179 86 5MB
English Pages 236 Year 2010
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
ONTOLOGIES AND SEMANTIC TECHNOLOGIES FOR INTELLIGENCE
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
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 biennial 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, N. Guarino, J.N. Kok, J. Liu, R. López de Mántaras, R. Mizoguchi, M. Musen, S.K. Pal and N. Zhong
Volume 213
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Recently published in this series Vol. 212. A. Respício et al. (Eds.), Bridging the Socio-Technical Gap in Decision Support Systems – Challenges for the Next Decade Vol. 211. J.I. da Silva Filho, G. Lambert-Torres and J.M. Abe, Uncertainty Treatment Using Paraconsistent Logic – Introducing Paraconsistent Artificial Neural Networks Vol. 210. O. Kutz et al. (Eds.), Modular Ontologies – Proceedings of the Fourth International Workshop (WoMO 2010) Vol. 209. A. Galton and R. Mizoguchi (Eds.), Formal Ontology in Information Systems – Proceedings of the Sixth International Conference (FOIS 2010) Vol. 208. G.L. Pozzato, Conditional and Preferential Logics: Proof Methods and Theorem Proving Vol. 207. A. Bifet, Adaptive Stream Mining: Pattern Learning and Mining from Evolving Data Streams Vol. 206. T. Welzer Družovec et al. (Eds.), Information Modelling and Knowledge Bases XXI Vol. 205. G. Governatori (Ed.), Legal Knowledge and Information Systems – JURIX 2009: The Twenty-Second Annual Conference Vol. 204. B. Apolloni, S. Bassis and C.F. Morabito (Eds.), Neural Nets WIRN09 – Proceedings of the 19th Italian Workshop on Neural Nets Vol. 203. M. Džbor, Design Problems, Frames and Innovative Solutions Vol. 202. S. Sandri, M. Sànchez-Marrè and U. Cortés (Eds.), Artificial Intelligence Research and Development – Proceedings of the 12th International Conference of the Catalan Association for Artificial Intelligence
ISSN 0922-6389 (print) ISSN 1879-8314 (online)
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence
Edited by
Leo Obrst The MITRE Corporation, McLean, Virginia, USA
Terry Janssen Lockheed Martin Corporation, Herndon, Virginia, USA
and
Werner Ceusters
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
The State University of New York at Buffalo, Buffalo, New York, USA
Amsterdam • Berlin • Tokyo • Washington, DC
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
© 2010 The authors and IOS Press. 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 978-1-60750-580-8 (print) ISBN 978-1-60750-581-5 (online) Library of Congress Control Number: 2010930895 Publisher IOS Press BV Nieuwe Hemweg 6B 1013 BG Amsterdam Netherlands fax: +31 20 687 0019 e-mail: [email protected]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
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
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved.
v
Preface This book had its origin at the Second International Ontology for the Intelligence Community (OIC) Conference, which was held on November 28–29, 2007, in Columbia, MD, USA. At that time, a volume was proposed by the editors that would feature chapters by selected authors from the conference, who could extend their OIC papers or write on related topics that fit the guidelines the editors established for this book. In addition, other authors were invited to submit chapters. This book represents a partial technology roadmap for government information technology decision makers for information integration and sharing, and situational awareness (improved analysis support) in the use of ontologies, and semantic technologies for intelligence. The general themes of both the OIC conferences and this book focus on intelligence community needs and the applications of ontologies and semantic technologies to assist those needs. Among the very many IC needs are the following: • •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
•
• • •
To increase the ability to meaningfully share information, within and among communities, across humans and machines To off-load some human cognitive functions and enable machines to assume these. By using ontologies and semantic technologies, machines come up to the human conceptual level, rather than humans having to go down to the machine level, which latter tack has largely defined information technology since its orgins up to this point. To increase the ability to automate some aspects of intelligence analysis, as for example, by supporting evidence-based reasoning, deductive (what logically follows, given the knowledge) and abductive (what is the best explanation, given the evidence) queries To provide assistance on probability of Hypothesis given the Evidence P(H|E), hypothesis generation, and analysis of competing hypotheses by using complex knowledge and logical mechanisms, and evaluating the consequences or ramifications of hypotheses To increase the capability to semantically integrate data from all intelligence disciplines To provide analytical tools that exploit the availability of semantically integrated information and knowledge To assist in semantic disambiguation, reference, co-reference/correlation of entities, relations, and events o o
Disambiguation: To determine the appropriate meaning for the given context Reference: To determine the actual entity in the world that the data refers to
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
vi
o •
Co-reference/correlation: To determine whether two entities are actually the same entity, and the properties and events those entities respectively possess and participate in To support reasoning over geospatial, temporal, and other data to infer additional information about the real world.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
The target audience of this book is the US and other intelligence communities (IC), including law enforcement and homeland security communities, along with other technical and budgetary decision makers and technologists working in intelligence. These technologists include ontologists and ontology developers, computer scientists, software engineers, and intelligence analysts who have a strong interest in semantic technologies and their applications. This book would not have been possible without the assistance, dedication, and patience of many generous individuals. We thank the IOS Press publisher and its dedicated representive Maarten Fröhlich for tolerance of delays in the editing of this book, while also providing constant and continuing encouragement. We thank the very many anonymous reviewers who helped improve the contributions of the authors by offering sound feedback and critical comments on multiple iterations of chapters. We thank the past organizers of the OIC conferences, for valuable suggestions and help on many issues, including in particular Barry Smith, Kathryn Blackmond Laskey, Duminda Wijesekera, and Paulo Cesar G. da Costa. We also thank Kevin Lynch and David Roberts, who provided governmental support for the OIC conferences and also feedback to the authors and editors on the impact of these technologies on the intelligence community, thereby serving to provide a pragmatic perspective to constrain the potential technological exuberance. Of course the editors also thank their friends and families, who have countenanced aggravation, missed social opportunities, and personal inattention, to enable the writing and editing of this volume. Finally, we underscore that the views expressed in this book are those of the authors alone and do not reflect the official policy or position of The MITRE Corporation, the Lockheed-Martin Corporation, or any other company or individual, nor that of any particular intelligence community, agency, organization, or government. Leo Obrst Terry Janssen Werner Ceusters
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
vii
Contents Preface Leo Obrst, Terry Janssen and Werner Ceusters
v
Chapter 1. Introduction: Ontologies, Semantic Technologies, and Intelligence Terry Janssen, Leo Obrst and Werner Ceusters
1
Chapter 2. How to Track Absolutely Everything Werner Ceusters and Shahid Manzoor
13
Chapter 3. Uses of Ontologies in Open Source Blog Mining Brian Ulicny, Mieczyslaw M. Kokar and Christopher J. Matheus
37
Chapter 4. A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support Terry Janssen, Herbert Basik, Mike Dean and Barry Smith Chapter 5. Ontologies for Rapid Integration of Heterogeneous Data for Command, Control, & Intelligence Leo Obrst, Suzette Stoutenburg, Dru McCandless, Deborah Nichols, Paul Franklin, Mike Prausa and Richard Sward Chapter 6. Ontology-Driven Imagery Analysis Troy Self, Dave Kolas and Mike Dean
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Chapter 7. Provability-Based Semantic Interoperability for Information Sharing and Joint Reasoning Andrew Shilliday, Joshua Taylor, Micah Clark and Selmer Bringsjord
57
71
91
109
Chapter 8. The Use of Ontologies to Support Intelligence Analysis Richard Lee
129
Chapter 9. Probabilistic Ontologies for Multi-INT Fusion Kathryn Blackmond Laskey, Paulo C.G. Costa and Terry Janssen
147
Chapter 10. Design Principles for Ontological Support of Bayesian Evidence Management Michael N. Huhns, Marco G. Valtorta and Jingsong Wang Chapter 11. Geospatial Ontology Trade Study James Ressler, Mike Dean and Dave Kolas
163 179
Chapter 12. Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future Leo Obrst, Werner Ceusters and Terry Janssen
213
Subject Index
225
Author Index
227
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-1
1
Chapter 1
Introduction: Ontologies, Semantic Technologies, and Intelligence Terry JANSSENa,1, Leo OBRSTb, Werner CEUSTERSc a Lockheed Martin Corporation, USA b The MITRE Corporation, c State University of New York at Buffalo
Abstract: In recent years ontologies and semantic technologies more generally have begun to be applied to assist the intelligence community, for information integration, information-sharing, decision-support, and in many other applications. This chapter introduces the topic of the book and provides background information concerning its rationale, historical perspective, a vision for the future, and briefly describes the chapters of the present volume. Keywords: Ontology, information-sharing, intelligence community, semantic technologies.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1. Why Ontologies: What do the Intelligence Community and Its Customers Actually Need? Probably the best way to address this question is look at what the Director of National Intelligence (DNI) said in 2008 about the near future of intelligence, in the report titled The DNI’s Vision 2015 [1]. We are unable to do justice to this insightful document in this Introduction so the reader is encouraged to read it in its entirety. Some important points are pulled from this report, not in perfect context, and are quoted here [1]: 2 In this [adversarial, terrorism] environment [worldwide] the key to achieving strategic advantage is the ability to rapidly and accurately anticipate and adapt to complex challenges… [p. 6] By 2015 we will need integrated and collaborative capabilities that can anticipate and rapidly respond to a wide array of threats and risks… [p. 7] To succeed in this fast-paced, complex environment, the Intelligence Community must change significantly. For example, our counterintelligence activities face an array of new and traditional adversaries, yet we must operate within a protected information-sharing environment that challenges existing notions of security [and] of risk… [p.7] Analytic precision and accuracy will be merely the minimum requirements expected by our customers; our accuracy must be clear, transparent, objective and intellectually rich… they will expect instantaneous support ‘on demand’. 1
Corresponding Author: Terry Janssen, Lockheed Martin Corporation, IS&GS/GSS, Herndon, VA 20171, USA; E-mail: [email protected]. 2 Terms in bold are in the original document.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
By 2015, a globally networked Intelligence Enterprise will be essential to meet the demands for greater forethought and improved strategic agility. [But is important to note that] the purpose of intelligence is not merely to determine truth, but to enable decision-makers to make better choices in dealing with forces outside of their control. Intelligence helps reduce the degree of uncertainty and risk when crucial choices are made. Our measure of success is simple: did our service result in a real, measurable advantage to our side… [p. 10] [We] also need to exploit commercial technologies to develop new ways of providing service. [p. 11] By 2015, the Intelligence Community will be expected to provide more details about more issues to customers. We anticipate different types of customers – with greater expectations – and new demands to change the basic engagement model by which we serve them… [p. 11] To engage customers effectively, we must use sophisticated techniques… to elicit ‘What do you want to accomplish?’ … [and intelligence collection and analysis] will become more of an relationship than an event… [p. 12] Our analytic products will increasingly resemble customized services with an emphasis on maximum utility rather than simple releasability. Under concepts such as effects-based analysis, we will engage customers with ‘What-if?’ considerations in addition to ‘What?’ conclusions. To do so, our analysts will leverage disparate data and analytic tools and services, working in mission-focused distributed analytic networks… [p. 12] To respond to the dynamic and complex threat environment of the 21st century, our operating model must emphasize mission integration – a networked knowledge sharing model that rapidly pulls together disperse and diverse expertise and resources against specific missions… [p. 13] [It will] require multiple integrated collection systems… [and a] fully integrated processing, exploitation and dissemination architecture… [p. 14] There will be more emphasis on multi-agency teams pursuing ‘multi-INT’ collection strategies… We envision a collection community capable of rapidly fielding technological innovations that contain needed information… Above all else is the demand that the information reach those who need it, when they need it, in a form that they can easily absorb. [p. 14] The analytic community will be expected to understand and make judgments on a broad spectrum of national security threats, support a more diverse customer set, and cope with unprecedented amounts and types of information. Information overload already presents a profound challenge… [and] the analytic community has no choice but to pursue major breakthroughs in capability. Applying the principle of Collaborative Analytics analysts will be freed to work in a fundamentally different way – in distributed networks focused on a common mission. [p. 15] Information overload will be averted through sophisticated data preparation and tools. In 2015, new information will be tagged so tools can trace our data across our holdings. Analysts will use such tools to mine the data, to test hypotheses, and to suggest correlations. [p. 15] By 2015 the focus should shift from information sharing (e.g., interoperable systems, information discovery and access) to knowledge sharing [using an automated approach to the extent possible, with the humans in the loop to understand and present it accurately, and end-users to make decisions and act upon that knowledge]. [p. 17] Although we will continue to rely on commercial best of breed technologies and best practices, the Intelligence Community will still need to research, development and field disruptive technologies to maintain a competitive advantage over our adversaries. We cannot evolve into the next generation ‘S curve’ incrementally; we need a revolutionary approach… [p. 18] Creating a culture of innovation will require greater focus on advanced concepts, technology and doctrine to enhance leadership, organization alignment, and resources. [p. 19]
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
3
2. Ontologies: An Enabling Technology Meeting the vision put forward by the DNI requires a fundamental shift in perspective from data sharing to knowledge sharing. It takes no more than a moment’s reflection to realize that knowledge cannot be shared unless it can be represented and communicated. In other words, interoperation at the knowledge level means more than syntactic interoperability and the sharing of data represented in standardized formats. The receiving system must interpret the data in the manner intended by the sending system. This means either that the communicating systems use common vocabularies with agreed-upon meanings for the terms, or that the interchange be mediated by an appropriate translation capability. It means that semantic information must be explicitly represented in a form accessible to all parties to a communication, so that a shared understanding of the knowledge being transmitted can be assured. This is precisely the purpose served by formal ontology. Ontologies represent the types of entities that exist in a domain, the relationships in which they can participate, and the attributes of entities of different types. Publicly available formal ontologies provide the basis for semantic interoperability by providing standardized representations to define the semantics of knowledge being exchanged. It is the thesis of this book that semantic technology used to address problems in the intelligence community is one of the “disruptive technologies” needed to maintain our competitive edge. Application of current-generation semantic technology can provide an immediate benefit. In addition, the process of developing applications will inevitably reveal important issues for which new research is needed, thus spurring advances in technology that will result in further practical benefits.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3. A Resource: The National Center for Ontological Research (NCOR) Recognizing the need for institutionalized leadership in semantic technology, the National Center for Ontological Research (NCOR) [2] was founded in October 2005 as a partnership of groups and institutions engaging in ontological research in the United States, with the State University of New York at Buffalo and Stanford as principal administrative sites. NCOR was established to serve as a vehicle to coordinate and to enhance ontological research activities, with a special focus on the establishment of tools and measures for quality assurance of ontologies, on training in and dissemination of good practices in the ontology field, and on the creation of strategies to advance the creation of federations of principles-based ontologies which work well together within the hub-and-spokes framework of the sort currently being advanced within the US Government’s Universal Core (UCore) and Command and Control (C2) Common Core (C2Core) initiatives [3, 4]. In 2008 NCOR was contracted by the US Army Net-Centric Data Strategy Center of Excellence to create a series of ontologies for use in the biometrics and C2 (command and control) domains, and also to create a standard Common Upper Ontology based on BFO and DOLCE, for the representation of entities in real-world domains. In 2009 NCOR worked with MITRE to develop UCore-SL, a Semantic Layer for UCore 2.0, an XML-based vocabulary resource designed to support data sharing sharing between agencies within the Department of Defense, Department of Homeland Security, Department of Justice and Department of Energy.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
4
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
UCore and C2Core (and possibly other “common cores” in the future, depending on need) are vocabularies that exist in the Community of Interest (COI) paradigm within the federal government [5, 6]. In the COI paradigm, a modular architecture, as depicted in Figure 1, acts as structure for the emerging range of vocabularies. UCore spans all vocabularies and in particular immediately spans all common core vocabularies such as C2Core. The common core vocabularies in turn span all appropriate, COI vocabularies. COI vocabularies involve narrower domains and can be hierarchically structured, as shown in the figure. A COI comes into existence when two communities ascertain that they need to share information. The two communities engage in a discussion of the kinds of data they have and wish to share, and the vocabularies they use to refer to that data. Then they evolve an agreed upon vocabulary for the data they wish to share, thus developing a specific COI vocabulary.
Figure 1. Universal Core, Common Cores, and COI Vocabularies
In addition some COIs will develop semantic models of their vocabularies, i.e., ontologies. Others will develop structural models in XML Schema. An example of an ontology developed for a vocabulary is that of UCore-Semantic Layer (UCore-SL), an ontology that provides a semantics for UCore [7, 8]. At this time, UCore-SL is not officially part of UCore, but is a module under UCore Affiliates. Also, see [9, 10] for an early advocacy for common Command and Control semantics. When Figure 1 is compared with Figure 2 [11], a typical rendition of the layers of ontologies, one notices that there is somewhat of a correspondence between the layered vocabularies and the layered ontologies. However, in actuality, UCore addresses objects that typically would reside in a mid-level or lower upper level ontology, i.e., person, organization, facility, location, etc., and their properties. It is expected that UCore-SL would itself be embedded under an upper (sometimes called “foundational”) ontology such as Basic Formal Ontology (BFO) [12, 13], which is discussed in more detail below.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
5
Figure 2. Ontology Architecture and Layers
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. A Resource: The Center of Excellence in Command, Control, Communications, Computing and Intelligence (C4I Center) C4I systems are essential to our national security. Recognizing the need for a strong intellectual base for C4I, and the need for comprehensive educational programs in C4I, George Mason University established the C4I Center in 1989 as the nation's first and only civilian university-based entity offering a comprehensive academic and research program in military applications of information technology. The Center performs research and supports educational programs in a wide variety of C4I areas. A central element in the C4I Center’s research is the formal representation of knowledge about the military and intelligence domains in both interoperable and machine processable forms. The Center is actively engaged in research to apply probabilistic ontologies to predictive naval situation awareness, and is developing an open source probabilistic ontology editor and reasoner. Another strong area of research is Battle Management Language (BML), a formalism to support reasoning about military doctrine and Command and Control (C2) processes, explicitly representing military task-based operations [14].
