hence in a better accessibility of the method knowledge. Evaluation: The Knowledge Entry Map (KEM) modeling technique has been applied in two case studies in two different IT organizations (section 4). By applying the technique the applicability of the technique in capturing the structure of method knowledge is validated. The applicability of the model was observed while participating in the process of creating the KEM. Besides observation, the people that were involved have been asked about their experiences in several interviews. Validation of the technique focused on the following three aspects: KEM should be able to model the complexity at the case study companies, KEM should be easy to learn and KEM should be easy to use (section 5). Conclusion: Evaluating the results of the case studies resulted in suggestions for further improvement of KEM. Moreover, the KEM’s of both case study companies have been compared to analyze the possible existence of a generic structure for structuring method knowledge in IT organizations (section 5). The conclusions and directions for further research are formulated in section 7.


207

3. Knowledge Entry Map: structuring method knowledge 3.1. Structuring knowledge using topic maps Before presenting the Knowledge Entry Map it is required to introduce topics maps a little bit further using [8], [9]. The basic elements of a topic map are: topics, occurrences, and associations. Topics can be any “thing” in the real world about which knowledge needs to be stored. In other words, it is the subject that is being referred to and that is of interest in a particular knowledge domain. Examples of topics in the method knowledge domain involve the products that a company delivers, the processes that are required to create the different products and deliverables that are the result of particular processes. These topics refer to things in the real world about which a company want to store and share knowledge. This knowledge is modeled as occurrences in a topic map and represent the knowledge resources that contain the knowledge. Examples of occurrences for a particular topic, e.g. a product, might involve a brochure in Word format (.doc), a user manual in Html format (.html) or an instruction video in Avi format (.avi). Finally, associations are used in a topic map to link topics to each other. These associations indicate that the topics are related to each other to some extent. The result of associating topics to each other is a network of related topics that is referred to as the knowledge structure. Besides linking topics, associations can also be used to link occurrences to one ore more topics. As such the knowledge resources are associated with the topics about which they contain knowledge. Topic D

Topic B knowledge structure lev el

Topic A Topic C

knowledge resource lev el

Occurence X

Occurence Y

Occurence Z

Fig. 1. Topics maps: Topics, occurrences, and associations

In figure 1 this is also shown schematically. It distinguishes between a knowledge structure level and a knowledge resource level. This separation in levels is used to stress that the knowledge structure is separated from the knowledge resources. This has the advantage that relations between topics can be changed/updated independently of the knowledge resources. Moreover, new occurrences can be easily added after which they are associated to one or more topics. By implementing topic maps in a knowledge repository system, knowledge is structured by creating a network of related topics that is optimized for navigation [11], [12]. Users of the knowledge repository system can locate method knowledge by browsing the topics and looking up related occurrences that contain a fragment of the total method knowledge of the company. As such, it provides a powerful aid in exploring knowledge resources in a knowledge repository that can be used in addition to keyword or taxonomy search.

208


3.2. Knowledge Entry Map To support the creation of a meta-level topic map we developed the Knowledge Entry Map. The topic map itself is an instantiation of such a meta-level topic map. The modeling technique used for KEM is based on UML class diagrams [13], which also has been for knowledge modeling as in [14]. However, not all the constructs of class diagrams are used. One of the differences with object-oriented approaches is that operations and methods are not included in the concept definitions. Conversely, it was necessary to introduce some new constructs that are required in the context of modeling knowledge structures. Product 0..1

= entry point

is created by 0..1 = zero or one 1..*

Process

1

has a 1

0..*

0..1

Deliv erable 0..1

0..* = zero or more n

is supported by

is created with

= one only

1..* = one or more

n..m

= explicitly def ined number = explicitly def ined range

0..* 0..*

Tool

Fig. 2. Example of knowledge entry map

The UML class diagram is used for modeling topic type definitions instead of the actual topics. Also the topic map standard distinguishes between topics and topic types, but it does not provide a schematic definition of the relation between the topic types. Figure 2 shows an example of a Knowledge Entry Map. The basic constructs in this Knowledge Entry Map are as follows: 1. Concept: A concept represents topic types in the method knowledge domain, which share similar properties. The notion of concept is similar to what is called ‘class’ or object class’ in other approaches. Properties of a topic type are defined as attributes of a concept. In the method knowledge domain a concept represents an element of a method about which a company wants to store knowledge. Concepts are typically described using a noun, examples are Product, Process, and Deliverable, and are represented using a box. 2. Association: An association between concepts tells something about the relationship between concepts. Associations are modeled using the binary relation construct that is known from UML class diagrams and E-R diagrams. Hence, relations are defined through a specification of arguments and for each the cardinality can be defined. Similar to concepts also associations can have attributes, which reflect the properties of a relation. Associations are typically described using a verb, examples are and , and are represented as lines between different concepts. Each association has two directions, for example, a Product is associated to a Process and vice versa. In the Knowledge Entry Map only one direction is mentioned and should be read from left-to-right or top-down. Therefore, the association between Product and Process is read as follows Product Process. When reading the association the other way around it reads as follows: Process Product. Cardinality defines restrictions on the relationships. Restrictions are needed in the method knowledge domain to indicate that some topics cannot be created without also creating another topic of another type. For example, it might not make sense to


209

introduce a new product in a company’s method without adding the processes and deliverables that are required to produce this product. But of course these restrictions depend on the requirements of an individual company. An example of the application of a restriction is: Product Process. The cardinality expression is included in the association after the ‘:’. In this case the restriction indicates that any number of tools can be associated to a single process. Formally an association is defined as follows (based on [Schreiber et al.]): BINARY RELATION is-created-by; INVERSE creates; ARGUMENT-1: product; CARDINALITY: 1..*; ARGUMENT-2: process; CARDINALITY: 0..1; ATTRIBUTES: effectivity: DATE; END BINARY RELATION is-created-by;

3. Entry point: So far, the KEM is similar to a traditional data model. However, entry points are an extension to data modeling. They represent the most common starting points for exploring a topic map. In taxonomy, for example, it is clear where to start a search: at the top. However, in the case of a topic map a search can start with any of the topics in the network. Hence, it is difficult to decide, especially for inexperienced users, where to start a search. Therefore, entry points are defined that indicate the most common starting point for a search. Only concepts in a Knowledge Entry Map can be defined as entry points. Entry points can be defined as an attribute to each concept which can have the value True or False. If the value of this attribute is true the concept serves as an entry point. Schematically an entry point is defined using a black triangle in the upper right corner of the concept as shown in figure 2. There is no limitation to the number of entry points. Although, from a user point of view it is desirable to keep the number of entry points limited. The following section presents the application of the schematic definition of the Knowledge Entry Map in two case studies. 4. Using KEM to structure method knowledge – case studies 4.1. Background of case studies In order to validate the Knowledge Entry Map technique, two case studies are conducted in two different types of companies in the professional IT industry. The first company is Baan, which is a major developer of ERP software. Around 2000 Baan was an independent company with about 5,000 employees, 8,000 customers, and 15,000 operational sites. After some downsizing it has recently been acquired by SSA Global. The Baan Research and Development unit develops the Baan products with offices in the Netherlands, India, and Germany. This department employs about 600 software engineers. The second company is Centric, which is a provider of IT services. Its headquarters are located in the Netherlands and there are offices in Belgium and Germany as well. In total Centric has 2,750 employees, but this case study was conducted in the Managed ICT Services (MIS) division that has about 1,000 employees. The MIS division is specialized in office automation products and services, such as consultancy, helpdesk services, migration services, ICT management and Outsourcing.

210


4.2. KEM of Baan The Knowledge Entry Map that is constructed by the team at Baan is presented in figure 3. Based on the explanation that is given in section 3, the KEM should not need detailed explanation. However, to get a better understanding the main concepts in the Knowledge Entry Map are explained in some more detail. For a more detailed description is referred to [21]. Work Product Ov erv iew

0..*

0..*

Project model

0..* 0..*

Process Ov erv iew

Roles 0..*

0..*

1..*

1

2..*

State

1 2..* 1

2..*

Work Product

2..* 1

1..*

1

0..* 1

0..1

State Flow

0..* 0..*

1

Template

QA checklist

1

0..* Tutorial

0..* 0..*

0..*

Work instruction

0..*

0..*

0..*

Example

Standard

Fig. 3. Knowledge Entry Map of Baan

The main concepts in the KEM are: Project Model: A Project Model involves all deliverables that are needed to complete a project milestone. Project models are available for such things as Release delivery, Software development, Knowledge transfer, Integration test, and System test. Process Overview: The Process Overview involves an overview of the main phases or steps in the software development process. Examples of phases are Feasibility study and Design and System Test. Work Product: Work Products are standards for documentation of deliverables. Examples of deliverables are project plan, version definition, definition study, software unit, and test design. Work Instructions: Work instructions are detailed procedures for the completion of well-defined smaller tasks, such as performing a software unit test, or a project audit, or risk management. Roles: A role is a set of responsibilities that is defined within the organization. Examples of roles are Software Engineer, Test Coordinator, and Project Leader Tutorials: Tutorials are training materials to learn how to execute certain tasks, e.g. a tutorial on how to conduct a risk assessment.


211

At Baan they build a knowledge repository, called DMethod, for managing their method knowledge that is based on the KEM. It is built using html and the homepage shows the 7 entry points as indicated in figure 3. Each of the entry points is implemented as a drop-down list and shows for instance all the instances of Work Product in the system (figure 4). When selecting an instance of Work Product, a description of that particular Work Product is shown including the associations to other instances.

Fig 4. Homepage of DMethod

In total approximately 140 instances have been defined of the concepts that are shown in the KEM. Each of these instances has its own web page in the system, which contains references to one or more knowledge resources that contain information concerning the instance. In total, the knowledge resources resemble approximately 5,000 pages of A4 at Baan Development. Moreover, each of the instances is associated to one or more other instances, which results in a network of associated web pages.

212


4.3. KEM of Centric The Knowledge Entry Map that is constructed by the team at Centric is presented in figure 5. Once again, only the main concepts are explained to get a better understanding. Product

1

is created by

1

is div ided in 0..*

Serv ice (sub-product)

1..* 1

is created by

1..*

Process 1 is composed of 1..* 1

Phase

has a

1..*

Deliv erable 0..*

1 is composed of 1..*

1 is created by 0..* 1

Activ ity 1

is 1..* supported by 0..*

1

Tool 1

1..*

1

Status

1..*

ref ers to

is conducted by

has a

Response

0..*

Ev aluation

is supported by

Role

1

Fig. 5. Knowledge Entry Map of Centric

The main concepts in the Knowledge Entry Map at Centric are: Product: These are the services that are sold to the customer, examples of services are ICT management and Helpdesk services. Centric has defined standard services that are altered for specific customers. Process: A process is a defined set of activities to produce a specified output for a particular customer or market. An example of a process is the Execution process that is part of the ICT management product. Phase: Phases are the defined stages or steps in a process. For example, the phases in the Execution process are Implementation preparation, Implementation and Exploitation,. Deliverable: A deliverable is the standardized end product(s) of a phase, examples are project plan, proposal, and service level agreement. Tool: A tool is an aid that is used to create a deliverable. Examples are a ROI spreadsheet, a proposal template, and a project plan guideline. Role: A role is a set of responsibilities that is defined within the organization. Examples of roles are account manager, product manager, and project leader. At Centric, the ‘old’ knowledge repository system, based on Lotus Notes, is being replaced by a new system. Instead of building a system based on html, like Baan, Centric is looking for commercial-off-the-shelf (COTS) software product that can support their requirements. In total, the new system will contain approximately 1,000 instances. Each of these instances refers to one or more knowledge resources that contain information on the associated instance. In practice, a knowledge resource is a document consisting of several pages A4. In other words, the system contains several thousands of documents, which resembles a multiple of this number in A4 pages. Initially, Microsoft Sharepoint Portal server was selected as knowledge repository system. However, during the prototype stage it seemed not feasible to support topic maps with this system. At the moment they are evaluating if there is another COTS software product that is capable of supporting topic maps as specified in section 3.


213

5. Discussion 5.1. Applicability of Knowledge Entry Map The first goal of the case studies is to study the applicability of the Knowledge Entry Map for modeling knowledge structures. Its applicability is evaluated on three aspects: - ability to model the complexity of knowledge structures - easy to learn - easy to use The first aspect of evaluation is the ability of KEM to model the complexity of the knowledge structures in both companies. In section 4 it has been demonstrated that both companies succeeded in applying KEM to model their knowledge structure on a meta-level. At Centric the basic constructs of KEM were sufficient. However, at Baan some additions were required that will be discussed in more detail. Variants of concepts: There can be different variants of the same concept, which is illustrated using the following example. The different product lines of Baan R&D used different development platforms, which gave necessity to accommodate variants of Templates, and therefore of Work Products and Instructions. In the design phase conventional 4GL development used Entity-Relationship modeling, whereas objectoriented approaches needed Class diagrams. This resulted in a Work Product for a Functional Design in two variants, i.e., Entity-Relationship and Class diagrams. The variants issue can be solved using the Scope construct, which is part of the topic maps standard [8]. A scope is basically a view on a topic map that only shows the topics (and its associations) that belong to a particular scope. In KEM scope should be defined as an attribute of a concept. In the example there are two possible values for the scope attribute of a Template: 4GL and OO. A concept is not necessarily part of only scope. If no scope is defined it is automatically part of all scopes. In the Knowledge Entry Map variants are indicated by underlining the concept name, as shown in figure 6. Work Product

1..*

1

0..1 Template

0..* Example

Fig. 6. Modeling variants of a concept in KEM

Conditional associations: There can be certain situations where the existence of an association depends on the existence or non-existence of another association, which is illustrated using the following example. At Baan we encountered the situation that an Example only needed to be associated with a Work Product if there was no Template associated with a Work Product and vice versa. In other words, the existence of an association between the concepts Example and Work Product was dependent on the nonexistence of an association between concepts Template and Work Product. To solve this issue an IF THEN ELSE construct is applied to the cardinality of the arguments in the binary relation construct. In this case, the conditional association affects the cardinality of

214


the Work Product argument in both binary relations. Formally, the example should be defined as follows: BINARY RELATION is-based-on; INVERSE for; ARGUMENT-1: workproduct; CARDINALITY: IF is-explained-by=exists THEN 0..1 ELSE 1; ARGUMENT-2: template; CARDINALITY: 1..*; ATTRIBUTES: ; END BINARY RELATION is-based-on; BINARY RELATION is-explained-by; INVERSE explains; ARGUMENT-1: workproduct; CARDINALITY: IF is-based-on=exists THEN 0..* ELSE 1..*; ARGUMENT-2: example; CARDINALITY: 1; ATTRIBUTES: ; END BINARY RELATION is-explained-by;

In KEM the logical condition is schematically indicated by connecting the involved associations with a circle. The text in this circle is used to indicate the type of logical condition that applies to the associations (see figure 7). Work Product 1..*

1 OR

0..1

Template 1

0..* 0..*

0..*

Example

Standard

Fig. 7. Modeling of conditional associations in KEM

The second aspect of evaluation is that KEM should be easy to learn. It was observed by the researchers and indicated by the people involved that it is rather easy to learn the basics of KEM. A possible explanation is that people in the IT industry are familiar with data modeling and therefore find it rather easy to master KEM. On the other hand, this knowledge of data modeling is actually a disadvantage. After all, KEM is not a data model, but a meta-level model of a knowledge structure. So after the initial (easy) introduction of KEM we had to discuss this difference in more detail with the team members so that they fully understand KEM. Also people that were not heavily involved in the project indicated that the KEM models are easy to understand. The third aspect of evaluation is that KEM should be easy to use. The people involved in creating the Knowledge Entry Map indicated that the technique was easy to use. The Knowledge Entry Map technique does not require sophisticated diagramming tools and


215

therefore the diagrams are easy to share with others. Moreover, the ease of use was also clear from the quality of the Knowledge Entry Map they produced, which complied with the syntax of KEM. But it was also indicated that it can be hard to identify concepts in the organization. The modeling technique itself does not prescribe how to elicit this information from the organization, except that a concept should be described using a noun. In both case studies it was decided to gather information on concepts using known techniques for business analysis and knowledge elicitation, e.g., interviews, value chain models, and process models. Although these are not primarily focused on identifying concepts it provided satisfactory input for identifying them. 5.2. Generic structures of method knowledge The second goal of the case studies is to compare the Knowledge Entry Maps of Baan and Centric to analyze if there is such a thing as a generic structure for method knowledge in the professional IT industry. Comparing both KEM’s shows that there are similarities because they contain to a large extent similar concepts and associations. For example, both contain concepts such as Process: Process versus Process Overview, and Deliverable: Deliverable versus Work Product. These similarities can be caused by the fact that the case studies were conducted sequentially. However, the influence of the outcome of the first case study on the second case study is low because the business requirements of each company have been the source for constructing the KEM. Despite the similarities that might support the possible existence of a generic structure, there are also some major differences. First of all, at Baan they use a generic description for all their projects. Consequently, all projects have the same phases, milestones and work products. Therefore, a key concept in their KEM is Project Model. Centric on the other hand uses a different approach for different products. Consequently, the phases in a project depend on the type of product. Therefore, a key concept in their KEM is Product. Secondly, studying the associations in both KEM’s reveals that for Centric most of the associations are mandatory (i.e. from the type 1 or 1..*). While for Baan most of the associations are optional (i.e. from the type 0..1 or 0..*). This results in a more rigid structure for Centric and a more flexible structure at Baan. Hence, at Baan it is easier to introduce a new instance of a concept because it is not necessary to check for mandatory associations to other instances. The fundamental difference between a project-oriented KEM at Baan and a productoriented KEM at Centric does not support the idea that there is one generic structure that applies to all organizations in the IT industry. On the other hand, it shows that the knowledge structure of a company is influenced by the way in which the company organizes its business, i.e. project-oriented vs. product-oriented. But it goes too far to say that we have found two generic structures because it is supported by only two case studies. Further research would be required to validate if we have found generic structures for project-oriented and product-oriented companies. 6. Related work Topic maps are a relatively new field in knowledge engineering. Examples of the application of topic maps in literature are [11], [23]. In both studies topic maps are applied to the clinical domain. In one study the application is limited to modeling the topic map according the ISO standard on topic maps. While in the other study a topic map is implemented using the Ontopia Knowledge Suite (OKS). But there are no studies on how to develop such a topic map.

216


There are several other fields in knowledge engineering that are closely related to topic maps. The related fields involve concept diagrams, associative networks, semantic networks, and ontology [6], [24], [25]. All focus on modeling knowledge domains by modeling concepts as well as associations between these concepts. The most formal one is ontology, which is defined as “a formal, explicit specification of a shared conceptualization” [18], [20], [22]. These techniques have much in common with our KEM, because they focus on modeling the main concepts in a knowledge domain. In that respect the KEM could be called a concept map or ontology, which identifies the main topic types and their associations. However, these techniques do not cover entry points and conditional associations as we discussed in this paper. Like topic maps, ontologies are also used as navigation aid for browsing information [26], [27], [28]. In the studies, ontologies are used to enable ontology based browsing for finding knowledge on the Internet. It enables the user to browse through the ontology and to find knowledge related to a specific concept. This option is offered next to full text search and attribute search, giving the user several options to find the knowledge that he needs. Another study shows how an ontology browser can be used on top of a document management system [29]. This is relevant for our research because much of the knowledge in the method domain is stored in documents. There is a fundamental difference, however, between using ontology for browsing and using a topic map. In the case of ontology browsing is limited to browsing concepts instead of browsing instances of concepts as is the case for topic maps. Hence, topic maps result in a more detailed level of structuring and consequently browsing. 7. Conclusion In this paper we introduced Knowledge Entry Maps as a meta-modeling technique for knowledge structures in the method domain of IT organizations. The technique supports the design process of a knowledge repository system that uses topic maps for structuring the knowledge resources in the system. Two case studies have been conduct to validate the applicability of KEM on three different aspects: KEM should be able to model the complexity at the case study companies, KEM should be easy to learn and KEM should be easy to use. The case studies showed that KEM is a useful aid in designing knowledge structures, which is easy to learn and use. However, to match the complexity confronted in the cases it was required to make minor changes to the existing constructs of KEM. Considering these minor changes we are convinced that KEM can also successfully applied at other companies, especially software developers and service providers. But the case studies also revealed that further research is needed to develop a technique for identifying concepts in the organization, because both companies experienced difficulties in doing this. Furthermore, the KEM’s of the two case study companies have been compared. The comparison did not provide evidence that there is a generic structure for method knowledge in professional IT organizations. Further research is needed to validate if we have found a generic structure for project-oriented organizations and a generic structure for productoriented organizations. Another direction for future research is the development of a knowledge repository system that uses topic maps to structure knowledge resources. Our experience in the case studies points in the direction of document management systems, because most of the method knowledge is stored in Office documents (i.e. Pdf, Word, Excel, and Powerpoint). However, such systems should support the idea of topic maps.


217

References [1] Kumar, K., Welke, R.J., Methodology Engineering: A Proposal for Situation-Specific Methodology Construction. In: W.W. Cotterman, J.A. Senn (Eds.), Challenges and Strategies for Research in Systems Development. Wiley, 1992. [2] Brinkkemper, S., Lyytinen, K., Welke, R.J. (Eds.), Method Engineering: Prin-ciples of Method Construction and Tool Support. Chapman and Hall, 1996. [3] Slooten, C. van, Brinkkemper, S., A Method Engineering Approach to Information Systems Development. In: Information Systems Development Process. Elsevier Science Publishers (A-30), pp. 167-186, September 1993. [4] Davenport, T., Prusak, L., Working Knowledge – How organizations manage what they now. Harvard Business School Press, 1998. [5] Dixon, N.M., Common Knowledge: How companies thrive by sharing what they know, Harvard Business School Press , 2000. [6] Becerra-Fernandez, I., Gonzalez, A., Sabherwal, R., Knowledge Management – Challenges, Solutions and Technologies. Pearson/Prentice Hall, 2004. [7] Weber, R.O., Aha, D.W., Intelligent delivery of military lessons learned. Decision Support Systems 34, pp. 287-304, 2002. [8] International Organization for Standardization, ISO/IEC 13250, Information Technology, SGML Applications, Topic Maps, Geneva, 1998, http://www.y12.doe.gov/sgml/sc34/document/0058.htm, 30-112004. [9] Pepper, S., The TAO of Topic Maps, presented at XML Europe, Paris, France, 2000, http://www.gca.org/papers/xmleurope2000/pdf/s11-01.pdf, 30-11-2004. [10] Park, J., Hunting, S., XML Topic Maps: creating and using topic maps for the web, Addison Wesley Professional, 2002. [11] Strauss, M., Lemke, H.U., Knowledge structure and templates for information elements in validation procedures for CARS, International Congress Series, Vol. 1256, pp. 577-583, 2003. [12] Dong, Y., Li, M., HyO-XTM: a set of hyper graph operations on XML Topic Map toward knowledge management, Future Generation Computer Systems, Vol. 20, pp. 81-100, 2004. [13] Schreiber, G., Akkermans, H., Anjewierden, A., Hoog, R. de, Shadbolt, N., Velde, W. van de, Wielinga, B., Knowledge engineering and management – The CommonKads methodology, MIT Press, 2000. [14] Rumbaugh, J., Jacobson, I., Booch, G., Unified Modeling Language Reference Manual, Addison-Wesley Professional, 2nd edition, 2004. [15] Guba, E. and Lincoln, Y., Competing Paradigms in Qualitative Research, in The Handbook of Qualitative Research by N. Denzin and Y. Lincoln, Thousand Oaks, CA, Sage: 105-117, 1994. [16] March, S. and Smith, G., Design and Natural Science Research on Information Technology. Decision Support Systems 15, pp. 251 – 266, 1995. [17] IS World, Design Research Methodology. http://www.isworld.org/Researchdesign/ drisISworld.htm, July 8, 2004. [18] Gruber, T.R., A translation approach to portable ontologies. Knowledge Acquisition 5, pp. 199-220, 1993. [19] Brinkkemper, S., Saeki, M., Harmsen, F., Meta-Modelling Based Assembly Techniques for Situational Method Engineering, Information Systems, Vol. 24, No. 3, pp. 209-228, 1999. [20] Uschold, M., Gruninger, M., Ontologies: principles, methods, and application, Knowledge Engineering Review, Vol. 11, Issue 2, pp. 93-155, 1996. [21] Brinkkemper, S., Method Engineering with Web-enabled Methods. In: Information Systems Engineering: State of the Art and Research Themes, Brinkkemper, S., Lindencrona, E., Sølvberg, A., (Eds.) pp. 123-133. Springer Verlag, London, June 2000. [22] Benjamins, V. R., Fensel, D., Gomez Perez, A., Knowledge Management Through Ontologies. In: Proceedings of the Second International Conference on Practical Aspects of Knowledge Management (PAKM'98), Basel, Switzerland, 1998. [23] Schwieger, R. Hoelzer, S., Rudolf, D., Riger, J., Dudeck, J., Linking clinical data using XML topic maps, Artificial Intelligence in Medicine, Vol. 28. pp. 105-115, 2003. [24] Gordon, J.L., Creating knowledge maps by exploiting dependent relationships, Knowledge-Based Systems, Vol. 13, pp.71-79, 2000. [25] Curtis, M.B., Davis, M.A., Assessing knowledge structure in accounting education: an application of Pathfinder Associative Networks, Journal of Accounting Education, Vol. 21, pp. 185-195, 2003. [26] Maedche, M., Web Information Tracking Using Ontologies. In: Proceedings of the 4th International Conference on Practical Aspects of Knowledge Management. Lecture Notes in Computer Science 2569, pp. 201-212, Springer Verlag, 2002.

218


[27] Spyns, P., Oberle, D., Volz, R., et al. OntoWeb - A Semantic Web Community Portal. In: Proceedings of the 4th International Conference on Practical Aspects of Knowledge Management. Lecture Notes in Computer Science 2569, pp. 189-200, Springer Verlag, 2002. [28] Edgington, T., Choi, B., Henson, K., Raghu, T.S., Vinze, A., Adopting ontology to facilitate knowledge sharing, Communications of the ACM, Vol. 47, No. 11, 2004. [29] Kim, H.-L., Kim, H.-G., Park, K.-M., Ontalk: Ontology-Based Personal Document Management System, In: Proceedings of 13th International WWW Conference, New York, May 17-22, 2004.


219

Building a Conceptual Skeleton for Enterprise Architecture Specifications a

Veikko HALTTUNENa, Antti LEHTINENa and Riku NYKÄNENb Department of Computer Science and Information Systems, University of Jyväskylä, Finland b Yomi Software Oy, Finland Abstract. This paper describes a way to build a conceptual model for diversified purposes of modelling Enterprise Architectures (EA). It is commonly known that, due to the complexity, Enterprise Architectures need to be considered from several viewpoints. This provokes an integration problem: how to ensure that parallel EA models are consistent. We believe that the best way to solve this problem is to build a generic conceptual model (or an ontology) that is based on the purpose and needs of EA modelling rather than on the metamodels or modelling techniques of the prevailing (viewpoint-specific) domains of EA modelling. In other words, instead of aggregating existing sub-domains of EA we should try to find the core concepts through analyzing the EA domain as a whole. We emphasize the importance of the process through which the conceptual model is produced. Therefore, besides the conceptual skeleton and its utilization we provide an in-depth description of the modelling process we have developed and applied.

Introduction Enterprise architectures (EA) are seen as an important means to convert ever changing business requirements into well-functioning information systems [1, 2, 3]. Managing enterprise architectures is, however, a complex task. The complexity of systems is typically managed by division of problem, i.e. by looking at the system from several viewpoints [4, 5, 6, 7]. Thus, an EA specification as a representation of a complex system comprises a number of different viewpoints (e.g. [8, 9]). This, in turn, provokes another problem: how to ensure that the diverse viewpoint-specific models form a consistent whole ([10], see also [11]). The problem could be solved or at least eased by an architectural language which conceptually ties up the loose viewpoints ([12, 13], see also [9]). We argue that such a language should be based on careful conceptual analysis resulting in exact definitions for the core concepts and a simple structure between the concepts. The viewpoint-specific languages can then be derived from the simple, generic conceptual base. Because the viewpoints are many and they evolve, it is necessary that the solution combines a fixed core with a means to modify and extend the core to emergent situations. During the past few years some attempts have been made to define concepts necessary for modelling enterprise architectures. The RM-ODP standard [14] and its Enterprise viewpoint in particular, is an important effort in this direction. RM-ODP is, however, very complex as a conceptual framework and even the definitions of the core concepts are difficult to understand [15]. In addition, we see that the enterprise viewpoint of RM-ODP is only a relatively small part of the whole framework, and most of the framework aims at building the architecture of an open, distributed system, not a model of an enterprise. Some applications of RM-ODP can be found in IS research literature [16, 17, 18, 19]. A covering presentation of RM-ODP can be found in [20] where one can also find a metamodel of the

220

V. Halttunen et al. / Building a Conceptual Skeleton for Enterprise Architecture Specifications

Enterprise viewpoint of RM-ODP (p. 468). In our opinion, however, it is too ambiguous as a basic model for EA modelling purposes. Recently a promising work has been published by Jonkers et al. [12, 13]. Their aim is to define an enterprise architecture language. We share this target for our research. There are, however, relatively important differences between the ways of developing the language1. Whereas Jonkers et al. [12, 13] consider enterprise architectures as a solution for the business-ICT alignment problem, we have adopted a more general approach. Despite seeing information systems as an important part of EA we emphasize that, since utilized by people of very different backgrounds, the generic EA concepts should not be biased by IT thinking and concepts. We found this problem real in the LARKKI project2 where EA tools were developed for the purposes of different stakeholders. Our work differs from that of Jonkers et al.’s [12, 13] in another respect, too: although both of us utilize existing frameworks like RM-ODP, Jonkers et al. [12, 13] have anchored their concepts on existing architectural domains (product, organization, process, information, data, application, technical infrastructure), while we have aimed at a conceptual model that is not built from the viewpoint-specific domains3 but for both existing and future viewpoint-specific domains. In other words, instead of building an EA language by aggregating the sub-domains of EA (i.e. by finding generalities from the concepts of the sub-domains), we analyze the purpose of EA domain as a whole. The former approach puts emphasis on which (existing) constructions the EA domain is built from, while the latter one considers the overall target of EA modelling. Therefore, the latter one is also more elastic for new ideas and developments. Consider, for example, a situation where a new viewpoint is added to the EA domain. In our way of building the conceptual model there are two sources for finding the core concepts. First, we can analyze the concepts of existing standards such as RM-ODP to find some starting point for the further analysis. Too narrow approaches (focusing on a specific viewpoint) should be avoided and only the most generic concepts should be selected into the core concepts. Second, we can discuss with practitioners to understand why to build EA specifications in practice and which components an EA specification consists of. The dialogue between researchers and practitioners is an important part of our modeling process. The paper is organized as follows. In the next chapter we define the basic concepts that are needed to understand our approach. In chapter 2 we describe how the conceptual model has been produced and what the current version of the model looks like. Chapter 3 provides some ideas of utilizing the model, and finally chapter 4 includes a brief discussion and conclusions. 1. Basic Concepts We consider an enterprise as a complex entity that 1. has an identifiable purpose, 2. can explicitly be separated from other entities, and 3. comprises: (a) real world entities like people, several kinds of artefacts like technologies and documents, and (b) more or less abstract entities like organization structures, processes, and information objects (closely connected with documents).

1

To language we prefer to use term conceptual model or ontology. The LARKKI research project (10/2001 - 4/2005) was funded by the National Technology Agency of Finland (TEKES) and three ICT companies: IBM, TietoEnator, and Yomi. 3 By viewpoint-specific domain we mean a sub-domain within the enterprise architecture domain. 2


221

As mentioned previously, enterprises as complex entities are usually modelled from several parallel perspectives. In our terminology such a viewpoint-specific model is referred to as an enterprise model. As Whitman et al. [10] put it, an enterprise model usually represents, due to the complexity of the modelling object a single perspective of the enterprise. Each model that represents any feature of an enterprise can, thus, be called an enterprise model. An enterprise architecture (EA) is generally seen as a “blueprint” or a “big picture” that assists in the (re-)design of an enterprise. According to Liles and Presley [7] it should define the following three things: what are the activities that an enterprise performs, how should these activities be performed, and how should the enterprise be constructed. We see that enterprise architecture something more: it holds all the necessary components of an enterprise, although a great majority of them may be hidden at the highest level of abstraction. An EA describes the structures through which the static or dynamic features of an enterprise are integrated into a consistent whole. In other words, whereas an individual enterprise model can describe either the dynamic or static features of an enterprise, the enterprise architecture combines these two into a structure that tells how the enterprise is built. Furthermore, it is important to include purpose statements for each part of the EA. This is often ignored in individual enterprise models that tell the “how” aspect instead of the “why” and “what” aspects (see [21]). By an enterprise architecture specification we mean a representation through which the EA can be analyzed and discussed by different stakeholders. The formality of the specifications may vary depending on the purpose of use. The generic conceptual model which is developed in this paper helps to interpret and understand EA specifications that are produced according to the principles of the conceptual model. It also helps to integrate the loose viewpoint-specific models into a whole. 2. The Process and Outcomes of Conceptual Modelling In this chapter we briefly describe how our conceptual model was produced and what the current version of the model looks like. We emphasize that understanding the conceptual modelling process plays an important role when estimating the acceptability of the resulting model. 2.1. The Modelling Context Our modelling effort was a result of the needs that came up during the first year of the LARKKI project. The aim of the project is to develop methods and other tools for the diversified purposes of EA modelling and management. In one of the three company parties, Yomi, there was a need to develop a method and/or a framework that complements their software development method with an EA modelling tool. Quite soon, it was realized that since the tool would be an important part of communication among different stakeholders it would be necessary to build the work upon a sound conceptual basis. Hence, we started the conceptual modelling that will be presented in the subsequent sections. There were two principles followed during the process: (1) to keep the conceptual structure as simple and generic as possible, and (2) to utilize a participatory approach (the users of the results i.e. representatives of the company were involved throughout the process). 2.2. The Modelling Process and its Outcomes in short The main principle we followed in our modelling process was that - despite using existing terminologies, ontologies, metamodels etc. – the prevailing conceptual structures have been

222


used as little as possible. In the other words, we have attempted to distil the plain concepts from complicated structures. The primary argument for this procedure was simple: in addition to a limited number of concepts we aimed at limited number of relationships. Most of the existing conceptual models seemed to include unnecessary relationships. In our model we have included only those relationships that are necessary and excluded those ones that are possible but not necessary. In our opinion this helps utilizing the conceptual model for the purposes of different domains, even those that occur in the future. Concerning our conceptual modelling process we owe much to various contributors of the ontology engineering field. Just to mention some, these contributions include the ontology learning process [22], ideas on ontology competence [23], steps for obtaining a global ontology [24], evaluation of ontologies [25], and heuristics based ontology creation methodology [26]. In producing the conceptual skeleton we have combined the top-down and bottom-up approaches. We agree with van der Vet et al. [27] who state that the bottom-up approach of ontology building complements the top-down approach. A good ontology consists of two mutually supplementary parts: a natural-language for explanatory purposes and a formal part for ambiguity reduction (see [28]). In our process this is aimed at by building a graphical model and by developing a complementary dictionary. The following tasks are the generic constituent of the modelling process: 1. Find concepts; 2. Define concepts; 3. Analyze and discuss concepts and relationships between them, then, if necessary, go back to (1) or (2). The actual process can be composed of different combinations of these tasks. The basic idea is to combine analytical/theoretical concept defining with validating discussions with practitioners who will use the modelling language in practice. In each discussion round the concepts are weighted to find the most relevant ones. To put the same thing more formally, assume that Task 1 produces three concepts, say X, Y, and Z. Then Task 2 attaches definitions x, y, and z to the concepts respectively. In Task 3 the concepts X, Y, and Z are analyzed by analyzing the definitions x, y, and z. If definition x, for example, refers to some already identified concept, say Y, that concept must be analyzed further in the same way, and so on. If x refers to concepts not yet identified, say A, B, and C, there is a need to estimate whether or not these concepts should be added to the “dictionary”. Furthermore, in discussions some concepts (I, J, K etc.) may appear that are not referred to by any of the already defined concepts. This “emergent” stuff of brainstorming is an important part of building the conceptual model. Common concepts are not included in the conceptual base, only those that are seen as specific for the area of interest (which in this case is EA). By this simple method the coverage and integrity of the conceptual model can be evaluated. In Figure 1 we depict an instance process that is build upon the generic tasks. This actual process was followed by in the LARKKI project as we built the conceptual model. The process is presented as a linear one, but the cyclic nature of the process can be seen as repetition of similar phases (e.g. analysis). Step I was carried out by researchers as a literature survey. Based on the survey, Steps II and III were organized as group work between the researchers and the representatives of the company. As a result, an extended list of EA concepts was identified. These concepts were seen as candidates for the conceptual model and they were discussed in another group session (Step IV). On the basis of this discussion the researchers analyzed the concepts and their relationships (Step V). The outcome of this step was the preliminary version of the conceptual skeleton. This version was intentionally left small: only the very basic concepts


223

were included in it. After that, the researchers analyzed the conceptual model further and extended the model by some additional concepts (Step VI). This version was discussed with the representatives of the company (Step VII), and some improvements were made (Step VIII). I Finding and defining the basic concepts

LIST OF CONCEPTS

II Discussing the basic concepts

III Introduction of new potential concepts

EXTENDED LIST OF CONCEPTS

IV Discussing the concepts V Defining the crucial concepts and their relatatiohships

VI Enlargement of the model and relating definitions

FIRST VERSION OF CONCEPTUAL MODEL

ENLARGED MODEL

VII Discussing the enlarged model

VIII Improvements to the model

IMPROVED MODEL

Figure 1. The development process

In the following we describe the steps in more detail. 2.3. A Detailed Description of the Steps Step I: Analysis of the basic concepts To agree on a single top-level ontology is a convenient way to integrate ontologies [29]. This principle was the starting point for our modelling efforts. In the first step, we attempted to find the very basic EA concepts that could form the backbone of the model. The selection of the concepts was based on a survey (reported in [30]) on the existing RMODP standard and ongoing development work by OMG (MDA and the efforts on Enterprise Distributed Object Computing). One strength of the RM-ODP standard is that it includes very generic concepts that could be utilized in a work like ours. On the basis of our survey we suggested that the following concepts should be included in the conceptual model: x entity x action x purpose x scope x policy

224


The concepts were preliminarily defined as follows: Entity: This concept is similar to the basic modelling concept ‘object’ of RM-ODP [31]. An entity can be anything that exists. Regarding information systems, entities are users, producers or carriers of information. Entities have their internal properties and particular behaviour. At a given point of time an entity is in a state possible to that entity. An entity can be, for instance, a person, a system or a part of it, an order, an invoice etc. Action: This concept is similar to the basic modelling concept ‘action’ of RM-ODP. Action is something which happens. Action is associated with at least one entity. Actions can be partitioned into internal actions and interactions. Purpose: This concept is partially similar to the concept ‘purpose’ described in an Enterprise Viewpoint document [32]. In that document purpose is defined as “the practical advantage or intended effect of the system”. We remark that a system should be understood in a system theoretical sense, i.e. the whole organization can be seen as a system. In our framework the term ‘purpose’ is defined as the reason why an (business) entity, which in this context usually means a business organization, exists. Because the reason tells the final objective of a business entity, we see that ‘purpose’ includes the concept ‘objective’ (which is presented as a separate concept in the previously mentioned ISO/IEC JTC7 document). Hence, ‘objective’ is seen as a sub-concept to ‘purpose’. Scope: By this concept we mean the distinction between (business) entity and its environment. Thus, in terms of enterprise modelling the scope deals with the boundaries of an enterprise or other business entity (see [20], p. 471). In [32] the scope is defined as “the behaviour that system is expected to exhibit”. This is in line with our definition keeping in mind that behaviour is always associated with entities and a system can be any system like an enterprise organization. Policy: This concept is in accordance with a set of the RM-ODP policy concepts ([31], pp. 10-11). Policies deal with the behaviour of an entity. They can be declared by specifying obligations, permissions and prohibitions. The concepts of obligations, permissions and prohibitions are often referred to as deontic statements about a system (see [18]). Step II Discussing the basic concepts and Step III Introduction new potential concepts (these steps were intertwined) At the next two steps the concepts selected in the previous step were discussed, part of them were re-named and/or re-defined and new concepts introduced. Step II was accomplished in two phases: first researchers and practitioners discussed the concepts in own groups and next the concepts were discussed together. Immediately after the group discussion Step III was carried out. Actually, the latter part of Step II and Step III were organized as one group session. Ideas from Step III were collected by the researchers. As a result the following list of concepts was formed for further analysis: x Business Action x Policy x Business Entity x Process x Domain x Prohibition x Environment x Purpose x Frequency x Relationship x Interface x Role x Objective x Scope x Obligation x Strategy x Permission


225

It is necessary to remind the reader that the concepts above can be at different level of ontological hierarchy. They were included in the list because they were potentially significant for modelling enterprise architectures. Step IV: Discussing the concepts In this step we aimed to attach preliminary definitions to the above concepts in order to work with the hierarchical structures and other relationships among the concepts. Some of the concepts seemed to be more relevant than the others and gained thus more attention by the development group. For some of the concepts it was difficult to find an exact purpose and meaning. Those concepts were ignored at this phase, but they could be added in the conceptual model later on. This step and its outcomes were very intuitive. This was, however, an intended way to work with the concepts, since the next steps were more formal and aimed to reveal inconsistencies and missing parts. Step V: Defining the crucial concepts and their relationships Next, on the basis of the group discussions, the researchers defined the selected concepts (Table 1) and developed the first version of the conceptual model (Figure 2). It was produced by using the GOPRR metamodelling language (GOPRR comes from Graph, Object, Property, Role, and Relationship, see e.g. [33]). The GOPRR language seemed very useful since we did not like to make any difference between entities and attributes at this phase. The GOPRR language feature of property link made it easy to modify the model. For example, the modelling concept Purpose was at first modelled as a similar modelling concept type to the modelling concept Business Entity (both modelled as GOPRR object type, represented as square boxes). As the relationship between Business Entity and Purpose was analyzed to be of attribute type, the symbols of the modelling concept could be untouched, the relationship was just modelled as property link. Table 1. Definitions of EA concepts at the first step of development EA Modelling Concept

Definition

Business Action

A functional whole that is important to run the business

Business Entity

An entity that is important to the object system. The object system of the highest level is an enterprise as a whole.

Objective

Defined state of things that is to be reached. Described as a goal that is both concrete and measurable. Can be attached as an attribute to Business Entity or Business Action.

Policy

Constrains the behaviour attached to a role. Policies are categorized into obligations, permissions and prohibitions

Purpose

An attribute for Business Entity. Defines why a Business Entity exists.

Role

Is a “container” for behaviour. A Business Entity behaves through roles. The same entity may play several roles in the object system.

Scope

Defines the boundary between an object system and its environment.

Strategy

Defines the means by which the purpose and objectives are to be reached. An attribute for a Business Entity.

226


Scope

contextualizes Purpose text

Purpose

Objective text

Objective

Strategy text

Strategy

relates

Policy text

Policy

Business Action

Objective text

Objective

Business Entity

consists of

Role

Business Action explode

Figure 2. The first version of the conceptual model

As one can see, only a subset of the previous list of concepts (produced in Steps II and III) was included in this version of the conceptual model. The aim was to find out the core concepts upon which the new versions could be built. Step VI: Enlargement of the model and relating definitions In this step the researchers analyzed the conceptual model further, especially concerning the modelling object ‘Business Action’ which was divided into three special cases: Behaviour, Business Process and Function. The result of this step is depicted in Figure 3. Step VII: Discussing the enlarged conceptual model In the next step the enlarged model was discussed with a representative of the company and some improvements were made to the model resulting in the current version of the conceptual model (Figure 4, Table 2). It is thoroughly elaborated on in the next paragraphs.


Scope

contextualizes

consists of

Role

Purpose text

Purpose

Objective text

Objective

Strategy text

Strategy

relates

Policy text

Policy

Business Action

Objective text

Objective

Business Entity

Behaviour

Business Process

Function

is formed by

is a set of

includes

Entity Operation

Step

Task

Figure 3. The enlarged conceptual skeleton

227

228


Scope

contextualizes

Purpose text

Purpose

Objective text

Objective

Strategy text

Strategy

Business Action

Objective text

Objective

Behaviour

Business Process

Function

is formed by

is a set of

includes

Entity Operation

Step

Task

Business Entity

consists of

constrains

Role

relates

Prohibition

Policy

Permission

Obligation

Figure 4. The current version of the conceptual skeleton


229

Table 2. Refined and new definitions of EA concepts EA Modelling Concept

Definition

Behaviour

Business Action can be considered from the viewpoint of the behaviour of a Business Entity.

Business Action

A functional whole that is important to run the business. Compare with RM ODP concept Action: “Something which happens”

Business Entity

An entity that is important to the object system. The object system is the highest level business entity.

Business Process

Business Action can be considered from the process viewpoint, where a Business Process is a set of consecutive operations or steps that contributes to the achievement or realisation of a business objective/purpose (or of a well defined business activity).

Entity Operation

The operations of a business entity form the behaviour of that entity.

Function

Business Action can be considered from a viewpoint, where functions are ordered according to the hierarchy of the organisation.

Objective

Defined state of things that is to be reached. Described as a goal that is both concrete and measurable. Can be attached as an attribute to Business Entity or Business Action.

Obligation

A category of policies that tells what are the obligatory features of a role.

Permission

A category of policies that tells what is permitted to do by a role.

Policy

Constrains the behaviour attached to a role. Policies are categorized into obligations, permissions and prohibitions

Prohibition

A category of policies which tells what is prohibited to do by a role.

Purpose

An attribute of a Business Entity, which defines why a business entity exists.

Role

Is a “container” for behaviour. A Business Entity performs Business Actions in certain Roles. A Business Entity may be in several roles in the object system.

Scope

Defines the boundary between an object system and its environment.

Step

A constituent of Business Process

Strategy

An attribute of a Business Entity that defines the means by which the purpose is pursued and objectives are to be achieved.

Task

A constituent of a function.

Objective

Defined state of things that is to be achieved. Described as a concrete and measurable statement. Can be attached as an attribute to Business Entity or Business Action.

consists of

A relationship: A Business Entity consists of other business entities.

contextualizes

A relationship: A Scope puts a Business Entity in a context. For an object system there should be one scope definition. The object system is one entity. The other entities of the scope description are environmental entities (belong to the environment of the object system)

constrains

A relationship: A Policy sets constraints for a Role.

includes

A relationship: A Function includes tasks.

is a set of

A relationship: A Business Process is an ordered set of tasks i.e. steps.

is formed by

A relationship: Behaviour is formed by Entity Operations.

relates

A relationship: A Business Entity relates to Business Action and a Role).

230


The conceptual skeleton described above is based on the idea that an enterprise architecture specification should express the purpose, scope and policies of the object system. This is in accordance with what is stated in the RM-ODP standard [34]. In the following we try to explain how this target is pursued in our model. The core of the conceptual model is the concept of Business Entity. A business entity is any real or abstract part of the object system (enterprise) and which due to its importance must be taken into account when trying to understand and model the system. Such an entity can be, for example, an organization, an actor of a business action, a user or a producer of products or information, an information system or part of it, a database, or a document. The enterprise itself is the highest level business entity. Business Entities are named things of the “universe of discourse” and together they form collections of things that exist in that context. Business Entities can be categorized in different ways and they can be set on different levels of abstraction. A Business Entity may have several attributes. At the moment we have included only those attributes that have been seen as the most relevant ones for enterprise modelling, namely those that tell the purpose, objective, and strategy of a business entity (see the definitions in Table 2). In our conceptual model two kinds of relationships between Business Entities are defined. The ‘contextualizes’ relationship consists of Scope and one Business Entity (object system) that is related to one or more other Business Entities (environmental entities). When modelling an enterprise the object system is the whole enterprise. That is the highest level entity we are interested in. The other entities (environmental entities) are entities that are (1) at the same conceptual/ontological level with the enterprise, (2) important to the enterprise, and (3) not parts of the enterprise. Scope can be seen as an abstract entity that holds both the enterprise and the environmental entities and tells what the enterprise is as well as its environment. Briefly, the scope puts the enterprise into its context. As an entity the scope can have attributes of its own, like description or name. There is a one-to-one relationship between the enterprise and the scope, i.e. one scope defines one enterprise. The other relationship between Business Entities is named ‘consists of‘. That relationship tells that there are hierarchical structures between business entities. A business entity can be decomposed into its parts which, in turn, can be further decomposed into their parts, and so on. It is a matter of practice how far this decomposition is extended. To put it simply, decomposition can be done until further decomposition gives no benefit. Another practical point which helps to manage hierarchies is that only one decomposition level is considered at a time. This means that each decomposition picture or other specification should usually cover only one hierarchical level. Besides Business Entity the second core concept in our model is Business Action. In our conceptual model Business Entity is related to Business Action by a simple ‘relates’ relationship. Since the concept Role also takes part in this relationship, the relationship is not a binary relationship. The ‘relates’ relationship should be read as follows: Business Entities relate Business Actions through Roles. It is necessary to keep the concepts Business Entity and Role apart from each other although they can sometimes look very much alike. The other concepts include Policy, Behaviour, Business Process, Function, Obligation, Permission, Prohibition, Entity Operation, Step, and Task (see definitions in Table 2). A Policy is a constraint for a business action, and it is attached to a certain role. So, since the intended behaviour of an enterprise is formed by roles, the policies for the whole enterprise can be aggregated from the roles included in the enterprise. Policies are divided in three categories: Obligations, Permissions and Prohibitions. They are statements that tell what the Business Entity must do, can do, and must not do (through roles).


231

In our model business actions can be considered from three different viewpoints: x as behaviour of business entities x as business processes x as organizational structures of tasks The categorization was influenced by practice. It was accomplished by analyzing the different (practical) purposes of modelling business actions. The first viewpoint (the behaviour of business entities) is based on the idea of object-oriented system modelling. The viewpoint is useful when an enterprise is seen as a system, especially when trying to integrate information systems at enterprise level (compare with component thinking). The second viewpoint (business processes) emphasizes the importance of business processes in business modelling. Business processes have been on focus during the last 10-15 years and many modelling practices have been developed around business process re-engineering. These practices have proved to be valuable and applicable. The third viewpoint relies on the fact that throughout the history business organizations have been organized into units which can be seen as groupings of business task reflecting the managerial and organisatory purposes. Sometimes, it is the most convenient and easiest way to model business actions. For example, directors may look at the business actions through organizational structure (hierarchy) which is the mechanism to divide power and responsibility. 3. Utilization of the Conceptual Skeleton As described before, we have set as our target to build a conceptual model that is simple and easy to use in different situations and for different purposes of EA modelling. In our conceptual modelling process we followed three principles: (1) the core concepts should be as general as possible (not biased by any existing viewpoint-specific domain), and (2) the number of core concepts and the number of relationships between the concepts should be minimized, and (3) there should not be any “circular” relationship between the core concepts (e.g. A refers to B that refers to A; instead, a relationship of a concept to itself is allowed); this helps “read” the model and find the core of the model (which in this case includes Business Entity and Business Action). As a result, we had a conceptual skeleton that, in our opinion, is flexible and comprehensive enough to be easily modified for the purposes of different EA modelling domains. The skeleton has two functions: (1) it serves as a basic model that guides in developing the viewpoint-specific metamodels and, furthermore, enterprise models, and (2) it provides a means for integrating diversified enterprise models, i.e. the skeleton remains the same although the muscles around can be different. We provide here an example of how to use the skeleton model for building enterprise models that are being integrated into a whole. There are some basic guidelines that are necessary to be followed when utilizing the skeleton model: 1. Because the concepts of the skeleton model are connected with each other in a very simple way (“the backbone plus the ribs”), integration of separate enterprise models is a natural outcome, if the partial (viewpoint-specific) conceptual models use or explicitly refer to the concepts of the skeleton model. 2. Selection of the concepts of the skeleton model to be utilized should be based on the purpose of the viewpoint-specific model. By analyzing the purpose against the skeleton model and the complementary dictionary the following issues should be addressed: x Which concepts of the skeleton model are necessary/useful for the viewpoint

232


x

Which additional concepts are needed for the viewpoint-specific model (specialization of the skeleton concepts) 3. Relationships between the concepts of the skeleton model can be specialized to make the relationships more explicit. 4. Attributes can be attached to the concepts of the skeleton model. 5. The integration mechanism should be decided. Basically, the integration is based on including the same concepts of the skeleton model in the viewpoint-specific conceptual models. If this is not the case, it must be decided on a reference system which is built upon the concepts of the skeleton model. The above mentioned principles are briefly considered through a simple example in the following sections. Let us assume that our aim is to model two features of an enterprise: the environment (context) and the basic processes of the enterprise. After a careful analysis we decided to select those parts of the skeleton model that are represented in Figure 5. Business Entity

Scope

Role

relates

contextualizes

Business Action

Business Entity

Business Process

is a set of

Step

(a) the basic modelling concepts of enterprise’s environment

(b) the basic modelling concepts of enterprise’s processes

Figure5. Selected parts of the skeleton model

Next, we considered what additional concepts or modifications were needed. First, we noticed that the concept ‘Business Entity’ is too general in terms of both context modelling and process modelling. Regarding context modelling Business Entity was divided into two subtypes, (1) Enterprise and (2) Environmental Business Entity. In the process modelling case, two subtypes were also been seen as necessary, they were (1) Actor (can be either Person or Organization) and (2) Information Object. The relationships were clarified so that in terms of context modelling the roles (the circle symbol; this refers to the GOPRR concept ‘role’ not to the skeleton model concept Role) were defined for the relationship ‘contextualizes’. On the process modelling side an additional relationship, transition, was identified. Attribute names were attached to some of the modelling concepts. The result of all these modifications can be found in Figure 6.