5. An Ontology Success Story: The Peculiar History of the Gene Ontology The most conspicuous successes in ontology technology thus far have been in the biomedical field, and they result especially from the fact that the Gene Ontology (GO) [15] has been so widely used as a resource for the integration of data in the domains of molecular biology, bio-chemistry, functional genomics, proteomics, and related fields now of increasing relevance to clinical research and treatment. It is noteworthy that, in the period 2000-2007 there has occurred a 17-fold increase in use of the term
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
6
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
‘ontology’ in the abstracts collected in the standard PubMed database of medical literature, yet almost all of this increase is associated with references to the Gene Ontology. There exist over 11 million annotations relating gene products described in major databases of molecular biology to terms in the GO. These annotations create linkages between genes and proteins to specific types of biological phenomena. Data related to some 180,000 genes have been manually annotated in this way, and the GO is hereby making the results of divergent kinds of life science research comparable and integratable. The GO is a founding partner of NCOR, and NCOR has played an important role in creating the OBO (Open Biomedical Ontologies) Foundry, a federation of the GO and its sister ontologies used in biomedical research [16]. The OBO Foundry is now serving as platform for the testing of NCOR strategies for quality assurance and ontology integration. The goal of the Foundry initiative is to create the conditions under which the data generated through biomedical research and clinical care will form a growing pool, to which algorithmic techniques can be applied in ways which serve the formulation and testing of clinical hypotheses at all levels. Efforts at ontology building are still standardly conceived in pragmatic terms, as projects motivated by the need to solve problems internal to the information technology needs of specific groups or organizations. The Foundry, in contrast, reflects a view of ontologies which sees them as lying outside the realm of software artifacts created to address specific local needs and sees them rather as part and parcel of the scientific enterprise [17]. Ontologies are from this perspective resources developed for the long term, freely available for use and subject to constant criticism and update.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6. An Example Initiative: The Basic Formal Ontology The hub, in the OBO Foundry and in a series of related initiatives, is Basic Formal Ontology (BFO), a top-level ontology building on lessons learned from the development of ontologies by logicians and philosophers over more than two millennia. BFO was developed initially to support the work of experimental scientists but is now increasingly gaining acceptance as a top-level ontology standard for general use. When ontologies are developed, like database schemas, simply to address local purposes, this will not only bring limited advantages in data integration but is indeed likely to intensify the very problems of forking which ontologies were designed to counteract. BFO is designed to serve as a top-level ontology standard that will constrain the developers of domain ontologies in such a way as to work against these effects. It is designed to be a very small, a true top-level ontology. This means that, in contrast to the foundational ontologies DOLCE [18] and SUMO [19], with which it otherwise has many features in common, BFO contains no terms which would properly belong within the domains of those spoke ontologies which extend it. Use of BFO in a hub-andspokes framework thereby establishes a clear division of expertise. It provides both a simple common starting point for scientists in creating their ontologies, and also a common set of guidelines for ensuring that ontologies are thereafter developed and maintained in tandem with each other. BFO and its associated guidelines for ontology development have been refined on the basis of experience in application in the context of the OBO Foundry, and they are now increasingly being used also outside the scientific domain.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
7
7. Training and Dissemination of Ontology Technology In addressing its training and dissemination roles, NCOR has organized a series of ontology tutorials and training workshops, working closely in this also with the National Center for Biomedical Ontology. Jointly with the Ontolog Forum, the National Institute for Standards and Technology (NIST), and the newly founded International Association for Ontology and its Applications (IAOA) [20], NCOR organizes the annual Ontology Summit, held at NIST in Gaithersburg, MD since 2005. The theme for Ontology Summit 2010 is “Creating the Ontologists of the Future,” with a focus on certification individual ontologists and accreditation of institutions for courses of study for ontologists [21]. Jointly with JCOR, the Japanese Center for Ontological Research, NCOR organizes the InterOntology (Interdisciplinary Ontology) conference series held in Tokyo since 2006. Most importantly for our purposes, here, however, is the Ontology for the Intelligence Community (OIC) conference, discussed below, which has become a vibrant yearly forum for exchange of ideas on the role of semantic technology for problems of interest to the intelligence community.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
8. Ontology for the Intelligence Community (OIC): A Forum for Knowledge Interchange The first two OIC conferences were organized by NCOR Director Barry Smith, and were held in Columbia, MD, in 2006 and 2007 [22] under the heading “Towards Effective Exploitation and Integration of Intelligence Resources.” The third and fourth OIC conferences were held in 2008 and 2009, respectively, at George Mason University under the auspices of the George Mason University C4I Center [23, 24]. The OIC series was established to support the work of those who are using ontologies to develop approaches to the analysis of intelligence that will enable greater flexibility, precision, timeliness and automation of analysis and thereby maximize valuable human resources in responding to fast-evolving threats. The OIC meetings have brought together researchers and intelligence analysts from major agencies, universities and other bodies involved in intelligence activities throughout the world. It provides an important venue for exchange of ideas and sharing of insights on the role and effective use of ontologies to problems in the intelligence domain. Speakers at the first two OIC conferences included Werner Ceusters, Director of the Ontology Research Group in the New York State Center of Excellence in Bioinformatics and Life Sciences in Buffalo, who described how BFO-based ontologies are being used to support the integration of instance data to enable tracking objects of all kinds in computer representations; representatives from the FBI, NSA, CIA, other USA agencies, the UK Defence Science and Technology Laboratory, and pivotal technologists involved in applying ontology and semantic technologies to intelligence needs; and Todd Hughes of the USA Defense Advanced Research Projects Agency (DARPA). OIC 2008 invited speakers were Deborah McGuinness of Rensselaer Polytechic Institute, one of the authors of the OWL Web Ontology Language; Michael Gruninger of the University of Toronto, ontology researcher and active participant in the ISO Common Logic standardization effort; and Leo Obrst, lead of the Information Semantics Group at MITRE Corporation. The invited speakers at OIC 2009 were Chris
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
8
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
Welty of IBM Watson Research Center, Doug Lenat of Cycorp, and a tutorial was presented by Leo Obrst.
9. Semantic Interoperability for DoD and IC Systems: BML and Beyond
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Modern military operations are increasingly becoming a joint endeavor, and it is hard to imagine any situation in which a single service would operate alone. Instead, most real world scenarios involve more than one country with multiple services, creating an urgent need for international standards to support coalition interoperability. BML is designed as an unambiguous language used to: 1) command and control forces and equipment conducting military operations; and 2) to provide situational awareness and a shared common operational picture. It can be seen as a standard representation of a digitized commander's intent to be used for real troops, for simulated troops, and for future robotic forces. BML is particularly relevant in a network centric environment for enabling mutual understanding. A BML development focus has always been conveying doctrinal knowledge among military forces with diverse C2 processes. Thus, the advantages are clear of evolving BML into a full ontology on military operations based on a strong formalism for reasoning about tasks and actions. Such a C2 ontology would be a useful complement to an intelligence community ontology. Achieving semantic interoperability among DoD and IC systems is essential, but one must recall that these systems will be always operating under the “fog of war”, where incomplete data is the rule and uncertainty is ubiquitous. Unfortunately, current ontology technologies provide no support for representing and reasoning in a principled way with uncertainty and incomplete data. PR-OWL, a Bayesian first-order logic extension to the ontology language OWL, was developed at the GMU C4I Center, and is an example of current efforts to build probabilistic-aware ontologies [25]. These and other similar efforts can be seen promising enabling technologies for the vision set forth in the DNI’s Vision 2015.
10. The Contributions There are a number of themes in the chapters of this book. Many of these themes span multiple chapters, and many chapters have multiple themes. For example, many of these chapters focus on supporting intelligence analysts using semantic technologies and develop proofs of concept, especially the earlier chapters. The Ceusters and Manzoor chapter, the Ulicny, Kokar, and Matheus chapter, Obrst et al, Lee, Shalliday et al, Self et al, Ressler et al primarily address this theme. Another theme, however, is that of dealing with uncertainty in ontology-based technologies, and hence addressing the interaction between ontology and epistemology. Chapters that focus on this theme include the Janssen et al, Laskey et al, and Huhns et al chapters. Finally, the first chapter (the current chapter) and the book’s final chapter focus primarily on the impact of ontologies and semantic technologies on intelligence collection and analysis – the former attending to past and current efforts, the latter addressing issues about and potential impacts on the future. The final chapter is that by Obrst, Ceusters, and Janssen.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
9
The individual chapters address significant issues for intelligence by focusing on particular aspects of ontologies and semantic technologies. In Chapter 2, “How to Track Absolutely Everything” Ceusters and Manzoor develop a referent tracking paradigm that tracks ontology-based unique real entities and events, but also information and data elements used in information systems to describe these. By using globally unique identifiers, knowledge and assertions about the real world, beliefs and cognitive representations based on observations of collectors, analysts, etc., and metadata including provenance, can be tracked over time, through many changes. Such a referent tracking facility would greatly extend the current capabilities of data and metadata repositories, for example. In Chapter 3, the “Uses of Ontologies in Open Source Blog Mining” (Ulicny, Kokara, Matheus), the authors make the case for using ontologies to mine blog entries, which are very dynamic and difficult to automatically interpret, aggregate, and report on. As with other structured and unstructured online data, an analyst cannot read everything, so if semantic tools can enable him/her to classifiy (i.e., bin in finer granular bins with certain topics and properties of interest) the content, the machine can better assist the analyst at finding truly relevant information. The need to find information in blogs is great enough that a number of blog-specific search engines have arisen in recent years, in including Technorati, BlogPulse. In Chapter 4, “A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support” (Janssen, Basik, Dean, Smith), a possible approach to the information overload that afflicts intelligence analysts is to augment human capabilities of winnowing, interpreting, and integrating data by enlisting machines that use ontologies and other semantic technologies. Software can thereby perform lower-level knowledge functions that are comparable to what a human would perform, draw reasonable human-like inferences over that knowledge, and then present the interesting subsets of knowledge that analysts would be most interested in, and establish linkages among those knowledge components. Because the knowledge needed for many intelligence problems is a fusion of information from many intelligence disciplines, the authors developed a multi-INT ontology and framework using the HighFleet (formerly known as Ontology Works, Inc.) knowledge server. In Chapter 5, “Ontologies for Rapid Integration of Heterogeneous Data for Command, Control, & Intelligence” (Obrst, Stoutenburg, McCandless, Nichols, Franklin, Prausa, Sward), the authors present a program that uses ontologies expressed in the Semantic Web Ontology Language OWL and rules expressed in the Semantic Web Rule Language (SWRL) to provide efficient runtime reasoning for situational awareness and course of action assistance. The authors’ prototype is focused on convoy movement in a theater of operation, where the convoy has a primary path (and possibly other paths) from its origin to its destination, and potentially encounters many hostile or unknown theater objects, which may impact the convoy’s mission. The ontologies focus on theater objects and intelligence information, while the rules focus on detecting theater objects, based on incoming intelligence, determining their impact on the convoy, and then alerting the convoy to their presence, and making some recommendations for their avoidance (for example, by changing course, or assuming a defensive posture if the theater object cannot be outrun). The ontologies and rules are transformed and then compiled into a logic programming engine. Using an enterprise service bus to link multiple instances of the logic programming reasoner and providing a terrain-oriented visualization based on Google Earth, the tool enables the convoy commander to employ high-level machine assistance for recognizing potentially hazardous situations and
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
10
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
thereby potentially improve decision making. Finally, the chapter shows how the initial set of ontologies was extended for other applications, such as for space objects and events and for unmanned aerial vehicle (UAV) flight deconfliction. Chapter 6, “Ontology-Driven Imagery Analysis” (Self, Kolas, Dean) addresses the improvement of imagery analysis by annotating images with terms from ontologies. A software imagery analysis environment is described which provides a way to record structured annotations to images based on their semantics, and so enabling much richer search and more powerful exploitation of imagery. For example, using semantic annotations of images may enable images to be combined more meaningfully, by determining if sets of observations are related. How are different images related over time? What has changed from a sequence of images over time? Are there correspondences or differences between related geospatial regions over time? How can one efficiently query semantically annotated image repositories? In Chapter 7, “Provability-Based Semantic Interoperability for Information Sharing and Joint Reasoning” (Shilliday, Taylor, Clark, Bringsjord), the authors describe their system for provability-based semantic interoperability that encodes a translation graph for the ontologies to be compared (including the same ontology as modified into different versions over time), using a many-sorted logic. They also situate their approach in the wider spectrum of approaches for addressing semantic interoperability, namely via the development of schema (ontology) mappings and schema (ontology) morphisms. By first creating signatures of the ontologies in a many-sorted logic, they are able to create a translation graph that visually depicts the incremental construction and interrelation of ontology signatures, showing the transformations that that map one ontology signature into another or into a different version of its prior self (as for versioning an ontology). The translation graph is therefore a directed graph with vertices of signatures and edges that represent the relationships between signatures. Finally, the authors provide an example of their framework in action, in the UAV domain, where multiple data-source ontologies need to be integrated. In Chapter 8, “The Use of Ontologies to Support Intelligence Analysis”, Richard Lee describes the Metadata Extraction and Tagging Service (METS) effort to support intelligence analysis by providing ontologies, i.e., semantic models, rather than simply XML structural models, for tagging data of interest to analysts, including entities obtained from information extraction over datasets. A multi-INT data fusion experiment is described, which highlights the limitations of XML-based, i.e., syntactic approaches. In Chapter 9, “Probabilistic Ontologies for Multi-INT Fusion” (Laskey, Costa, Janssen), the authors focus on multi-INT fusion of heterogeneous information sources using semantic resources such as ontologies, but primarily those extended with probabilities expressed in the Probabilistic OWL (PR-OWL) formalism. PR-OWL extends the Web Ontology Language OWL with probabilistic support from Bayesian semantics, so that complex patterns of evidential relationships among uncertain hypotheses can be represented and reasoned over by machine. An early version of a reasoning engine, UnBBayes-MEBN (where MEBN stands for Multi-Entity Bayesian Networks), is described, which support PR-OWL querying and inferencing. Chapter 10, “Design Principles for Ontological Support of Bayesian Evidence Management” (Huhns, Valtorta, Wang) takes the approach that indeed Bayesian semantics should be wedded to ontologies for management of evidence. Ontologies provide knowledge about the domain, events, and causality, and then Bayesian reasoning provides evidential reasoning in their Magellan system using fragments of
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
11
situations, which can be combined to provide sets of possible situations that are consistent with the evidence. The combination, according to the authors, provide analysts with a way to gauge competing hypotheses and reduce the uncertainty of their potential outcomes. Magellan uses XMLBIF (eXtensible Markup Language Bayesian Interchange Format), along with OWL, RDF, and the SPARQL query language. Chapter 11, “Geospatial Ontology Trade Study” (Ressler, Dean, Kolas) is focused on the ontologies and other semantic standards that are valuable to geospatial analysis, of particular interest to those involved with the GEOINT discipline, but also others who require geospatial reasoning. Geospatial semantics ranges over representation of geometry (i.e., points, lines, spaces, spheres, etc., involved in addressing regions), geopolitics (i.e., borders, locations defined politically), temporal notions (how geospatial notions change over time), and geographical and topographical knowledge. As part of their trade study, the authors provide a matrixed view of the semantic standards that are possibly most applicable and important to geospatial analysis. In the concluding chapter, Chapter 12: “Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future” (Obrst, Ceusters, Janssen), the view described is that of the future and the potential promise of ontologies and semantic technologies for addressing intelligence analysis and collection. A discussion is presented on the use cases for the applications of these technologies vs. their complexity, as gauged by the required expressiveness of the semantic models needed to provide those applications. Cost must be addressed too, and measured against potential benefits, as some emerging ontology cost models intend to provide. In addition, emerging standards and technologies are discussed, from the SPARQL query language, triple stores (that are repositories of OWL/RDF instance data, structured in graphs), and rules and rule languages such as the Rule Interchange Format (RIF) – all of which are supported by rapidly emerging tools. A prospective lesson is given for intelligence that draws on the experience of using realist ontologies in biomedicine and healthcare, in the hope that some notion of the value of ontologies and semantic technologies can be indicated for intelligence collection and analysis. A distinction is then made between ontology (the ways things are) and epistemology (the ways things are believed to be or that we have current evidence for). Both technical disciplines and their tools are crucial for intelligence, and provide complementary value. A human being has only one birth date (ontology), but which of several ascribed to a particular person is correct (epistemology)? Finally, the authors express optimism about the emerging convergence of intelligence analysis and semantic technologies, and the potential value of that convergence for intelligence.
REFERENCES [1] Vision 2015: A Globally Networked and Integrated Intelligence Enterprise. Office of the Director of National Intelligence (ODNI), July 22, 2008. http://www.dni.gov/reports/Vision_2015.pdf. [2] National Center for Ontological Research (NCOR). http://ncor.us/. [3] Universal Core (UCore). https://ucore.gov. [requires authorized login] [4] Command and Control Common Core (C2Core). https://www.us.army.mil/suite/page/473883. [requires authorized login] [5] Communities of Interest (COI) – Home, Assistant Secretary of Defense (Networks and Information Integration) Department of Defense Chief Information Officer (ASD(NII)/DoD CIO). http://www.defenselink.mil/cio-nii/sites/coi/coi.shtml.htm.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
12
T. Janssen et al. / Introduction: Ontologies, Semantic Technologies, and Intelligence
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[6]
DoD Directive 8320.2. Data Sharing in a Net-Centric Department of Defense. December 2, 2004, Certified Current as of April 23, 2007. http://www.defenselink.mil/cionii/sites/coi/Governance/832002p.pdf. [7] Smith, Barry; Lowell Vizenor; and James Schoening. 2009. Universal Core Semantic Layer. In Proceedings of the Ontologies for the Intelligence Community (OIC) conference, October 20-22, 2009, George Mason University, Fairfax, VA. [8] Smith, Barry. 2009. Universal Core Semantic Layer (UCore SL): An Ontology-Based Supporting Layer for UCore 2.0, UCore Conference, MITRE, McLean, VA. September 23, 2009. [9] Chaum, Erik; Richard Lee. Command and Control Common Semantic Core Required to Enable Netcentric Operations. AFCEA-George Mason University Symposium "Critical Issues in C4I" May 20-21, 2008. [10] Winters, Leslie and Andreas Tolk. 2009. C2 Domain Ontology within Our Lifetime. In Proceedings of the 14th International Command and Control Research and Technology Symposium (ICCRTS), Jun 1517, 2009, Washington, DC. [11] Semy, S.; Pulvermacher, M.; L. Obrst. 2005. Toward the Use of an Upper Ontology for U.S. Government and U.S. Military Domains: An Evaluation. MITRE Technical Report, MTR 04B0000063,November 2005. http://www.mitre.org/work/tech_papers/tech_papers_05/04_1175/index.html. [12]. Basic Formal Ontology (BFO). http://www.ifomis.org/bfo. [13] P Grenon, B Smith. 2004. SNAP and SPAN: Towards dynamic spatial ontology. Spatial Cognition and Computation: An Interdisciplinary Journal, Vol. 4, No. 1. (2004), pp. 69-104. [14] Battle Management Language (BML). http://c4i.gmu.edu/BML.php. [15] Gene Ontology. http://www.geneontology.org. [16] Open Biomedical Ontologies (OBO) Foundry. http://obofoundry.org. [17] Smith, Barry. “Ontology (Science)”, in C. Eschenbach and M. Gruninger (eds.), Formal Ontology in Information Systems. Proceedings of the Fifth International Conference (FOIS 2008), Amsterdam: IOS Press, 21-35. http://precedings.nature.com/documents/2027/version/2. [18] Descriptive Ontology for Linguistic and Cognitive Engineering . DOLCE. http://www.loacnr.it/DOLCE.html. [19] Suggested Upper Merged Ontology (SUMO): http://www.ontologyportal.org/. [20] International Association for Ontology and its Applications (IAOA). http://www.iaoa.org/. [21] Ontology Summit 2010: Creating the Ontologists of the Future. http://ontolog.cim3.net/cgibin/wiki.pl?OntologySummit2010. [22] Hornsby, Kathleen Stewart, ed. 2007. Proceedings of the Second International Ontology for the Intelligence Community Conference (OIC) 2007, November 28-29, 2007, Columbia, MD, USA. CEUR Workshop Proceedings, Volume-299. http://ftp.informatik.rwth-aachen.de/Publications/CEURWS/Vol-299/. [23] Laskey, Kathryn Blackmond; Duminda Wijesekera. 2008. Proceedings of the Third International Ontology for the Intelligence Community Conference. Fairfax, VA, USA, December 3-4, 2008. CEUR Workshop Proceedings Volume-440. http://sunsite.informatik.rwth-aachen.de/Publications/CEURWS/Vol-440/. [24] Costa, Paulo; Kathryn Laskey; Leo Obrst, eds. 2009. Proceedings of the 2009 International Conference on Ontologies for the Intelligence Community. Fairfax, VA, USA, October 21-22, 2009. CEUR Workshop Proceedings Volume-555. http://sunsite.informatik.rwth-aachen.de/Publications/CEURWS/Vol-555/. [25] Costa, Paulo; Kathryn Blackmond Laskey. 2006. PR-OWL: A Framework for Probabilistic Ontologies. Proceedings of the Fourth International Conference on Formal Ontology in Information Systems. November 2006. http://ite.gmu.edu/~klaskey/papers/FOIS2006_CostaLaskey.pdf.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-13
13
Chapter 2
How to Track Absolutely Everything Werner CEUSTERS1, Shahid MANZOOR Ontology Research Group, New York State Center of Excellence in Bioinformatics and Life Sciences, University at Buffalo, USA
Abstract: The analysis of events prior to and during September 11 revealed that a smooth execution of the intelligence process is hampered by inadequate information sharing. This caused a rethinking of the intelligence process and a transition towards a ‘Globally Networked and Integrated Intelligence Enterprise’ with the goal that more detailed, tagged, and, therefore, traceable, information will reach those who need it, when they need it, and in a form that they can easily absorb. We present the referent tracking paradigm and its implementation in networks of referent tracking systems as an enabling technology to make this vision come true. Referent tracking uses a system of singular and globally unique identifiers to track not only entities and events in first-order reality, but also the data and information elements that are created to describe such entities and events in information systems. By doing so, it meets the requirements of the Nation’s Information Sharing Strategy. Keywords: referent tracking
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1. Introduction Intelligence, as defined by the Central Intelligence Agency (CIA), is ‘the information our nation’s leaders need to keep our country safe’ [1]. This information is produced by the US Intelligence Community (IC), i.e. the departments and agencies cooperating to fulfil the goals of Executive Order 12333 which stipulates that ‘The United States intelligence effort shall provide the President and the National Security Council with the necessary information on which to base decisions concerning the conduct and development of foreign, defense and economic policy, and the protection of United States national interests from foreign security threats’ [2]. This is achieved through the performance of what is called the ‘intelligence process’ which consists of five steps: (1) the determination of the information requirements, (2) the collection of raw data, (3) the processing of the raw data into forms that are more usable for intelligence analysts or other consumers, (4) the integration, evaluation and analysis of the data in order to generate reports satisfying the requirements, and (5) the dissemination of the results to the appropriate level [3]. This last step, typically, leads to new information requirements which initiate a new cycle of the intelligence process.