Name of the scope string

Name of business entity string

Actor

Business Entity

Scope

Information Object

Description of the scope text

context part

env. part

233

Role

relates

contextualizes Business Action

object part

Business Process


Business Process

explode

Business Entity

is a set of

Enterprise

Environmental Entity to

Step

transition

from

(a) the modified conceptual model of modelling enterprise’s environment

(b) the modified conceptual model of modelling enterprise’s processes

Figure 6. The modified conceptual models

Finally, the integration mechanism(s) needed to be decided. In this simple example it is not problematic since there is a common concept, Business Entity, in both conceptual models. This should be a means to integrate the models into a whole. In other words, referencing from one model to another happens through Business Entities. In practice, it is a slightly more problematic task, because identifying and naming the business entities (instances of Business Entities) vary in different models. The solution would be to decide on an unambiguous identifying system. This is, however, a problem that is always present in modelling efforts. If there were no concept shared by the partial models the integration would be less straightforward. In this case the skeleton model should be used for finding a “connecting concept” which would build a bridge between the partial models. Since the two models of our example deal with Business Entities at different levels, it is obvious that additional models would be needed for full integration of the two types of enterprise models. Such a model could be a decomposition model the coverage of which is depicted in Figure 7. The decomposition model would describe how the context level entities would be decomposed into smaller entities that would be relevant in process modelling aspects.

whole

Business Entity

consists of


part

Figure 7. A conceptual model for a decomposition type enterprise model

234


4. Discussion and Conclusions Fox et al. [22] argue that the key criterion for evaluating ontologies is that what they call competence. Briefly, competence tells how well an ontology supports the task in which it is utilized. Every task sets different requirements for the ontology. We apply this idea for evaluating our conceptual model. The primary aim of our work is to provide a simple conceptual basis for diversified EA modelling purposes so that the resulting enterprise models could form a consistent whole. The competence questions can be put as follows: Is the model easy to understand and use? How does the model support in integrating partial enterprise models? Is the model easy to understand and use? Our model is based on a semantic analysis on the key concepts of EA modelling and relationships between them (compare the approach with the ideas by Huhns et al. [11]). It has been knowingly developed as a skeleton model rather than a big picture of loosely coupled individual models. We have intentionally avoided representing relationships that could be biased by a viewpoint-specific domain. All relationships are seen as generic and, thus, necessary in all circumstances where the concepts, which are part of the relationship, are used. We have also avoided representing circular relationships between modelling concepts. In reality, circular relationships are common, but at a more abstract level it is useful to avoid them in order to realize what the core concepts are - where to start to read the model. As compared with the existing frameworks, such as the extended Zachman framework [9] and RM-ODP (e.g. [20]), our model has far less concepts and much easier conceptual structure. Furthermore, our model as a generic conceptual model for EA is not biased/charged by IT-related concepts (compare with the EA language by Jonkers et al. [12] the generic concepts of which include concepts like data object, message, and document). How does the model support in integrating partial enterprise models? Since the partial conceptual models are instantiations and modifications of the skeleton model, integration will be a natural outcome. To obtain a good result requires, however, that (1) the purpose of the modelling effort is carefully analyzed against the skeleton model and the complementary dictionary, (2) the concepts of the skeleton model is used as such, or they are explicitly referred to, (3) practical decisions are made on the integration mechanisms through which integration is implemented in practice (for example, how modelling elements in different models are named). If these principles are followed, the skeleton model provides an unambiguous way to integrate parallel enterprise models into a consistent EA specification. In other words, our aim was not to aggregate existing viewpoints (compare with [9]) or enterprise model types (compare with [35]) but to find a conceptual basis (the most relevant concepts and relationships between them) that helps building situational viewpoint-specific architectural languages having a common core. This is somewhat different from the majority of current approaches. To sum up, EA modelling is such a large field that the applicability and usefulness of the skeleton model cannot be fully assessed until it has been applied for several purposes of EA modelling. It is possible, even obvious, that our skeleton model evolves in time. We, however, are convinced that our way of creating the conceptual skeleton is durable and can be applied to many other situations where one is trying to catch the conceptual basis of a complex, multi-view task similar to EA modelling.


235

References [1] Armour, F.J., Kaisler, S.H., & Liu, S.Y. (1999). A big-picture look at enterprise architectures. IT Professional. 1(1), 35-42. [2] Armour, F.J., Kaisler, S.H., & Liu, S.Y. (1999). Building an enterprise architecture step by step, IT Professional, 1(4), 31-39. [3] Chalmeta, R., Campos, C., & Grangel, R. (2001). Reference architectures for enterprise integration. The Journal of Systems and Software. 57(3), 175-191. [4] Deubler, H-H. (2001). Employing multiple views to separate large-scale software systems. The Journal of Systems and Software. 56(2), 105-113. [5] Groznik, A., & Kovacic, A. (2002). Business renovation: from business process modelling to information system modelling, In Proceedings of the 24th International Conference on Information Technology Interfaces, 2002. ITI 2002. (Vol. 1, pp. 405-409). [6] Kateel, G. P. Kamath, M., & Pratt, D.B. (1996). An overview of CIM Enterprise Modeling Methodologies. In WSC '96: Proceedings of the 28th conference on Winter simulation (pp. 1000-1007): ACM Press. [7] Liles, D. H., & Presley, A. R. (1996). Enterprise modeling within an enterprise engineering framework. In WSC '96: Proceedings of the 28th conference on Winter simulation (pp. 993-999): ACM Press. [8] Zachman, J. A. (1987). A framework for information systems architecture. IBM Systems Journal, 26(3), 276-292. [9] Sowa J. F., & Zachman J. A. (1992). Extending and formalizing the framework for information systems architecture, IBM Systems Journal, 31(3), 590-616. [10] Whitman, L., Ramachandran, K., & Ketkar, V. (2001). A taxonomy of a living model of the enterprise. In WSC '01: Proceedings of the 33nd conference on Winter simulation (pp. 848-855): IEEE Computer Society. [11] Huhns, M., Jacobs, N., Ksiezyk, T., Singh, M., Shen, W.M., & Cannata, P. (1993). Integrating enterprise information models in Carnot. In Proceedings of International Conference on Intelligent and Cooperative Information Systems, Rotterdam, Holland. [12] Jonkers, H., van Buuren, R., Arbab, F., de Boer, F., Bonsangue, M., Bosma, H., ter Doest, H., Groenewegen, L., Scholten, J. G., Hoppenbrouwers, S., Iacob, M.-E., Janssen, W., Lankhorst, M., van Leeuwen, D., Proper, E., Stam, A., var der Torre, L., & van Zanten, G. V. (2003). Towards a language for coherent enterprise architecture descriptions. In EDOC '03: Proceedings of the 7th International Conference on Enterprise Distributed Object Computing (pp. 28): IEEE Computer Society. [13] Jonkers, H., Lankhorst, M., van Buuren, R., Hoppenbrouwers, S., Bonsangue, M., & van der Torre, L. (2004). Concepts for modelling enterprise architectures. International Journal of Cooperative Information Systems, 13(3), 257-287. [14] ISO/IEC JTC1/SC07. (1996-98). Information technology - Open Distributed Processing - Reference Model. ISO/IEC 10746-1, 2, 3, 4. [15] Genilloud, G., & Wegmann, A. (2000). A foundation for the concept of role in object modelling. In Proceeding for Conference EDOC 2000 (pp. 76-85). [16] Blanc, X., Gervais, M. P., & Le-Delliou, R. (1999). Using the UML language to express the ODP enterprise concepts. In Proceedings of the Third International Enterprise Distributed Object Computing Conference (pp. 50-59). [17] Tanaka, A., Nagase, Y., Kiryu, Y., & Nakai, K. (2001). Applying ODP Enterprise Viewpoint Language to hospital information systems. In Proceedings of Fifth IEEE International Enterprise Distributed Object Computing Conference, EDOC '01 (pp. 188-192). IEEE Computer Society. [18] Linington, P. Milosevic, Z., & Raymond, K. (1998). Policies in communities: extending the ODP enterprise viewpoint. In Proceedings of the 2nd International Conference on Enterprise Distributed Object Computing, Workshop (pp. 14-24). [19] Wegmann, A., & Naumenko, A. (2001). Conceptual modeling of complex systems using an RM-ODP based ontology. In Proceedings of Fifth IEEE International Enterprise Distributed Object Computing Conference, EDOC '01 (pp. 188-192). IEEE Computer Society. [20] Putman, J. (2001). Architecting with RM-ODP. Upper Saddle River, N.J. Prentice Hall PTR . [21] Gordijn, J., Akkermans, H., & van Vliet, H. (2000). Business modelling is not process modelling. In Conceptual Modeling for e-Business and the Web. Retrieved May 27, 2005, from http://www.cs.vu.nl/~gordijn/ecomo-gordijn.pdf [22] Maedche, A., & Staab, S. (2001). Ontology learning for the Semantic Web. Intelligent Systems, IEEE, 16(2), 72-79. [23] Fox, M. S., Barbuceanu, M., & Gruninger, M. (1995). An organisation ontology for enterprise modelling: preliminary concepts for linking structure and behaviour, Computers in Industry, 29, 123-134.

236


[24] Lopez de Vergara, J.E., Villagra, V.A., Berrocal, J., Asensio, J.I., & Pignaton, R. (2003). Semantic management: application of ontologies for the integration of management information models, In Proceedings of the Eighth IFIP/IEEE International Symposium on Integrated Network Management (pp. 24-28). [25] Gómez-Pérez, A. (1995). Some ideas and examples to evaluate ontologies. In Proceedings of the 11th Conference on Artificial Intelligence for Applications (299-305). [26] Sugumaran, V., & Storey, V. C. (2002). Ontologies for conceptual modeling: their creation, use and management. Data & Knowledge Engineering, 42(3), 251-271. [27] van der Vet, P.E., & Mars, N.J.I. (1998). Bottom-up construction of ontologies. IEEE Transactions on Knowledge and Data Engineering, 10 (4), 513-526. [28] Gruber, T. R. (1991). The role of common ontology in achieving sharable, reusable knowledge bases. In J. F. Allen & R. Fikes & E. Sandewall (Eds.), KR'91: Principles of Knowledge Representation and Reasoning (pp. 601-602): Morgan Kaufmann. [29] Guarino, N. (1998). Formal ontology and information systems. In N. Guarino (Ed.), Proceedings of the 1st International Conference on Formal Ontologies in Information Systems, FOIS'98 (pp. 3-15): IOS Press. [30] Halttunen, V. (2002). Conceptual framework for enterprise architecture modelling, Internal project report of the LARKKI-project, University of Jyväskylä. [31] ISO/IEC JTC1/SC07. (1996). Information technology - Open Distributed Processing - Reference Model. Part 2: Foundations. ISO/IEC 10746-2:1996(E). [32] ISO/IEC JTC1/SC7 N2359. (2000, September 9). FCD 15414: Information Technology - Open Distributed Processing – Reference Model - Enterprise Viewpoint (A Draft). [33] Tolvanen, J.-P. (1998). Incremental method engineering with modeling tools, PhD Thesis, University of Jyväskylä, Jyväskylä. [34] ISO/IEC JTC1/SC07. (1996). Information technology - Open Distributed Processing - Reference Model. Part 3: Architecture. ISO/IEC 10746-3:1996(E). [35] Bubenko, J., Persson, A., & Stirna J. (2001). EKD user guide, IST-2000-28401.


237

Fixed-Point Quantifiers in Higher Order Logics Klaus-Dieter Schewe1 , José Mar´ıa Turull Torres2 Massey University, Information Science Research Centre & Department of Information Systems 1 Private Bag 11222, Palmerston North, New Zealand 2 PO Box 756, Wellington, New Zealand [k.d.schewe|j.m.turull]@massey.ac.nz Abstract. We add inflationary and non-inflationary fixed-points to higher-order logics. We show that, for every order, it is sufficient to increase the order of the given logic by one to capture inflationary fixed-points and by two to capture non-inflationary fixed-points. In the two cases, restricting to the existential fragment of the corresponding logic turns out to be enough. This also holds for non-deterministic fixed-points.

1

Introduction

Fixed-points play a decisive role in database theory [1]. On one hand the addition of fixed-point operators to a simple query language such as the relational calculus increases the expressiveness of these languages significantly. On the other hand various efficient ways to implement these languages had been developed, which have led to practical prototypes of languages of the DATALOG family. We distinguish between inflationary fixed-points (IFP) and non-inflationary (or partial) fixed-points (PFP). For the former ones it is always guaranteed that they exist, while the latter ones are more expressive. In terms of logic the addition of fixedpoints to a simple query language such as the relational calculus corresponds to a switch from first-order logic to fixed-point logics [5]. Thus, the expressiveness of query languages can be studied entirely within a logical framework. In doing so, we obtain interesting links to the complexity of queries. Known results from descriptive complexity include that IFP captures the complexity class P T IM E when added to first-order logic in the presence of a total order relation in the input structures, i.e. the databases [9, 11], while under the same conditions PFP captures P SP ACE [11]. There is also a non-deterministic version of these fixed-points. When added to first-order logic in the presence of a total order relation, the inflationary fixed-point quantifier captures the complexity class N P , whilst the non inflationary fixed-point captures P SP ACE [2]. The decisive insight from these results is that logics beyond first-order logic may be considered not much worth for computing in general, but they can be turned into very powerful tools when applied to finite structures, i.e. relational databases. However,

238

K.-D. Schewe and J.M. Turull Torres / Fixed-Point Quantifiers in Higher Order Logics

none of the fixed-point extensions of first-order logic captures all computable queries [4]. In fact, the gap between the class of all computable queries and the class obtained by fixed-point extensions to first-order logic, i.e. the so-called “deductice databases”, is far from being fully understood. It is not too difficult to see that the fixed-point operators are just special Lindstrøm quantifiers. In particular, it is well known that first-order logic plus IFP can be expressed in existential second-order logic, while first-order logic plus PFP can be expressed in existential third-order logic. This justifies to investigate higher order logics on finite structures. Preliminary work in this area has been sone in [7, 10, 8]. Here, we are interested in adding fixed-point quantifiers to higher-order logics. Distinguishing again between inflationary fixed-points (IFP) and partial fixed-points (PFP) we show that the known extensions to first-order logic can be generalised to higher order logics. That is, adding IFP to a logic HOi of order i can be equivalently expressed in ∃HOi+1 , i.e. by taking the existential fragment of the logic of order i + 1. Similarly, adding PFP to a logic HOi of order i can be expressed in ∃HOi+2 , i.e. by taking the existential fragment of the logic of order i + 2. Thus, increasing the order by one or two and restricting to the existential fragment is enough to capture inflationary or non-inflationary fixed-points, respectively. We also obtain the analogous result for non-deterministic fixed-points. In Section 2 we introduce some preliminaries concerning the syntax and semantics of higher-order logics. We also state the main result from [8] on the expressiveness of these logics. In Section 3 we provide direct constructions for formulae to implement partial and inflationary fixed-point operators, in combination with higher-order logic – increasing the order by two or one, respectively. Finally, in Section 4 we investigate nondeterministic fixed points. We sketch a direct construction for expressing inflationary and non-inflationary non-deterministic fixed-points in the existential fragment of logics of order increased by one or two, respectively.

2

Preliminaries

We will regard a relational database schema, as a relational signature, and a database instance or simply database as a finite structure of the corresponding signature [1, 5]. If A is a database or structure of some schema σ, we denote its domain by dom(A). If R is a relation symbol in σ of arity r, for some r ≥ 1, we denote by RA the (second order) relation of arity r which interprets the relation symbol R in A, with the usual notion of interpretation. We denote as Bσ the class of finite σ-structures, or databases of schema σ. In this paper, we will consider total queries only. That is, if σ is a schema, a computable Boolean query of schema σ ([4]), is a total recursive function q : Bσ → {0, 1} which preserves isomorphisms. We denote the class of computable queries by CQ. If ϕ is a formula of signature σ of some logic and A ∈ Bσ , we denote as ϕA the relation defined by ϕ in the structure A, and its arity is given by the number of free variables in ϕ. Formally, we say that a sentence ϕ of signature σ expresses a Boolean query q if for every database A of schema σ, is q(A) = 1 iff A |= ϕ. Syntax: For every i ≥ 2, in the alphabet of a Higher Order Logic of order i, HOi , besides the usual logical and punctuation symbols, we have a countably infinite set of


239

individual variables, and for every arity, and for every order 2 ≤ j ≤ i, a countably infinite set of relation variables. We will use caligraphic letters like X and Y for relation variables, and lower case letters like x and y for individual variables. The class of atomic formulae is as follows: 1) X (x0 , . . . , xr−1 ), where X is a relation variable of order 2 and of arity r, and x0 , . . . , xr−1 are individual variables; 2) X (Y0 , . . . , Yr−1 ), where X is a relation variable of order j, for some 3 ≤ j ≤ i, and of arity r, and Y0 , . . . , Yr−1 are relation variables of order j − 1, and of arity r; 3) X = Y, where X , Y are relation variables of the same order and of the same arity. To define the set of well formed formulae (wff), on the vocabulary σ, we add the following rule to the set of usual rules in F O: If ϕ is a wff, and X is a relation variable, then the following are also wff’s: ∃X (ϕ), ∀X (ϕ). Semantics: Let r ≥ 1. A second order relation of arity r is a relation in the classical sense, i.e., a set of r-tuples of elements of the domain of a given structure. For an arbitrary i ≥ 3, a relation of order i of arity r or an i-th order relation of arity r is a set of r-tuples of relations of order i − 1. In general by higher order relations we will mean relations of order i, for some i ≥ 2. Without loss of generality, and for the sake of simplicity, we assume that the arity of a higher order relation is propagated downwards, i.e., the relations of order i − 1 which form the r-tuples for a relation of order i, are themselves of arity r, and so on, all the way down to the second order relations, which are also of arity r. A valuation v on a σ-structure A, is a function which assigns to each individual variable x an element in dom(A), and to each relation variable X of order j, for some 2 ≤ j ≤ i, and of arity r, for some r ≥ 1, a relation of order j and of arity r on dom(A). Let v0 , v1 be two valuations on a σ-structure A, and let V be a variable of whichever kind, we say that v0 and v1 are V -equivalent if they coincide in every variable of whichever kind, with the possible exception of variable V . We will also use the notion of equivalence with respect to sets of variables. Let A be a σ-structure, and let v be a valuation on A. To define inductively the notion of satisfaction, we add the following rules to the usual rules for F O: 1) A, v |= X (x0 , . . . , xr−1 ), where X is a relation variable of order 2 and of arity r, and x0 , . . . , xr−1 are individual variables, iff the r-tuple (v(x0 ), . . . , v(xr−1 )) belongs to the second order relation v(X ); 2) A, v |= X (Y0 , . . . , Yr−1 ), where X is a relation variable of order j, for some 3 ≤ j ≤ i, and of arity r, and Y0 , . . . , Yr−1 are relation variables of order j − 1 and of arity r, iff the r-tuple of relations of order j − 1, (v(Y0 ), . . . , v(Yr−1 )) belongs to the relation of order j v(X ); 3) A, v |= X = Y, where X , Y are relation variables of order j, for some 2 ≤ j ≤ i, and of arity r, iff v(X ) = v(Y), i.e., v assigns to the two relation variables the same relation of order j and of arity r on dom(A); 4) A, v |= ∃X (ϕ), where X is a relation variable, and ϕ is a wff, iff there is a valuation v , which is X -equivalent to v, such that A, v |= ϕ; 5) A, v |= ∀X (ϕ), where X is a relation variable, and ϕ is a wff, iff for every valuation v , which is X -equivalent to v, A, v |= ϕ. For later use use let f (n) be a function defined on the set of natural numbers. Then we define exp(0, f (n)) = f (n), and for i ≥ 1 it is exp(i, f (n)) = 2exp(i−1,f (n)) . For every i ≥ 2, the expressive power of the existential fragment of higher order logic of order i, i.e., ∃HOi , is characterized by the complexity class N T IM E(exp(i − 2, nc )), with some constant c ([8]).

240

3


Fixed-Points in Higher Order Logics

Let the following be the syntax of the fixed-point quantifier for higher order logics, for both non inflationary, or partial (PFP) and inflationary (IFP): [F PR,R1 ,...,Rk ψ(R, R1 , . . . , Rk )](V1 , . . . , Vk ) where R is a k-ary relation variable of order (i + 1), for some k ≥ 1, R1 , . . . , Rk are k-ary relation variables of order i, V1 , . . . , Vk are k-ary relations of order i, and ψ is a formula in HOi , for i ≥ 2, whose (relational) vocabulary does not contain the relation symbol R, and which has exactly R1 , . . . , Rk as its only free relation variables. In general, a fixed-point operator (either inflationary or non inflationary) in HOi , for i ≥ 1, induces on a given structure, an iteration or a sequence of relations of order (i+1) and of arity k, for some k ≥ 1, as values of a k-ary relation variable R of order (i + 1), and binds that relation variable and a k-tuple (R1 , . . . , Rk ) of relation variables of order i. In each step of the iteration, a new k-ary relation is defined as the set of k-tuples of relations of order i which make ψ true when the relation variable R is interpreted as the relation defined in the previous step. Initially, the relation variable R is interpreted as the empty relation. If it so happens that the relations defined in two consecutive steps are equal, then that relation is said to be the “fixed-point” of ψ, otherwise the fixedpoint of ψ does not exist and is assumed to be the empty relation. Then the expression above is true iff the tuple of relations (V1 , . . . , Vk ) belongs to the fixed-point of ψ. As to the expression of the two fixed-point operators for the logics HOi , for i ≥ 2, we follow the same strategy as in F O. In the quantification of a relation variable we will write as a super-index of the variable symbol a pair of numbers, the first meaning the order and the second meaning the arity. We will use upper case caligraphic letters for variables of the highest order, and upper case regular letters for variables of all the other orders. Partial fixed-points can be expressed by the following formula. i+2,1 ∃ ≤i+2,2 “≤i+2 is a total order in S” ∧ ∃S ∀T i+1,k S(T ) → (“T is the first relation in ≤i+2 ” ∧ “T is the empty relation”∨ (∃X i+1,k (“X is the predecessor of T in ≤i+2 ”∧

∀W1i . . . Wki (T (W1i , . . . , Wki ) ↔ ψ(X, W1i , . . . , Wki )))) ∧ ∃U i+1,k S(U ) ∧ “U is the last relation in ≤i+2 ”∧ ∀W1i

. . . Wki (U (W1i , . . . , Wki )

↔

ψ(U, W1i , . . . , Wki ))

∧ U (V1 , . . . , Vk )

The informal expressions used in this formula and in the next one regarding inflationary fixed-points, are easily expressible in (HOi ). Then we have the following result. Theorem 3.1. Non inflationary (deterministic) fixed-point on HOi , for i ≥ 2, is expressible in ∃HOi+2 (i.e., in existential (i + 2)-th order logic).


241

We now consider Inflationary Fixed-Points in (HOi ). In the general case, i.e. in partial fixed-point (PFP), a fixed-point of the given formula (ψ) might not exist in a given database. We can force a formula to have always a fixed-point if we make it inflationary. A formula ψ(R, R1 , . . . , Rk ) is inflationary in R if for every database A of the corresponding schema, and for every k-ary relation T of order (i + 1) on the domain of A, the relation defined by ψ on A when the relation symbol R is interpreted by the relation T , includes T . In particular, ψ is inflationary in R if it is of the form R(R1 , . . . , Rk ) ∨ ψ (R1 , . . . , Rk , R). If ψ is inflationary in R, then the iteration induced by ψ is always increasing. Thus, it must have a “limit” which is the fixed-point. In the following, we assume without loss of generality that the formula ψ is of the form R(R1 , . . . , Rk ) ∨ ψ (R1 , . . . , Rk , R). Regarding HOi , for arbitrary i ≥ 2, the maximum cardinality of the set of relations of order (i + 1) and of arity k, for some k ≥ 1, which are possible values of a k-ary relation variable R of order (i + 1), is O(exp(i, nk )), i.e., it is the number of all possible relations of order i + 1 and of arity k, in a domain of n elements [8]. Now, if the fixed-point is inflationary, as before by definition the sequence of values for the k-ary relation symbol is smaller, since each relation in the sequence includes the previous one. Hence the cardinality of the set for HOi , with arbitrary i ≥ 2, is O(exp(i − 1, nk )), instead of O(exp(i, nk )). As in F O, we can take advantage of this fact by encoding the whole sequence of relations as a set of indexes (i.e., tuples) to the last tuple included at every stage, on some total order of the tuples. That is, the indexes will “point” to tuples in the full relation of the corresponding order and arity, i.e., the relation which has all the tuples of relations of one lower order. ∃ ≤i+1,2k “≤i+1 is a total order in the set of all k-tuples of order i”∧ i+1 i i i i ∃RFi+1,k P ∀X1 . . . Xk RF P (X1 , . . . , Xk ) → ”∧ (“(X1i , . . . , Xki ) is the first k-tuple in the order ≤i+1 restricted to RFi+1 P i+1 i i [∀A1 . . . Ak ((A1 , . . . , Ak ) ≤ (X1 , . . . , Xk ) ↔ ψ(φ, A1 , . . . , Ak )] ∨ ∃Y1i . . . Yki (“(Y1i , . . . , Yki ) is the predecessor of (X1i , . . . , Xki ) in ≤i+1 restricted to RFi+1 P ”∧ ∃S i+1,k T i+1,k (∀Z1i . . . Zki Ai1 . . . Aik ([(Z1i , . . . , Zki ) ∈ RFi+1 P∧ (Z1i , . . . , Zki ) ≤i+1 (Y1i , . . . , Yki )] ↔ S i+1 (Z1i , . . . , Zki )

i i i+1 (X1i , . . . , Xki )] ↔ T (Ai1 , . . . , Aik )) ∧ [(Ai1 , . . . , Aik ) ∈ RFi+1 P ∧ (A1 , . . . , Ak ) ≤

∧ ∀B1i . . . Bki (ψ(S, B1i , . . . , Bki ) ↔ T (B1i , . . . , Bki ))))∨

“(X1i , . . . , Xki ) is the last k-tuple in the order ≤i+1 restricted to RFi+1 P ”∧ ∃S i+1,k (∀Z1i . . . Zki ([(Z1i , . . . , Zki ) ∈ RFi+1 P∧

(Z1i , . . . , Zki ) ≤i+1 (X1i , . . . , Xki )] ↔ S(Z1i , . . . , Zki ) ∧ ∀B1i . . . Bki (ψ(S, B1i , . . . , Bki ) ↔ S(B1i , . . . , Bki )) ∧ S(V1 , . . . , Vk )))

Then we have the following result.

242


Theorem 3.2. Inflationary (deterministic) fixed-point on HOi , for i ≥ 2, is expressible in ∃HOi+1 (i.e., in existential (i + 1)-th order logic). 3.1

An Example of Fixed-Point on a Higher Order Logic

Let σ = G2 be the vocabulary of digraphs. Let us consider the query “the input graph is an n-cube for some n” in the vocabulary σ, and let the digraphs be symmetric and loop-free, that is, let the digraphs represent undirected graphs. Next we will present an HO3 formula which expresses that query and which uses an inflationary fixed-point on a second order sub-formula. Recall that K2 is the complete undirected graph of two nodes. ∃S 3,1 R3,2 “S 3 is a set of symmetric loop-free digraphs whose sets of nodes are included in the domain of the input graph G” ∧ R3 ⊆ S 3 × S 3 ∧ ∀H12,2 H22,2 ((S 3 (H12 ) ∧ S 3 (H22 ) ∧ R3 (H12 , H22 )) ←→ “the graph H2 is the result of multiplying the graph H1 with K2 ” ∧ [F PT 3,2 ,R2,2 ,R2,2 (R3 (R1 , R2 ) ∨ ∃R32,2 (R3 (R1 , R3 ) ∧ T 3 (R3 , R2 )](K2 , G)) 1

2

in which “the graph H2 is the result of multiplying the graph H1 times the graph K2 ” is, informally, expressed by the formula ∃F12,2 F22,2 s.t. “F1 is a total injective function from the domain of H1 to the domain of H2 ”, and “F2 is a total injective function from the domain of H1 to the domain of H2 ”, and “the range of F1 and the range of F2 form a partition on the domain of H2 ”, and “the subgraph induced in H2 by the range of F1 is isomorphic to the graph H1 through F1 ”, and “the subgraph induced in H2 by the range of F2 is isomorphic to the graph H1 through F2 ”, and “for every node x in the graph H1 there is a pair of symmetric edges in the graph H2 between the two nodes in H2 which are the images of the node x by F1 and F2 ”, and “there is no other edge in the graph H2 between nodes of the two isomorphic copies of H1 mapped through F1 and F2 in H2 ”. Note that the formula above can be generalized in a straightforward way. Let the third order set S be a k-ary set, for some k ≥ 1, of databases of some schema ρ = R1 , . . . , Rk , whose domain is included in the domain of the input database, and let the third order relation R be a 2k-ary relation which for a given kind of updates t, contains the pairs of ρ-databases (I1 , I2 ) such that database I2 is the result of applying one single transaction of kind t to database I1 . Then, the modified formula on the vocabulary σ = R11 , . . . , R1k , R21 , . . . , R2k , would express the query given the database I I I I I I , . . . , R1k , R21 , . . . , R2k , the sub-database I2 = DI , R21 , . . . , R2k can be I = DI , R11 obtained by only applying a sequence of updates of kind t to the sub-database I1 = I I , . . . , R1k . In the example above, the only transaction of kind t is multiplication DI , R11 by K2 .


4

243

Non-Deterministic Fixed-Points

Following [2] we define next the non-deterministic fixed-point quantifier for F O. Let the following be the syntax of NFP (inflationary and non inflationary): [N F PR,x1 ,...,xk (φ(R, x1 , . . . , xk ), ψ(R, x1 , . . . , xk ))](v1 , . . . , vk ) where R is a k-ary relation variable R, for some k ≥ 1, x1 , . . . , xk are individual variables, and φ, ψ are first order formulae whose (relational) vocabulary does not contain the relation symbol R, and which have exactly x1 , . . . , xk as their only free relation variables. The formulae φ(x1 , . . . , xk , R) and ψ(x1 , . . . , xk , R) generate sequences of stages that are obtained by successively evaluating φ or ψ. That is, the pair generates sequences of k-ary relations of the form S0 , S1 , . . . , Sm , where S0 = ∅, either Si+1 = φ(Si ) or Si+1 = ψ(Si ), and φ(Sm ) = ψ(Sm ) = Sm . We call Sm a local non-deterministic fixed-point of the pair φ, ψ. Note that the pair φ, ψ can have more than one local non-deterministic fixed-point or none. We define the non-deterministic fixed-point of the pair φ, ψ as the union of all local non-deterministic fixed-points of the pair φ, ψ. If no local non-deterministic fixed-point exits, then we define the nondeterministic fixed-point to be the empty set. Non-deterministic fixed-point logics are obtained by augmenting 1st-order logic with the non-deterministic (inflationary and non-inflationary) fixed-points formation rules, under the restriction that negation cannot be applied to non-deterministic fixed-points. In [2] it was proven that on ordered structures non-deterministic non-inflationany fixed-point on F O captures P SP ACE, same as deterministic non-inflationary fixed-point, and non-deterministic inflationany fixed-point on F O captures N P . That is, on ordered structures and for the case of non-inflationary fixed-points, non-determinism does not lead to an increase in expressiveness. By the results in [8] this implies that non-inflationary NFP is expressible in ∃HO3 (since ∃HO3 = N T IM E(2n ), which includes PSPACE) and by the well-known equivalence ∃SO = N P inflationary NFP should be expressible in ∃SO. This means that there are formulae in those logics to express these fixed-point quantifiers. For the non-inflationary case note that, as the tuple must be in the union of all local FP, it is equivalent to say that “there is a sequence of relations (i.e., an ordered set of order 3) s.t. there is a local fixed-point in that sequence and the tuple (v1 , . . . , vk ) is in that local fixed-point. This can be written with an existential HO3 formula. As to inflationary NFP, we can use the same approach, and as we can represent an inflationary sequence with a single relation, as it is usually done in the deterministic case, an existential SO formula suffices. As to the expressibility of non-deterministic fixed-point in Higher Order logics, note that we can use exactly the same approach which we used for F O. We only need to consider that the syntax will include higher order relation variables. Let the following be the syntax of NFP (inflationary and non inflationary): [N F PR,R1 ,...,Rk (φ(R, R1 , . . . , Rk ), ψ(R, R1 , . . . , Rk ))](V1 , . . . , Vk ) where R is a k-ary relation variable R of order (i + 1), for some k ≥ 1, R1 , . . . , Rk are k-ary relation variables of order i, V1 , . . . , Vk are k-ary relations of order i, and

244


φ, ψ are formulae in HOi , with i ≥ 2, whose (relational) vocabulary does not contain the relation symbol R, and which have exactly R1 , . . . , Rk as their only free relation variables. As to the formulae to express the two fixed-points, for the non inflationary case, note that as the tuple must be in the union of all local FP, then it is equivalent to say that “there is a sequence of relations (i.e., an ordered set of order (i + 2)) s.t. there is a local FP in that sequence and the tuple of relations of order i, (V1 , . . . , Vk ) is in that local FP”. This can be written with an existential HOi+2 formula. As to inflationary NFP, we can use the same approach, and as we can represent an inflationary sequence of relations of order (i + 1) with a single relation of order (i + 1), as we did in the deterministic case, an existential HOi+1 formula suffices. Then we get the following result. Theorem 4.1. (i) Non inflationary non-deterministic fixed-point on HOi , for each i ≥ 2, is expressible in ∃HOi+2 (i.e., in existential (i + 2)-th order logic). (ii) Inflationary non-deterministic fixed-point on HOi , for each i ≥ 2, is expressible in ∃HOi+1 (i.e., in existential (i + 1)-th order logic).

References 1. Abiteboul, S., R. Hull, and V. Vianu, Foundations of Databases, Addison-Wesley, 1994. 2. Abiteboul, S., Vardi, M. and Vianu, V., Fixed-Point Logics, Relational Machines, and Computational Complexity, Journal of ACM 44(1), pp. 30-56, 1997. 3. Balcázar, J., J. D´ıaz, and J. Gabarró, Structural Complexity I, Springer, 2nd. ed., 1995. 4. Chandra, A. K. and D. Harel, Computable Queries for Relational Data Bases, Journal of Computer and System Sciences 21(2), p. 156-178, 1980. 5. Ebbinghaus, H., and J. Flum, Finite Model Theory, Springer, 2nd. ed., 1999. 6. Fagin, R., Generalized First Order Spectra and Polynomial Time Recognizable Sets, in “Complexity of Computation, SIAM AMS Proceedings”, ed. R. Karp, 7, 1974. 7. Hull, R., and J. Su, On the Expressive Power of Database Queries with Intermediate Types, Journal of Computer and System Sciences 43(1), 219-267, 1991. 8. Hella, L., Turull Torres, J. M., Expressibility of Higher Order Logics, Electronic Notes in Theoretical Computer Science, Vol. 84, 2003. 9. Immerman, N., Relational Queries Computable in Polynomial Time, Information and Control, 68: 86-104, 1986. 10. Leivant, D., Descriptive Characterizations of Computational Complexity, Journal of Computer and System Sciences 39(1), p. 51-83, 1989. 11. Vardi, M., The Complexity of Relational Query Languages, Proc. 14th ACM Symp. on Theory of Computing, p. 137-146, 1982.


245

Discovery of Narrativity on the WWW based on Perspective Information Access Mina AKAISHI*, Nicolas SPYRATOS†, Koichi HORI* and Yuzuru TANAKA‡ *RCAST, University of Tokyo, Japan †Laboratoire de Recherche en Informatique, Université de Paris-Sud, France ‡Meme Media Laboratory, Hokkaido University, Japan E-mail:[email protected], [email protected], [email protected], [email protected] Abstract. We propose a framework for discovering narrative relationships between objects on the WWW based on perspective information access. Today people can access Web resources easily. However, it is difficult to reach information that satisfies users’ requests. Search engines are the most popular tools to find information, if a user can specify appropriate keywords to express the concrete contents of the required information. However, our target users are persons looking for new information concerned with other information, even when the relationship among them is not yet known. To support such users, we propose a perspective information access framework that shows paths to reach the required information. Such perspective path corresponds to narrativity between the source and the destination information.

1

Introduction

The rapid proliferation of information sources in recent years and the advent of the Internet have created a world wide web of interconnected information resources. Today, the Web represents the largest collection of information resources to which individuals have ever had access - and it continues to grow at a furious pace. Several kinds of tools have been developed in recent years to help individual users access Web resources. So-called search engines are the most popular such tools, as they allow users to access indexed resources using a very simple search mechanism: namely, keywords or combinations of keywords. However, the answers obtained from search engines usually contain large amounts of information unrelated to what the user had in mind when asking the question. To solve such situations, software is emerging that analyzes search results, sorts them automatically into categories, and presents them visually to provide far more information than the typical textual list. For example, the Vivisimo [1] Clustering Engine automatically categorizes search results into hierarchical folders. Filing is done through a combination of linguistic and statistical analysis. Another example is Grokker [2], which not only sorts search results into categories but also ”maps” the results in a holistic way, showing each category as a colorful circle. Within each circle, subcategories appear as more circles that can be clicked on and zoomed in. These tools show all the possible categories of information the Internet offers. Such results help users understand the overview of retrieved results at a glance.

246 M. Akaishi et al. / Discovery of Narrativity on the WWW Based on Perspective Information Access

However, it is still difficult to specify the appropriate keywords to get the necessary information.Therefore, we propose a perspective information access approach. Retrieval keywords specified by users do not only represent the content parts of target information but also describe retrieval purposes, roles of target information, and how to use the information. Current search engines regard keywords as parts of the retrieved results. On the other hand, we regard keywords as information on a path that leads to results. In this paper, we treat the huge collection of Web pages as a large information base where anchors play the role of descriptors for distant web pages and the web pages themselves play the role of described objects. While accumulating a large collection of Web pages, a user may define a variety of descriptors of Web pages. The same Web page is referred to by various descriptors based on its context. A path is a sequence of descriptors that lie on the concatenation of hyper links from any Web page to another. The path explains the relationship between the start and the end pages. Each path describes different information in which the desired information is embedded, helping users select correct ways to reach the desired information. It also suggests new contexts for the target information of which a user may not be aware. Therefore, we contend that the paths from specified viewpoint to the target object lead users to the necessary information. In addition, a path suggests unknown relations between result pages and the viewpoint page. We regard the path as a narrative between two objects. The perspective information access method gives several ways to reach the object and the meanings of paths, that is narrative, defined dynamically based on the current collection of Web pages. The goal of this paper is to propose a framework for discovering paths to reach the information that satisfies retrieval purposes and to show the effectiveness of perspective information access systems using some examples. The remainder of this paper is organized as follows. In section 2, we explain the primitive concepts of the Perspective Information Access Approach. In section 3, we describe the framework that discovers narrative paths based on a Perspective Information Access Approach. Finally, in section 4 we offer some concluding remarks. 2

Perspective Information Access Approach

When we start information access, sometimes we do not know exactly what we want. Most of the existing systems - knowledge base, database, and web applications - only work when requirements are articulated in advance; we cannot search for information using web search engines if we cannot express what we want in keywords. Generally, those who know part of the content of the target information can describe their requests as queries. They are looking for already known information, which means that they have already accessed the target information or that they are somehow convinced of its existence. For example, if looking for a book written by Goethe, one can just ask a database to select books written by Goethe. However, for persons who do not know the content of the necessary information, it is difficult to describe a query; They are looking for new information. Perhaps the result information is known; however, the relationship connecting the search with the obtained information is unknown. Perspective information access suggests possible relationships from the search purpose to new information by chain reactions. Let us consider a person who wants to improve her skiing ability: she would like to learn new effective ways to practice skiing; such an image result is vague even when the retrieval purpose is clear. Perspective information access returns ”new and effective” information

M. Akaishi et al. / Discovery of Narrativity on the WWW Based on Perspective Information Access 247 URL0 URL1

Japanese

URL2

………. ………. i and repeat the process, for the second case take i < i and redo the process. It is reasonable to choose the ordering of the selection predicates according to the data volume generated by the fragment. Each query qi requires a particular portion sif from fragment f . So, vol(f ) = N i=1 fi · sif , where the sum runs over all queries and fi is the frequency of query qi indicates the total data volume due to fragment f . Therefore, order the selection predicates according to decreasing data volume.

6

Conclusion

We continued our work on distribution design for complex value databases emphasising a detailed analysis of the effects of horizontal fragmentation on total query costs. We argued that besides reliability with respect to non-available sites a major reason for distribution is the expected increase of performance. Therefore, we argued that query optimisation must be considered as an intrinsic part of distribution design. We developed a query processing cost model based on a query algebra for complex value databases. If query trees are optimised and an optimal allocation of intermediate

296

H. Ma and K.-D. Schewe / Query Optimisation as Part of Distribution Design

results to locations in a network is given, a horizontal fragmentation will only impact on the allocation of a leaf and at most two of its predecessors. That is, the optimisation of allocation of locations to nodes in a query tree and the allocation of locations to fragments can be considered as two largely orthogonal problems. We concentrated on horizontal fragmentation and the allocation of fragments. For this we presented an allocation heuristics based on the needs of queries from the fragments. We showed that a horizontal fragmentation into two parts either has no effect at all or reduces the total query costs. In the former case the fragmentation can be discarded. This observation then gives rise to a binary search procedure to detect an “optimal” fragmentation and allocation. We are currently investigating a larger experimental case study to further consolidate these results.

References 1. Abiteboul, S., Buneman, P., and Suciu, D. Data on the Web: From Relations to Semistructured Data and XML. Morgan Kaufmann Publishers, 2000. 2. Bellatreche, L., Karlapalem, K., and Simonet, A. Algorithms and support for horizontal class partitioning in object-oriented databases. Distributed and Parallel Databases 8, 2 (2000), 155–179. 3. Chinchwadkar, G. S., and Goh, A. An overview of vertical partitioning in object oriented databases. The Computer Journal 42, 1 (1999). 4. Ezeife, C. I., and Barker, K. A comprehensive approach to horizontal class fragmentation in a distributed object based system. Distributed and Parallel Databases 3, 3 (1995), 247–272. 5. Ezeife, C. I., and Barker, K. Distributed object based design: Vertical fragmentation of classes. Distributed and Parallel Databases 6, 4 (1998), 317–350. 6. Ma, H., and Schewe, K.-D. Fragmentation of XML documents. In Proceedings XVIII Simp´ osio Brasileiro de Bancos de Dados (SBBD 2003) (Manaus, Brazil, 2003), pp. 200–214. 7. Ma, H., and Schewe, K.-D. A heuristic approach to horizontal fragmentation in object oriented databases. In Proceedings of the 2004 Baltic Conference on Databases and Information Systems (Riga, Latvia, 2004). 8. Malinowski, E., and Chakravarthy, S. Fragmentation techniques for distributing object-oriented databases. In Conceptual Modeling - ER ’97 (1997), D. W. Embley and R. C. Goldstein, Eds., vol. 1331 of Lecture Notes in Computer Science, Springer, pp. 347–360. 9. Schewe, K.-D. Fragmentation of object oriented and semi-structured data. In Databases and Information Systems II (2002), H.-M. Haav and A. Kalja, Eds., Kluwer Academic Publishers, pp. 1–14. 10. Schewe, K.-D., and Thalheim, B. Fundamental concepts of object oriented databases. Acta Cybernetica 11, 4 (1993), 49–84. 11. Thalheim, B. Entity-Relationship Modeling: Foundations of Database Technology. Springer-Verlag, 2000.


297

Position Paper: Digital Media Archive for Academic Resources Takanari HAYAMA, Eiji KAWAGUCHI, Yasushi KIYOKI Research Institute for Digital Media and Content Keio University

Today, there is high demand for digital media archives that can archive wide variety of academic resources and deliver customized contents upon audiences’ needs. Key challenges of the digital media archives for academic resources are a) a methodology to embed academically meaningful interpretation of resources, i.e. metadata, to digital media data themselves b) search mechanism to such academic metadata, and c) sophisticated access methodology to such digital archives. In this position paper, we describe our approach to these challenges, and how we’re going to implement the digital media archives in our project.

1.

Introduction

Digital multimedia contents became the first citizens today. Many appealing multimedia contents are available on the net. Any consumer level PCs are able to browse sophisticated video and audio media. However, although the tools for multimedia contents authoring are available for anyone, it is not yet for anyone. One of the reasons is that there is no ecosystem to support authoring and distribution of multimedia contents that could be used by any individual. Digital multimedia contents are valuable for academic purpose as well. Audio and video academic resources that help audiences to understand topics are very important. Not only contents themselves, but also additional descriptive information is very valuable to let audiences to get background knowledge of topics. Also, if customizations of such contents are possible easily and on-the-fly, then contents could be generated based on audiences’ needs on demand. To support such use of academic resources, an ecosystem for academic multimedia contents, which support creation, archiving, searching, editing and distribution of contents, is essential (Figure 1).

298

T. Hayama et al. / Position Paper: Digital Media Archive for Academic Resources

When we say “academic multimedia contents”, not only data themselves, but descriptive information and interpretation of academic resources, i.e. metadata, are more crucial. Each academic resource is often given many metadata based on different views and aspects. In this paper we describe our approach to realize a digital media archive for academic resources. Key challenges of such archive are a) a methodology to embed academically meaningful interpretation of resources, i.e.

Creation Creation

Distribution Distribution

Archiving Archiving

Digital Digital Media Media Archive Archive forAcademic forAcademic Resources Resources

Editing Editing

Searching Searching

metadata, to digital media data themselves, b) Figure 1 Digital Media Archive for search mechanism to such academic metadata, Academic Resources and c) sophisticated access methodology to such digital archives. Our approach for a) is to use steganographic scheme that allows㩷 embedding of metadata directly to the media data themselves. For b), we are using the mathematical model of meaning to support “Kansei” information retrieval. Finally for c), we are developing new system to access the archive using the special terminal device.

2.

Embedded Metadata using Steganographic Scheme

2.1. Problems in metadata handling in most database systems Most database systems have troublesome problems in metadata handling. When someone creates and edits metadata of media data, he/she must use a dedicated metadata-editing subsystem of the database systems. This means the workplace for metadata handling is limited within the system. In other words, the editing operations cannot be done “off-line.” Another problem is a migration of the metadata, as every media data and its metadata must be migrated together from one system to another. It creates a possible isolation of metadata and media data.

2.2. A solution using a steganographic technique One solution to the issues is to embed metadata to media data itself, i.e. turn the media data into “multi-layered information structure.” The external layer carries the original view of the media data (e.g., the view of the photo picture), and the internal layer carries its metadata (e.g., the profile of the photo). The metadata can be easily extracted from the

T. Hayama et al. / Position Paper: Digital Media Archive for Academic Resources Original

Target

External

299

Multi-layered

internal layer by using a Metadata Internal predefined data-extracting Figure 2 A Structure of Multi-Layered Image Data algorithm when needed. An external-and-internal layer is not a physical structure, but a conceptual structure of the media data. Information on the internal layer is invisible. Our approach employs a steganographic method to create multi-layered information structure, typically by way of Bit-Plane Complexity Segmentation (BPCS) Steganography[1]. 2.3. Outline of the BPCS algorithm The BPCS algorithm for getting a multi-layered information structure of the image data has two procedures, a metadata embedding procedure and an extracting procedure. The embedding procedure is as follows: 1) Decode an image data and bit-slice to create multiple binary images from the image data. 2) Lists all “noise-like” 8x8 pixels blocks in each bit-plane. 3) Compress and fragment the given metadata and generate a series of 8x8 square blocks. 4) Replace all noise-like blocks with the fragmented metadata in a block by block manner. 5) Re-encode the image data in the original file format. The metadata-embedded image, i.e. multi-layered photo data, has no visual degradation compare to the original image. Figure 2 illustrates this structure. The internal layer is unrecognizable to human. Additionally, the size of the image data doesn’t increase even after metadata is embedded.

3.

Search Mechanism for Academic Resources

3.1. Importance of “kansei” information retrieval In the design of multimedia database systems, one of the important issues is how to deal with “kansei” of human beings. The concept of “kansei” includes several meanings on sensitive recognition, such as “human senses,” “feelings,” “sensitivity,” “psychological reaction” and “physiological reaction”[2]. The field of kansei was originally introduced as the word “aesthetic” by Baumgrarten in 1750. The aesthetics of Baumgrarten had been established and succeeded by Kant㩷 with his ideological aesthetics. In the research field of multimedia㩷 database systems, it is becoming important to deal with kansei㩷 information for human beings for defining and extracting media data㩷 according to impressions and senses of individual users.

300


3.2. A methodology for kansei information retrieval Our approach to such kansei information retrieval is to use mathematical model of meaning [3,4]. The mathematical model of meaning provides semantic functions for computing specific meanings of keywords which are used to perform unambiguous and dynamic retrieval of images. The key of this model is that the semantic associative search is performed in the orthogonal semantic space. This space is created for dynamically computing semantic equivalence or similarity between the metadata items of the images and keywords. A learning mechanism is very important for database systems dealing with kansei information to adapt retrieval results according to individual’s sense and improving accuracy of the retrieval results. Our approach includes also mechanism to support the learning mechanism [5]. 4.

Ubiquitous Digital Media Archive Access

4.1. Importance of ubiquitous access Legacy database access requires users to access database using mouse and keyboards, the input device to give instructions. Users are required give keywords or select menu to find the contents, and they need to find information from the long list of candidates. The way users want to find information they need, they may not be always having access to PC, or they may not know the appropriate keywords to get more information related to the topic interests. For instance, when a user sees a painting on the wall and wants to know more about it, the ideal way to access to such information is directly from the painting itself. If we can embed URI by RFID, NFC or 2-D code directly to the painting, then what the user should do is to holds a device to read the URI from the painting and get access to more information they he/she needs. This allows anywhere and anytime access to academic resources or knowledge per se. Academic institute holds or owns resources that cannot be digitized completely, i.e. they cannot be stored in digital media archive. For example, a piece of stone that helped an artist to come up with new inspiration, and or a texture used as a material for art work cannot be completely digitized. It could be store as images or 3D model data, but not as objects themselves. Therefore, connecting real world and digital world is crucial when our goal is to create digital media archive for academic resources. There should be the way to access from real objects to its digitized metadata and vice versa. Ubiquitous access methodology can help to realize such diverse ways of accessing to academic resources.