1 Corresponding Author: Werner Ceusters, Ontology Research Group, New York State Center of Excellence in Bioinformatics and Life Sciences, 701 Ellicott street, Buffalo NY, 14203, USA; E-mail: [email protected]
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
14
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1.1. Challenges and barriers Ideally, the information that is finally disseminated is (1) reliable, thus corresponding faithfully to what is the case in reality, (2) complete, such that nothing what is essential or required for the consumer to make adequate decisions is missing, (3) relevant, such that decisions can be made efficiently, and (4) timely, guaranteeing that decisions can be made early enough for the resulting actions to have the desirable effect. Unfortunately, this ideal is very hard to achieve because of many barriers and challenges [4]. A large number of these challenges are brought about by the multiplicity of agencies, organizational levels within these agencies and information consumers that are involved. Although each step in the intelligence process comes with its own challenges, the multiplicity of involved actors affects primarily the information requirements assessment and the data-integration and analysis steps. So do the information requirements that a specific organizational level has to take into account not only consist of the external requirements put forward by the consumers to whom intelligence reports of a specific nature and content need to be delivered, but also of the internal requirements which determine what sorts of detailed information elements are required and accessible to provide high quality reports. The integration, evaluation and analysis step can be hampered by insufficient lower-level data (both quantitatively and qualitatively), wrong information, and lack of meaningful data linkage. The net effect is that the reliability, completeness and relevancy of the resulting conclusions suffer considerably. Although these three notions are intuitively straightforward, they can be defined in various ways and for each such way, objective quantification is hard, if possible at all. Furthermore, these notions are not entirely independent from each other. Reliability, for instance, relates to accuracy which itself relates to relevancy: the more a measurement is accurate, the more reliable it seems to be, yet, the relevancy of it might diminish depending on the objectives of the intelligence effort: whereas providing information on the duration of intercontinental flights in minutes to compare the performance of foreign carriers with that of national ones seems reliable, accurate and relevant, doing so in hours is hardly reliable, while in seconds for sure not relevant. Redundancy of information elements within a collection of information will not harm the completeness and relevancy of that collection as a whole, but for sure the relevancy of the redundant elements themselves. At the other hand, from a second order perspective, the presence of redundant information, if obtained from various independent sources, might be an indication for the reliability of the collection. 1.2. Intelligence and Security Informatics The analysis of events prior to and during September 11 revealed that a smooth execution of the intelligence process is hampered by inadequate information sharing [5]. Not only are there legal and cultural barriers to information sharing – the ‘need-toknow’ culture during the Cold War is now recognized to be a handicap in dealing with terrorism and other asymmetric threats [6] – it is also technically very difficult to integrate and combine data that are stored in different database systems running on different hardware platforms and operating systems [7]. Although the Office of Homeland Security, in 2002, identified information sharing across jurisdictional boundaries of intelligence and security agencies as one of the key foundations for
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
15
ensuring national security [8], the appropriate infrastructure is not yet there. This recognition led to the development of a new science: ‘Intelligence and Security Informatics’ (ISI) [9], which is commonly defined as ‘the study of the use and development of advanced information technologies, systems, algorithms, and databases for national- and homeland-security-related applications through an integrated technological, organizational, and policy-based approach’ [10]. ISI tries to overcome the barrier that data which reside in distinct data sources are organized in different schemas, and therefore are difficult to integrate. But still, once some sort of integration has been achieved, it remains often very hard to determine, for instance, whether two distinct pieces of information are about the same entity or which piece of information is correct when several pieces about the same entity can’t be true at the same time. As an example, a case study in a local police department revealed that more than half of the suspects had either a deceptive or an erroneous counterpart existing in the police system: 42% of the suspects had records alike due to various types of unintentional errors, while about 30% had used intentionally a false identity [11]. Deception is in the context of ISI a very hard problem indeed; it is not limited to providing false identities, but includes also ‘cognitive hacking’ which involves disinformation attacks on the mind of the end user of a networked computer system such as a computer connected to the Internet [12]. Identifying such attacks is crucial in an era in which the Intelligence Community seeks to make better use of Open Source Information (OSINT) [13].
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1.3. Vision 2015 To further advance the modernization of the information technology within the Intelligence Community, the Office of the Director of National Intelligence [14] published in February 2008 its ‘Information Sharing Strategy’ report [6], followed in July by the ‘Vision 2015’ document [15]. They key idea, first introduced in the National Intelligence Strategy [16], is the move towards a ‘Globally Networked and Integrated Intelligence Enterprise’ with the goal that more detailed, tagged, and, therefore, traceable, information will reach those who need it, when they need it, and in a form that they can easily absorb. Efforts in these directions are expected to create the ability to develop, digest, and manipulate vast and disparate data streams ‘about the world as it is today’ by means of tags that enable the use of tools that can ‘trace related data across our holdings, to mine the data, to test hypotheses and to suggest correlations’ in addition to ‘measuring performance’ [15]. The key characteristics of the new information sharing model are [6]: C1. C2. C3. C4. C5. C6.
‘responsibility to provide’: sharing intelligence data while still addressing the need to protect privacy, civil liberties, and sources and methods; enterprise-centric: providing services across agencies, partners, and international borders for multiple mission use; mission-centric: able to adapt rapidly to changing needs and new partners; information-centric: security built into the data and environment using tags; attribute-based: access based on attributes that go beyond security classification (e.g. environmental, affiliation, mission focus, etc.); data ‘stewardship’ (rather than data ‘ownership’), focusing on quality and reusability of data rather than, but not excluding, protection.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
16
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
1.4. Tagging, indeed, but what and how? Because ‘tagging’ seems to be an important part of the proposed solution to make this vision come true, the issues that we address here are (1) where the tags should come from, (2) what it is that should be tagged, and (3) according to what sort of logical schema data and tags should be organized in order for the data to track faithfully what is going on in the world. We argue, in response to each of the issues just mentioned, (1) that the tags should correspond to the terms (or codes) which are used as representations for universals and defined classes in realism-based ontologies, thus covering what is generic, (2) that what is tagged should not only be the data about firstorder entities (persons, vehicle movements, parcels, disease outbreaks, …), but also how and by whom (and what) these data are generated and manipulated, and (3) that the data should be organized in a structure which mimics the structure of that part of reality that is described by the data and that is capable to reflect all sorts of changes that reality undergoes in the course of history.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2. Naive tagging Today, information is primarily maintained in information systems which consist of data repositories that contain data in either unstructured form (such as free text or digital multi-media objects) or structured form, the latter being such that numerical information is expressed by means of numbers, and non-numerical information by means of codes or terms associated with what is commonly called ‘concepts’, taken from different sorts of terminologies (such as vocabularies, nomenclatures, concept systems, and so forth) as they are offered in terminology servers. Since data in structured form are better suited to provide software agents with a deep understanding of what the data represent, considerable efforts are spent to turn unstructured data into structured data, at least partially. However, whether data are captured in structured form when entered, or rendered as such afterwards using text and image analytics software which add tags corresponding to concepts, current information systems exhibit at least two major shortcomings as far as concept-based tagging is concerned: (1) formal impreciseness about what is tagged, and (2) incompatibility of distinct tagging systems. 2.1. Missing the point(ers) Mainstream information systems do not offer a mechanism to unambiguously determine in each individual case what entity in reality a concept from a terminology server is used to relate to. As a consequence, information systems thus conceived work with instances of data, but algorithms working on such data have no clue what the data are about, i.e. about what specific entity in reality each specific data-element contains the information. If, for example, a driving license number is used in an information system, it is often not formally clear whether the number is used to denote the driving license of a person or that person itself. As a further example, if in an information system the gender of a person is stated to be ‘unknown’, then it is often not formally clear whether this means either (1) that the person does have a gender which is one of the scientifically known gender types
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
17
such as female, male, mosaic, etc., but that information of the precise gender of that person is not available in that information system, or (2) that the gender of that person is known to be of a type which scientifically has not yet been determined. Another example is that if at a certain time the gender of a specific person is registered in some information system as ‘male’, and at a later time as ‘female’, then there is, under existing data storage paradigms, no way to derive from this change whether the change in the information system reflects (1) a change in reality, for instance, because the person underwent transgender surgery, (2) a change in what became known about reality: the person’s gender might because of a congenital disorder not have been determinable at the time of birth, but only later after several investigations, or (3) that there was no change in reality or what we know about it, but that at the time of the first entry a simple mistake was made. One can even imagine a fourth possibility, namely that the meaning of the word ‘female’ would have been changed. The latter might seem to be too far fetched – in fact, this did never happen for the words ‘male’ and ‘female’ – but there are several examples in the past that come close. The title ‘Chief Executive Officer’, for instance, was introduced in Europe in the late eighties, replacing titles such as ‘Director General’ or ‘Managing Director’. A change in title, in those days, for sure did not entail a change in position or power of the person to whom the new title was attributed. These types of issues are insufficiently addressed in modern Semantic Web applications because they are not yet generally recognized: attempts to address them are sparse.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2.2. Missing semantics The most recent hype in information system networking is semantic interoperability. By ‘semantic interoperability’, it is meant the ability of two or more computer systems to exchange information and have the meaning of that information automatically interpreted by the receiving system accurately enough to produce useful results, as defined by the end users of both systems. Current attempts to achieve semantic interoperability rely on agreements about the meaning of so-called concepts stored in terminology-systems, such as nomenclatures, vocabularies, thesauri, or ontologies, the idea being that if all computer systems use the same terminology, they can understand each other perfectly. The reality is, however, that, rather than one such terminology being generally adopted, the number of terminology-systems with mutually incompatible definitions or non-resolvable overlap amongst concepts grows exponentially, thereby contributing more to the problem of semantic noninteroperability than solving it. Of course, ontologies developed for different purposes can only reasonably be expected to have partial overlap, but more efforts should be conducted to exploit overlap when resolvable.
3. Fundamentals of realism-based ontologies and data repositories In contrast to traditional terminology approaches, the realist orientation in terminology and ontology is based on the view that terms in terminologies are to be aligned not on concepts but rather on entities in reality [17]. Central to this view are three assumptions [18]. The first is that reality exists objectively in itself, i.e. independent of the perceptions or beliefs of cognitive beings. Thus not only do a wide variety of entities
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
18
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
exist in reality (human beings, terrorists, guns, attacks, countries, ...), but also how these entities relate to each other (that human beings are citizens of countries, that in most attacks guns are used, and so forth) is not a matter of agreements made by scientists or database modellers but rather of objective fact. The second assumption is that reality, including its structure, is accessible to us and can be discovered: it is scientific research that allows human beings to find out what entities exist and what relationships obtain between them. It is intelligence analysis that allows analysts to find out which specific human beings are terrorists. The third assumption is that an important aspect of the quality of an ontology or terminology is determined by the degree to which the structure according to which the terms are organized mimics the pre-existing structure of reality. In the context of information systems, it means that an important aspect of the quality of an information system is determined by the degree to which (1) its individual representational units correspond to entities in reality, and (2) the structure according to which these units are organized mimics the corresponding structure of reality.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3.1. Faithful representations The above assumptions form the basis for distinguishing between three levels of reality which have a role to play wherever ontologies are used as artifacts for annotation and tagging, and wherever automated or semi-automated reasoning is required to be able to deal with an overload of information, parts of which can be expected to be wrong. Ontologies and data repositories for the intelligence community are no exception to this. The three levels are [18]: • Level 1: the (first-order) reality ‘in the field’: the persons that are tracked, the events that are monitored, the users of the information system, and so forth; • Level 2: the beliefs and cognitive representations of this reality embodied in observations and interpretations on the part of observers, data collectors, analysts and others; • Level 3: the publicly accessible concretizations of such cognitive representations in representational artifacts of various sorts, of which ontologies, terminologies and data repositories are examples. Ontologies contain typically representations for what is generic, thus representing entities such as person, weapon, war, and so forth. Repositories cover what is specific, thus holding representations for entities such as President George W. Bush Jr., the gun that killed John F. Kennedy, The Gulf War, etc. In line with the theory of granular partitions [19] we argue that complex representations should be composed in modular fashion of sub-representations built out of representational units that are assumed to correspond to portions of reality (POR). Some characteristics of the units in a representation created for intelligence purposes are: • each such unit is assumed by the authors of the representation to be veridical, i.e. to conform to some relevant POR as conceived on the best understanding (which may, of course, rest on errors). Thus if in a data repository a representational unit standing proxy for a specific person is associated with the name ‘George Bush’, then, under the realist paradigm, we assume that a person with this name exists or has existed (that on the basis of the name only
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
•
•
19
it cannot be determined which specific person is meant, does not make the unit non-veridical); several units may correspond to the same POR by presenting different though still veridical views or perspectives, for instance at different levels of granularity (one thing may be described both as being brown and as reflecting light of a certain wavelength, or one event as an event of administering and of consuming drugs); what units are included in a representation depends on the purposes which the representation is designed to serve.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3.2. Keeping track of changes The real world is subject to constant change, and so also is our knowledge thereof. To keep track of these two sets of changes, any representation concerning a relationship between entities should be associated with at least the following pieces of information: (P1) an index for the time period during which the relationship obtains, (P2) an index for the time at which the representation is made, i.e. the time at which the relationship is (believed to be) known, (P3) an index for the time that piece of information is made available in the system, and (P4) an identifier standing proxy for the author of the representation. Keeping track of these various types of information makes it possible not only to track reality faithfully from an individual analyst or agency perspective, but also to preserve the knowledge about what was known by whom and at what time after information which was residing originally in distinct systems becomes merged. It also allows to assess whether information is disclosed in a timely fashion. Suppose, for instance, that at time t10 it is known by analyst A1 that suspect S was since t9 member of group G of possible terrorists, but that an entry to that effect in the information system of his agency is made available not earlier than at t11. Thus between t10 and t11, that information was not accessible. Furthermore, in reality, it might be that S was already member of G at t5. That information might have been known in another agency since t6, and made available at that time in their information system. When the information in the two systems becomes merged, for instance after the Vision 2015 situation becomes reality, it can still be assessed what was known at each point in time in each agency.