301

4.2. Experimental system for ubiquitous access For experimental purpose, we’re planning to develop a device which has built-in RFID reader and high-speed wireless network as a digital media archive terminal. RFID tag with unique ID will be attached to original real-world resources. Information related to resources that could be digitized is stored that are somehow digitized and stored in digital media archives. Such information includes images of resources and their metadata. As the result, Users will have access to digital media archives to know interpretation or background, i.e. metadata, of real world resources.

5.

Conclusion

In this paper, we have presented our approach to the digital media archive for academic resources. Key challenges have been addressed and explained our solution to them. The authors’ group in Research Institute for Digital Media and Content at Keio University (Keio DMC Institute) started a project to construct several metadata embedded photo archive systems. We have started the research project in 2004Q3, and are planning to develop the first prototype system by 2005Q2. Our goal is to develop a complete global ecosystem for such academic resources.

References [1] Kawaguchi, E. and Eason, R., “Principle and applications of BPCS Steganography”, Proceedings of SPIE: Multimedia Systems and Applications Vol.3528, pp.464-473, 1998. [2] Harada, A. (eds.), “Report of modeling the evaluation structure of KANSEI 1997”, University of Tsukuba, 1997. [3] Kitagawa, T. and Kiyoki, Y., “A Mathematical Model of Meaning and its Application to Multidatabase Systems,” Proc. 3rd IEEE International Workshop on Research Issues on Data Engineering: Interoperability in Multidatabase Systems, pp.130-135, 1993. [4] Kiyoki, Y., Kitagawa, T. and Hayama, T., “A Metadatabase System for Semantic Image Search by a Mathematical Model of Meaning,” Multimedia Data Management - using metadata to integrate and apply digital media -, Chapter 7, A. Sheth and W. Klas (eds.), McGraw Hill (book), 1998. [5] Kiyoki, Y., Miyagawa, A. and Kitagawa, T., “A multiple view mechanism with semantic learning for multidatabase environments,” Information Modeling and Knowledge Bases, Vol. IX, P.-J. Charrel and H. Jaakkola (eds.), IOS Press, 1998.


302

An Ontology Learning and Reasoning Framework Hele-Mai HAAV Institute of Cybernetics at Tallinn University of Technology, Akadeemia tee 21, 12618 Tallinn, ESTONIA

Abstract. Most of current ontology languages and methodologies lack of learning support combined with reasoning mechanisms. This motivates us creating an ontology learning and reasoning framework in order to guarantee accuracy, transparency and consistency of ontology representation by automatic or semiautomatic methods of ontology learning and reasoning. Novelty of our approach is in combining ontology learning and reasoning into one framework. For ontology extraction and learning, we use Formal Concept Analysis. For ontology reasoning purposes, we require that learned ontology should be automatically mapped to some logic language. We use predicate logic as a target ontology language in this paper.

1. Introduction and motivation In recent years, ontology has become a concept in computer science. In most widely used understanding, an ontology is constituted by a specific vocabulary used to describe a certain domain of interest and a set of explicit assumptions regarding the intended meaning of the vocabulary [11]. Assumptions are usually represented in a logic language. As analysis of existing ontologies and ontology design methodologies shows, it is very difficult for a designer to develop accurate, transparent and consistent ontology [14, 3, 11]. In order to help experts in ontology design process, an initial domain ontology can be automatically or semi-automatically built from domain-specific knowledge captured in domain-specific texts, documents, database schema or data as demonstrated in [4, 12, 6]. This new approach is called ontology learning. The approach combines methods from Machine Learning (ML) and manual ontology description as well as NLP. There is another important issue concerning ontology development languages and methodologies: their reasoning capabilities. Ontology reasoning is important for checking concept satisfiability, subsumption (is-a), instances, and classification. According to recent study by Corcho and Gomez-Perez [1] most of well-known ontology specification languages do not have sound and complete inference engine implemented for them. Only F-logic [9] and OIL [8] from given list of languages supported sound and complete inference mechanism. Analysis and evaluation of ontology development methodologies [3] result in conclusion that high level design methods do not support well nontrivial ontology-based reasoning. Recent study of DAML ontology library by Tempich and Volz [14] shows that semantic web ontologies are designed in rather heterogeneous way and many semantic web ontologies fail in being usable for inference. This indicates that quality of ontology representation is not high enough. Consequently, most of current ontology languages and methodologies lack of learning support combined with reasoning mechanisms.

303

H.-M. Haav / An Ontology Learning and Reasoning Framework

This motivates us in current research to create an ontology learning and reasoning framework in order to provide assistance to ontology designer in ontology development process. The framework should guarantee accuracy, transparency and consistency of ontology representation by automatic or semiautomatic methods of ontology learning and reasoning. Novelty of our approach is in combining ontology learning and reasoning into one framework. In this paper, we restrict our research to domain ontologies that capture the essential concepts and relationships among those concepts in particular field of interest (application area). This work is an extension of our previous work on semi-automatic extraction and expression of domain-specific ontologies using Formal Concept Analysis (FCA) [6, 7]. For ontology extraction and learning, we use Formal Concept Analysis as a solid mathematical basis. For ontology reasoning purposes, we require that learned ontology should be automatically transformed (mapped) to some logic language. We use in this paper predicate logic, but we plan to use also description logics as a target ontology language in our further work. The rest of the paper is structured as follows. Section 2 gives an overview of Formal Concept Analysis. General ontology learning and reasoning framework is presented in section 3. Section 4 discusses related works and section 5 concludes the paper. 2. Learning Conceptual Structures by using FCA A reader is referred to [5] for detailed knowledge about Formal Concept Analysis (FCA). In the following we give a very basic introduction to the main principles of FCA by using examples. For example, a context K(O,C,R) of real estate domain can be as shown as in Table 1. The set of objects O is a collection of real estate domain specific texts (ads) about real estate items denoted by references like A1, A2, etc. in the table 1. The set of attributes C is the set of noun-phrases chunked from these texts by using NLP. If a text describing a real estate item contains certain noun-phrase, then the relationship R holds and we denote it by number 1 in the table below. Table 1. Real estate domain context Objects

A1 A2 A3 A4 A5 A6

Real estate 1 1 1 1 1 1

Family house 1 1

Attributes (noun-phrases) Country Summer Blockhouse house house

1 1

Skyscraper

1

1 1

1

1

A formal concept of the context K(O, C, R) is defined as pair (A, B), where A O, B C, A´= B and B´=A, where A´ is the set of attributes common to all the objects in A and B´ is the set of objects having the attributes in B. The extent of the concept (A, B) is A and its intent is B. For concepts (A1, B1) and (A2, B2) in the set S of all concepts of the context K(O,C,R) we have (A1, B1) d (A2, B2) A1 A2 B1 B2. The relation d is an order on S. It is shown in [5] that (S(K), d) is a complete lattice and this lattice is known as the concept lattice of the context K(O, C, R).

304


For example, the following concept lattice corresponds to the context presented in the table 1. Real estate A1, A2, A3, A4, A5, A6 Real estate Summerhouse

Real estate Blockhouse

Real estate, Family house

A2, A3

A4, A5

A2, A6

Real estate Country house

A1, A2 Real Estate, Summerhouse, Country house, Family house

Real estate, Blockhouse, Skyscraper

A4

A2

Real estate, Summerhouse, Blockhouse, Country house, Family house, Skyscraper

{}

Fig. 1. The concept lattice of a real estate domain

Each node in this lattice (denoted by black circle) is a formal concept. For example, one of the formal concepts of the context described in Table 1 is {A2, A6} u {Real estate, Summerhouse}, where the set {A2, A6} is the extent of the concept and the set {Real estate, Summerhouse} is its intent. Sub and super-concept relationships between the formal concepts are represented by edges in the Hasse diagram in Fig. 1. 3. An Ontology Learning and Reasoning Framework The proposed semi-automatic ontology learning and reasoning framework is based on the two concepts introduced in the following sections: concept lattice based ontology representation and its Horn logic model. 3.1 General Schema of the Framework The framework comprises the following steps: 1. Extracting formal context of a domain from domain-specific texts or data 2. Computing initial ontology as the concept lattice from the context using FCA and reduction procedures 3. Transforming initial ontology to a set of expressions in first order language 4. Extending initial ontology with additional rules and facts General schema of this method is drawn in the Fig. 2 as follows:


305

Domain specific texts or data

NLP based context extraction

Formal context

FCA and reduction

Concept lattice based ontology expression

Transformations

Set of rules describing initial ontology

More rules and facts

Inference

Complete set of rules and facts representing ontology

Fig. 2. General schema of ontology learning and reasoning framework

In the following we describe each step of the method in more detail. 3.2 Extraction of a Formal Context According to the method, first task is to produce a formal context K(O,C,R) for extractable domain-specific ontology. For our approach, objects of formal context can be textual descriptions of domain entities written in some natural language. We assume that those descriptions use domain specific vocabulary and are rather short. For example, suitable descriptions can be ads of real estate items in the real estate web catalogues, descriptions of products in product catalogues, technical descriptions of components. Attributes of an object for FCA are noun-phrases present in the domain-specific text describing a given domain-specific entity. Binary relationship R between descriptions (texts) of domain entities and noun phrases is discovered during the NLP process of text sources. A resulting set of noun phrases together with references to the domain-specific text sources are stored into the database table, which represents a context for the application domain in the form of binary relationship between descriptions of entities and noun phrases. A reader is referred to an example provided in the previous section, where the table 1 represented context of a sample domain. 3.3 Computing Initial Ontology As a result of FCA, concept lattice corresponding to the given context is computed. The next step is to reduce the lattice and to perform certain naming procedure as described in the following. There is redundant information in concept lattice. The two kinds of redundancy can be eliminated from concept lattice without loosing any information: redundant elements in formal concepts intents and redundant objects in formal concepts extents. Our reduction procedure has 2 steps: elimination of redundant elements from formal concepts intents and

306


elimination of lattice of extents. We call the resulting lattice LCR of reduction procedure as ontology lattice. Fig. 3 shows the lattice LCR of our example. Naming of concepts is done as follows: 1. A concept gets a unique name that is the name of the attribute(s) of formal concepts, which are left after reduction procedure. 2. After the previous naming procedure, there might be nodes that do not get names. In principle, the names for these nodes need to be provided by domain expert or ontology designer. It is possible automatically generate formal names (e.g. c1, c2…) for those nodes and then ask advice from human expert. The Fig. 3 depicts real estate domain ontology produced from concept lattice shown in Fig.1 using reduction and naming procedures. Real estate

Summerhouse

Blockhouse

Country house

Skyscraper

Family house

C1

C2

Fig. 3. Ontology lattice

One may notice that 2 nodes in the lattice are denoted by generated concept names C1 and C2. Ontology designer may analyse the lattice above and give the appropriate names to the concepts. It is also interesting that those nodes really denote new unknown (discovered) concepts, because the domain-specific texts did not include any noun-phrases for denoting these concepts. There exist several algorithms for FCA and construction of line diagrams of concept lattices. Excellent comparison of performance of those algorithms can be found in [10]. 3.4 Transforming Initial Ontology to Horn Logic Next step is to provide means for transforming ontology lattice into Horn logic. This process enables to produce logical expression of ontology lattice and specify intended semantics of the descriptions in first order logic. As our approach uses first order language, then it is possible to attach different ontology inference engines for practical applications by translating ontology lattice logical expression to any inference engine rule language. This was one of the reasons behind choosing Horn logic based rule language. In the following, we provide first order logic model for LCR. The model was first provided by us in [7]. 3.4.1 Language Constructs We use standard syntax for first order logic and define a simple rule language based on Horn clauses as follows.


307

An alphabet of the rule-language is defined as follows: 1. Set of constants N that consists of the set of concept names C, names of properties A, and special names any (lattice top, empty top is always True) and nil (lattice bottom, empty bottom is always False). 2. Set of variable names V. Uppercase letters denote the variables in V. 3. Set of predicate symbols P An interpretation for the rule-language is defined as a set of ground atoms constructed from predicate names in P and constants in N. 3.4.2 Mapping of Concept Lattice to Rule Language The mappings from lattice to rule language are defined as follows. Mapping concepts Concepts are represented using their names in ontology as constants in rule language. For example, house, summerhouse etc. If we like to refer to extents, then concepts can also be represented by predicates like house(X). At the moment, we are not interested in extents of concepts. Mapping taxonomic relationships The predicate subconcept(Concept1,Concept2) is used to denote that Concept1 is an immediate subconcept of Concept2. Subconcept predicates are automatically generated according to the given lattice LCR. Predicate isa is used to represent partial order relationship between concepts. Rules for lattice axioms As LCR is complete lattice, then the rules for lattice axioms are as follows: Reflexivity: isa(Concept, Concept) Transitivity: isa(Concept1, Concept2)msubconcept(Concept1, Concept2) isa(Concept1, Concept2) m isa(Concept1, Concept3), isa(Concept3, Concept2) Predicate subconcept denotes that Concept1 is an immediate subconcept of Concept2. Antisymmetry: equal(Concept1,Concept2) m isa(Concept1,Concept2), isa(Concept2,Concept1) Rules for lattice operations As LCR is a complete lattice, then for each set of concepts, there exists always a greatest lower bound (glb or greatest common subconcept) and a least upper bound (lub or a least common superconcept). Lattice meet is used to calculate glb and join is operation to calculate lub. We define these lattice meet operation using the following set of rules. Lattice meet operation c_subconcept(C, C1,…,Ck)misa(C, C1),…, isa(C, Ck) g_c_subconcept(C, C1,…,Ck) m c_subconcept(C, C1,…,Ck), c_subconcept(T, C1,…,Ck), isa(T, C) The predicate c_subconcept(C, C1,…,Ck) means that the concept C is a common subconcept of the set of concepts {C1,…,Ck} The predicate g_c_subconcept(C, C1,…,Ck) means that the concept C is the greatest common subconcept of the set of concepts {C1,…,Ck}. Symmetrically, we define predicates and rules for join operation, as provided in [7]. Logical model of a given lattice LCR can automatically be generated on the basis of mappings presented above. The following is a subset of a set of automatically generated ground subconcept atoms for ontology lattice shown in Fig. 3.

308


subconcept(familyhouse, real-estate) subconcept(c1, familyhouse) subconcept(c2, c1) 3.5 Extending Initial Ontology with Additional Rules and Facts As we have seen, taxonomic relationships between concepts can be automatically generated from a given lattice based ontology expression using logic-based formulation of concept lattice. In order to define non-taxonomic relationships the corresponding groups of predicates and rules are defined. Properties of concepts For defining properties of concepts, the following predicate can be used: hasproperty(Conceptname, Propertyname). Inheritance of properties Inheritance of properties can be represented by the following rule: hasproperty(C1, X)misa(C1,C2), hasproperty(C2, X) Ontological relationships Ontological relationships like part-of, related-to etc can be easily represented via predicates. The following predicates demonstrate opportunities adding other ontological relationships: partof(C1,C2) synonyms(C1,C2), etc. Non-taxonomic relationships give additional possibilities for ontology representation and reasoning about ontology. 3.6 Reasoning about Ontology Reasoning is important to ensure the quality of design of ontology. It can be used to find contradictory concepts, to derive implied relationships, etc. Inference rules for lattice axioms and operations can be used to decide taxonomic relationships between concepts as well as perform lattice operations. For example, to find the least common superconcept of the set of concepts {summerhouse; countryhouse}, we define the following query: l_c_superconcept(X, summerhouse, countryhouse). The answer is the concept real-estate. We may be interested in all the superconcepts of the concept skyscraper, for example. The query isa(skyscraper, X) gives the list of ground atoms as an answer. In our example, this is the list of the following atoms isa(skyscraper, real-estate) and isa(skyscraper, blockhouse). Inference about non-taxonomic relationships is made possible due to additional rules. For example, a designer may add the fact hasproperty(blockhouse,no_of_floors). Using inference, we may ask query hasproperty(X, no_of_floors) and receive an answer that also the fact hasproperty(skyscraper, no_of_floors) holds. 4. Remarks on Related Works There are several attempts on ontology learning and ontology extraction [16, 13, 2, 14, 15]. For example, in KRAFT [13], local ontology is extracted from shared ontology. In the system Infosleuth [16] ontologies are constructed semi-automatically from textual databases. They use ontology-learning approach, where human experts provide a set of initial words denoting concepts from high-level ontology. In [12] discovering non-


309

taxonomic relationships from texts using shallow text processing methods is presented. Their technique is integrated into ontology learning tool TextToOnto. The method presented in [15] uses manually built initial ontology to build conceptual hierarchy semiautomatically from texts. Resulting ontology is presented in Description Logics. Similarity of our approach and those discussed above is in using texts as descriptions of conceptualisation of domain and learning formal ontology from the given texts. To contrast our approach to the research mentioned above, we like to express that we combine automatic extraction of taxonomic relationships from domain-specific texts with semi-automatic expression of full ontology (including also non-taxonomic relationships) in first-order language for further reasoning and search. 5. Conclusion We have provided an ontology learning and reasoning framework in order to help ontology designer automatically to extract initial ontology from given set of domain specific texts, to map it automatically to rule language and to use rule language for adding non-taxonomic relationships to the ontology representation. Our main contribution resides in learning the ontology from the NL texts by using FCA and transforming it to Horn logic for reasoning. References [1] Corcho O.and Gómez-Pérez, A., A Roadmap to Ontology Specification Languages, In Lecture Notes on artificial Intelligence. 2nd International Conference on Knowledge Engineering and Knowledge Management. Springer-Verlag. October 2000, 80-96. [2] Decker S., Erdmann M., Fensel D., and Studer R., Ontobroker: Ontology Based Access to Distributed and SemiStructured Information, Meersman R. et al. (eds.): Semantic Issues in Multimedia Systems. Proceedings of DS-8. Kluwer Academic Publisher, Boston, 1999, 351-369. [3] Fernandez-Lopez M. and Gomez-Perez A., Overview and Analysis of Methodologies for Building Ontologies, The Knowledge Engineering Review, Vol. 17:2, 129-156, Cambridge University Press, 2002 [4] Gangemi A, et al, An Overview of the ONIONS Project, Data and Knowledge Engineering, 31, 1999 [5] Ganter, B. and Wille, R., Formal Concept Analysis, Mathematical Foundations, Springer, (1999) [6] Haav, H.-M.: Learning Ontologies for Domain-specific Information Retrieval, W. Abramowicz (Ed), KnowledgeBased Information Retrieval and Filtering from the Web, Kluwer Academic Publishers, (2003), 285-300 [7] Haav, H.-M., Combining FCA and a Logic Language for Ontology Representations, J. Barzdins (Ed), Databases and Information Systems, Sixth International Baltic Conference on Databases and Information Systems, June 6-9, Riga, 2004, Scientific papers of University of Latvia, Vol 672, (2004), 436-451 [8] Horrocks, I.,et al.: The Ontology Inference Layer OIL http://www.ontoknowledge.org/oil/oilhome.shtml [9] Kifer, M., Lausen, G., and Wu, J., Logical Foundations of Object-Oriented and Frame-Based Languages. Journal of ACM, 1995, 42(4), pp 741-843 [10] Kuznetsov, S. O., and Obiedkov, S. A., Comparing performance of Algorithms for Generating Concept Lattices, Journal of Experimental & Theoretical Artificial Intelligence vol 14(1), 2002, Nos. 2-3, pp 189-216 [11] Maedche, A.: Ontology Learning for the Semantic Web, Kluwer Academic Publishers, (2002) [12] Maedche, A. and Staab, S.: Discovering Conceptual Relations from Text. ECAI 2000. Proceedings of the 14th European Conference on Artificial Intelligence, IOS Press, Amsterdam, (2000) [13] Preece, A. D., Hui, K-J., Gray, W. A., et al., The Kraft arhitecture for knowledge fusion and transformation, In Proc. th of 19 SGES Int. Conference, Springer 1999 [14] Tempich, C. and Volz, R.: Towards a benchmark for Semantic Web reasoners-an analysis of the DAML ontology library, Sure Y (ed) Proceedings of Workshop of Evaluation of Ontology-based Tools (EON 2003) at 2nd Int. Semantic Web Conference (ISWC 2003), USA, (2003) [15] Todirascu, A. and Beuvron, F., Using Description Logics for Ontology Extraction, In: Proceedings of Conference TALN 2000, Lausanne, October 16-18, 2000 [16] Woelk, D. and Tomlinson, C., The Infosleuht project: Intelligent Search Management via Semantic Agents. In Second World Wide Web Conference ’94:Mosaic and the Web, 1994


310

Strong Relevant Logic as the Universal Basis of Various Applied Logics for Knowledge Representation and Reasoning Jingde CHENG Department of Information and Computer Sciences Saitama University, Saitama, 338-8570, Japan [email protected]

Abstract. In many applications in computer science and artificial intelligence, in order to represent, specify, verify, and reason about various objects and relationships among them, we often need a right fundamental logic system to provide us with a criterion of logical validity for reasoning as well as a formal representation and specification language. Although different applications may require different logic systems, the fundamental logics must be able to underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete reasoning, and paraconsistent reasoning. Based on our experiences, this paper shows that strong relevant logic can be used as the universal basis to construct various applied logics to satisfy the requirements. The paper discusses why any of the classical mathematical logic, its various classical conservative extensions, and its non-classical alternatives is not a suitable candidate for the universal basis to construct various applied logics, shows that strong relevant logic is a more hopeful candidate for the purpose, and presents our experiences on constructions of temporal relevant logics, deontic relevant logics, spatial relevant logics, and spatial-temporal relevant logics.

1. Introduction In many applications in computer science and artificial intelligence, in order to represent, specify, verify, and reason about various objects and relationships among them, we often need a right fundamental logic system to provide us with a criterion of logical validity for reasoning as well as a formal representation and specification language. The question, “Which is the right logic?” invites the immediate counter-question “Right for what?” Only if we certainly know what we need, we can make a good choice. It is obvious that different applications may require different characteristics of logic. However, can we have a set of essential requirements as the universal core requirements for various applications and a logic system satisfying the essential requirements as the universal core logic such that we can construct various applied logics by extending the logic? The present author considers that we should consider the following three essential requirements for the universal core logic. First, as a general logical criterion for the validity of reasoning as well as proving, the logic must be able to underlie relevant reasoning as well as truth-preserving reasoning in the sense of conditional, i.e., for any reasoning based on the logic to be valid, if its premises are true in the sense of conditional,

J. Cheng / Strong Relevant Logic as the Universal Basis of Various Applied Logics

311

then its conclusion must be relevant to the premises and must be true in the sense of conditional. Second, the logic must be able to underlie ampliative reasoning in the sense that the truth of conclusion of the reasoning should be recognized after the completion of the reasoning process but not be invoked in deciding the truth of premises of the reasoning. From the viewpoint to regard reasoning as the process of drawing new conclusions from given premises, any meaningful reasoning must be ampliative but not circular and/or tautological. Third, the logic must be able to underlie paracomplete reasoning and paraconsistent reasoning. In particular, the so-called principle of Explosion that everything follows from a contradiction cannot be accepted by the logic as a valid principle. In general, our knowledge about a domain as well as a scientific discipline may be incomplete and sometime even inconsistent in many ways, i.e., it gives us no evidence for deciding the truth of either a proposition or its negation, and it directly or indirectly includes some contradictions. Therefore, reasoning with incomplete and inconsistent knowledge is the rule rather than the exception in our everyday lives and almost all scientific disciplines. Based on our experiences, this paper shows that strong relevant logic can be used as the universal basis to construct various applied logics to satisfy the three essential requirements. The paper discusses why any of the classical mathematical logic, its various classical conservative extensions, and its non-classical alternatives is not a suitable candidate for the universal basis to construct various applied logics, shows that strong relevant logic is a more hopeful candidate for the purpose, and presents our experiences on constructions of temporal relevant logics, deontic relevant logics, spatial relevant logics, and spatial-temporal relevant logics. 2. Basic Notions Reasoning is the process of drawing new conclusions from given premises, which are already known facts or previously assumed hypotheses to provide some evidence for the conclusions (Note that how to define the notion of ‘new’ formally and satisfactorily is still a difficult open problem until now). Therefore, reasoning is intrinsically ampliative, i.e., it has the function of enlarging or extending some things, or adding to what is already known or assumed. In general, a reasoning consists of a number of arguments in some order. An argument is a set of statements (or declarative sentences) of which one statement is intended as the conclusion, and one or more statements, called “premises,” are intended to provide some evidence for the conclusion. An argument is a conclusion standing in relation to its supporting evidence. In an argument, a claim is being made that there is some sort of evidential relation between its premises and its conclusion: the conclusion is supposed to follow from the premises, or equivalently, the premises are supposed to entail the conclusion. Therefore, the correctness of an argument is a matter of the connection between its premises and its conclusion, and concerns the strength of the relation between them (Note that the correctness of an argument depends neither on whether the premises are really true or not, nor on whether the conclusion is really true or not). Thus, there are some fundamental questions: What is the criterion by which one can decide whether the conclusion of an argument or a reasoning really does follow from its premises or not? Is there the only one criterion, or are there many criteria? If there are many criteria, what are the intrinsic differences between them? It is logic that deals with the validity of argument and reasoning in a general theory. A logically valid reasoning is a reasoning such that its arguments are justified based on some logical validity criterion provided by a logic system in order to obtain correct conclusions (Note that here the term ‘correct’ does not necessarily mean ‘true’). Today,

312


there are so many different logic systems motivated by various philosophical considerations. As a result, a reasoning may be valid on one logical validity criterion but invalid on another. For example, the classical account of validity, which is one of fundamental principles and assumptions underlying classical mathematical logic and its various conservative extensions, is defined in terms of truth-preservation (in some certain sense of truth) as: an argument is valid if and only if it is impossible for all its premises to be true while its conclusion is false. Therefore, a classically valid reasoning must be truth-preserving. On the other hand, for any correct argument in scientific reasoning as well as our everyday reasoning, its premises must somehow be relevant to its conclusion, and vice versa. The relevant account of validity is defined in terms of relevance as: for an argument to be valid there must be some connection of meaning, i.e., some relevance, between its premises and its conclusion. Obviously, the relevance between the premises and conclusion of an argument is not accounted for by the classical logical validity criterion, and therefore, a classically valid reasoning is not necessarily relevant. Proving is the process of finding a justification for an explicitly specified statement from given premises, which are already known facts or previously assumed hypotheses to provide some evidence for the specified statement. A proof is a description of a found justification. A logically valid proving is a proving such that it is justified based on some logical validity criterion provided by a logic system in order to obtain a correct proof. The most intrinsic difference between reasoning and proving is that the former is intrinsically prescriptive and predictive while the latter is intrinsically descriptive and non-predictive. The purpose of reasoning is to find some new conclusion previously unknown or unrecognized, while the purpose of proving is to find a justification for some specified statement previously given. Proving has an explicitly given target as its goal while reasoning does not. Unfortunately, until now, many studies in Computer Science and Artificial Intelligence disciplines still confuse proving with reasoning. Logic deals with what entails what or what follows from what, and aims at determining which are the correct conclusions of a given set of premises, i.e., to determine which arguments are valid. Therefore, the most essential and central concept in logic is the logical consequence relation that relates a given set of premises to those conclusions, which validly follow from the premises. To define a logical consequence relation is nothing else but to provide a logical validity criterion by which one can decide whether the conclusion of an argument or a reasoning really does follow from its premises or not. Moreover, to answer the question what is the correct conclusion of given premises, we have to answer the question: correct for what? Based on different philosophical motivations, one can define various logical consequence relations and therefore establish various logic systems. In logic, a sentence in the form of ‘if ... then ...’ is usually called a conditional proposition or simply conditional which states that there exists a relation of sufficient condition between the ‘if’ part and the ‘then’ part of the sentence. In general, a conditional must concern two parts which are connected by the connective ‘if ... then ...’ and called the antecedent and the consequent of that conditional, respectively. The truth of a conditional depends not only on the truth of its antecedent and consequent but also, and more essentially, on a necessarily relevant and conditional relation between them. The notion of conditional plays the most essential role in reasoning because any reasoning form must invoke it, and therefore, it is historically always the most important subject studied in logic and is regarded as the heart of logic [1]. When we study and use logic, the notion of conditional may appear in both the object logic (i.e., the logic we are studying) and the meta-logic (i.e., the logic we are using to study the object logic). In the object logic, there usually is a connective in its formal language to represent the notion of conditional, and the notion of conditional, usually


313

represented by a meta-linguistic symbol, is also used for representing a logical consequence relation in its proof theory or model theory. On the other hand, in the meta-logic, the notion of conditional, usually in the form of natural language, is used for defining various meta-notions and describing various meta-theorems about the object logic. From the viewpoint of object logic, there are two classes of conditionals. One class is empirical conditionals and the other class is logical conditionals. For a logic, a conditional is called an empirical conditional of the logic if its truth-value, in the sense of that logic, depends on the contents of its antecedent and consequent and therefore cannot be determined only by its abstract form (i.e., from the viewpoint of that logic, the relevant relation between the antecedent and the consequent of that conditional is regarded to be empirical); a conditional is called a logical conditional of the logic if its truth-value, in the sense of that logic, depends only on its abstract form but not on the contents of its antecedent and consequent, and therefore, it is considered to be universally true or false (i.e., from the viewpoint of that logic, the relevant relation between the antecedent and the consequent of that conditional is regarded to be logical). A logical conditional that is considered to be universally true, in the sense of that logic, is also called an entailment of that logic. Indeed, the most intrinsic difference between various different logic systems is to regard what class of conditionals as entailments, as Diaz pointed out: “The problem in modern logic can best be put as follows: can we give an explanation of those conditionals that represent an entailment relation?” [14] 3. Strong Relevant Logics As we have mentioned, the universal core logic must be able to underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete and paraconsistent reasoning. Classical mathematical logic (CML for short) was established in order to provide formal languages for describing the structures with which mathematicians work, and the methods of proof available to them; its principal aim is a precise and adequate understanding of the notion of mathematical proof. CML was established based on a number of fundamental assumptions. Among them, the most characteristic one is the classical account of validity that is the logical validity criterion of CML by which one can decide whether the conclusion of an argument or a reasoning really does follow from its premises or not in the framework of CML. However, since the relevance between the premises and conclusion of an argument is not accounted for by the classical validity criterion, a reasoning based on CML is not necessarily relevant. On the other hand, in CML the notion of conditional, which is intrinsically intensional but not truth-functional, is represented by the notion of material implication, which is intrinsically an extensional truth-function. This leads to the problem of ‘implicational paradoxes’ [1, 2, 15, 19-21, 23]. CML cannot satisfy any of the above three essential requirements because of the following facts: a reasoning based on CML is not necessarily relevant; the classical truthpreserving property of a reasoning based on CML is meaningless in the sense of conditional; a reasoning based on CML must be circular and/or tautological but not ampliative; reasoning under inconsistency is impossible within the framework of CML [6]. These facts are also true to those classical conservative extensions or non-classical alternatives of CML where the classical account of validity is adopted as the logical validity criterion and the notion of conditional is directly or indirectly represented by the material implication [6]. Traditional relevant (or relevance) logics ware constructed during the 1950s in order to find a mathematically satisfactory way of grasping the elusive notion of relevance of

314


antecedent to consequent in conditionals, and to obtain a notion of implication which is free from the so-called ‘paradoxes’ of material and strict implication [1, 2, 15, 19-21, 23]. Some major traditional relevant logic systems are ‘system E of entailment’, ‘system R of relevant implication’, and ‘system T of ticket entailment’. A major characteristic of the relevant logics is that they have a primitive intensional connective to represent the notion of conditional (entailment) and their logical theorems include no implicational paradoxes. The underlying principle of the relevant logics is the relevance principle, i.e., for any entailment provable in E, R, or T, its antecedent and consequent must share a propositional variable. Variable-sharing is a formal notion designed to reflect the idea that there be a meaning-connection between the antecedent and consequent of an entailment. It is this relevance principle that excludes those implicational paradoxes from logical axioms or theorems of relevant logics. Also, since the notion of entailment is represented in the relevant logics by a primitive intensional connective but not an extensional truth-function, a reasoning based on the relevant logics is ampliative but not circular and/or tautological. Moreover, because the relevant logics reject the principle of Explosion, they can certainly underlie paraconsistent reasoning. In order to establish a satisfactory logic calculus of conditional to underlie relevant reasoning, the present author has proposed some strong relevant (or relevance) logics, named Rc, Ec, and Tc [6]. The logics require that the premises of an argument represented by a conditional include no unnecessary and needless conjuncts and the conclusion of that argument includes no unnecessary and needless disjuncts. As a modification of traditional relevant logics R, E, and T, strong relevant logics Rc, Ec, and Tc rejects all conjunction-implicational paradoxes and disjunction-implicational paradoxes in R, E, and T, respectively. What underlies the strong relevant logics is the strong relevance principle: If A is a theorem of Rc, Ec, or Tc, then every propositional variable in A occurs at least once as an antecedent part and at least once as a consequent part. Since the strong relevant logics are free of not only implicational paradoxes but also conjunctionimplicational and disjunction-implicational paradoxes, in the framework of strong relevant logics, if a reasoning is valid, then both the relevance between its premises and its conclusion and the validity of its conclusion in the sense of conditional can be guaranteed in a certain sense of strong relevance. The logical connectives, axiom schemata, and inference rules of relevant logics are as follows: Primitive logical connectives: (entailment), ¬ (negation), (extensional conjunction). Defined logical connectives: (intensional conjunction, AB =df ¬(A¬B)), (intensional disjunction, AB =df ¬AB), (intensional equivalence, AB =df (AB)(BA)), (extensional disjunction, A B =df ¬(¬A¬B)), (material implication, AB =df ¬(A¬B) or AB =df ¬A B), (extensional equivalence, AB =df (AB)(BA)). Axiom schemata: E1: AA, E2: (AB)((CA)(CB)), E2: (AB) ((BC)(AC)), E3: (A(AB))(AB), E3: (A(BC))((AB)(AC)), E3: (AB)((A(BC))(AC)), E4: (A((BC)D))((BC)(AD)), E4: (AB)(((AB)C)C), E4: ((AA)B)B, E4: (AB)((BC) (((AC)D)D)), E5: (A(BC))(B(AC)), E5: A((AB)B), N1: (A(¬A))(¬A), N2: (A(¬B))(B(¬A)), N3: (¬(¬A))A, C1: (AB)A, C2: (AB)B, C3: ((AB)(AC))(A(BC)), C4: (LALB)L(AB), where LA =df (AA)A, D1: A(A B), D2: B(A B), D3: ((AC)(BC))((A B)C), DCD: (A(B C))((AB) C), C5: (AA)A, C6: (AB)(BA), C7: ((AB)(BC)) (AC), C8: (A(AB))B, C9: ¬(A¬A), C10: A(B(AB)), IQ1: x(AB)


315

(xAxB), IQ2: (xAxB)x(AB), IQ3: xAA[t/x] (if x may appear free in A and t is free for x in A, i.e., free variables of t do not occur bound in A), IQ4: x(AB) (AxB) (if x does not occur free in A), IQ5: x1 ... xn(((AA)B)B) (n0). Inference rules: E: from A and AB to infer B (Modus Ponens), I: from A and B to infer AB (Adjunction), I: if A is an axiom, so is xA (Generalization of axioms). Thus, various relevant logic systems may now defined as follows, where we use ‘A | B’ to denote any choice of one from two axiom schemata A and B: T =df {E1, E2, E2, E3 | E3} + E, E =df {E1, E2 | E2, E3 | E3, E4 | E4} + E, E =df {E2, E3, E4} + E, E =df {E1, E3, E4} + E, R =df {E1, E2 | E2, E3 | E3, E5 | E5} + E, T,¬ =df T + {N1, N2, N3}, E,¬ =df E + {N1, N2, N3}, R,¬ =df R + {N2, N3}, T =df T,¬ + {C1~C3, D1~D3, DCD} + I, E =df E,¬ + {C1~C4, D1~D3, DCD} + I, R =df R,¬ + {C1~C3, D1~D3, DCD} + I, Tc =df T,¬ + {C3, C5~C10}, Ec =df E,¬ + {C3~C10}, Rc =df R,¬ + {C3, C5~C10}, TQ =df T + {IQ1~IQ5} + I, EQ =df E + {IQ1~IQ5} + I, RQ =df R + {IQ1~IQ5} + I, TcQ =df Tc + {IQ1~IQ5} + I, EcQ =df Ec + {IQ1~IQ5} + I, RcQ =df Rc + {IQ1~IQ5} + I. Here, T, E, and R are the purely implicational fragments of T, E, and R, respectively, and the relationship between E and R is known as R = E + ALA; T,¬, E,¬, and R,¬ are the implication-negation fragments of T, E, and R, respectively; Tc, Ec, Rc, TcQ, EcQ, and RcQ are strong relevant (relevance) logics. The strong relevant logics can underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete and paraconsistent reasoning. Therefore, they can satisfy all three essential requirements for the universal core logic.

4. Temporal Relevant Logics Time is one of the most fundamental notions in our cognition about the real world. The ability of representing and reasoning about temporal knowledge conceptually is one of the most intrinsic characteristics of human intelligence. Therefore, no account of reasoning can properly be considered to be complete if it does not say something about how we reason about change. Classical temporal logics was established in order to represent and reason about notions, relations, and properties of time-related entities within a logical framework, and therefore to underlie temporal reasoning, i.e., reasoning about those propositions and formulas whose truth-values may depend on time. As a conservative extension of CML, they have remarkably expanded the uses of logic to reasoning about human (and hence computer) time-related activities [5, 16, 25, 27, 28]. However, because any of classical temporal logics is a classical conservative extension of CML in the sense that it is based on the classical account of validity and it represents the notion of conditional directly or indirectly by the material implication, no classical temporal logic can satisfy the three essential requirements. The present author has proposed temporal relevant logics which are obtained by introducing the following temporal operators and related axiom schemata and inference rules into strong relevant logics [7]. Temporal operators: G (future-tense always or henceforth operator, GA means ‘it will always be the case in the future from now that A’), H (past-tense always operator, HA means ‘it has always been the case in the past up to now that A’), F (future-tense sometime or eventually operator, FA means ‘it will be the case at least once in the future from now that A’), P (past-tense sometime operator, PA means ‘it has been the case at least once in the past up to now that A’). Note that these temporal operators are not

316


independent and can be defined as follows: GA =df ¬F¬A, HA =df ¬P¬A, FA =df ¬G¬A, PA =df ¬H¬A. Axiom schemata: T1: G(AB)(GAGB), T2: H(AB)(HAHB), T3: AG(PA), T4: AH(FA), T5: GAG(GA), T6: (FAFB)F(AFB)F(AB)F(FAB), T7: (PAPB)P(APB)P(AB)P(PAB), T8: GAFA, T9: HAPA, T10: FAF(FA), T11: (AHA)F(HA), T12: (AGA)P(GA). Inference rules: TG : from A to infer GA and HA (Temporal Generalization) We can obtain minimal or weakest propositional temporal relevant logics as follows: T0Tc = Tc + {T1~T4} + TG, T0Ec = Ec + {T1~T4} + TG, T0Rc = Rc + {T1~T4} + TG. Note that the minimal or weakest temporal classical logic Kt = all axiom schemata for CML + E + {T1~T4} + TG. Other characteristic axioms such as T5~T12 that correspond to various assumptions about time can be added to T0Tc, T0Ec, and T0Rc respectively to obtain various propositional temporal relevant logics. Various predicate temporal relevant logics then can be obtained by adding axiom schemata IQ1~IQ5 and inference rule I into the propositional temporal relevant logics. For examples, minimal or weakest predicate temporal relevant logics are as follows: T0TcQ = T0Tc + {IQ1~IQ5} + I, T0EcQ = T0Ec + {IQ1~IQ5} + I, and T0RcQ = T0Rc + {IQ1~IQ5} + I. The temporal relevant logics can underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete and paraconsistent reasoning, and temporal reasoning (in particular, anticipatory reasoning). Therefore, they can be used as the fundamental logic to underlie reasoning about dynamics of a knowledge-based system where not only truth-vules of propositions and/or formulas but also relevant relationships between them may depend on time. Probably, the most important application of temporal relevant logics is to underlie anticipatory reasoning in prediction [7, 17, 24].

5. Deontic Relevant Logics In general, the actual behavior (as it is) of a computing system in its running is somewhat different from the ideal (or normative) behavior (as it should be) of the system which is specified by requirements of the system. Therefore, to distinguish between ideal behavior and actual behavior of a computing system is important to defining what behavior is illegal and specifying what should be done if such illegal but possible behavior occurs. On the other hand, information security and assurance problems are intrinsically concerning the following questions: Who is authorized (permitted, or has right) to access and modify what information by what way from what machine or site in what time or state, and who does not be authorized to do so. Therefore, information security and assurance engineering is intrinsically an engineering discipline to deal with normative requirements and their implementation and verification (validation) techniques for secure information systems. Any approach to specifying and reasoning about information security and assurance cannot be considered to be adequate if it does not deal with normative notions explicitly, soundly and completely. Deontic logic is a branch of philosophical logic to deal with normative notions such as obligation (ought), permission (permitted), and prohibition (may not) for underlying normative reasoning [3, 4, 18, 22]. Informally, it can also be considered as a logic to reason about ideal versus actual states or behaviour. It seems to be an adequate tool to specify and verify information security and assurance requirements. However, classical deontic logic has the problem of deontic paradoxes as well as the problem of implicational paradoxes. No classical deontic logic can satisfy the three essential requirements. Tagawa and the present author have proposed deontic relevant logics based on strong relevant logics to remove those deontic paradoxes from classical deontic logic [26]. The


317

deontic relevant logics are obtained by introducing the following deontic operators and related axiom schemata and inference rules into strong relevant logics. Deontic operators: O (obligation operator, OA means “It is obligatory that A”), P (permission operator, PA =df ¬O(¬A), PA means “It is permitted that A”). Axiom schemata: DR1: O(AB)(OAOB), DR2: OAPA, DR3: ¬(OAO¬A), DR4: O(AB)(OAOB), DR5: P(AB)(PAPB). Inference rules: O-necessitation: “if A is a logical theorem, then so is OA” (Deontic Generalization) We can obtain propositional deontic relevant logics as follows: DTc = Tc + {DR1~DR5} + O-necessitation, DEc = Ec + {DR1~DR5} + O-necessitation, DRc = Rc + {DR1~DR5} + O-necessitation. Various predicate deontic relevant logics then can be obtained by adding axiom schemata IQ1~IQ5 and inference rule I into the propositional deontic relevant logics. The deontic relevant logics provide a formal language with normative notions which can be used as a formal specification language for specifying information security and assurance requirements. The deontic relevant logics also provide a sound logical basis for proving systems, and therefore it can be used as a analysis and verification tool for verifying information security and assurance requirements. The deontic relevant logics also provide a sound logical basis for reasoning systems, and therefore it can underlie reasoning about ideal versus actual, present versus potential states or behaviour of a computing system. Finally, the deontic relevant logics provide a unified, explicit, logical criterion for all the three parties of users, developers, and managers of information systems when they are involved in some legal actions or cases.

6. Spatial Relevant Logics Space is another one of the most fundamental notions in our cognition about the real world. The ability of representing and reasoning about spatial knowledge conceptually is another one of the most intrinsic characteristics of human intelligence. Classical spatial logics was proposed in order to deal with geometric and/or topological entities, notions, relations, and properties, and therefore to underlie spatial reasoning, i.e., reasoning about those propositions and formulas whose truth-values may depend on a location [10-13, 16, 25]. However, these existing spatial logics are classical conservative extensions of CML in the sense that they are based on the classical account of validity and they represent the notion of conditional directly or indirectly by the material implication. Therefore, these spatial logics cannot satisfy the three essential requirements for the fundamental logics. We have propose a new family of relevant logic systems, named spatial relevant logic [9]. The logics are obtained by introducing region connection predicates and axiom schemata of RCC [10-13], point position predicates and axiom schemata, and point adjacency predicates and axiom schemata into predicate strong relevant logics. Let {r1, r2, r3, …} be a countably infinite set of individual variables, called region variables. Atomic formulas of the form C(r1, r2) are read as ‘region r1 connects with region r2.’ Let {p1, p2, p3, …} be a countably infinite set of individual variables, called point variables. Atomic formulas of the form I(p1, r1) are read as ‘point p1 is included in region r1.’ Atomic formulas of the form Id(p1, p2) are read as ‘point p1 is identical with p2.’ Atomic formulas of the form Arc(p1, p2) are read as ‘points p1, p2 are adjacent such that there is an arc from point p1 to point p2, or more simply, points p1 is adjacent to point p2.’ Note that an arc has a direction. Atomic formulas of the form Path(p1, p2) are read as ‘there is a directed path from point p1 to point p2.’ Here, C(r1, r2), I(p1, r1), Id(p1, p2),

318


Arc(p1, p2), and Path(p1, p2) are primitive binary predicates to represent relationships between geometric objects. Note that here we use a many-sorted language. The logical connectives, region connection predicates, point position predicates, point adjacency predicates, axiom schemata, and inference rules are as follows: Primitive binary predicate: C (connection, C(r1, r2) means ‘r1 connects with r2’), I (inclusion, I(p1, r1) means ‘p1 is included in r1’), Id (the same point, Id(p1, p2) means ‘p1 is identical with p2’), Arc (arc, Arc(p1, p2) means ‘p1 is adjacent to p2’), Path (path, Path(p1, p2) means ‘there is a directed path from p1 to p2’). Defined binary predicates: DC(r1, r2) =df ¬C(r1, r2) (DC(r1, r2) means ‘r1 is disconnected from r2’), Pa(r1, r2) =df r3(C(r3, r1)C(r3, r2)) (Pa(r1, r2) means ‘r1 is a part of r2’), PrPa(r1, r2) =df Pa(r1, r2)(¬Pa(r2, r1)) (PrPa(r1, r2) means ‘r1 is a proper part of r2’), EQ(r1, r2) =df Pa(r1, r2)Pa(r2, r1) (EQ(r1, r2) means ‘r1 is identical with r2’), O(r1, r2) =df r3(Pa(r3, r1)Pa(r3, r2)) (O(r1, r2) means ‘r1 overlaps r2’), DR(r1, r2) =df ¬O(r1, r2) (DR(r1, r2) means ‘r1 is discrete from r2’), PaO(r1, r2) =df O(r1, r2)(¬Pa(r1, r2))(¬Pa(r2, r1)) (PaO(r1, r2) means ‘r1 partially overlaps r2’), EC(r1, r2) =df C(r1, r2)(¬O(r1, r2)) (EC(r1, r2) means ‘r1 is externally connected to r2’), TPrPa(r1, r2) =df PrPa(r1, r2)r3(EC(r3, r1)EC(r3, r2)) (TPrPa(r1, r2) means ‘r1 is a tangential proper part of r2’), NTPrPa(r1, r2) =df PrPa(r1, r2)(¬r3(EC(r3, r1)EC(r3, r2))) (NTPrPa(r1, r2) means ‘r1 is a nontangential proper part of r2’). Axiom schemata: RCC1: r1r2(C(r1, r2)C(r2, r1)), RCC2: r1(C(r1, r1)), PRCC1: p1r1r2((I(p1, r1)DC(r1, r2))¬I(p1, r2)), PRCC2: p1r1r2((I(p1, r1)Pa(r1, r2)) I(p1, r2)), PRCC3: r1r2(O(r1, r2)p1(I(p1, r1)I(p1, r2))), PRCC4: r1r2(PaO(r1, r2) p1(I(p1, r1)I(p1, r2))p2(I(p2, r1)¬I(p2, r2))p3(¬I(p3, r1)I(p3, r2))), PRCC5: r1r2(EC(r1, r2)p1(I(p1, r1)I(p1, r2)p2(¬Id(p2, p1)¬I(p2, r1)¬I(p2, r2))), PRCC6: p1r1r2((I(p1, r1)TPrPa(r1, r2))I(p1, r2)), PRCC7: p1r1r2((I(p1, r1) NTPrPa(r1, r2))I(p1, r2)), APC1: p1p2(Arc(p1, p2)Path(p1, p2)), APC2: p1p2p3((Path(p1, p2)Path(p2, p3))Path(p1, p3)). We can now obtain some spatial relevant logics as follows: RTcQ =df TcQ + {RCC1, RCC2, PRCC1~PRCC7, APC1, APC2}, REcQ =df EcQ + {RCC1, RCC2, PRCC1~PRCC7, APC1, APC2}, RRcQ =df RcQ + {RCC1, RCC2, PRCC1~PRCC7, APC1, APC2} The spatial relevant logics can underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete and paraconsistent reasoning, and spatial reasoning. Therefore, they can be used to underlie reasoning about geometric and/or topological entities, notions, relations, and their properties. 7. Spatio-temporal Relevant Logics To represent, specify, verify, and reason about spatial objects and relationships among them that may change over time, e.g., mobile agents, a right fundamental logic to underlie both spatial reasoning and temporal reasoning is indispensable [16, 25]. The present author has proposed a new family of relevant logic systems, named spatiotemporal relevant logic [8]. The logics are obtained by introducing region connection predicates and axiom schemata of RCC, point position predicates and axiom schemata, and point adjacency predicates and axiom schemata into various predicate temporal relevant logics. Therefore, they are conservative extensions of strong relevant logics, temporal relevant logics, and spatial relevant logics. For examples: ST0TcQ = T0TcQ + {RCC1, RCC2, PRCC1~PRCC7, APC1, APC2}, ST0EcQ = T0EcQ + {RCC1, RCC2, PRCC1~PRCC7, APC1, APC2}, ST0RcQ = T0RcQ + {RCC1, RCC2, PRCC1~PRCC7,


319

APC1, APC2}. The spatio-temporal relevant logics have the following characteristics. First, as conservative extensions of strong relevant logics satisfying the strong relevance principle, the logics underlie relevant reasoning as well as truth-preserving reasoning in the sense of conditional, ampliative reasoning, paracomplete reasoning, and paraconsistent reasoning. Second, the logics underlie spatial reasoning and temporal reasoning. We can select any one of them according our purpose in an application from various aspects of relevance, temporality, and spatiality. The spatio-temporal relevant logics themselves have various possible applications. For example, because the logics can underlie relevant, truth-preserving, ampliative, paracomplete, paraconsistent, spatial, and temporal reasoning, they provide us with criteria of logical validity for reasoning about behavior of mobile agents with incomplete or even inconsistent knowledge acting concurrently in spatial regions changing over time. On the other hand, the spatio-temporal relevant logics provide us with a foundation for constructing more powerful applied logics in order to deal with various notions and issues in application systems. For examples, we can add epistemic operators and related axiom schemata into the logics in order to reason about interaction among agents as well as epistemic states of agents; we can also add deontic operators and related axiom schemata into the logics in order to reason about information security and assurance in mobile multiagent systems. 8. Concluding Remarks We have shown that strong relevant logic can be used as the universal basis to construct various applied logics to underlie truth-preserving and relevant reasoning in the sense of conditional, ampliative reasoning, paracomplete and paraconsistent reasoning, and other special reasoning. We presented our experiences on constructions of temporal relevant logics, deontic relevant logics, spatial relevant logics, and spatial-temporal relevant logics. The work presented in this paper is our first step for establishing a fundamental logic system as a universal basis to construct various applied logics. There are many challenging theoretical and technical problems that have to be solved in order to apply the strong relevant logics and their various extensions to practices in the real world. A challenging open problem, named ‘NRT problem’ by the present author, is as follows: Although it is necessary to deal with all notions of normativeness, relevance, and temporality explicitly and soundly in specifying, verifying, and reasoning about information systems, until now there is no formal logic system can be used as a unified logical basis for the purpose. Can we establish a unified formal logic system, which takes all notions of normativeness, relevance, and temporality into account, to underlie specifying, verifying, and reasoning about information systems satisfactorily? References [1] [2] [3] [4] [5]

A. R. Anderson and N. D. Belnap Jr., “Entailment: The Logic of Relevance and Necessity,” Vol. I, Princeton University Press, 1975. A. R. Anderson, N. D. Belnap Jr., and J. M. Dunn, “Entailment: The Logic of Relevance and Necessity,” Vol. II, Princeton University Press, 1992. L. Aqvist, “Introduction to Deontic Logic and the Theory of Normative Systems,” Bibliopolis, 1987. L. Aqvist, “Deontic Logic,” in: D. Gabbay and F. Guenthner (Eds.), “Handbook of Philosophical Logic, 2nd Edition,” Vol. 8, pp. 147-264, Kluwer Academic, 2002. J. P. Burgess, “Basic Tense Logic,” in D. Gabbay and F. Guenthner (Eds.), “Handbook of Philosophical Logic, 2nd Edition,” Vol. 7, pp. 1-42, Kluwer Academic, 2002.