4. Fundamentals of Referent Tracking Referent Tracking (RT) is a paradigm for information management that is distinct from other approaches in that each data element has to point to a portion of reality in a number of predefined ways (Figure 1). It has been introduced in the context of Electronic Health Record keeping [20], but its applicability is wider than that, examples being digital rights management [21] and corporate memories [22]. By ‘portion of reality’ is meant any individual entity or configuration of entities standing in some relation to each other. By ‘entity’ is meant anything that exists or has existed in the past, whatever its nature. A ‘configuration’ is a portion of reality which is not an entity in its own right. Whereas a specific person, his or her activities, the social network he belongs to, the analyst examining information about that person, and that
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
20
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
examination itself are each individual entities, the configuration that the activities of this person are being monitored by an intelligence agency, or his or her being part of that social network, is not. Another example of a configuration is the being of an engine in a car. Both that car and that engine are entities, but the fact that that engine is in that car, is not. If that engine would not be in the car, but, for instance be placed by a mechanic outside the car for repair purposes, still the very same entities (the car and the engine) would be involved, but there would be another configuration. Within the RT paradigm, configurations are referred to by means of a data type called a ‘RT-tuple’, whereas entities are represented by means of a data type called ‘representation’. Both data types come in several forms depending on the nature of the portion of reality they carry information about (see section 6). RT, through its data types, allows also for the drawing of an explicit distinction made in Basic Formal Ontology (BFO) [23] between specific entities called ‘particulars’ from generic entities called ‘universals’. Particulars are specific and unique entities, unique in the sense that they each occupy specific regions of space and time, and that nothing other than a specific particular can be that particular. Examples are concrete persons such as George W. Bush Jr. and George W. Bush’s heart. Some particulars, such as each of four tanks in a specific squadron, may exactly look the same, but they are still distinct particulars. One can be destroyed, while the other three remain intact. For particulars of specific interest, such as persons, ships, and hurricanes, proper names are used to mark the importance of their individual identity. For other particulars, such as cars or pieces of complex equipment, serial numbers are used for unique identification purposes. Portion of Reality Entity
Configuration
represents
Universal Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Particular
contains
Non-referring particular
Information bearer Representation
is about
RT-tuple
corresponds-to
Defined class Representational unit Denotator denotes
CUI
IUI
UUI
denotes denotes
Figure 1: Reality and representations
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
21
Universals, in contrast, are such that they are (1) generic and (2) expressed in language by means of general terms such as ‘person’, ‘ship’, and ‘car’, and (3) represent structures or characteristics in reality which are exemplified in an open-ended collection of particulars in arbitrarily disconnected regions of space and time. Through yet other data types, RT makes explicitly the distinction between two sorts of particulars: those that are ‘information bearers’, and those that are not; the latter called ‘non-referring particulars’. Whereas non-referring particulars belong exclusively to the first level of reality – they are pure first-order entities – information bearers play a role in both levels 1 and 3. Examples of information bearers are a piece of paper containing a text about a person’s educational background, and a digital object, such as an image of a person in an information system. Information bearers are about something else, while nonreferring particulars are not about something else. Information bearers can be about not only non-referring particulars, an example being the driving license card of a person which is about its driving rights, but also about other information bearers, an example being a textual description of a specific person’s driving license, stating, for instance, that the name of the driver is almost not readable. A copy of such a driving license can be at the same time about both the card and the rights enjoyed by the license holder.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4.1. Relations between information bearers and portions of reality RT distinguishes explicitly and formally between various relations that obtain between information bearers and the various types of portions of reality it is capable of describing. These relations are: • is-about, which obtains between an information bearer and a portion of reality, such as, for example, a book about George W. Bush Sr. (the book being an information bearer) being about parts of the life of George W. Bush Sr. and his environment (a combination of several configurations in which figure, besides George W. Bush Sr., various other entities such as his advisors, friends, trips, speeches, and so forth). • corresponds-to, which obtains between an RT-tuple and a configuration; • represents, which obtains between a specific subtype of information bearer, namely what we call a ‘representation’, and some further entity (or collection of entities). A representation is thus such that (1) the information it contains is about an entity, and not a configuration, external to the representation and (2) it stands for or represents that entity. Examples are an image, record, description or map of the United States. Note that a representation (e.g. a description such as ‘the man over there on the corner’) represents a given entity even though it leaves out many aspects of its target. • denotes, which obtains between data-elements expressed by means of a data type that we call ‘denotator’ (see further) and an entity. • contains, which obtains between information bearers and can be used to express what pieces of information of a specific data type are parts of other pieces of information. An example is a digital message which contains RTtuples describing configurations of entities in which a specific person figures.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
22
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
4.2. Denotators A denotator is a representational unit which denotes directly an entity in its entirety without providing a description. An example of a denotator is the string ‘Bush’ in the sentence ‘President Bush visited Europe several times’ when, whether or not known to the reader of the sentence in question, the writer had in mind a particular Bush, whether George Bush Jr. or George Bush Sr. The sentence itself is an information bearer according to our terminology. Because many representations are built out of constituent sub-representations as their parts, in the way in which paragraphs are built out of sentences and sentences out of words, RT uses the data type called ‘representational unit’ to represent such smallest part. Examples are: icons, names, simple word forms, or the sorts of alphanumeric identifiers found in digital records. Note that many images are not composite representations since they are not built out of smallest representational units in the way in which molecules are built out of atoms (Pixels are not representational units in the sense defined.) [18]. RT distinguishes explicitly and formally between three types of denotators, referred to respectively as ‘IUI’, ‘UUI’ and ‘CUI’. An IUI – abbreviation for ‘Instance Unique Identifier’ – is a denotator in the form of a persistent, globally unique and singular identifier which denotes (or is believed to denote) a particular and which is managed in a referent tracking system. A UUI – for ‘Universal Unique Identifier’ is a denotator which denotes a universal within the context of a realism-based ontology. A CUI – abbreviation for ‘Concept Unique Identifier’ – is a denotator for entities of a type that is commonly and ambiguously called a ‘concept’ [17], but which in BFO is called a ‘defined class’, and defined as a subset of the extension of a universal which is such that the members of this subset exhibit an additional property which is (a) not shared by all instances of the universal, and (b) also might be exhibited by particulars which are not instances of that universal.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5. Referent Tracking System A referent tracking system (RTS) is a special kind of digital information system which keeps track of (1) what is the case in reality and (2) what is expressed in other information systems about what is believed to be the case in reality. It does this unambiguously by means of the data types just sketched – in the first place resorting to IUIs – using principles and methods that assure – modulo the occurrence of errors, the resolution of which is also covered by the RT paradigm – that an IUI is (1) persistent because once created in a RTS it is never deleted, (2) globally unique because an IUI denotes only one entity within an RTS, and (3) singular because within an RTS, there is only one IUI for a specific entity. Figure 2 shows the various components of an RTS and how an RTS can be used in association with external information systems and terminology (or ontology) servers. The direction of the arrows depicted therein shows the processing of service requests, the communication, however, being bi-directional to accommodate responses to the requests.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
23
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
5.1. Components of a referent tracking system Figure 2: Components of a referent tracking system User
User
External Information System
Referent Tracking System
IUI Component
RTS Proxy Peer
Referent Tracking System User Interface(s)
Referent Tracking Server (Peers) Reasoning Server Referent Tracking Data Access Server
RTS Server Proxy Peer
Internal Ontology Referent Tracking Data Store
Terminology Server
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Vocabulary or
Thesaurus
or
Nomenclature
or
Concept System
or
Realism-based Ontology
An RTS includes at least four types of components: (1) one or more referent tracking servers, (2) one or more referent tracking system user interfaces, (3) an RTS Proxy Peer, and (4) an RTS Server Proxy Peer. The components execute on one or more processors, computers or computing devices. Further, all of the components of an RTS can run on one computing unit; one or more components can run on one computing unit, while others run on one or more other computing units; or the components may be distributed among various computing units. Each referent tracking server includes a data access server [24], which manages service requests coming from an RTS Proxy Peer or RTS Server Proxy Peer and which performs data manipulation on the server’s main component: a referent tracking data store thereby assisted by a reasoning server. The latter performs various sorts of reasoning functions by combining data from the data store with information coming from external terminology servers. The type of reasoning that can be performed depends on whether the terminology server contains nomenclatures, vocabularies, thesauri, and so forth. The referent tracking server comes also with an internal ontology which is a repository dedicated, for instance, to store information obtained during the initialization process, access control information about authorized users and usages, and so forth. The referent tracking system user interfaces allow direct users of the RTS to perform (1) a variety of management functions such as registering new external information systems, configuring a referent tracking server, adding additional referent
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
24
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
tracking servers, and so forth, and (2) content functions such as running patternmatching algorithms on the data in the referent tracking data store to detect inconsistencies, invoke triggers and alerts, perform population-based studies, and so forth. 5.2. Layered architecture Figure 3 provides further details regarding the four-layered architecture of a RTS. The outer layer is a client side layer which connects to a RTS client which is typically a third party information system or a middleware component. The latter send a query to a Proxy Peer in the network layer that forwards the request to the appropriate RTS server in the network. During execution of the query, the RTS server calls the services of the RTS core API to retrieve the results from the Database Management System databases (DBMS) that constitute the data source layer. A referent tracking data store includes, for instance, two parts: an IUI-repository and a referent tracking database (RTDB). The IUI-repository includes, as explained in section 6, the A-tuples and D-tuples which provide meta-information about information about first-order entities. The IUI-repository thus manages the statements about the assignment of IUIs to particulars, and provides a central repository of IUIs to the RTS. The RTDB is a database of statements representing the detailed information about particulars, examples being ‘#IUI-1 instantiates the universal Person’ and ‘#IUI-1 has the name ‘John’’. The RTS Core layer implements the business logic of RT, namely, the insertion and retrieval of RT-tuples in any of its databases. The IUI-repository and RTDB components are implemented through a series of application programming interfaces (APIs). The IUI-repository includes services to search particular representations and to insert new ones in its corresponding DBMS. Similarly, the RTDB components provide API get methods to search and create methods to insert tuples in its database. Information System
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Client side layer
Referent Tracking System RTS Proxy Peer
Network layer Referent Tracking Data Access Server RTS Services Server
RTS Services Factory
Referent Tracking Data Store Referent Tracking Database
IUI Repository
RTS core layer Database Managing System
Database Managing System
RTDB Tables
IUI repository Tables
Data source layer
Figure 3: Layered implementation of a referent tracking system
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
25
The IUI-repository and RTDB components are implemented independently of any specific DBMS (e.g. MYSQL, HSQL). DBMS support is controlled by DBMS specific driver components, such as for MYSQL and HSQL. Insertion services allow inserting a new RT tuple into the repository. The RTtuples are inserted in a transaction, which is an information unit. As an example, entering a patient's blood pressure could involve a couple of RT statements which could include one or more RT-tuples. All tuples in a transaction are guaranteed to be committed in the data store. In case where either a system breaks down (by power failure or other means) or a user aborts the operation (e.g. a user closes/cancels the data entry screen while entering data), no partial information is stored in the data store. This service marks the start of a transaction for a specific session of a user. The RT paradigm does not allow any deletion operation in order to be able to always return to a state of the database as it was at a certain time in history. To prevent mistakes in creating new tuples in the IUI-repository, the tuples are cached right after the create operation. The client can remove or modify the tuples from the cache, as long as the commit service has not been called.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5.3. Networks of Referent Tracking Systems Since referent tracking is to make reference to entities in reality by means of singular and globally unique identifiers, an ideal setup is one in which only one RTS is used worldwide. More realistic, however, is the adoption of the RT paradigm in a step-wise fashion: each organization first installs its own RTS, and afterwards connects them in expanding networks. To support this evolution, as shown in , the RTS is built upon Peer to Peer (P2P) technology, enabling data sharing in such a way that a search query can be executed concurrently over distributed RTS servers (peers). In an RTS P2P network, a client thus sends a query to an RTS server which besides executing the query itself can forward it to other connected RTS servers for subsequent execution. Each peer then collects the results and sends them to the requesting peer. Finally, the RTS server who received the initial request returns the aggregated results to the client. Furthermore, an RTS P2P application is capable of database load sharing over multiple RTS server peers such that the network behaves as a singular database. This capability is useful in cases where a very large database cannot be hosted on a single machine, for instance because of computational limits. It includes also capabilities for discovering a new peer in a network, for authenticating users, and for ensuring secure communication. shows an example of an RTS network in which three organizations, A, B and C, are running their own RTS peers. The peers are installed so that they are not directly known outside their corresponding organization’s environment. In organization A, the Server Peers are alike in all respects and implement the objective of distributing a very large database load. When Information System A sends a search query to the RTS Proxy Peer within organization A, the latter forwards the query to all available Server Peers (A1, A2, …) in the organization which concurrently execute the query and return the results to the Proxy Peer that finally sends the results to the Information System. Each organization can form its own local group of servers whose membership is not known outside the organization. This protects against unauthorized access to the peers
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
26
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
in the group. Controlled public access to each organization’s data is offered through the Proxy Server peers. The separation of local peer advertisement within an organization from public (outside the host organization) contexts is the basis for the Information System A
Information System C
Referent Tracking System A
Referent Tracking System C
Referent Tracking Server A1
RTS Proxy Peer
Referent Tracking Server C1
RTS Proxy Peer
Referent Tracking Server A2
Referent Tracking Server C2 RTS Server Proxy Peer
Referent Tracking Server A3
…
Referent Tracking System B
Referent Tracking Server B1
RTS Server Proxy Peer
Referent Tracking Server B2
RTS Server Proxy Peer
Referent Tracking Server C3
…
Information System B
RTS Proxy Peer
Referent Tracking Server B3
… Figure 4: Peer-to-Peer implementation of Referent Tracking Systems
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
implemented security layer. The peers which are known locally provide full access to the local database, and the peers which are known publicly provide very restricted access to the database (they might, for instance, allow only searches over certain sorts of RT-tuples as explained further). 5.4. Reasoning services Reasoning is a part of the RTS and its purpose is double. The first one is to prevent inconsistent data from being entered. By ‘inconsistent data’, we mean here data that cannot be true at the same time under the ontologies in whose terms the data are expressed. It is of course plausible that some analysts might be under the impression that, say ‘John is in Paris’ while others think that ‘John is in London’. That analysts think different things is not inconsistent, but clearly they cannot both be right. The second purpose for having reasoning services is to draw inferences during the execution of the search queries using the generic knowledge expressed in the terminology and ontology servers used to annotate the data and by exploiting the reasoners that operate on them. Various third party reasoners exist, some being specific to a particular knowledge source, some coming with a public DIG (Description Logic Implementation Group) interface for description logic representations, while others use directly OWL-DL (Web Ontology Language-Description Logics).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
27
In order to be able to deal with terminology servers and the various sorts of knowledge sources they offer (nomenclatures, thesauri, ontologies, ...), the RTS includes a Reasoning API which helps in sending reasoning queries uniformly to different terminology servers. The Reasoning API has an abstract class called OntologyConnector, which provides an interface to the external terminology systems by means of services. The interpretations of the OntologyConnector services are specific to a particular terminology server; therefore, a separate implementation of the OntologyConnector is required for each terminology server which is used to annotate the particulars in the RTS. Description logics are widely used for building ontologies. The reasoners for such ontologies may take from 1 second to a day to compute inferences over the ontology classes depending on their size and definitional complexity. Therefore, instead of always directly communicating with the reasoners for each ontology when a specific query is launched, the RTS is able to store these queries and the results that have been returned by these reasoners as an inference graph in a database [24]. Thus, because the execution time of the OntologyConnector services can range from milliseconds to minutes depending on the query execution time in the external terminology system, the OntologyConnector caches the results returned from these systems. The cache is stored, for instance, in a RDBMS. During the execution of any of the OntologyConnector services, it first searches in the cache.
6. Referent Tracking Data Elements: RT-tuples RT-tuples, although all corresponding to portions of reality, come in various flavors depending on the sort of information they contain.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.1. A-tuples A-tuples correspond to the assignment by some agent of an IUI to a particular. For the typical case, that particular is a pure first-order entity such as a specific person or a specific building about which information is to be stored in the RT system. However, by storing tuples, the RT system itself acts as an agent that assigns IUIs to the tuples itself. Indeed, for each insertion of an A-tuple, there is a corresponding insertion of a D-tuple that contains information about the corresponding A-tuple. To prevent infinite regress, the assignment of these IUIs does not involve the generation of an additional A-tuple, but is implemented through the use of these tuple-IUIs as an internal annotation to the tuple itself. Three factors can be distinguished as structural elements involved in such an assignment act: (1) the generation of the relevant alphanumeric string, (2) its attachment to the relevant object, and (3) the publication of this attachment [20]. A-tuples are of the form < IUIp, IUIa, tap > where IUIp is the IUI of the particular in question, IUIa is the IUI of the author of the assignment act, and tap is a time-stamp indicating when the assignment was made. 6.2. D-tuples In light of the need or desire to resolve mistakes [25], RT includes the use of D-tuples, which are to be created whenever (1) a tuple other than a D-tuple is added to the RTS Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
28
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
Data Store, in which case it includes meta-data about by whom and at what time the corresponding tuple was deposited or (2) a tuple, including D-tuples, is declared invalid in the system, in which case it includes additional info concerning the type of mistake committed and the reason therefore. D-tuples are of the form < IUId, IUIT, td, E, C, S >, where: • IUIT is the IUI of the tuple about which the D-tuple contains information. • IUId: is the IUI of the entity annotating IUIT by means of this D-tuple, • E is either the symbol ‘I’ (for insertion) or any of the error type symbols as discussed further, • C is a symbol for the applicable reason for change as discussed further, • td is the time the tuple denoted by IUIT is inserted or ‘retired’, and • S is a list of IUIs denoting the tuples, if any, that replace the retired one.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.3. PtoP-tuples Descriptions which express configurations amongst particulars have the form of PtoP – particular to particular – tuples. Here again a number of structural elements can be distinguished: (1) an authorized user observes one or more objects which have already been assigned IUIs in the referent tracking system (RTS) in hand, (2) the user recognizes or apprehends that these objects stand in a certain relation, which is represented in some realism-based ontology, (3) the user asserts that this relation holds and publishes this assertion by entering corresponding data which are then published in the referent tracking data store. This relationship data will then take the form of an ordered sextuple , where • IUIa is the IUI of the author asserting that the relationship referred to by r holds between the particulars referred to by the IUIs listed in P; • ta is a time-stamp indicating when the assertion was made; • r is the denotator in IUIo of the relationship obtaining between the particulars referred to in P; • IUIo is the IUI of the ontology from which r is taken; • P is an ordered list of IUIs referring to the particulars between which r obtains; and • tr is a time-stamp representing the time at which the relationship was observed to obtain. P contains as many IUIs as are required by the arity of the relation r. In most cases, P will be an ordered pair which is such that r obtains between the particulars represented by its first and second IUIs when taken in this order. 6.4. PtoU-tuples Another type of information that can be provided about a particular concerns what universal within an ontology it instantiates. Here, too, time is relevant, since a particular, through development, growth or other changes, may cease to instantiate one universal and start to instantiate another: thus George W. Bush Sr. changed from foetus to newborn, and from child to adult. Descriptions of this type (which we will refer to as PtoU-tuples – for: particular to universal) are represented by ordered tuples of the form , where
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
29
• • • •
IUIa is the IUI of the author asserting that IUIp is an instance (inst) of UUI; ta is a time-stamp indicating when the assertion was made; inst is the denotator in IUIo of the relationship of instantiation; IUIo is the IUI of the realism-based ontology from which inst and UUI are taken; • IUIp is the IUI referring to the particular whose inst relationship with the universal denoted by UUI is asserted; • UUI is the denotator of the universal in IUIo with which IUIp enjoys the inst relationship; and • tr is a time-stamp representing the time at which the relationship was observed to obtain. Note that it is specified from which ontology inst and UUI are taken (and precisely which inst relationship in those cases where an ontology contains several variants). Such specifications not only ensure that the corresponding definitions can be accessed automatically, but also facilitate reasoning in the RTS Reasoning Server across ontologies that are interoperable with the ontology specified.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.5. PtoC-tuples Whereas for PtoU-tuples their denotators of relationships and universals are taken from realism-based ontologies rather than from other knowledge repositories in terminology servers, PtoC-tuples do allow CUIs to be used instead of UUIs. Of course, the relationship to be used is not to be some variant of ‘inst’ since the standard definitions in use for ‘concept’ (such as ‘unit of knowledge’ or ‘unit of thought’) disallow most particulars from being declared as instances of concepts. PtoC-tuples (for particular to concept code) have the form , where: • IUIa is the IUI of the author asserting that terms associated to CUI may be used to describe IUIp; • ta is a time-stamp indicating when the assertion was made; • IUIc is the IUI of the concept-based system from which CUI is taken; • IUIp is the IUI referring to the particular which the author associates with CUI; • CUI is the CUI in the concept-system referred to by IUIc which the author associates with IUIp; and • tr is a time-stamp representing a time at which the author considers the association appropriate. Such tuples are to be interpreted as providing a facility equivalent to a simple index of terms in a work of scientific literature. 6.6. PtoU(-) – tuples Since the RT paradigm requires that only entities that exist or have existed are to be assigned an IUI, a capability is provided that deals with what is called ‘negative findings’ or ‘negative observations’ as captured in expressions such as: ‘no criminal history’, ‘membership of terrorist organization ruled out’, ‘absence of imminent danger’, and ‘attack prevented’. Such statements seem at first sight to present a problem for the referent tracking paradigm, since they imply that there are no entities in reality to which appropriate unique identifiers could be assigned. We therefore defined the relationship ‘p lacks u with respect to r at time t’ such that there obtains a relation
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
30
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
between the particular p and the universal u at time t, which is such that p stands to no instance of u in the relationship r at t [26, 27]. This ontological relation can be expressed by means of a ‘PtoU(-) tuple’ which is a lacks-counterpart of the PtoU-tuple and has the form , expressing that the particular referred to by IUIa asserts at time ta that the relation r of ontology IUIo does not obtain at time tr between the particular referred to by IUIp and any of the instances of the universal UUI at time tr. 6.7. PtoN-tuples Important particulars such as persons, ships, hurricanes, and so forth are often given proper names which function as denotators in reality outside the context of a referent tracking system. This sort of information is stored in an RTS by means of one or more ‘PtoN-tuples’ where ‘N’ stands for ‘name’. These tuples have the form < IUIa, ta, nt, n, IUIp, tr , IUIc >, where • IUIa is the IUI of the author asserting that n is a name of type nt used by IUIc to denote IUIp; • ta is a time-stamp indicating when the assertion was made; • IUIc is the IUI for the particular that uses the name n (this can be a person, a community of persons, an organization, an information system, ...); • IUIp is the IUI referring to the particular which the author associates with n; • n is the name which the author associates with IUIp; • nt is the nametype (examples being first name, last name, nick name, social security number, and so forth); and • tr is a time-stamp representing a time at which the author considers the association appropriate.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7. Discussion 7.1. Referent Tracking and action-oriented formalisms RT, at first sight, might look similar to other approaches. For instance, the need to track objects through time as they change, and to reason (and to have machines sometimes reason) over information that describes such changes, is what motivated calculi such as the situation calculus, the event calculus, and the fluent calculus, as well as some Knowledge Representation and Reasoning Systems. These approaches seek an efficient solution to the projection problem [28]: given an action theory that specifies the preconditions and effects of actions (including sensing), and a knowledge base about the initial state of the world, determine whether or not some condition holds after a given sequence of actions has been performed [29]. The situation calculus is a logic formalism that was first introduced by John McCarthy in 1963 [30] and since then underwent a few modifications [31]. The basic elements of situation calculus are: (1) actions that can be performed in the world, (2) fluents that describe the state of the world, each fluent thus being the representation of some property, and (3) situations. McCarthy and Hayes considered a situation to be ‘a complete state of the universe at an instant of time’ [32], a position which is also maintained in fluent calculus [33], whereas others redefined situations as finite
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
31
sequences of actions, thus a history of actions [31]. Event calculus does without situations, and uses only actions and fluents, whereby the latter are functions – rather than predicates as is the case in situation calculus – which can be used in predicates such as HoldsAt to state at what time which fluents hold [34]. RT differs in substantial ways from these logical formalisms. First of all, the goal of RT is not just to represent actions and changes, but all entities that exist in reality. Furthermore, these sorts of logics focus on computational aspects, but do not provide an integrated ontological characterization of entities such as actions, plans, and, because of their four-dimensionalist nature, for sure not of objects. It has been shown that it pays off to add more ontological rigor to formalisms such as situation calculus, for instance by using it only as one component for causal reasoning within a more elaborate, multi-component system [35]. RT, in contrast, is not in the first place a computational framework, but rather a representational one anchored in the realist view adhered to in Basic Formal Ontology (BFO) [23]. BFO distinguishes, for instance, continuants (such as George W. Bush) from occurrents (such as George W. Bush’s life or his last trip from Washington to New York). These distinctions, including BFO’s treatment of locations, positions and location schemes, was deemed essential in building a robot navigation model on top of situation calculus as embedded in Kuipers’ Spatial Semantic Hierarchy [36]. Relationships of the sort expressed by, for instance, RT’s PtoP- and PtoU-tuples hold only during certain time-periods [37, 38], and when they hold is expressed in the corresponding tuples themselves. In addition, PtoU-tuples express what universals a particular instantiates, thus also whether the entity described is an action or an object. Although no attempt has been made thus far, it seems plausible to assume that it is possible to express part of an RT database in terms of situation or event calculus.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.2. Facts versus beliefs The requirements within RT that tuples must make direct and explicit reference to that what they are about, and that this can only be done for entities that exist or have existed, would seem to make it very difficult to represent uncertain, or possibly deceptive knowledge. One can wonder if, for example, an intercepted communication contains ‘Cain will strike down Abel’ and it is believed that ‘Cain’ and ‘Abel’ are code words for non-personal entities, whether this belief can be recorded in this system. Similar questions can be asked about things in the future: isn't it important for a representational framework to be able to state knowledge about future happenings and entities that might not exist until the future, such as tomorrow’s sunset or Al-Qaeda’s next attack? It is here that the distinction between three levels of reality as discussed in section 3.1 and the assignment of IUIs to RT-tuples themselves play a role. If a PtoP-tuple to which IUI-457 is assigned states that George W. Bush was president of the US in 2007, then the latter is taken to be a representation of reality – which of course may be a mistake – whereas IUI-457 is the proposition that the latter is the case. That this proposition is entertained (or not) by a specific person can be expressed by additional PtoP-tuples that relate the tuple in question to that person by referring also to adequate belief-related relations or processes depending on what sort of ontology is used. As in the case of action logics, RT itself does not come with a logic of beliefs, but from the representations, so we believe, secondary representations in terms of a belief logic can be generated.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
32
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
For entities in the future, RT offers the possibility to reserve IUIs, rather than to assign IUIs [20]. Thus it is possible to assign an IUI to the plan to see and enjoy next Sunday’s sunset, whereas the detailed RT representation of that plan itself would contain a reserved IUI for that particular sunset.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.3. Maintaining integrity There are several challenges in maintaining the representational integrity of an RT system, specifically with respect to the requirements that an IUI within an RTS should denote only one entity, and that there is only one IUI for a specific entity. If, for instance, one doesn’t know that ‘Usama bin Ladin’ and ‘Osama bin Laden’ denote the same individual, how could one possibly know to relate both names to the IUI denoting that individual? Here responsibility for faithful representation is shared between the user and the user interface. Whereas the former must devote enough effort to find out in each specific case what individual a name denotes, the latter, assisted by additional applications, must make it possible to reduce the effort required. Term comparison algorithms might be used to inform a user that a name similar to the one entered is already registered. Triggers and alerts can be implemented to warn a user that distinct individuals have the same name, and so forth. All this, however, does not guarantee that the right decision will be made in every case, and errors will very likely occur. So there have to be procedures to detect and correct mistakes. It is here that the D-tuples play an important role [25]. Easy to solve, once detected, are mistakes in which a particular has been assigned more than one IUI. In this case, only one of these IUIs would be used in future tuples, whereas all tuples in which the other IUIs are used will be replaced by tuples in which that one IUI will replace the redundant ones. This mechanism guarantees that it still remains known that during some period in the past, information concerning one particular was believed to be about two or more particulars. More work would be required in the opposite case, i.e. when the same IUI is used to denote distinct particulars. Here it might be necessary to perform a manual revision of the tuples in which that is used. To detect mistakes, the ontologies in whose terms RT-tuples are expressed can be used to guide integrity-checking routines that run over the RTDB. Because, for instance, persons (or any material continuant) cannot be at two distinct places in the same time, the presence of RT-tuples in the RTS that suggest this to be the case, indicates a mistake of the type ‘one IUI for distinct particulars’. Logically, because two distinct material continuants cannot occupy the same spatial region, any collection of RT-tuples representing that this would be the case must contain an error of the type ‘distinct IUIs for the same particular’. 7.4. RT and the Semantic Web The various types of tuples enumerated in section 6 are expressible using standard Semantic Web technologies, though with some additional formalisms implemented at the data-base storage level. This is indeed the approach that has been taken in implementing the system [24]. The Resource Description Framework (RDF) [39] was used as the basic representation language. Our RDF representations of the RT-tuples are treated as resources themselves: each resource is therefore prefixed with the RTS name space
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
33
URI and the prefix ‘rts:’ such that, for instance, the resource rts:IUI-1 is the same as http://org.buffalo.edu/RTS#IUI-1. To declare properties for resources, we used RDFS and mapped the RT-tuples to RDFS classes, thereby ensuring that the class names are identical to the template names, with the exception of PtoU-, which, because of restrictions in the RDFS naming conventions, has been mapped to PtoLackU. Our implementation of the RTS is accessible through Web services which are invoked through SOAP messages [40] containing the procedure information (procedure name, parameters and return type) and port type (location of the procedure). The RTS uses Axis for Java [41] to host the web services thereby taking advantage of the native support of the Web Services Definition Language (WSDL) [42] that Axis provides. The RTS has been build to be independent of any data source technology. To achieve this goal, we have defined the RTRepository class as an abstract Java class. This class provides all necessary services for managing the data based on the principles defined in the RT paradigm. To manage the RT data in a specific data source technology, an extension of the RTRepository for that specific technology is required. We have decided to develop the RTRepositorySesameImp class by extending the RTRepository such that it targets the SAIL Sesame API for manipulating RDF graphs as a data source [43]. Because the RT data are expressed in RDF, RDF query languages such as RQL [44], SPARQL [45] and SeRQL [43] can be used for retrieval. To this end, the RTRepository comes with the service ‘repository.query(querystring, language)’ which has an argument for the query string and a second one for the name of the query language in which the first argument is expressed. The SeRQL query language is implemented with the help of the Sesame SeRQL query language module, and the SPARQL query language is implemented with the help of the ARQ query module (a SPARQL processor for Jena) [46].
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
8. Conclusion: meeting the new intelligence criteria When set up in appropriate ways, a network of referent tracking systems is able to meet all the requirements identified for the envisioned Globally Networked and Integrated Intelligence Enterprise (see section 1.3). The requirement to share intelligence data while still addressing the need to protect privacy, civil liberties, and sources and methods (C1), can be met by using the IUIs, typically the ones that stand proxy for persons, as pseudonyms. It would even be possible to go much further, for instance that all the information collected by credit card companies, banks, department stores, telecom providers and so forth would be pooled. Most citizens would find it unacceptable if that information were used for intelligence purposes without there being any reason to do so. But with the appropriate setup of IUIRepositories and RTDBs in such a way that, for instance, one specific agency has the means to link IUIs to persons, but otherwise no access to other RT-data, while other agencies would be able to do data-mining and pattern analysis on the pseudonymized data, no privacy or civil liberties would be violated. When analysts would detect suspicious patterns in the pseudonymized data pool, similar mechanisms as search warrants can be used to obtain re-identification of the data. The requirements to provide services across agencies, partners, and international borders for multiple mission use (C2) and to be able to adapt rapidly to changing needs
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
34
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
and new partners (C3) are supported by the possibility for referent tracking systems to cooperate in growing networks. The C4 requirement, i.e. to have security built into the data and environment using tags, together with the C5-requirement that access should be based on attributes that go beyond security classification, is met by the specific ways in which RT-tuples are set up: they contain in every case an indicator for the provenance of the data and all data are coded by means of ontologies or terminologies. Furthermore, each RT-tuple can be treated as a first-order entity, thereby receiving its own IUI, and that IUI can be used in other RT-tuples, for instance to describe to what type of entities or specific entities it may be disclosed. The same IUI can be used to track the flow of the data-element throughout the intelligence network. Data stewardship, finally, focusing on quality and reusability of data rather than, but not excluding, protection (C6) is a natural feature of the paradigm. One reason are the principles for IUI assignment which require that before an IUI is assigned to an entity, it should be checked whether that entity has already an IUI assigned to it. Mistakes will happen, of course, but they are traceable over time; if, for instance, when data accumulate, two IUIs start to appear repeatedly in the same configuration, then they may stand proxy for the same entity. Or, if the database at some stage contains a PtoP-tuple stating that the entity with IUIx was in some place at a given point in time, while in a completely different place a bit later, then it is likely, modulo other types of mistakes, that IUIx is denoting different things. A problem, at first sight, might be the amount of work required to represent information in this way. But here again, other types of software such as natural language processing applications, might assist. Furthermore, as shown in [47, 48], it is in many cases possible to translate structured information into a form that is RTcompatible automatically. We argue that the effort to make systems of this kind acceptable is not greater than the effort to bring about the change in mindset to realize Vision 2015.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
9. References [1]
Central Intelligence Agency. What is Intelligence? 2007 June 20, 2008 [cited 2008 August 12]; Available from: https://www.cia.gov/news-information/featured-story-archive/2007-featured-storyarchive/what-is-intelligence.html [2] Reagan R. Executive Order 12333--United States intelligence activities. 1981. [3] United States Intelligence Community. The Intelligence Process. 2008 [cited 2008 August 12]; Available from: http://www.intelligence.gov/2-business.shtml [4] Travers R. A Blueprint for Survival; The Coming Intelligence Failure. Studies in Intelligence. 1997;Semiannual Edition, No. 1:35-43. [5] Chen H. Intelligence and Security Informatics for International Security. Information Sharing and Data Mining. New York: Springer-Verlag 2006. [6] Office of the Director of National Intelligence. United States Intelligence Community Information Sharing Strategy. 2008. [7] Hasselbring W. Information system integration. Communications of the ACM. 2000;43(6):33-8. [8] Office of Homeland Security. National Strategy for Homeland Security. Washington D.C.: Office of Homeland Security 2002. [9] H. Chen, R. Miranda, D. Zeng, T. Madhusudan, C. Demchak, Schroeder J. Intelligence and Security Informatics: Proceedings of the First Symposium on Intelligence and Security Informatics (ISI’03). New York: Springer-Verlag 2003. [10] Chen H, Wang F-Y, Zeng D. Intelligence and Security Informatics for Homeland Security: Information, Communication, and Transportation. IEEE Transactions on Intelligent Transportation Systems. 2004 December;5(4):329-41.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
35