320


[6] J. Cheng, “A Strong Relevant Logic Model of Epistemic Processes in Scientific Discovery,” in E. Kawaguchi, H. Kangassalo, H. Jaakkola, and I. A. Hamid (Eds.), “Information Modelling and Knowledge Bases XI,” pp. 136-159, IOS Press, 2000. [7] J. Cheng, “Temporal Relevant Logic as the Logical Basis of Anticipatory Reasoning-Reacting Systems,” in D. M. Dubois (Ed.), “Computing Anticipatory Systems: CASYS 2003 - Sixth International Conference, Liege, Belgium, 11-16 August 2003,” AIP Conference Proceedings, Vol. 718, pp. 362-375, American Institute of Physics, 2004. [8] J. Cheng, “Spatio-temporal Relevant Logic as the Logical Basis for Specifying, Verifying, and Reasoning about Mobile Multi-agent Systems,” in S. Wang, K. Tanaka, et al. (Eds.), “Conceptual Modeling for Advanced Application Domains, ER 2004 Workshops CoMoGIS, CoMWIM, ECDM, CoMoA, DGOV, and eCOMO, Shanghai, China, November 2004, Proceedings,” Lecture Notes in Computer Science, Vol. 3289, pp. 470-483, Springer-Verlag, 2004. [9] J. Cheng and Y. Goto, “Representing and Reasoning about Spatial Knowledge Based on Spatial Relevant Logic,” in S. Wang, K. Tanaka, et al. (Eds.), “Conceptual Modeling for Advanced Application Domains, ER 2004 Workshops CoMoGIS, CoMWIM, ECDM, CoMoA, DGOV, and eCOMO, Shanghai, China, November 2004, Proceedings,” Lecture Notes in Computer Science, Vol. 3289, pp. 114-126, SpringerVerlag, 2004. [10] A. G. Cohn, B. Bennett, J. Gooday, and N. M. Gotts, “Qualitative Spatial Representation and Reasoning with the Region Connection Calculus,” Geoinformatica, Vol. 1, pp. 1-46, 1997. [11] A. G. Cohn, B. Bennett, J. Gooday, and N. M. Gotts, “Representing and Reasoning with Qualitative Spatial Relations About Regions,” in O. Stock (Ed.), “Spatial and Temporal Reasoning,” pp. 97-134, Kluwer Academic, 1997. [12] A. G. Cohn, N. M. Gotts, Z. Cui, D. A. Randell, B. Bennett, and J. M. Gooday, “Exploiting Temporal Continuity in Qualitative Spatial Calculi,” in M. J. Egenhofer and R. G. Golledge (Eds.), “Spatial and Temporal Reasoning in Geographic Information Systems,” Oxford University Press, pp. 5-24, 1998. [13] A. G. Cohn and S. M. Hazarika, “Qualitative Spatial Representation and Reasoning: An Overview,” Fundamenta Informaticae, Vol. 45, pp. 1-29, 2001. [14] M. R. Diaz, “Topics in the Logic of Relevance,” Philosophia Verlag, 1981. [15] J. M. Dunn and G. Restall, “Relevance Logic,” in D. Gabbay and F. Guenthner (Eds.), “Handbook of Philosophical Logic, 2nd Edition,” Vol. 6, pp. 1-128, Kluwer Academic, 2002. [16] M. J. Egenhofer and R. G. Golledge (Eds.), “Spatial and Temporal Reasoning in Geographic Information Systems,” Oxford University Press, 1998. [17] Y. Goto, S. Nara, and J. Cheng, “Efficient Anticipatory Reasoning for Anticipatory Systems with Requirements of High Reliability and High Security,” International Journal of Computing Anticipatory Systems, Vol. 14, pp. 156-171, CHAOS, 2004. [18] R. Hilpinen, “Deontic Logic,” in L. Goble (Ed.), “The Blackwell Guide to Philosophical Logic,” pp. 159182, Blackwell, 2001. [19] E. D. Mares, “Relevance Logic,” in D. Jacquette (Ed.), “A Companion to Philosophical Logic,” pp. 609627, Blackwell, 2002. [20] E. D. Mares, “Relevance Logic: A Philosophical Interpretation,” Cambridge University Press, 2004. [21] E. D. Mares and R. K. Meyer, “Relevant Logics,” in L. Goble (Ed.), “The Blackwell Guide to Philosophical Logic,” pp. 280-308, Blackwell, 2001. [22] D. Nute (Ed.), “Defeasible Deontic Logic,” Kluwer Academic, 1997. [23] S. Read, “Relevant Logic: A Philosophical Examination of Inference,” Basil Blackwell, 1988. [24] F. Shang and J. Cheng, “Anticipatory Agents Based on Anticipatory Reasoning,” in M. S. Hacid, N. Murray, Z. W. Ras, S. Tsumoto (Eds.), “Foundations of Intelligent Systems, 15th International Symposium, ISMIS 2005, Saratoga Springs, USA, May 25-28, 2005, Proceedings,” Lecture Notes in Artificial Intelligence, Vol. 3488, pp. 445-455, Springer-Verlag, 2005. [25] O. Stock (ed.), “Spatial and Temporal Reasoning,” Kluwer Academic, 1997. [26] T. Tagawa and J. Cheng, “Deontic Relevant Logic: A Strong Relevant Logic Approach to Removing Paradoxes from Deontic Logic,” in M. Ishizuka and A. Sattar (Eds.), “PRICAI 2002: Trends in Artificial Intelligence, 7th Pacific Rim International Conference on Artificial Intelligence, Tokyo, Japan, August 18-22, 2002. Proceedings,” Lecture Notes in Artificial Intelligence, Vol. 2417, pp. 39-48, Springer-Verlag, 2002. [27] J. van Benthem, “Temporal Logic,” in D. M. Gabbay, C. J. Hogger, and J. A. Robinson (Eds.), “Handbook of Logic in Artificial Intelligence and Logic Programming, Vol. 4,” pp. 241-350, Oxford University Press, 1995. [28] Y. Venema, “Temporal Logic,” in L. Goble (Ed.), “The Blackwell Guide to Philosophical Logic,” pp. 203-223, Blackwell, 2001.


321

Concept Theory in the Context of Information Systems Jari Palomäki* & Momir Radiþeviü** * Massey University, Department of Information Systems Private Bag 756, Wellington, New Zealand E-mail: [email protected] ** Waiariki Institute of Technology, School of Business and Computing Private Bag 3028, Rotorua, New Zealand E-mail: [email protected] Abstract: In this paper the role and the proper place of a concept theory in the context of information systems is proposed. We define the basic modelling situation consisting of objects to be modelled, a modeller, i.e. a person who is modelling, a model, and the relationships between these. We continue this process by describing an information system. Having described information systems we consider a conceptual modelling process, the goal of which is to develop a conceptual schema of the domain of application. In this context the place of a concept theory is found as the foundation for all conceptual modelling. Keywords: Concept, concept theory, conceptual modelling, information systems, system.

1. Introduction In this paper the role and the proper place of a concept theory in the context of information systems is proposed. Firstly, we define the basic modelling situation consisting of objects to be modelled, a modeller, a model, and relationships between these. Secondly, we describe an information system as an instance of basic modelling situation. Thirdly, having described information systems we consider a conceptual modelling process, the goal of which is to develop a conceptual schema of the domain of application. Lastly, we consider a concept theory and its role in conceptual modelling.

2. Basic Modelling Situation Modelling starts when for some special purposes something is to be modelled. This creates a basic modelling situation, which consists of 1) objects to be modelled, 2) a modeller, who is doing the modelling, 3) a model, which is to be a result of the modelling, and 4) different relationships between these. The basic modelling situation is then a three-place relation: M(object(x), modeller(y), model(z)), which is to be read: “an object x is modelled by y as z.” The objects to be modelled form the object domain, i.e. the universe of discourse. The objects in the object domain do not have necessarily to be concrete things in space and time of which we are to have immediate sense perception, but can as well be abstract objects consisting e.g. expert’s knowledge. Moreover, the objects to be

322

J. Palomäki and M. Radiˇcević / Concept Theory in the Context of Information Systems

modelled do not have to exist before the model of it is created. This situation happens, for example, when designing or planning something. A modeller is the subject of a modelling situation. In most cases a modeller consists of the group of persons having different expertises. The most important task for a modeller is to consider those features of the objects to be modelled which are relevant for the purpose of modelling. A model is a result of an abstraction that is used to represent the objects in the object domain. Abstraction is an epistemological process, where some relevant aspect of the objects in the object domain is separated. There are also other epistemological processes connected with creating a model, for example, classification, i.e. grouping the objects into classes on the basis of some of common properties, generalization, i.e. arriving at some general notion from the instances, axiomatisation, i.e. by giving the basic propositions (truths) from which we can deduce other propositions (truths) the result of which is called an axiom system, etc. Modelling relations consist of logical and epistemological relationships between the modeller, model, and object domain.

Figure 1. Basic Modelling Situation

3. Information System Having defined the basic modelling situation, we shall proceed to an information system. Firstly, any system may be characterised as a set of interrelated components that interact with each others to form a whole, and work together toward a common goal. The four major components of a system are (1) inputs, (2) processes, (3) outputs, and (4) feedback control. An input is anything that enters the system. A system may have one or many inputs. An output is anything leaving the system. Outputs are the goal of the system, the purpose for which the system exists. The outputs of one system can become the inputs of another system. A process performs the work: it transforms an input into an output. The system exchanges inputs and outputs with its environment. Changes in the environment affect the system. To adopt itself to changing conditions the system have a feedback. A feedback reintroduces a portion of the output of a system as an input into the same system, and thus, a feedback is


323

actually a special kind of input. It is used as a control mechanism within the system to indicate the difference between the goals and actual performance of a system, (see e.g. Zwass 1998, Whitten et al. 2001).

Input

Output

Process

Feedback

Figure 2. System. In information system inputs are of two kinds: firstly, they are result of gathering and capturing data to be stored in databases and, secondly, they consist of queries concerning the data stored in databases. Information processing involves converting or transforming queries into valuable information, and it can be done manually or with the assistance of a computer. Processing performed by a computer is actually performed by the computer’s Central Processing Unit (CPU) which, in doing that, interplays with databases and a conceptual schema, sometimes also called the logical schema in relational data models. Resulting information, as an answer to a query, may be used as a feedback to make changes to input or processing activities. Information System

0 V c c1 1

a1

b1

a2

b2

2

5

Conceptual Schema

6

CPU 3

4

7 a3

b3

a4

b4

8

GND 0

Database

Figure 3. Information System 1. An information system is not isolated from its environment, and so it has also a larger sense. An information system is a set of people, data, and processes that work together to achieve the common goal of information management. A computer-based information system includes also hardware, which consists of everything in the physical layer, and software, which consists of systems programs that control the hardware and application programs used to produce the required information.

324


Information System

0

Answer

Vcc1 1

2

a1

b1

a2

b2

5

Conceptual Schema

6

CPU 3

User

4

a3

b3

a4

b4

7

8

GND 0

Database Query

Object to be Modelled

Figure 4. Information System 2.

4. Conceptual Model and Conceptual Modelling A conceptual modelling consists of analysing and synthesising concepts which are “drawn” from a given object domain, and constructing a conceptual model of it by means of those concepts, (Kangassalo 1990). The result of this process is presented in a form of conceptual schema, which provides a conceptual description of the data under inspection. Accordingly, databases are used to store data representing a certain section of the reality, and the classification and organization of that data is described by means of a conceptual schema. A conceptual schema of the object domain, in turn, is used as an input to a support system, e.g. CASE tool, which generates a database of it. The database implementation includes update and query functionalities as well. In the construction of the conceptual schema at least the following four principles would serve as guidelines (van Griethuysen 1982, Kangassalo 1983, Marjomaa 1997): 1. The conceptualization principle, according to which only conceptual aspects should be taken into account when constructing the conceptual schema. 2. 100% -principle, according to which all the relevant aspects of the domain should be described in the conceptual schema. 3. The formalization principle, according to which conceptual schemata should be formalisable in order to be implementable. 4. The semiotic principle according to which conceptual schemata should be easily interpretable and understandable. From the principles 1. and 2. it follows that all the relevant aspects of the domain appearing in the conceptual schema are to be conceptual. The principle 3. is essentially a restrictive principle, and it is partly opposite to the principle 4. The motivation for the principle 4. is that the language for representing conceptual schemata should be common for everyone taking part in the modelling process, and be appropriate for fulfilling the task of modelling. A reason why the principle 4. may sometimes be opposite to the principle 3. is that not everyone is familiar with the formalisms used in this area.


325

In order to express the result of conceptual modelling as a conceptual schema, we need a modelling language. The three most popular notations for information modelling are Object-Role Modelling (ORM), Entity-Relationship (ER) diagrams, and Unified Modelling Language (UML), (see e.g. Halpin 2001). However, many accounts of conceptual modelling emphasise an intensional aspect concept as opposed to real world things in the domain of objects, which belong to the extensions of those concepts. This means that intensionality should be visible in the modelling language as well, (Kangassalo 1992/93).

Figure 5. Conceptual Modelling 1. Let us note that there are two different inquires each having a right to the name “conceptual modelling”. The first one is when we are actually doing conceptual modelling, and the second one is when we are describing the process of conceptual modelling. The first one is more basic, since in the second one conceptual modelling is conceptually modelled, and hence, it has different ontological status than the first one. In any given discussion, it is easy to fall into confusion through failure to determine to which of the two inquiries the discussion is intended to belong.

5. Concepts and Concept Theory A conceptual model is composed of particular concepts and particular relations between them. A concept theory, in turn, is studying concepts and the relations and operations between them in general. There are numerous theories of concepts, for example, concepts can be thought of as any of the following: supersensible entities as universals, meanings, abstract objects, definitions, or predicates and relations; mental entities or states as composite images, ideas, thoughts, conceptions, or innate ideas; neutral entities joining e.g. words, thoughts, and things; human or animal skills or abilities, etc. (see e.g. Weitz 1988). In order to condense this great number of theories without sacrificing too many of their indispensable properties, the following more abstract classifications can be made:

326


1. The entity theories of concepts, according to which concepts are entities such as ideas, images, thought, abstract objects, etc. 2. The dispositional theories of concepts, according to which concepts are understood to be the abilities or skills to do something, especially the ability to use language correctly. We may go even further. One of the major doctrines of both the entity and the dispositional theories of concepts is that a concept, in order to be a concept, must be characterised by a definite set of necessary and sufficient criteria. This is called a closed concept, and the requirement that all concepts are to be closed is called the closure principle of concepts. Thus we have abstracted the particular conceptions of concepts to the point where there are just concepts, and whatever form these take, they must satisfy the closure principle of concepts. The advantage of this approach is that developing a concept theory it does not confine us to any particular theory of concepts with its specific accompanying restrictions and special characteristics. Relations between concepts enable us to establish conceptual structures. The most essential relation between concepts is the intensional containment relation, which is based on the intensions of concepts, (see Kauppi 1967, Kangassalo 1992/93, Palomäki 1994). Another important relation between concepts is the conceptual containment relation between concepts, which is based on the conceptual constituents of concepts, and is thus a different relation than the intensional containment, (Palomäki 1997, 2003). In current literature, the relations between concepts are mostly based on the set theoretical relations between the extensions of concepts. For example, according to Nebel & Smolka (1990), the conceptual intersection of the concepts of ‘man’ and ‘woman’ is the empty-concept, and their conceptual union is the concept of ‘adult’. However, intensionally the common concept which contains both the concepts of ‘man’ and of ‘woman’, and so is their intensional conceptual intersection, is the concept of ‘adult’, not the empty-concept, and the concept in which they both are contained, and so is their intensional conceptual union, is the concept of ‘androgyne’, not the concept of ‘adult’. Similarly, in the formal concept analysis (Ganter & Wille 1996) concepts are defined as ordered pairs of sets of attributes and sets of things, which are connected by the incidence relation. The relation between concepts, called the hierarchical order, is then defined by means of the set-theoretical subset relation between a set of things (or a set of attributes) of the concepts. The resulting conceptual structure forms a complete lattice, and is called a concept lattice. Thus, although they are talking about concepts, they are dealing with them only in terms of extensional set theory, not intensional concept theory. One of the crucial differences between extensionality and intensionality is that in extensionality a collection is determined by its elements, whereas in intensionality a collection is determined by a concept, a property, an attribute, etc. Now intensional notions (e.g. concepts) are not strictly formal notions, and it would be misleading to take these as subjects of study for logic only, since logic is concerned with the forms of propositions as distinct from their contents. Perhaps only part of the theory of intensionality which can be called formal is pure modal logic and its possible worlds semantic. However, in concept theories based on possible worlds semantic, (see e.g.


327

Hintikka 1969, Montague 1974, Palomäki 1997, Materna 1998), intensional notions are defined as (possibly partial, but indeed set-theoretical) functions from the possible worlds to extensions in those worlds. In these approaches intensional notions are once more either reduced to extensional set-theoretic constructs in diversity of worlds or as being non-logical notions left unexplained. So, when developing an adequate presentation of a concept theory it has to take into account both formal (logic) and contentual (epistemic) aspects of concepts and their relationships.

Figure 6. Conceptual Modelling 2. Thus, the foundation of conceptual modelling is based on a concept theory, where the idea of the concept of a concept and the relations between concepts serves as the basis for all conceptual modelling, and all conceptual structures are ultimately reducible to concept theory, which is to be understood in the similar way as “all mathematics is ultimately reducible to set theory”.

Conclusion In this paper the role and the proper place of a concept theory in the context of information systems was proposed. First, we defined the basic modelling situation consisting of objects to be modelled, a modeller, i.e. a person who is modelling, a model, and the relationships between these. In this context we described an information system. Having described information systems we were considering conceptual modelling process, the goal of which is to develop a conceptual schema of the object domain. A conceptual schema is a physical part of an information system, and creating a conceptual schema presupposes a conceptual model, which it represents and is based on. In this context the place of a concept theory was found as the foundational basis for all conceptual modelling. Acknowledgement: This paper was inspired by the “simple” question put forward by Professor Eiji Kawakuchi at the 12th European-Japanese Conference on Information Modelling and Knowledge Bases, Krippen, Germany, 27.-30.05.2002, which concerned the status of a concept theory in the context of information systems.

328


References: Ganter, B. & Wille, R., 1996: Formale Begriffsanalyse – Mathematische Grundlagen. Berlin, Heidelberg, New York: Springer-Verlag. van Griethuysen, J. J., (ed.), 1982: Concepts and Terminology of the Conceptual Schema and the Information Base. A Report ISO/TC97/SC5-N695. Halpin, T., 2001: Information Modelling and Relational Databases – From Conceptual Analysis to Logical Design. San Fransisco: Morgan Kaufmann Publishers. Hintikka, J., 1969: Models for Modalities. Dordrecht: D. Reidel. Kangassalo, H., 1983: “Structuring Principles of Conceptual Schemas and Conceptual Models”. Information Modeling. Ed. Bubenko, J. A., Lind: Studentlitteratur, 223-307. Kangassalo, H., 1990: “Foundations of Conceptual Modelling: A Theory Construction View”. Information Modelling and Knowledge Bases. Eds. H. Kangassalo, S. Ohsuga, and H. Jaakkola. Amsterdam, Washington, Tokyo: ISO Press, 19-35. Kangassalo, H., 1992/93: “COMIC: A system and methodology for conceptual modelling and information construction”, Data and Knowledge Engineering 9, 287-319. Kauppi, R., 1967: Einführung in die Theorie der Begriffssysteme. Acta Universitatis Tamperensis. Ser. A. Vol. 15. Tampere: Tampereen yliopisto. Marjomaa, E., 1997: Aspects of Relevance in Information Modelling: Methodological Principles and Conceptual Problems. Filosofisia tutkimuksia Tampereen yliopistosta, Vol. 63, Tampere: Tampereen yliopisto. Materna, P., 1998: Concepts and Objects. Acta Philosophica Fennica. Vol. 63. Helsinki: Societas Philosophica Fennica. Montague, R., 1974: Formal Philosophy. Ed. R. Thomason. New Haven and London: Yale University Press. Nebel, B. & Smolka, G., 1990: “Representation and Reasoning with Attributive Descriptions.” Sorts and Types in Artificial Intelligence. Eds. Bläsius, K. H., Hedstück, U., and Rollinger, C. R. Lecture Notes in Computer Science 418. Berlin, etc.: Springer-Verlag, 112-139. Palomäki, J., 1994: From Concepts to Concept Theory. Acta Universitatis Tamperensis. Ser. A. Vol. 416. Tampere: University of Tampere. Palomäki, J., 1997: “Three Kinds of Containment Relations of Concepts”. Information Modelling and Knowledge Bases VIII. Eds. H. Kangassalo, J.F. Nilsson, H. Jaakkola, and S. Ohsuga. Amsterdam, Berlin, Oxford, Tokyo, Washington, DC.: IOS Press, 261-277. Palomäki, J., 2003: “Intensional vs. Conceptual Content of Concepts”. Information Modelling and Knowledge Bases XIV. Eds. H. Jaakkola, H. Kangassalo, E. Kawaguchi and B. Thalheim. Amsterdam, Berlin, Oxford, Tokyo, Washington, DC.: IOS Press, 128-134. Weitz, M., 1988: Theories of Concepts: A History of the Major Philosophical Tradition. London and New York: Routledge. Zwass, V., 1998: Foundations of Information Systems. Boston, etc.: McGrawHill/Irwin. Whitten, J., Bentley, L., & Dittman, K., 2001: Systems Analysis and Design Methods, Boston, etc.: McGraw-Hill/ Irwin.


329

Describing Knowledge Architectures Enn TYUGU Estonian Business School Lauteri 3, 10114 Tallinn, Estonia Abstract. A conceptual framework for specifying knowledge architectures is represented. Knowledge architecture is considered independently of its implementation; abstraction of knowledge system as an interpreted free deductive system is used for analysis of knowledge-based systems. Visual notation for knowledge modules and their connections is introduced and used in analysis of known knowledge based tools. Keywords: Knowledge architecture; Transparency; Modularity.

The things we deal with in practical life are usually too complicated to be represented by neat, compact expressions. Especially when it comes to understanding minds, we still know so little that we can't be sure our ideas about psychology are even aimed in the right directions. In any case, one must not mistake defining things for knowing what they are. Marvin Minsky, The Society of Mind, 1985

1. Introduction Our aim is to present a conceptual framework and a related visual notation for describing architecture of intelligent systems on knowledge level, without referring to implementation details. In order to validate the definitions we are using them in the analysis of properties of knowledge-based systems. This paper is based on long-time experiences with various knowledge-based tools, in particular, with large heterogeneous knowledge bases and knowledge based software tools used in engineering domains [9,10]. A motivation for the paper has been an observation that it is misleading to describe a knowledge based system in terms of implementation details such as data structures and algorithms. One may try to use UML for this purpose, but UML has been developed under the strong influence of data modelling and object-oriented paradigm. Let us consider an example of intelligent software environment that performs automatic program synthesis. A well-known system of this kind is Amphion [5] -- a deductive program synthesis system developed in NASA for solving planetary navigation problems. Ignoring completely all aspects of implementation, one can see two mechanisms of knowledge handling in it: logic and computational knowledge system. The first is used to prove solvability of the given task. The second for performing computations according to the program extracted from the proof. They are tightly coupled in a very clean way – logical formulas have computational meaning. One does not need any implementation details in order to understand the knowledge handling principles of the system. However, we have no means for representing this simple architecture on knowledge level and without speaking about data structures and procedures.

330

E. Tyugu / Describing Knowledge Architectures

Another observation is that to use one and the same knowledge handling mechanism for solving various knowledge-intensive tasks is inefficient, hence advanced knowledge-based tools include several knowledge handling mechanisms, and possess a kind of modularity on the knowledge level. Also this is demonstrated by the given example of Amphion that includes at least two knowledge systems: logic and computational system. Observations in cognitive science support the idea of this kind of modularity of knowledge also in human thinking [1]. We are going to consider modularity of knowledge where knowledge modules are of various kinds. The question is, how to combine different knowledge representation and handling techniques to achieve a common goal. The first idea could be to use a multi-agent architecture. This is suitable for loosely coupled systems, and does not work for tightly bound knowledge systems constituting one single tool. We investigate ways of binding various knowledge systems, often hierarchically, into a single knowledge-based tool. There are abstract descriptions of hierarchical knowledge systems. An example is the US patent No. 4,591,983 [11] that has even earned some money to Teknowledge, and has been a case in courts. However, this patent is about hierarchies in a single knowledge representation framework. We separate knowledge architecture from implementation aspects. For this purpose we first introduce a unifying and quite abstract approach to knowledge system and knowledge handling. To demonstrate feasibility of the definitions we use them to prove some theorems about properties like consistency and transparency of knowledge systems. Then we introduce visual notations for knowledge modules and apply them in describing various configurations of knowledge architectures. In particular, we describe the knowledge architecture of intelligent software tools and systems that we have developed for engineering domains.