[11] Wang GA, Atabakhsh H, Petersen T, Chen H. Discovering Identity Problems: A Case Study. IEEE international conference on intelligence and security informatics. Atlanta, GA, USA Springer 2005. [12] Cybenko G, Giani A, Thompson: P. Cognitive Hacking: A Battle for the Mind. IEEE Computer. 2002;35(8):50-6. [13] Mercado S. A Venerable Source in a New Era: Sailing the Sea of OSINT in the Information Age. CIA Studies in Intelligence. 2004;48(3):45-55. [14] Office of the Director of National Intelligence. Office of the Director of National Intelligence. 2008 August 22 [cited 2008 August 28]; Available from: http://www.dni.gov/index.html [15] Office of the Director of National Intelligence. Vision 2015: A Globally Networked and Integrated Intelligence Enterprise. 2008. [16] Office of the Director of National Intelligence. The National Intelligence Strategy of The United States of America. Office of the Director of National Intelligence 2005. [17] Smith B. Beyond concepts: ontology as reality representation. Proceedings of the third international conference on formal ontology in information systems (FOIS 2004). Amsterdam: IOS Press 2004:73-84. [18] Smith B, Kusnierczyk W, Schober D, Ceusters W. Towards a Reference Terminology for Ontology Research and Development in the Biomedical Domain. KR-MED 2006, Biomedical Ontology in Action. Baltimore MD, USA 2006. [19] Bittner T, Smith B. A Theory of Granular Partitions. In: Duckham M, Goodchild MF, Worboy MF, eds. Foundations of Geographic Information Science. London: Taylor & Francis Books 2003:117-51. [20] Ceusters W, Smith B. Referent Tracking in Electronic Healthcare Records. In: Engelbrecht R, Geissbuhler A, Lovis C, Mihalas G, eds. Connecting Medical Informatics and Bio-Informatics Medical Informatics Europe 2005. Amsterdam: IOS Press 2005:71-6. [21] Ceusters W, Smith B. Referent Tracking for Digital Rights Management. International Journal of Metadata, Semantics and Ontologies. 2007;2(1):45-53. [22] Ceusters W, Smith B. Referent Tracking for Corporate Memories. In: Rittgen P, ed. Handbook of Ontologies for Business Interaction. New York and London: Idea Group Publishing 2007:34-46. [23] Grenon P, Smith B, Goldberg L. Biodynamic Ontology: Applying BFO in the Biomedical Domain. In: Pisanelli DM, ed. Ontologies in Medicine. Amsterdam: IOS Press 2004:20-38. [24] Manzoor S, Ceusters W, Rudnicki R. Implementation of a Referent Tracking System. International Journal of Healthcare Information Systems and Informatics. 2007;2(4):41-58. [25] Ceusters W. Dealing with Mistakes in a Referent Tracking System. In: KS H, ed. Proceedings of Ontology for the Intelligence Community 2007 (OIC-2007). Columbia MA 2007:5-8. [26] Ceusters W, Elkin P, Smith B. Referent Tracking: The Problem of Negative Findings. In: Hasman A, Haux R, Lei Jvd, Clercq ED, Roger-France F, eds. Studies in Health Technology and Informatics Ubiquity: Technologies for Better Health in Aging Societies - Proceedings of MIE2006. Amsterdam: IOS Press 2006:741-6. [27] Ceusters W, Elkin P, Smith B. Negative Findings in Electronic Health Records and Biomedical Ontologies: A Realist Approach. International Journal of Medical Informatics. 2007 March;76:326-33. [28] Reiter R. Knowledge in Action. Logical Foundations for Specifying and Implementing Dynamical Systems. Boston: MIT Press 2001. [29] Vassos S, Levesque H. Progression of Situation Calculus Action Theories with Incomplete Information. In: Veloso M, ed. Proceedings of IJCAI-07 2007. [30] McCarthy J. Situations, actions and causal laws. Stanford, CA: Stanford University Artificial Intelligence Laboratory; 1963. [31] Reiter R. The frame problem in the situation calculus: a simple solution (sometimes) and a completeness result for goal regression. In: Lifshitz V, ed. Artificial intelligence and mathematical theory of computation: papers in honour of John McCarthy. San Diego, CA, USA: Academic Press Professional, Inc 1991:359-80. [32] McCarthy J, Hayes PJ. Some philosophical problems from the standpoint of artificial intelligence. Machine Intelligence. 1969;4:463-502. [33] Thielscher M. Introduction to the Fluent Calculus. Electronic Transactions on Artificial Intelligence. 1998;2(3-4):179-92. [34] Kowalski R. Database updates in the event calculus. Journal of Logic Programming. 1992;12(1-2):12146. [35] Kuipers B. The spatial semantic hierarchy. Artificial Intelligence. 2000 May;119(1-2):191 - 233. [36] Bateman J, Farrar S. Modelling Models of Robot Navigation Using Formal Spatial Ontology. Spatial Cognition IV Reasoning, Action, and Interaction. Berlin / Heidelberg: Springer 2005:366-89. [37] Smith B, Ceusters W, Klagges B, Köhler J, Kumar A, Lomax J, et al. Relations in biomedical ontologies. Genome Biology. 2005;6(5):R46. [38] Smith B, Grenon P. The Cornucopia of Formal-Ontological Relations. Dialectica. 2004;58(3):279-96. [39] Manola F, Miller E. RDF Primer. 2004 [cited; Available from: http://www.w3.org/TR/rdf-primer/
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
36
W. Ceusters and S. Manzoor / How to Track Absolutely Everything
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[40] Mitra N. SOAP Version 1.2 Part 0: Primer. W3C Recommendation 2003. [41] The Apache Software Foundation. Axis: A Webservices toolkit. 2005 [cited 25 January, 2007]; Available from: http://ws.apache.org/axis/ [42] Christensen E, Curbera F, Meredith G, Weerawarana S. Web Services Description Language (WSDL) 1.1. W3C Note 2001. [43] Broekstra J, Kampman A, Harmelen Fv. Sesame: A Generic Architecture for Storing and Querying RDF and RDF Schema. Lecture Notes in Computer Science - International Semantic Web Conference ISWC2002. Heidelberg: Springer 2002:54-68. [44] Foundation for Research and Technology – Hellas. The RDF Query Language (RQL). 2003 July 18 [cited 25 January 2007]; Available from: http://139.91.183.30:9090/RDF/RQL/ [45] Prud'hommeaux E, Seaborne A. SPARQL Query Language for RDF. W3C Working Draft 2006 October 4th [cited January 22, 2007]; Available from: http://www.w3.org/TR/rdf-sparql-query/ [46] RDF Data Access Working Group. ARQ - A SPARQL Processor for Jena. 2007 [cited 15th Febuary, 2007]; Available from: http://jena.sourceforge.net/ARQ/ [47] Rudnicki R, Ceusters W, Manzoor S, Smith B. What Particulars are Referred to in EHR Data? A Case Study in Integrating Referent Tracking into an Electronic Health Record Application. In: Teich JM, Suermondt J, C H, eds. American Medical Informatics Association 2007 Annual Symposium Proceedings, Biomedical and Health Informatics: From Foundations to Applications to Policy. Chicago, IL 2007:630-4. [48] Manzoor S, Ceusters W, Rudnicki R. A Middleware Approach to Integrate Referent Tracking in EHR Systems. In: Teich JM, Suermondt J, C H, eds. Proceedings of the American Medical Informatics Association 2007 Annual Symposium Biomedical and Health Informatics: From Foundations to Applications to Policy. Chicago IL 2007:503-7.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-37
37
Chapter 3
Uses of Ontologies in Open Source Blog Mining Brian Ulicnya, Mieczyslaw M. Kokara,b, Christopher J. Matheusa a VIStology, Inc. b Northeastern University
Abstract: The blogosphere provides a novel window into an important segment of public opinion, but its dynamic nature makes it an elusive medium to analyze and interpret in the aggregate, where it is most informative. We are developing a new open-source blog mining technology that employs ontologies to solve this problem by fusing the signals of the blogosphere and zeroing in on issues that are most likely to migrate offline. This technology is designed to enable analysts to anticipate the threats or opportunities these issues represent in a timely and efficient fashion. Keywords: Blog mining, Malaysia, human terrain, situation awareness
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction Although much, perhaps even the majority, of what is discussed in the blogosphere is of little consequence and fleeting interest, blogs continue to emerge as powerful organizing mechanisms, giving momentum to ideas that shape public opinion and influence behavior. There are nearly 16 million active blogs [17] on the Internet with more launched every day, and bloggers have increasingly made an impact politically in a range of places and situations. For example, Malaysian bloggers have recently become quite assertive in confronting perceived corruption in their national government despite strict governmental control of the major media [1]. Although one must be careful not to extrapolate from the population of bloggers to the population as a whole, clearly blogs provide unparalleled access to an important segment of public opinion about events of the day. The perspective blog mining provides is much more complete than that provided by the letters to the editor section of a newspaper or magazine, if it has one. Even premier print newspapers such as the New York Times publish only 15 to 20 of the 1000 letters they receive daily in reaction to their reports [5]; by contrast, there are typically over 3,000 blog posts that cite New York Times stories for any particular day, including many posts not in English. By mining the unfiltered reactions of bloggers to the day’s events, we can clearly evaluate and quantify the reaction to the days’ news reports of a highly motivated group of users. This is particularly useful where the local press is tightly controlled.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
38
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
Mining a particular subset of a local blogosphere requires careful data aggregation because the data is relatively sparse. In the blogosphere as a whole, it is not as difficult to detect trends because there is substantial data redundancy. In a specific local blogosphere, such as the focus of our case study, the Malaysian social/political blogosphere, the data is sparse enough that ontologies (or ontology-like structures) are useful to accurately aggregate different ways of referencing the same thing, among other relations. That is, in a specific local blogosphere, because there are not massively redundant occurrences of various entities (e.g. named entities, URLs, events), care must be taken to aggregate counts of the same entity referenced in different ways. Ontologies can encode this information, along with much more. Several commercial blogmining services have begun operating in the last few years.1 These companies offer sophisticated analyses of mentions of (typically, brandrelated) entities in blogs, producing analysis of the volume of mentions in social media such as blogs, trends in quantity of mentions, some sentiment analysis (whether a post seemed favorable or unfavorable toward the entity), and some demographic analysis (who is doing the mentioning, by age and other demographic categories). These systems presuppose that there is a set of known entities or issues one wishes to track, and they do not provide an estimate of the number of bloggers within a demographic category that have not mentioned an entity in that particular category. Thus, these services can tell how many bloggers who mentioned X were Malaysian, but not what percentage of the Malaysian blogosphere mentioned X. Further, they cannot list the top entities or issues that are important to the Malaysian blogosphere as a whole. Similarly, blog search engines such as Technorati, BlogPulse, and the like track the most popular URLs cited by all blogs. BlogPulse can also rank the most cited phrases, blogs, or personal names. These services allow one to search blogs worldwide, and identify the most popular news articles, persons mentioned, or URLs cited worldwide, but they do not aggregate the most popular topics or stories for blogs within a particular geographic area. They are interested only in the global blogosphere, not local blogospheres. VIStology's International Blogs (IBlogs) project is a three-year effort funded by the Air Force Office of Scientific Research’s Distributed Intelligence program to develop a platform for automatically monitoring foreign blogs. The purpose of this technology is to provide blog analysts a tool for monitoring, evaluating, and anticipating the impact of blogs by clustering posts by news event and ranking their significance according to novel metrics of information quality. This technology will enable analysts to discover the issues that are important in a local blogosphere, by providing accurate measurements particular to that locale alone. Blog posts are not standalone documents; therefore, information retrieval metrics must take into account the articles they cite as well as the commentary they add. In particular, because of the exophoric and quotational nature of blogs, it is important to identify links to news articles that posts cite and analyze them. In contrast to current commercial blog search engines, the IBlogs search engine addresses the intricate nature of blogs by ranking blog posts according to their relevance to a query, as well as their timeliness, specificity and credibility. We briefly summarize these metrics here (see [20] for details). 1
These include TMS Media Intelligence/Cymfony (relevantnoise.com), Nielsen BuzzMetrics (buzzmetrics.com), (umbrialistens.com), Attentio (attentio.com) and others.
(cymfony.com), Relevant Noise Techdirt (techdirt.com), Umbria
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
39
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
RELEVANCE: What a blog post is about is determined not only by the text of a post, but also by the text of any news article it references. Terms in news articles and blog posts are not ranked by the familiar term frequency divided by document frequency metric standard in information retrieval in light of the clumpiness of the corpus and journalistic conventions. • TIMELINESS: The timeliness of a blog post is determined by comparing the timestamp of a blog post with the publication date of a news article that it cites. Timeliness, as distinguished from recency, is about proximity to the relevant event. (Locality is the spatial correlate of timeliness; it is computed by comparing the location of a blog author to the event a post is about.) • SPECIFICITY: The number of unique individual entities mentioned in a blog post and any news article it cites determines the specificity of a blog post. This is approximated as the number of unique proper nouns and their variants. Attention should also be paid to depth in a domain ontology. • CREDIBILITY: The credibility of a blog author’s posts is determined by the presence of various credibility-enhancing features that we have validated as informing human credibility judgments [19],[8]. These include blogging under one’s real name, linking to reputable news outlets, attracting non-spam comments, and so on. These metrics must be computed for each author, since blogs can have multiple authors. The number of inbound hyperlinks alone does not determine blog credibility. Architecturally, the IBlogs system includes an RSTC (Relevance, Specificity, Timeliness, Credibility) Ranking module, an indexer (Lucene), a crawler (a modified version of the open-source Nutch crawler), an ontology reasoner (BaseVISor) and a consistency checker (ConsVISor). See Figure 1. The IBlogs engine outputs information annotated according to an ontology of news events and participants. Our goal is to cluster blog posts by the news events that they are about, where any given news event may have more than one news story that reports it, and each of those stories may be published at one or more URLs. A news event is thus typically two levels removed from a blog post that references it. 1. Background: The Malaysian Sopo Blogosphere The IBlogs project has focused on analyzing the Malaysian “sopo” blogosphere (“sopo”, for social/political, as they characterize it in Malaysia) as a case study in monitoring a localized political blogosphere. Malaysia is a nation of approximately 26 million in an ethnically and religiously plural constitutional monarchy in Southeast Asia comprised of thirteen sultanates. Malaysia is a former British colony, and Malaysian English is the second-most popular language, used widely in business as well as in blogging. The official religion is Islam, with Muslims constituting roughly 60% of the population. Ethnic Malays form a plurality of the society, constituting 52% of the society, with substantial Chinese- and Indian-descendent populations, as well. Longstanding political tensions exist between these ethnic and sectarian groups.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
40
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 1. IBlogs Components
The Malaysian blogosphere provides an interesting set of properties for analysis. The Malaysian government tightly controls the press. In 2007, Reporters Without Borders (Reporters Sans Frontieres (RSF), in French), the respected journalism watchdog group, ranked Malaysia 124th out of 169 countries surveyed in its 2007 Press Freedom Index [15]. This was Malaysia’s worst rank since the Index was inaugurated in 2002. However, the Malaysian government has encouraged the use of Web technologies and the growth of Internet-based enterprise. As a result, the Malaysian blogosphere has become important as a place to air dissenting views and publish incriminating information. The Malaysian sopo blogging community has managed to wage a campaign against perceived governmental corruption. Since January 2007, the Malaysian government has attempted to intimidate bloggers by means of defamation suits, police interrogations and arrests [1], [13]. The Malaysian law enshrines free speech as a matter of principle, but Malaysia’s sedition law makes restrictions on what can be said about the government or about ethnic groups. The government must license print newspapers in Malaysia, and it has sometimes repealed publishing licenses to newspapers for printing controversial stories. In August, 2008, the Malaysian government blocked access to the online news source Malaysia Today, which often published critical remarks,2 after police interrogated the website publisher the previous month. Malaysian defamation law allows prosecution for astronomically large sums. Malaysian assembly laws require permits for any gathering in public places (see [21] for details). Organizations, even affiliations of bloggers, must have public, named officers. According to the RSF 2008 Annual Report on press freedom in Malaysia [16], the stakes continue to be high for Malaysian sopo bloggers. Malaysian government officials often issue dismissive comments about bloggers. The Malaysian prime 2
“Malaysia blocks anti-government news Web site”, Associated Press, August 28, 2008.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
41
minister has called bloggers “liars”, and the law minister has suggested that bloggers were liable to prosecution under the Internal Security Act. More seriously, a defamation suit was filed in January 2007, against two of the most prominent Malaysian bloggers. Another blogger was detained in 2007 under the Official Secrets Act, allegedly for posting a doctored photo of the deputy prime minister. Police complaints have been filed against blogs for allegedly insulting Islam and the King of Malaysia, the elected monarch, and bloggers have been accused of fomenting ethnic unrest by failing to police comments on their blogs adequately. Nevertheless, in the recent BERSIH rally held in Kuala Lumpur on November 10, 2007, tens of thousands of Malaysians potentially risked prison sentences of up two years in order to participate in a rally for electoral reform, in anticipation of the general election of March 2008, during which the ruling party eventually lost the supermajority they had enjoyed since the country’s independence. This legislative dominance had allowed them to modify the country’s constitution at will. The BERSIH rally could not have been mobilized except through the Internet, since the government tightly controls the media in Malaysia. The depiction of this event in the Malaysian sopo blogosphere has been a focus of our research. In order to predict or monitor the behavior of a local social/political blogging community, one must be able to delimit some reasonable approximation of that community. To estimate the size of the Malaysian blogger community, we examined blogging platform statistics and derived metrics from blogging networks. One online study produced by MSN Live in Malaysia estimated that 46% of those online in Malaysia have a blog [4], which would amount to approximately 6.3 million bloggers given then current estimates of Malaysian Internet users 3. This study was based on an online survey at a blog hosting service’s website, however, and seems far too large. Fortunately, the most popular blogging platform in Malaysia, Blogger.com, enables geographic search of bloggers (although one cannot query blog posts geographically). A search on Blogger.com returns 152,000 Malaysia blog profiles (sopo and non-sopo). Similar queries on the fee-based Typepad blog platform identify only about 40 additional blogs; Wordpress reveals another 90 bloggers. 2,930 are listed on The Star Online’s Malaysian blog directory at AllMalaysia.info4. The number of Malaysian bloggers is much smaller than the number of Malaysians using social networking sites, which also usually provide blogging-like capabilities, although these blogs may be semi-private (i.e. one must be part of someone’s social network in order to read the posted information). MySpace profiles reveal about 293,000 identifiable Malaysian profiles (including organizations and bands as well as individuals), as well as 0.5 million profiles on Facebook. The social networking site Friendster is much more popular in Malaysia, with approximately 5.4 million profiles. 5 Including the social networking sites, the potential number of sopo bloggers in Malaysia is somewhere in the range of hundreds of thousands to a few million. Our research indicates that only a few thousand blogs constitute the Malaysian sopo blogosphere, however, with a small cadre of leading bloggers being most influential 3
Malaysian Communications and Multimedia Commission, 2006. http://www.internetworldstats.com/asia/my.htm 4 The Star (Kuala Lumpur) is an official Malaysian newspaper, with a blog directory at AllMalaysia.info/ambp. Site accessed October, 2008. 5 All figures as of Summer, 2008.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
42
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
[18]. All of the most influential bloggers use public blogging platforms, not the semiprivate blog-like capabilities of social networking sites.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2. Ontologies in Blog Mining IBlogs uses ontologies and ontological relations in three ways. First, IBlogs uses a domain ontology for query expansion and term normalization. Second, IBlogs uses an ontology of the blogosphere to represent and normalize blog data. Third, IBlogs outputs data expressing an ontology of events. Resource Description Framework6 (RDF) and Web Ontology Language7 (OWL) are World-Wide Web Consortium (W3C) standards for encoding and sharing ontological knowledge. RDF and OWL represent subject-predicate-object logical statements, called “triples”, that can be encoded using one of two syntactic notations. RDF and OWL can both be used to model ontological relations between resources; OWL layers a richer expression of semantics over RDF and RDF Schema (RDFS). A resource, the subject of an RDF or OWL subject-predicate-object triple, is usually represented by a Universal Resource Indicator (URI), which is similar to a URL, or Web address, but which doesn’t necessarily correspond to a document that a Web browser can render. This scheme allows various organizations to encode information using the same terms, since the URI http://www.abc.com/Bersih corresponds to a potentially different entity than the URI http://www.def.org/Bersih. Statements about the former resource may be applied to the latter resource (and vice versa) by asserting that the resources are identical via the use of an rdfs:sameAs predicate. On the other hand, statements about the former resource may be explicitly stated not to apply to the latter (and vice versa) by means of the rdfs:differentFrom predicate. At present, as the subject-matter ontology, we are using a subset of Malaysian data from dbPedia8. dbPedia is data encoded in RDF that has been automatically extracted and converted into triples by parsing Wikipedia9 to identify structured elements such as Wikipedia “infobox” elements 10 , geo-coordinates and other simple facts. Recently, dbPedia announced that over 100 million triples had been extracted from Wikipedia in a range of languages (see dbpedia.org for current statistics and details). The dbPedia project includes 3,900 resource pages containing the term “Malaysia”. Of these, 32 pages are marked as belonging to the category Malaysia, and 13 are notated as being in category Politics_of_Malaysia. 11 Each resource page contains multiple properties associated with the head term. Table 1, for example, presents the list of properties associated with the current Prime Minister of Malaysia, Abullah Ahmad Badawi12. 6
http://www.w3.org/RDF/ http://www.w3.org/OWL/ 8 http:// dbpedia.org 9 http://www.wikipedia.org 10 http://en.wikipedia.org/wiki/Help:Infobox 11 Unfortunately, dbPedia’s retention of the Wikipedia convention of using “Category:” for categories introduces a systematic syntactic problem in translation to RDF or OWL. The use of the colon-delimited category names allows this resource to be interpreted as a qualified name (Qname), a convention intended to allow for the simplification of resource names by abbreviating common prefixes via a namespace declaration, in addition to its intended interpretation. 12 http://dbpedia.org/page/Abdullah_Ahmad_Badawi 7
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
43
dbPedia uses a language tag in RDF elements to incorporate the same information in different languages. For example, the p:abstract 13 property contains a summary paragraph from Wikipedia in several languages, based on the initial paragraph in the Wikipedia page for a subject in the language. Because some of this data is not in UTF8 encoding, we have had to eliminate it prior to processing. An important property for our system is the dbPedia redirect property, which is used by Wikipedia to redirect variant names to a canonical name for a topic, so that there are not multiple pages on the same topic. We use the dbPedia redirect property to implement term expansion and term normalization. Several ontologies, schemas and namespaces are incorporated in dbPedia, including Friend of a Friend14 (FOAF), for describing persons and their relationships, Simple Knowledge Organization System15 (SKOS), for describing topics, and YAGO,16 for annotating standard concepts. FOAF is a well-known RDF and OWL-readable ontology for describing persons and their relationships that has gradually been evolving since 2000. SKOS is an ontology for sharing controlled vocabularies as thesauri, taxonomies, classification headings and the like. YAGO is a manually-verified set of facts and conceptual relations constituting over 14 million triples that is based on a project at the Max-Planck-Institute, Saarbrücken. The rdf:type property for Badawi (Table 1), for example, asserts that Badawi is an instance of YAGO categories LivingPeople, CurrentNationalLeaders and PrimeMinistersOfMalaysia, among others.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Table 1. Properties for Malaysian PM Badawi in dbPedia p:abstract p:birthPlace