2. Abstract representation of a knowledge system Marvin Minsky has warned us from defining things related to mind we do not understand sufficiently well. However, here we have to clarify the concepts of knowledge and knowledge system before using them. Formal representation of knowledge as texts has been thoroughly discussed in Lorents [3]. We define a knowledge system in a way that enables us to use this concept for any system that can store and use knowledge for performing some tasks, e.g. obtaining answers to some given questions. We have extended the definition given in [3] by adding a knowledge handling mechanism, although some knowledge systems may be used for transfer and storing of knowledge only. We consider a knowledge system as a knowledgeprocessing module, and a knowledge-based system as a composition of modules of this kind. The aim of this section is to outline the concept of knowledge system in a way that is independent of any particular implementation, and is usable in the analysis and explanation of architectures of knowledge-based applications. First, we assume that knowledge is always represented either 1) by one or more knowledge objects which can be, for instance formulae, or just texts in a suitable language, or 2) by a state of some system. At a closer look one can detect that the second form of knowledge representation is reducible (can be transformed) to the representation in the form of knowledge objects. Second, we assume that to use knowledge means performing some operations on knowledge objects in such a way that new knowledge objects appear. We do not


331

see any other ways of knowledge usage. In the case of knowledge representation as a state of a system, knowledge usage means transforming the state. Also this can be represented as changing the set of knowledge objects that encode the state. These two assumptions lead us to the concept of deductive system. For detailed discussion of this formalism we refer to Maslov [7]. Third, we assume that knowledge objects have some meaning that may be objects or maybe even some effects on the environment where they exist. To make it precise, one has to consider a collection of possible meanings and a mapping from knowledge object to meanings. This can be formalized as an interpretation of knowledge objects. One can, instead of the mapping, consider a relation that binds knowledge objects with their meanings. This is a relation of notation-denotation intuitively well known in philosophy and logic. Also this is a part of the formalism presented in Lorents [3, 4]. Conventional concept of a deductive system is defined as a set of initial objects and derivation rules for generating new objects from given objects. (See, for instance, Maslov [7].) We are going to use a slightly different concept -- a free deductive system, where initial objects are not fixed. Giving some initial objects in the language of a free deductive system makes it a conventional deductive system. Definition 1. Free deductive system is a language of objects and rules for generating new objects from given objects. A good example of a free deductive system is a Post’s system without initial objects. Let us have an alphabet A (i.e. a set of symbols) and a set of variables P. Objects are words built from symbols of the alphabet A. Inference rules are given in the following form: V1 … Vm ________ V0 where V0, V1, ... , Vm are words in the alphabet A ∪ P, and V0 doesn't contain variables which do not occur in V1, ... ,Vm. A word W0 is derivable from the words W1, ... ,Wm by an inference rule in the calculus iff there exists an assignment of words in A to variables occurring in the rule such that it transforms the words V0, ... , Vm into W0, ... , Wm. Post’s systems are in some sense minimal and universal deductive systems. As they do not have syntactic restrictions on objects, they are a general form of deductive systems with textual objects. They have quite interesting general properties. For instance, it is known that there exists a universal Post’s system that can represent behaviour of any other Post’s system [7]. We are not going to use Post’s systems here. They can be used for theoretical investigation of representability of knowledge, of derivability etc., but not for practical representation of knowledge. They are like Markov’s normal algorithms that are, due to their syntactic simplicity, useful in theory of algorithms, but useless in programming. Definition 2. Interpretation of a (free) deductive system is a set M of entities that are possible meanings of objects of the deductive system, and a relation R that binds at least one meaning with its every object, and binds at least one object (a notation) with every meaning included in M. The relation R is called a relation of notation-denotation (see Lorents [3]).

332


Remark. A widely used method to get a meaning of an object generated by a rule is to construct it from meanings of other objects used in the rule. From the other side, we do not know any general way to find objects that are bound to a meaning by the relation of notation-denotation. Definition 3. Knowledge system is an interpreted free deductive system. Examples of free deductive systems are natural languages with one derivation rule that generates a new sequence of sentences ST for any correct sequences of sentences S and T of a language. This rule can be written as follows: S T ST If we are able to get a meaning for any sequence of sentences, and can find a sentence denoting a given meaning, we can say that we have a knowledge system. In other words – we a have a knowledge system, if and only if we understand the language. One more example of a knowledge system is a calculus of computable functions -- CCF. This can be any representation of computable functions together with suitable computation rules, e.g. lambda calculus, Turing machine, a programming language etc. In this paper we shall use CCF every time we have to represent computations as the principal way of knowledge handling, ignoring the actual computational device, i.e. the implementation. Sometimes we are interested only in knowledge representation, and not in making inferences, i.e. we need only a language of objects with interpretation. Formally we have then a knowledge system with empty set of derivation rules. We call a knowledge system with empty set of rules a closed knowledge system. Sometimes we are unable (or are just not interested) to find a notation for an arbitrary given meaning, but still can find a meaning for any notation. Then we can use a mapping from notations to denotations instead of the relation of notationdenotation that binds a notation with every denotation. Weak interpretation of a deductive system is a set M of entities that are possible meanings of objects of the deductive system, and a mapping P that binds at least one meaning with every object of the deductive system. Weak knowledge system is a free deductive system with weak interpretation. We call a knowledge system normal, if we wish to stress that it is not weak. Besides the notation-denotation relation we are going to use some mappings for investigating a connection between notations and their meanings. A relation R between the sets A and B induces two natural mappings: a mapping from A to B that associates with any element x of A all such elements of B that satisfy R together with x, and another mapping from B to A that associates with any element y of B all such elements of A that satisfy R together with y. If there is no danger of misunderstanding, we denote the natural mappings also by R and R-1 respectively. We are also going to denote by R(x) an image of x in B, where x is either an element or a subset of A, and use R-1(y) for denoting an image of y in A.

3. Modularity of knowledge In this section we consider knowledge systems as modules of larger knowledge-based tools, discuss how knowledge systems interact with one another,

333


and how knowledge architecture can be represented by knowledge systems connected in various ways.

3.1 Hierarchical connection Let us have two knowledge systems K1, K2 with sets of notations (knowledge objects) S1, S2, sets of meanings M1, M2 and notation-denotation relations R1, R2 respectively. We say that knowledge systems K1 and K2 are hierarchically connected, if there is a relation R between the set of meanings M1 and the set of notations S2, and strongly hierarchically connected, if there is a one-to-one mapping C between the elements of a subset of M1 and a subset of S2. A hierarchical connection of knowledge systems can be observed quite often in real life. This situation arises, for instance, when we have one knowledge system that deals with shallow knowledge on a layman’s level, and another knowledge system that deals with deeper expert knowledge, and layman’s meanings are expert’s objects. This connection exists also between two knowledge systems, when one is a metasystem to another.

K1

objects meanings

K1

M1 K2

R

K2

C

S2

(a)

(b)

(c)

Figure 1.Knowledge system (a), hierarchical connection of knowledge systems (b) and strong hierarchical connection of knowledge systems (c) Having a sequence of knowledge systems K1, K2,..., Kn, one can build a hierarchy (a tower) of knowledge systems – a knowledge tower where each pair Ki, Ki+1 are hierarchically connected. Towers of closed knowledge systems have been investigated by Lorents in [3]. Practical examples of knowledge towers can be found in networking -- stacks of network protocols are examples of towers of knowledge systems. Fig. 1a shows graphical notation for a knowledge system. We have decided to denote explicitly the notations and the meanings of a knowledge system. This allows us to show different ways of connecting knowledge systems. Fig. 1b shows us hierarchical connection of knowledge systems and Fig. 1c shows strong hierarchical connection.

3.2 Semantic connection Several knowledge systems can have one and the same set of meanings. Then we say that they are semantically connected, and use the graphical notation as in Fig. 2a. This

334


is the case, for instance, with classical logic systems that have different sets of inference rules, or programming languages for one and the same computer, and even natural languages for identical cultures. (However, strictly speaking, the cultures depend to some extent on the language, hence there are no absolutely identical cultures with different languages.) If we wish to express the properties of a knowledge system visually more precisely, then we can add graphical notations as it is shown for weak and closed knowledge systems in Fig. 2b and Fig. 2c. K1

K2

S1

S2

M (a)

(b)

(c)

Figure 2. Semantically connected (a), weak (b) and closed (c) knowledge systems

3.3 Union Like for formal languages, one can define a variety of operations (e.g. union and intersection) between semantically connected knowledge systems. For instance, union of knowledge systems is defined as follows. Union of semantically connected knowledge systems K1, K2 over a set of meanings M is a knowledge system over the same set of meanings M with the set of object S1 ∪ S2, the notation-denotation relation R1∪R2 and derivation rules of both knowledge systems K1 and K2. It is desirable to build a union of knowledge systems K1, K2 even when their sets of meanings are different. We can build a new set of meanings M = M1 ∪ M2, and take it as a set of meanings for both knowledge systems. Now we have to add new denotations into sets S1 and S2 to denote the added meanings. We have to be careful, so that the derivation rules could not be applied to the new denotations. This is more precisely expressed by the following definitions. Conservative extension of a knowledge system K with set of notations S and set of meanings M on a set of meanings M’, such that M⊂ M’, is a knowledge system with 1) set of meanings M’, set of notations S∪ {something}, where something is a new knowledge object that denotes every element of the set M’\M, 2) derivation rules of K, and 3) notation-denotation relation of K extended by new pairs (something,x) for every x∈M’\M. Theorem 1. An object z is derivable from objects x,...,y in a conservative extension K’ of a knowledge system K iff it is derivable from the same objects in K. Proof. The derivation rules of K are neither applicable to something, nor can they produce something, hence the derivability in K’ is the same as in K. Usage of the notation something enables one to reason about an open world, where one can reason only about its part, and not to care about other meanings, indistinguishable in the knowledge system.

335


Union of knowledge systems K1, K2 is union of semantically connected conservative extensions of K1, K2 on M1 ∪ M2. It is not our aim to investigate the properties of knowledge systems built by means of operations like union, intersection etc. that is an interesting area of research in formal languages and systems. We are not going to look deep inside the knowledge systems, but describe architectures where they appear as components, and investigate only some very general architectural properties here.

3.4 Operational connection Knowledge system K1 is operationally dependent on a knowledge system K2, if some of its derivation rules uses K2 in deriving a new object, i.e. the result of derivation in K1 depends on knowledge processing in K2, Fig. 3 a. Knowledge systems K1, K2 are operationally connected, if K1 is operationally dependent on K2, and K2 is operationally dependent on K1. Fig. 3b shows operationally connected knowledge systems. It is easy to see that union of knowledge systems can be used to specify their operational connection. It can be done, if their sets of meanings have common elements. However, it is not clear how to realize an operational connection specified in this way – one needs a constructive way to map meanings to denotations. K1

K1

K2

(a)

K2

(b) K1

K2

(c) Figure 3. Operationally dependent (a), operationally connected (b) and connected (c) knowledge systems Sometimes it is possible to detect that knowledge systems interact, but it is impossible or difficult to make precise the way in which they interact. Then we say that knowledge systems are connected, but will not apply any attribute to specify how they are connected. The respective graphical notation is shown in Fig. 3c.

4. Consistency and transparency Consistency and transparency are desirable properties of modular systems. Here we use our abstract definitions of knowledge systems and their connections in order to investigate these notions.

336


Several meanings may be considered equivalent in some cases. Let H be an equivalence relation on a set of meanings M of a knowledge system. We denote a set of sets of equivalent (with respect to H) elements by F(M,H). This set of equivalence classes (called also a factor set) plays an important role in the analysis of knowledge architectures. It can be used to describe consistency and transparency of knowledge systems. Let us have a knowledge system with notation-denotation relation R and an equivalence relation H on the set of meanings M. The knowledge system is called consistent with respect to H, iff for its any object q all meanings of the object are equivalent. This is expressed by the formula ∀q∈S ∃u∈F(M,H) R(q)⊆ u. Remark. Equivalence relation H on set of meanings M of a consistent knowledge system with notation-denotation relation R induces an equivalence relation H’ on its set of notations S such that the factor set F(S,H’) consists of images of elements of F(M,H) produced by the mapping R-1. Having two connected knowledge systems K1, K2, we define a set of induced meanings of a meaning m of K1 as the set of elements reachable by composition of the mapping C from M1 to S2, natural mappings of notation-denotation relation R2 of K2 , and the inverse mapping C-1 of C. More precisely, it is the set C-1(R2-1 (R (C(m)))), see Fig. 4. induced meaning of m m.

C-1

C

R2

R2-1

Figure 4. Defining the induced meaning We say that a hierarchical connection is transparent with respect to an equivalence relation H given on the set of meanings M of the knowledge system K1, iff for any meaning m of K1 all its induced meanings are equivalent to m. The transparency condition can be extended on a tower of knowledge systems in an obvious way – one has to take into account natural mappings of all knowledge systems of lower levels and all connecting mappings. The transparency is an important feature of knowledge-based systems that is used quite often intuitively. We are used to speak about the transparency of protocols, etc. Theorem 2. A hierarchical connection of knowledge systems K1 and K2, strongly connected by a mapping C such that C(M1 ) = S2, is transparent with respect to an equivalence relation H given on the set of meanings of the knowledge system K1, iff there exists an equivalence relation H’ given on the set of meanings M2 such that the knowledge system K2 is consistent with respect to H’, and the equivalence


337

relation H´´ induced by H´ on S2 is the same as the relation induced on S2 by H through the mapping C. Proof. By definition of H´´ and properties of the mapping C, every equivalence class of H is mapped on a separate equivalence class of H´´, and vice versa -- equivalence classes of H´´ - are mapped in the same way on the equivalence classes of H. Hence the transparency depends only on the notation-denotation relation of K2, and the consistency condition of K2 is precisely the condition required for transparency. Theorem 3. A hierarchical connection of knowledge systems K1 and K2, strongly connected by a mapping C is transparent with respect to an equivalence relation H given on the set of meanings of the knowledge system K1, if the knowledge system K2 is consistent with respect to an equivalence relation H´ such that the equivalence relation induced by H´ on S2 coincides on C(M1) with the relation induced on S2 by H through the mapping C. Proof. That the consistency is a sufficient condition is proved in the same way as for theorem 2. But this is not a necessary condition, because there may exist knowledge objects of K2 not satisfying consistency, and not belonging to C(M1). These meanings are unreachable by natural mappings defining the transparency (because they are not in C(M1) ), hence they do not influence the transparency.

5. Examples of knowledge architectures 5.1 Program synthesis tools In this section we consider knowledge systems as architectural components of knowledge-based tools. A simple example of efficient usage of connected knowledge systems is given by deductive program synthesis. There are different methods of deductive program synthesis. Well known is the approach of Manna and Waldinger [6] implemented, for instance, in the Amphion system [5]. We have used structural synthesis of programs [8] in several software tools [2]. In all these cases a knowledge system of logic is used for deriving a proof of the goal. Another knowledge system is CCF for performing the computations. A code generator transforms a proof into an executable goal and connects these systems. It implements a mapping C from interpretations of logical formulas into objects of CCF, as shown in Fig. 5. Another software tool where logic as a knowledge system is used is Prolog. It includes a knowledge system of logic – calculus of Horn clauses (CHC) as well as a calculus of computable functions for terms used in atomic formulas and for implementing some predicates. The two knowledge systems of Prolog are operationally connected as shown in Fig. 6. logic

C CCF

Figure 5. Knowledge architecture of deductive program synthesis software

338


CCF

CHC

Figure 6. Knowledge architecture of Prolog Our first experience of working with knowledge-based systems was obtained with the PRIZ family [10]. These systems used intuitionistic propositional calculus (IPC) for deductive program synthesis. Specifications for synthesis were presented in a high-level language for convenience of users. Meaning of a specification was a set of formulas of IPC. These formulas were knowledge objects of the second calculus. Formulas of IPC have a computational meaning that is mapped into knowledge objects of the calculus of computable functions that represents computations performed by synthesized programs. Three knowledge systems constitute a knowledge tower shown in Fig. 7. The mappings C1and C2 are identity mappings. specification calculus

C1

intuitionistic propositional calculus

C2 calculus of computable functions

Figure 7. Calculi of PRIZ system

New knowledge systems have been added to another software tool with program synthesis features -- NUT that has been used in several large applications [2]. It has a visual knowledge system for developing and using visual specifications. Meaning of a visual specification is a specification in a language similar to the specification language of PRIZ. Besides that, a rule-based knowledge system that uses geometric logic has been added for expanding a specification, if a synthesis problem appears unsolvable. Hence, the main synthesizer working in the intuitionistic logic is operationally dependable on the rule-based system. The latter adds deeper meaning to specifications and it is in this way hierarchically connected with the specification calculus. This gives us the architecture shown in Fig. 8.


339

visual knowledge system

specification calculus

intuitionistic propositional logic

geometric logic

calculus of computable functions

Figure 8. Knowledge architecture of NUT system

6. Concluding remarks We have discussed examples of knowledge architectures on an abstract level, deliberately ignoring all aspects of realization. Architectural components are various knowledge systems. Some of them are precisely defined, for instance, Horn clause logic, , classical logic used in the deductive synthesis. In order to understand exact construction of other systems, one has to look more closely at their sets of notations and denotations, as well as at the notation-denotation relation. This does not necessarily mean looking at the implementation. Unfortunately, a common practice even in knowledge-based systems research has been for a long time to present the essence of a system by describing its realization in a form of data structures and algorithms. Modelling languages like UML are more and more used in software engineering, but not much in knowledge-based system research. We are looking forward to see more abstract and knowledge-oriented notations in architectural descriptions on intelligent systems as well.

Acknowledgements. This work has been supported by Estonian Business School and by grant No. 5400 of Estonian Science Foundation.

References 1. F. Gobet (2001) Is Experts' Knowledge Modular?. In Proceedings Proceedings of the 23rd Meeting of the Cognitive Science Society, 336-431.

340


2. V. Kotkas et al. (2000) Ontology-based design of surveillance systems with NUT. Proc. 3rd International Fusion 2000 Conference. Paris, 3 - 9. 3. P. Lorents (2001) Formalization of data and knowledge based on the fundamental notation-denotation relation. IC-AI´2001 International Conference. Proceedings. v. III.. CSREA Press, 1297 – 1301. 4. P. Lorents, D. Lorents (2003) Intelligenc e and the Notation-Denotation Relation. IC-AI´2003 v. III.. CSREA Press, 5. M. Lowry, A. Philpot, T. Pressbürger, I. Underwood (1994) Amphion: Automatic Programming for Scientific Subroutine Libraries. LNCS v. 869, Springer Verlag, 16 – 19. 6. Z. Manna, R. Waldinger (1993) The Deductive Foundations of Computer Programming. Addison-Wesley. 7. S. Maslov. (1987) Theory of Deductive Systems and its Application. The MIT Press. 8. M. Matskin, E. Tyugu (2001) Strategies of Structural Synthesis of Programs and Its Extensions. Computing and Informatics. v.20,1 -25. 9. E. Tyugu (1994) Large Heterogeneous Knowledge Bases. Machine Intelligence, v.13, Clarendon Press, Oxford, 269 – 288. 10. E. Tyugu (1991) Knowledge-Based Programming Environments. KnowledgeBased Systems, v. 4, 4 – 15. 11. Hierarchical Knowledge System. US Patent #: 4,591,983

341


Author Index Akaishi, M. Al-Fedaghi, S.S. Brinkkemper, S. Chen, X. Cheng, J. de Nijs, F. Delvecchio, T. Dessaigne, N. Di Lecce, V. Ďuráková, D. Duží, M. Fiedler, G. Fukunaga, K. Funyu, Y. Gajdoš, P. Gordillo, S. Haav, H.-M. Halttunen, V. Hausser, R. Hayama, T. Heimbürger, A. Helms, R. Henno, J. Higuchi, M. Hori, K. Ito, K. Jaakkola, H. Kangassalo, H. Kawaguchi, E. Kirikova, M. Kitahashi, T. Kiyoki, Y.

245 94 205 50, 260 310 205 260 1 260 38 20 94 172 180 38 191 302 219 280 297 20 205 v 172 245 83 v, 268 v 297 156 172 v, 50, 68, 297

Kojima, A. Lehtinen, A. Ma, H. Martinez, J. Mäkinen, T. Niemi, T. Niinimäki, M. Nikiforova, O. Nykänen, R. Ohashi, H. Palomäki, J. Pertino, G. Radičević, M. Sasaki, H. Sasaki, J. Schewe, K.-D. Soyama, Y. Spyratos, N. Suzuki, T. Tanaka, Y. Thalheim, B. Tocho, S. Toivonen, S. Tokuda, T. Turull Torres, J.M. Tyugu, E. van Oosterum, J. Varkoi, T. Wojtkowski, W. Yoneda, T. Zambrano, A. Zarri, G.P.

172 219 289 1 268 112, 127 112, 127 156 219 50 321 191 321 68 180 237, 289 253 245 253 245 94, 268 191 112, 127 253 237 329 205 268 156 180 191 138

Information Modelling and Knowledge Bases XVII: Volume 136 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XIX: Volume 166 Frontiers in Artificial Intelligence and Applications (Frontiers in Artificial Intelligenece and Applications)

Information Modelling and Knowledge Bases XXII - Volume 225 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XVIII - Volume 154 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XX - Volume 190 Frontiers in Artificial Intelligence and Applications

Artificial Intelligence in Education (Frontiers in Artificial Intelligence and Applications)

Information modelling and knowledge bases XIV

Information Modelling and Knowledge Bases Xiii

Artificial Intelligence Research and Development (Frontiers in Artificial Intelligence and Applications, Vol. 146) (Frontiers in Artificial Intelligence and Applications)

Artificial Intelligence Research and Development: Volume 163 Frontiers in Artificial Intelligence and Applications

Artificial Intelligence Research and Development: Volume 163 Frontiers in Artificial Intelligence and Applications

Ontology and the Semantic Web: Volume 156 Frontiers in Artificial Intelligence and Applications (Frontier in Artificial Intelligence and Applications)

Ontology and the Semantic Web: Volume 156 Frontiers in Artificial Intelligence and Applications (Frontier in Artificial Intelligence and Applications)

Advances in Logic, Artificial Intelligence and Robotics: Laptec 2002 (Frontiers in Artificial Intelligence and Applications, 85)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications)

Modular Ontologies (Frontiers in Artificial Intelligence and Applications, WoMO 2011)

Active Mining (Frontiers in Artificial Intelligence and Applications, 79)

Active Mining (Frontiers in Artificial Intelligence and Applications)

Algorithms and Architectures of Artificial Intelligence (Frontiers in Artificial Intelligence and Applications)

Algorithms and Architectures of Artificial Intelligence (Frontiers in Artificial Intelligence and Applications)

Formal Ontologies Meet Industry (Frontiers in Artificial Intelligence and Applications)

New Trends in Multimedia and Network Information Systems (Frontiers in Artificial Intelligence and Applications)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications) (Frontiers in Artificial Interlligence and Applications)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications) (Frontiers in Artificial Interlligence and Applications)

Self-Organization and Autonomic Informatics (I): Volume 135 Frontiers in Artificial Intelligence and Applications

Ontology Learning and Population: Bridging the Gap between Text and Knowledge - Volume 167 Frontiers in Artificial Intelligence and Applications

Applications of Data Mining in E-Business and Finance (Frontiers in Artificial Intelligence and Applications)

Knowledge Transformation for the Semantic Web (Frontiers in Artificial Intelligence and Applications, 95)

Design Problems, Frames and Innovative Solutions - Volume 203 Frontiers in Artificial Intelligence and Applications (Subseries: Knowledge-Based Intelligent Engineering Systems)

Multi-Relational Data Mining: Volume 145 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XVII: Volume 136 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XIX: Volume 166 Frontiers in Artificial Intelligence and Applications (Frontiers in Artificial Intelligenece and Applications)

Information Modelling and Knowledge Bases XXII - Volume 225 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XVIII - Volume 154 Frontiers in Artificial Intelligence and Applications

Information Modelling and Knowledge Bases XX - Volume 190 Frontiers in Artificial Intelligence and Applications

Artificial Intelligence in Education (Frontiers in Artificial Intelligence and Applications)

Information modelling and knowledge bases XIV

Information Modelling and Knowledge Bases Xiii

Artificial Intelligence Research and Development (Frontiers in Artificial Intelligence and Applications, Vol. 146) (Frontiers in Artificial Intelligence and Applications)

Artificial Intelligence Research and Development: Volume 163 Frontiers in Artificial Intelligence and Applications

Artificial Intelligence Research and Development: Volume 163 Frontiers in Artificial Intelligence and Applications

Ontology and the Semantic Web: Volume 156 Frontiers in Artificial Intelligence and Applications (Frontier in Artificial Intelligence and Applications)

Ontology and the Semantic Web: Volume 156 Frontiers in Artificial Intelligence and Applications (Frontier in Artificial Intelligence and Applications)

Advances in Logic, Artificial Intelligence and Robotics: Laptec 2002 (Frontiers in Artificial Intelligence and Applications, 85)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications)

Modular Ontologies (Frontiers in Artificial Intelligence and Applications, WoMO 2011)

Active Mining (Frontiers in Artificial Intelligence and Applications, 79)

Active Mining (Frontiers in Artificial Intelligence and Applications)

Algorithms and Architectures of Artificial Intelligence (Frontiers in Artificial Intelligence and Applications)

Algorithms and Architectures of Artificial Intelligence (Frontiers in Artificial Intelligence and Applications)

Formal Ontologies Meet Industry (Frontiers in Artificial Intelligence and Applications)

New Trends in Multimedia and Network Information Systems (Frontiers in Artificial Intelligence and Applications)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications) (Frontiers in Artificial Interlligence and Applications)

Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms (Frontiers in Artificial Intelligence and Applications) (Frontiers in Artificial Interlligence and Applications)

Self-Organization and Autonomic Informatics (I): Volume 135 Frontiers in Artificial Intelligence and Applications

Ontology Learning and Population: Bridging the Gap between Text and Knowledge - Volume 167 Frontiers in Artificial Intelligence and Applications

Applications of Data Mining in E-Business and Finance (Frontiers in Artificial Intelligence and Applications)

Knowledge Transformation for the Semantic Web (Frontiers in Artificial Intelligence and Applications, 95)

Design Problems, Frames and Innovative Solutions - Volume 203 Frontiers in Artificial Intelligence and Applications (Subseries: Knowledge-Based Intelligent Engineering Systems)

Multi-Relational Data Mining: Volume 145 Frontiers in Artificial Intelligence and Applications

Recommend Documents