p:after p:constituencyMp
p:before p:deputy
p:honorificP refix p:parliament p:religion
p:monarch
p:name
p:birthDate p:hasPhotoCollect ion p:order
p:party p:spouse
p:predecessor p:start
p:reference p:termEnd
p:termStart
p:title
p:wikiPageUsesTemp late
p:wordnet_ type rdfs:label foaf:img p:before p:leader p:relations
p:years
rdf:type
p:wikipagede|es|fi|ja|no|pl|ru| zh rdfs:comment
skos:subject foaf:page p:beforeElection p:leaderName p:spouse
foaf:depiction p:after p:incumbent p:predecessor p:successor
foaf:homepage p:afterElection p:keyPeople p:redirect owl:sameAs
13
“p:” indicates a property in the dbPedia namespace http:// foaf-project.org 15 http://www.w3.org/TR/skos-reference/ 16 http://www.mpi-inf.mpg.de/~suchanek/downloads/yago/ 14
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
44
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
VIStology’s BaseVISor inference engine is used to query the domain ontology and normalize output. BaseVISor [12] is a forward-chaining inference engine that can process ontological information expressed in OWL or RDF. Its inference procedure uses a subset of axioms that define the semantics of OWL. Moreover, the ontology and the data annotation can be expanded with “IF – THEN” rules that express what additional statements hold true (the THEN part) when the antecedent conditions are satisfied. BaseVISor allows the system to expand queries based on the domain ontology as input to the underlying Lucene index constructed by the customized Nutch crawler. Suppose, for example, one wanted to query all blog posts containing the term “Malaysia” and one or more variant names for the Prime Minister, Abdullah Badawi, using “Pak Lah” as the (non-canonical) name variant for which to find all other variants. (Pak Lah is a nickname for the Prime Minister, roughly meaning “Uncle (Abdul)Lah”.) One would issue this request to IBlogs using the following syntax: Malaysia aliases:Pak+Lah
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
BaseVISor processes this query and return a set of Lucene queries, each element of which will correspond to a query for blog posts containing “Malaysia” and one variant name of Badawi, making necessary adjustments in the query syntax such as replacing underscores with spaces and adding quotation marks indicating phrase searches: Malaysia “Abdullah Ahmad Badawi” Malaysia “Datuk Seri Abdullah Ahmad Badawi” Malaysia “Pak Lah” Malaysia “Ahmad Badavi” Malaysia “Abdullah Badawi” Malaysia “Datuk Seri Abdullah bin Haji Ahmad Badawi” Malaysia “Dato' Seri Abdullah bin Haji Ahmad Badawi” Malaysia “Datuk Abdullah Ahmed Badwi” Malaysia “Datuk Abdullah Ahmad Badawi” Malaysia “Badawi, Abdullah Ahmed” Admittedly, some of these search queries are rather unlikely to be productive in practice since they contain a much more formal version of the name than is likely to be used in informal blog post text, although it may appear in cited news articles. However, retaining these name variants in the ontology costs only a minimal amount of query and processing time, and they are useful for normalizing named entity output to a canonical form in the reverse direction. It would be possible, of course, to canonicalize names at the indexing stage, so that one indexes a normalized version of the text, where all name variants such as “Pak Lah” have been indexed as “Abdullah Ahmad Badawi”, and so on. Obviously, however, this requires a more complicated indexing structure, since three tokens, as in this example, would be indexed in the place of two tokens in the original. Lucene is capable of handling such complications, however. In our case, though, we have chosen to keep the index and domain ontology separate and more weakly linked because this allows us to change domain ontologies at will without reindexing the corpus, and because it is necessary for queries involving ontological relations other than identity or
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
45
subclass. For example, one could ask for all blog posts containing the term “Malaysia” and mentioning the spouse of Badawi in this way:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Malaysia spouse:Abdullah+Ahmad+Badawi This will, in turn, result in the Lucene index being queried for posts that contain “Malaysia” and “Jeanne Abdullah”, the name of his wife. It would certainly do no good to insert “Jeanne Abdullah” into the index every time there was a mention of her husband, since she is not necessarily mentioned in the same blog post. Therefore, in order to facilitate querying arbitrary ontological relations, separating the domain ontology from the index is necessary. VIStology’s ConsVISor is a rule-based system for checking consistency of ontologies represented in RDF, OWL, or DARPA Agent Markup Language (DAML). We use ConsVISor to identify conflicts and inconsistencies between multiple domain ontology resources. ConsVISor can be used to determine whether two entities (with or without the same name) are coreferential [11]. ConsVISor does this by determining whether the addition of a rdfs:sameAs triple for two entities with the same name allows the derivation of a contradiction. If so, then the administrator of the domain ontologies can be notified of the inconsistency. dbPedia data contains rdfs:sameAs properties for entities that provide a basis for consistency checking. An ontology of the blogosphere is implicit in the system as the schema of the IBlogs index. This schema is realized by adding fields to the default Lucene index schema. The Semantically-Interlinked Online Communities (SIOC) ontology 17 provided a useful starting place for this work, but we found it necessary to extend it, for example, to assign a type to an outlink from a blog post (possible values are: news, trackback, blogroll, self-citation, other), and so on. Although blogs normally publish feeds that express the most recent updates, and these feeds are encoded in RDF, using the Really Simple Syndication18 (RSS) standard, or in Atom,19 a standard that is translatable into RDF, indexing the feeds alone is not sufficient. Blog feeds require further analysis because the feed itself is not guaranteed to contain the entire blog post, blog comments, images and profile information relevant to determining blog credibility. All this requires parsing and analyzing HTML blog pages that are designed for human consumption. Further, it is not possible to reason about the blog post relevance, specificity, or credibility directly in terms of the RDF representation of a local blogosphere, because RDF reasoning is purely logical, usually based on whether a set of triples meets or doesn’t meet some logical condition based on logical relations such as subclass. Reasoning about relevance, credibility, and other information quality metrics requires a quantitative answer, however. Relevance and other such notions, that is, are graded notions, based on intrinsic qualities of the post such as term counts, weights and information quality. Pure RDF or OWL reasoning does not produce graded or quantitative answers. Typically, the blogosphere is modeled as simple network of blog-blog links. However, more insight can be gleaned from the “fine structure” of the blogosphere, as represented in Figure 2. 17
http://www.sioc-project.org http://www.rssboard.org/ 19 http://tools.ietf.org/html/rfc4287 18
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
46
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 2. The "Fine Structure" of the Blogosphere
Dyads of post/outlink & blog/profile are the crucial units of analysis in our system. Because of this, the system must assign a type to a link; it is not sufficient to simply identify the hypertext links by their syntax. Links from blog to blog (called blogrolls) are often stale, and thus an unreliable guide to a blogger’s current concerns or interests. On the other hand, links from one blog post to another blogger’s post are actually quite rare, and it is not clear from just a link between posts that the linking post is citing the linked post approvingly or disapprovingly. Therefore, links between posts is an unreliable indicator of blog communities or shared interests. IBlogs outputs information annotated according to an ontology of news events and participants. Our goal is to cluster blog posts by the news events that they are about, where any given news event may have more than one news story that reports it, and each of those stories may be published at one or more URLs. A news event is thus typically two levels removed from a blog post that references it. Our system outputs results in the OpenSearch 1.1 RSS20 standard, which we have extended with concepts from the Dublin Core21 metadata standard and with our own namespace elements for news event representations. NewsML 22 and the associated EventML standard represent industry-originated attempts to standardize representations of news articles and the events they report. These standards can be readily converted to OWL ontologies. Reuters and Agence France Press (AFP) currently use these standards, and we plan to adapt these emerging standards as they develop in order to standardize the representation of news articles and news events in the hope that we will be able to directly use output in these formats produced by news providers in the future.
20
http:// opensearch.org http:// dublincore.org 22 http:// newsml.org 21
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
47
3. Ontological Issues
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Named Entities An important aspect of the IBlogs system is the identification of named entities in the text of the blog. Named entity extraction is required to determine blog post specificity (aggregate count of named entities), timeliness (difference between blog post time and news citation time), locality (difference between blog locale and news citation locale), credibility (identification of affiliations in blog profile), as well as the clustering of events (two accounts depict the same event to the extent that they encode the same participants, time and location). As such, we have evaluated the performance of a state of the art named entity software package against a much simpler pattern matching and filtering approach. The software package evaluated is the Gnosis named entity extractor from ClearForest, now a Thomson Reuters company. 23 Gnosis identifies several types of named entities including geographical entities (city, continent, country, region, etc.), persons, organizations, products, and industry terms (e.g. “airline”, “telephone conversation”, “tear gas”), among others. Gnosis does not identify temporal reference entities such as dates or even temporal elements such as months and years. In an evaluation of 215 named entities encoded in the post text and outgoing hyperlink texts of one of the most central sopo bloggers, Jeff Ooi’s, account of the BERSIH rally, we found that the Gnosis name entity extractor did far worse than a much simpler extractor of our own that simply identified named entities on the basis of orthography (capitalized initial letters) and deleted string-initial initial stop words (e.g. The, Before, Everyone) and phrases that were entirely participles (e.g. Leading, Branded). See Table 2. Recall is defined as the number of correctly identified named entities divided by the number of total named entities in the text. Precision is defined as the number of correctly identified named entities divided by the total number of identified named entities. F-measure is the harmonic mean of the two, evenly weighted scores. Because the two identifiers were designed for different purposes, we did not penalize Gnosis for not identifying dates; nor did we penalize the pattern matcher for not identifying “technical terms”. We did count positive identifications of either in their favor. Table 2. Named Entity F-measure comparison System
F (Recall, Precision)
ClearForest Gnosis
17.6% (21.4%, 100%)
Pattern-Match and Filter
33.5% (61.9%,72.9%)
Pattern-Match and Filter, Including partial matches
44.5% (80.5%, 100%)
Although Gnosis did identify some non-capitalized phrases as important entities (e.g. “tear gas” and “police”), it more often neglected to identify crucial capitalized terms such as “Yang di-Pertuan Agong”, the title of the Malysian monarch, and “BERSIH”, the name of the coalition sponsoring the rally.
23
http://gnosis.clearforest.com
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
48
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
On the other hand, while our pattern-match and filter technique does not associate entity types with named entities, the domain ontology can be used to make such associations. Of the most frequent named entities identified in the corpus of blog posts and news articles associated with the BERSIH Rally, the most frequent are well represented in dbPedia. By means of the presence of geo-coordinates in the dbPedia triples, it is possible to distinguish place names from other types of named entities. Similarly, one could identify persons from the use of FOAF properties and so on. See Table 3. Table 3. Most Frequently Identified Named Entities Term Malaysia Allah Bersih Malaysian / Malaysians Islam Dalam Yang di-Pertuan Agong (King) Istana Negara (Palace) Malaysiakini (Newspaper) Kuala Lumpur (City)
In dbPedia? Yes Yes Yes Yes / Redirect Yes No Yes Yes Yes Yes
Geocoordinates? Yes No No No No No No No Yes
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Event Identity NewsML and EventML provide no criterion for saying that two depictions of events are depictions of the same event, but as W.V.O. Quine urged, ontologies should follow the maxim “no entity with out identity” [14]. That is, entities are not legitimately included in an ontology unless a criterion has been provided for specifying when two ways of picking out such entities actually correspond to the same thing. This is crucial in a computational setting like the present blog mining context where aggregation and counting of like entities is essential. As we just described, for atomic entities like persons, criteria of identity are encoded in the domain ontology. As such, we can identify the referent of different names as the same person, or identify the referent of a particular name as being identical to someone in a particular relationship to that person, e.g. their spouse. For personal names and other individual entities, we rely on dbPedia to provide variant names for individuals, such as persons, organizations, locations, and so on. In IBlogs, events are modeled as complex entities consisting of a set of named participants at a location and a day. To a first approximation, we use news stories as primarily encoding events, in that sopo blogging (to a greater degree in the US context than in the Malaysian context, it turns out) is mostly a second-order commentary on events as reported in the news somewhere, not as observed by the participants at first hand. Thus, the event associated with a blog post in IBlogs is partially determined by the news article or articles that it cites. We use days as the relevant time period for event granularity in conformance with the standard daily news cycle of newspaper distribution and daily news broadcasts. Therefore, whatever the same set of participants does at a location on a particular day counts as a single event, for our purposes. This is an obvious oversimplification in metaphysical terms. For instance, this entails that there are no events that last longer than a day. Secondly, this model aggregates the actions of a set of participants at a place and time into a single event in a counterintuitive way. For example, the same
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
49
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
group of people may eat lunch at a particular place on a particular day and then play volleyball, but the lunch would intuitively be considered a different event than the volleyball playing. In our scheme of event individuation, however, the lunch and the volleyball playing do not individuate different events. Perhaps, more charitably, we can say that the model doesn’t distinguish between these events, which may be considered as parts of the same event at the granularity of a day. We adopt this approach to the granularity of events because it is useful for our purposes in practice. That is, we are assuming that it is rare for the same set of participants to make news in two different ways at the same place and on the same day, and it is this notion of news event that we are trying to capture. Consider the following news story: Malaysia blocks anti-government news Web site By SEAN YOONG The Associated Press Thursday, August 28, 2008; 8:02 AM KUALA LUMPUR, Malaysia -- Malaysia has blocked access to a popular news Web site that often runs afoul of authorities for its sensational political reporting, sparking complaints Thursday that the government has reneged on its pledge to keep cyberspace uncensored. The Malaysian Communications and Multimedia Commission, the government's industry regulator, ordered local Internet service providers on Wednesday to cut off access to the Malaysia Today site, said a commission official who requested anonymity because of the sensitivity of the issue….
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In our system of event individuation it is taken to encode the event: Event: place: Kuala Lumpur, Malaysia, day: Wednesday, August 27, 2008 participants: Malaysia, Malaysian Commission, Internet, Malaysia Today.
Communications
and
Multimedia
We assume an inverted pyramid approach to news stories, and primarily consider the named entities in the lede (or “lead”) paragraph, which expresses the who, what, when, where and why of the event in general terms, often using common nouns (here, “a news web site”, “the government”) and the second paragraph, which provides the specifics of the story, usually by identifying the individuals involved, here the Malaysian Communications and Multimedia Commission and Malaysia Today. Named entities that occur in successive paragraphs are given a discounted weight based on the paragraph in which they occur (for details, see [20]). Sameness of events in IBlogs is a graded notion based on clustering events with the same time and place based on their weighted participants, where the participant names have been normalized based on dbPedia, using BaseVISor. This model of event individuation is based on the example of the US blogosphere, where at least 50% of all political blog posts cite a news article, almost always a major domestic news source [1]. This model is somewhat less apt in the Malaysian context
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
50
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
where the domestic news sources are more likely to suppress or fail to report stories of interest to the local blogosphere, and news stories depicting the events of interest, are more likely to be from foreign sources [18]. Foreign sources and local sources are likely to encode the same event differently. For example, the Malaysian English-language daily The Star reported the BERSIH rally of November 2007 as follows24: Sunday November 11, 2007 Teargas and water cannons used on illegal assembly, 245 held By LOONG MENG YEE, PAUL CHOO and RASHITHA A.HAMID KUALA LUMPUR: Police used teargas and water cannons to disperse thousands of people who tried to march from Jalan Tun Perak to Dataran Merdeka for an Opposition-led illegal assembly yesterday. Inspector-General of Police Tan Sri Musa Hassan said the police were forced to use teargas and water cannons because the marchers refused to disperse when instructed to do so at 2.30pm. He put their number at 4,000. He said 245 people were detained for questioning but were released after their statements were recorded. “We exercised restraint in our approach and only resorted to using teargas and water cannons to disperse the gathering,” Musa said, adding that no teargas or water was sprayed into the Masjid Jamek mosque in Jalan Tun Perak. The gathering was organised by The Coalition for Clean and Fair Election (Bersih), a group of 60 non-governmental organisations supported by five Opposition parties. …
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
The BBC reported the rally as follows25: Saturday, 10 November 2007, 16:12 GMT Malaysia police break up protest By Robin Brant BBC News, Kuala Lumpur Malaysian police have used tear gas and water cannon to disperse thousands of protesters who were marching in Kuala Lumpur to demand electoral reform. The event was organised by a group called Bersih which is made up of opposition parties and dozens of non-governmental organisations. The marchers were prevented from entering central Merdeka Square because police said they did not have a permit. Police estimated the crowd at between 10-30,000 people. As we can see from Table 4, the two encodings of the event are quite different. The Star, aligned with the Malaysian government, focuses on the police and doesn’t 24
http://thestar.com.my/news/story.asp?file=/2007/11/11/nation/19443759&sec=nation. 2008 25 http://news.bbc.co.uk/2/hi/asia-pacific/7088877.stm Retrieved August, 2008.
Retrieved August,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
51
mention the sponsor of the rally until the 4th paragraph and doesn’t mention the purpose of the rally. The BBC account, on the other hand, mentions the sponsoring coalition in the second paragraph. The BBC account uses the Anglicized place name “Merdeka Square”, which unfortunately does not redirect to the Malay place name “Dataran Merdeka” in dbPedia. The Star uses the full name of the Coalition for Clean and Fair Elections and its nickname (BERSIH), while the BBC uses only the nickname. Finally, and problematically for our purposes, the BBC story and the Star stories assign a different day to the event because the BBC essentially stamped the story with its filing date while the Star timestamps the story with its publication date. Thus, the challenge is to find a metric of event similarity that is clusters event depictions in this way in a reasonable way. Table 4. Comparison of Two Event Encodings for BERSIH Rally Star encoding of BERSIH rally Place: Kuala Lumpur Day: Sunday, November 11, 2007 Jalan Tun Perak Dataran Merdeka Opposition Inspector-General of Police Tan Sri Musa Hassan Masjid Jamek (* discount/3) Coalition for Clean and Fair Elections (Bersih) (* discount/4)
BBC encoding of BERSIH rally Place: Kuala Lumpur Day: Saturday, November 10, 2007 Malaysian Kuala Lumpur BERSIH Merdeka Square (* discount/2)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Non-textual Representations of Events Sites like Flickr (photos) and YouTube (videos) have become important repositories of depictions of such social disruptions as demonstrations, protests, riots, fights, and accidents. Users make depictions of these events discoverable primarily by means of tagging. Tagging a media object (text, video, photo, etc.) is providing a set of keywords that the user feels would describe that depiction for others seeking depictions of that event. These tag phrases are not selected from any preauthorized set of phrases for depicting events or other things. They are simply whatever the user thinks would be useful to others looking for a similar depiction. As such, they display a high level of variety. In some cases, a tagging site will suggest tags for a particular item, either based on what others have tagged the same item (e.g. bookmark URLs in the case of del.icio.us) or through processing the text of the item (e.g. Twine.com). As manifested by posts about the BERSIH rally in Malaysia, digital photography and video and sites on which such media are shared have become an important outlet for sharing depictions of local events with a worldwide audience. A crawl of all the outgoing hyperlinks from blog posts about the BERSIH rally, for example, produced 373 URLs on the Flickr and YouTube sites26. These provide a further challenge in identifying depictions of the same event. These present a further challenge for clustering blog posts by the events they reference. Rather than using named entities and discourse structure to identify the place, participant and day, one would have to rely on metadata like the folksonomic tags applied, system timestamps, and mediaspecific metadata such as camera timestamps and geo-coordinates. 26
This is out of a total of 19,426 outlinks from 1,271 blog posts about BERSIH in the week following the event, including blog rolls and other static links.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
52
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
Most frequently, a tag simply asserts the name or kind of thing depicted: people are tagged with their names; scenes are tagged with their location and visible features and setting (e.g. Paris, Eiffel tower, rain, evening). Events make a challenging case for processing tagged items, however, because public events often do not have an obvious description. For example, the set of tags in Table 5 come from a set of 54 YouTube unique videos linked to by various bloggers discussing the BERSIH rally. Table 5. Tag sets for top YouTube videos. BERSIH rally, Nov. 2007 politics, BERSIH, Reformasi, Malaysia, KL, Kuala, Lumpur Kuala, Lumpur, gathering, Bersih Malaysia, Kuala, Lumpur, demonstrations, rally, Hamish, McDonald, Al, Jazeera, Aljazeera, grassroots, outreach BERSIH, demonstration, Malaysia, rally, 10, november, dataran, merdeka 101, East, AlJazeera, Malaysia, protest, democracy, government, tear, gas, political, reform, south, east, asia, asian, human, rights BERSIH, malaysia, 10, November, SPR, demo, demonstrasi, protest, pkr, pas, dap, ngos, dataran, merdeka
In general, different YouTube users almost never tag videos of the same event in the same way. Further, each set of tags is too precise for use as a query to find videos of the same event: almost all of them contain too many keywords. Thus, since using all of the tags as keywords would direct the system to look for the boolean conjunction of those terms, the recall associated with each set of tags as a whole is very low, where recall is the number of relevant videos returned as a fraction of the entire set of videos depicting that event. The key to identifying videos of the same event here is noticing that the term ‘BERSIH’ is common to nearly all of the tag sets (except the third and fifth). Table 6. Flickr tag sets for top Bersih rally photos bersih malaysia, bersih, hindraf Copyright © 2010. IOS Press, Incorporated. All rights reserved.
hijab kuala lumpur, bersih, farispis farispis lagi, bersih Traffic jam, my life, sungaibesi, bersih
Flickr, unlike YouTube, allows multiword tags, facilitating tokenization. In addition, Flickr exposes camera metadata representing the geocoordinates and time when the image was taken, if the camera records such information and the user makes it visible, although it is possible, for instance, to find camera timestamps on Flickr that are off by many hours, presenting photos as having been taken before or after the event actually occurred. Nevertheless, correlating Flickr depictions of events (Table 6) with news accounts (and secondarily, with blog posts or online videos about the same event, in our sense) may be possible with some additional effort and some laxness concerning calendar attributions.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
53
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. Related Work In “Mining Political Blog Networks” [7], Wojciech Gryc and a team from IBM Research’s Predictive Modeling Lab outlines the use of Topic-Link Latent Dirichlet Allocation analysis, incorporating both network structure and blog content, in order to predict links between new posts in the (US) political blogosphere. This work is part of IBM’s Blog Analysis of Network Topology and Evolving Responses (BANTER) project, which aims to discover coherent subregions of the blogosphere, identify authoritative blogs there, detect emerging topics and identify the sentiment associated with the post. In [9], John Kelly (Columbia) and Bruce Etling (Harvard) use a clustering algorithm to discover similarities in Persian language blogs, independent of the location of their authors, based on the URLs to which they link. This mapping is said to reveal four major subsections of the Persian blogosphere, which they label Secular/Reformist, Conservative/Religious, Persian Poetry and Literature, and Mixed, a grab bag of topics in sports, popular culture, and specialized interests. This work categorizes Persian blogs by topic, and thus allows comparison of blogs by topic matter. The distribution of a set of terms extracted from Persian Wikipedia is used to model topic. Unlike the IBlogs system, it does not aim to identify what events the Persian blogosphere as a whole, or any particular subregion of it is focused on at any time in a more specific way. In [10], Jason Kessler outlines a system that uses pattern matching over the output of a shallow parser to identify “stances” (positive, negative, or neutral) towards a proposition, e.g. the content of a that-clause as in “I agree that cats are fickle.” Kessler then uses these techniques to identify the relative proportion of bloggers by stance towards a particular proposition. Kessler’s system is indicative of the power that relatively simple techniques can have. Kessler’s techniques may be more applicable to the global blogosphere than to measuring directly the support for or against a particular proposition in a local political blogosphere, like our Malaysian example. We found that none of the blog posts we examined concerning the BERSIH rally expressed support for or against the same deontic statement (e.g. “the police should…”, “the government ought to…”) except via posting the same message in multiple blog post comment sections. In “Sentiment and affect analysis of Dark Web forums: Measuring radicalization on the internet” [3], Hsinchun Chen and his colleagues in the “Dark Web” project at the University of Arizona presented an automated approach to the sentiment and affect analysis of radical Jihadist web forums. Chen and his colleagues were able to show empirically that one forum had significantly more violent and hateful content than the other and that the level of violent expressions increased in one forum across a specified period while remaining relatively constant in the other forum. Finally, in “The DoD Encounters the Blogosphere” [6], Rebecca Goolsby identifies several “illusions” and “delusions” to be avoided in defense or intelligence analysis of the blogosphere. According to Goolsby, the first illusion is the idea that blogs and other social media are vast untapped reservoirs of information which can inform decision-making but that don’t first require the development of new tools and methodologies for analysis. The second is that machine translation of blogs and social media is adequate for analytic purposes. Goolsby claims that machine translation cannot convey the cultural, religious, or historical context of messages. Finally, Goolsby’s third illusion is that terrorists conduct their business on a parallel “Dark
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
54
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
Web” which is rife with actionable information about operations; in contrast, Goolsby believes that, at most, propagandists, radicals and terrorist sympathizers attempt to spread their ideology and recruit sympathizers via forums on the ordinary Web. Fortunately, we believe that the work described here does not fall victim to any of these illusions. 5. Conclusions The IBlogs project demonstrates that ontologies are useful for aggregating information concerning the elements of the blogosphere, topical subject matter and semantic relations between posts. First, a domain ontology is useful to allow query expansion and normalization for named entities; secondly, a domain ontology facilitates querying by ontological relations instead of by keywords. An ontology of the blogosphere is necessary for accurately aggregating information across disparate blogging platforms, despite the existence of standardized blog feed formats as RSS or Atom. Finally, an ontology of events with a granularity appropriate to the blogosphere is necessary for clustering blog posts by topic to facilitate the analysis of shifts in attention in a particular local blogosphere over time. We have outlined the use of ontologies in the system we are developing and illustrated some of the ontological issues we have encountered.
Acknowledgements
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This material is based upon work supported by the United States Air Force Office of Scientific Research under Contract No. FA9550-06-C-0023. The views and findings expressed here do not necessarily reflect the views of the United States Air Force. We would also like to acknowledge the Nutch, Wikipedia and dbPedia open source communities for their efforts.
References [1] [2] [3]
[4] [5] [6] [7] [8]
Adamic, L, and Glance, N. 2005. “The Political Blogosphere and the 2004 U.S. Election: Divided They Blog”. Proceedings of 3rd International Workshop on Link Discovery. Chicago, IL. Analysis: Tension Between Malaysian Bloggers, Authorities Appears To Intensify FEA20070914318786. OSC (Open Source Center) Feature - Malaysia -- OSC Analysis 13 Sep 07 Chen, H. and the Dark Web Team. 2008. "Sentiment and Affect Analysis of Dark Web Forums: Measuring Radicalization on the Internet, " in Proceedings of the IEEE International Intelligence and Security Informatics Conference (Taipei, Taiwan, July 17-20, 2008). Springer Lecture Notes in Computer Science. Colette, M. 2006. Blogging Phenomenon Sweeps Asia. MSN Press Release. November 27, 2006. Feyer, T., 2004. Editors' Note. The Letters Editor and the Reader: Our Compact, Updated. New York Times, May 23, 2004 Goolsby, R. 2008. The DoD Encounters the Blogosphere. Presentation slides. 1st Int’l Conference on Social Computing, Behavioral Modeling and Prediction, Phoenix, AZ. Gryc, W., Liu, Y., Melville, P., Perlich, C., Lawrence, R.D. 2008. Mining Political Blog Networks. Harvard Networks in Political Science Conference. Presentation Slides. Kaid, L.L, Postelnicu, M. 2007. Credibility of Political Messages on the Internet: A Comparison of Blog Sources. In M. Tremayne (ed.), Blogging, Citizenship, and the Future of Media. Routledge: New York.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
B. Ulicny et al. / Uses of Ontologies in Open Source Blog Mining
[9]
[10] [11]
[12]
[13] [14] [15] [16] [17] [18]
[19] [20]
Kelly, J., Etling, B. 2008. Mapping Iran’s Online Public: Politics and Culture in the Persian Blogosphere. Berkman Center for Internet & Society Publication Series. Research Publication No. 2008-01 Cambridge, MA. Kessler, J. 2008. Polling the Blogosphere: a rule-based approach to belief classification. International Conference on Weblogs and Social Media (ICWSM 2008), Seattle, Washington. Kokar, M.M., Matheus, C., Letkowsk, J., Baclawski, K. and Kogut, P., 2004. Association in Level 2 Fusion. In Proc of SPIE Conference on Multisensor, Multisource Information Fusion, Orlando, FL., April 2004 Matheus,C., Baclawski, K. and Kokar, M.M. 2006. BaseVISor: A Triples-Based Inference Engine Outfitted to Process RuleML and R-Entailment Rules. In Proceedings of the 2nd International Conference on Rules and Rule Languages for the Semantic Web, Athens, GA, Nov. 2006. BaseVISor is freely downloadable. (vistology.com/basevisor). Ooi, J. 2007. “Bloggers Sued in Malaysia.” Screenshots (blog). http://www.jeffooi.com/2007/01/ bloggers_sued_in_malaysia.php. Quine, W.V. 1986. Theories and Things. Harvard University Press. Cambridge, MA. Reporters Without Borders, 2007 Press Freedom Index. http://www.rsf.org/article.php3?id_article=24025 Reporters Without Borders. 2008. Annual Report – Malaysia. http://www.rsf.org/article.php3?id_article=25659&Valider=OK Sifry, D. 2007. State of the Live Web, April, 2007. http://www.sifry.com/alerts/archives/000493.html Ulicny, B. 2008. Modeling Malaysian Public Opinion by Mining the Malaysian Sopo Blogosphere. In H. Liu, J. Salerno, M. Young (eds), Proc. Of 1st Int’l Conference on Social Computing, Behavioral Modeling and Prediction, Phoenix, AZ. Ulicny, B., and Baclawski, K. 2007. New Metrics for Newsblog Credibility, Proceedings of 1st International Conference on Weblogs and Social Media (ICWSM'07). Boulder, CO Ulicny B., Baclawski, K., and Magnus, A. 2007. New Metrics for Blog Mining. Proceedings of SPIE Defense & Security Symposium ’07 Vol. #6570. Orlando, FL. Wu, T.H. 2005. Let a Hundred Flowers Bloom: A Malaysian Case Study on Blogging Towards a Democratic Culture. Paper presented at the 3rd Annual Malaysian Research Group In UK and Eire at Manchester Conference Centre, 4 Jun 2005.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[21]
55
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-57
57
Chapter 4
A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support Terry JANSSENa, Herbert BASIKa, Mike DEANb, Barry SMITHc a Lockheed Martin Corporation b BBN Technologies Inc., cUniversity of Buffalo
Abstract: Lockheed Martin Corp. has funded research to generate a framework and methodology for developing semantic reasoning applications to support the discipline of Intelligence Analysis. This chapter outlines that framework, discusses how it may be used to advance the information sharing and integrated analytic needs of the Intelligence Community, and suggests a system / software architecture for such applications. Keywords: intelligence analysis, semantic technology, reasoning, common logic
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction The volume of data available to intelligence agencies and the complexity of the national security environment are increasing so rapidly as to overwhelm a finite workforce of analysts. Machines, running knowledge-based applications, are needed to augment human cognitive capacity in order to achieve the levels of situational awareness desired by decision-makers and commanders. We describe the state of the art in semantic approaches (i.e., ontology-based solutions) to this problem, in both the intelligence domain and in other domains with similar information and knowledge requirements. We also outline the results of Lockheed Martin Corp. research to address some of the specific challenges confronting the integration and fusion of data generated by multiple intelligence agencies.
1. State of the Ontological Art Over time, each intelligence agency has developed its own mechanisms for representing data, information, and knowledge. The divergence of resulting representations and standards poses severe obstacles to the automated integration and management of data, etc. This gave rise to a US Government Executive Order [1] of August 27, 2004, which expressed a mandate to the effect that intelligence agencies must strengthen their mechanisms for the sharing of terrorist information, for example through more widespread and systematic use of XML and similar markup standards.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
58
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
Experience in other domains (especially in bioinformatics) suggests that ontologies can play a crucial role in supplementing such standards and in guiding their development and use. The following paragraphs discuss the Gene Ontology, from which the reader can observe parallels with similar issues faced by intelligence agencies. The Gene Ontology (http://www.geneontology.org) is the most successful ontology initiative thus far when measured by numbers of users and of supporting software tools, and by the diversity of problems to which it has been applied. The Gene Ontology is a controlled vocabulary for describing the attributes of gene products. It has been used to annotate data derived from biomedical experiments. It makes these data, which would otherwise exist in multiple separate silos maintained by multiple separate research communities (i.e. akin to the current structure of the various intelligence agencies), more easily integratable and comparable. Lessons from experiments on the effects of given drugs or toxins or pathogens on mice, for example, can be drawn to help in the understanding of what the comparable effects might be in human beings. This success has led to the creation of other ontologies covering other domains of biology, including the cell ontology or the phenotype ontology, which now form part of a higher-level ontology-based integrating framework that is being realized already within the context of the Open Biomedical Ontologies (OBO) Foundry initiative [2, 3]. A plurality of ontology modules is being created by different community groups using both Web Ontology Language (OWL) and OBO-specific ontology formats against a background of common development principles designed to ensure their interoperability. The OBO Foundry ontologies contain terms designed to capture in an algorithmically reasonable form the qualitative aspects of biomedical phenomena. We focus on qualitative data here, similarly, in the security area, pertaining for example to intention or threat, to religion and family relationships, or to relative spatial location, as expressed for example in observation reports [4]. In this we go beyond traditional methods of what is called “information fusion” having been developed primarily for integration of quantitative data. Experience with OBO Foundry has shown that a combination of semantic technologies is appropriate for capturing such qualitative data. Our goal here, more specifically, is to advance the needs of intelligence agencies in integrating and interpreting very large bodies of qualitative data through the application of semantic technology. Little et al. describe elsewhere in this volume those aspects of our project which pertain to the use of ontologies to support multi-INT data fusion when enhanced through the consideration of probabilities [5]. Here we confine ourselves to describing the general approach and to giving some outline of the results.
2. Premise of Current Research The premise of this research is that a system can be built using ontological models which enables multi-INT data to be semantically fused; that is, to be integrated in such a manner as to allow automated reasoning to be performed over the combined dataset. This is by virtue of the fact that the data is associated with terms for which reasoning rules are defined in ontologies. The semantic benefits of ontologies result from the fact that they capture certain aspects of the meanings of those terms, in addition to their taxonomic and other structural relationships.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
59
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
For intelligence data to be semantically fused, it must exist as “stand-alone” facts or assertions. That is, extracted from, or tagged within, intelligence products, such as images, reports, cables, etc. They must also be able to be uniquely accessed from relational and other types of databases. Part of this research investigated the integration of data extraction tools in a semantic architecture. Considerable effort has been expended on the entity extraction problems from free text, and other unstructured data formats. In this work, we focused on entity extraction from imagery. For this, we used an image semantic mark-up tool, GARCON™ [6] which will be discussed in greater detail in a later section. In particular, our research focused on using COMINT (Communications Intelligence) data, such as telephone transactions, to illustrate and demonstrate our conclusions. The problem addressed was the automated generation and analysis of social networks. The work of associating persons-of-interest with telephone transactions, and of investigating degree-of-separation-type problems, had been performed previously. In this effort, we attempted to demonstrate the integration and use of multi-INT data (i.e., data provided by two or more intelligence agencies or disciplines) to improve the quality and utility of the social networks generated. Specifically, the integration of IMINT (Imagery Intelligence) information such as geolocation and geo-temporal relationships, with social relationships. For example, an ontological model may allow the system to conclude (on the analyst’s behalf) that if a cell phone has been used in a communications transaction and that that cell phone is geo-located to a particular place and time, then the owner of that cell phone may be similarly located. In this case, simple rules allow the performance of inferential steps linking communication events to specific telephone numbers and thus to people; thereby adding intelligence value to the social networks involved. The ultimate goal of this research is to extend such reasoning to include different data types, in order to advance our total situational knowledge. For example, to data types associated with intelligence questions dealing with activity, behavior, and intent. Our assumption at this point is that entity extraction and markup will have previously occurred (e.g., from natural language text), and that the “facts” extracted have been entered into a knowledgebase, perhaps virtual. 2.1. Differing Agency Missions and Vocabularies There are many efforts in the DNI/IC/DoD/DHS to develop common vocabularies for particular agencies and their missions. Some examples are DISA’s Directive to establish a common vocabulary and define a set of services and interfaces common to DoD information systems [7], and DoD’s Directive including Unique Identification (UID) Standards for a Net-Centric Department of Defense, with common vocabulary and definitions [8]. An example is the information and vocabulary associated with the domain of “terrorism.” This is defined as all information, whether collected, produced, or distributed by intelligence, law enforcement, military, homeland security, or other activities relating to [9]: a)
The existence, organization, capabilities, plans, intentions, vulnerabilities, means of finance or material support, or activities of foreign or international terrorist groups or individuals, or of domestic groups or individuals involved in transnational terrorism,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
60
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
b) Threats posed by such groups or individuals to the United States, United States persons, or United States interests, or to those of other nations, c) Communications of or by such groups or individuals, or d) Groups or individuals reasonably believed to be assisting or associated with such groups or individuals. Each agency and mission area dealing with a particular subset or aspect of terrorism and may use differing terminology and interpretations. A key conclusion of our research is that these differences must be respected (from a practical perspective), and we do not advocate attempting to define and enforce a single vocabulary. We intend the concept of semantic fusion to overcome these obstacles. 2.2. Overview of the Concept of Semantic Fusion
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
The solution proposed by this research is to integrate the conceptual knowledge models of the various intelligence agencies by constructing a formal comprehensive conceptual model (CCM) which spans the Intelligence Community and DoD. Our proposed concept is illustrated in Figure 1. The purpose of the CCM is twofold. First, it models those aspects of the real world which are not contained in INT-specific models. Second, it provides an integrating framework in which INT-specific concepts can be understood.
Figure 1. Integration Via Comprehensive Ontological Framework
The first conclusion reached by this project is the recognition that conceptual models (i.e., ontologies) written in the same system of logic cannot be effectively Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
61
integrated. This is true because small differences in “common” concepts cannot be detected (or resolved) if both versions are represented in languages with the same “expressivity.” That is, the difference itself may exceed the language’s ability to represent it. The relevance of this conclusion to the problem at hand is that the Intelligence Agencies and DoD need to develop (or already have developed) INT-specific conceptual models, which define terms and relationships in similar, but not exact, ways as noted above. They have behaved as stove-pipes. In general, the most advanced of these models are represented in the Web Ontology Language / Description Logic (OWL-DL). Description Logic (DL) refers to a system of logic intended to provide a basis for “describing” information, entities, and relationships, and for supporting inference within these descriptive models. The full CCM proposed by this project will need to be written in a more expressive language than OWL-DL. It will be written in the dialect of Common Logic (CL). Being a first order logic, CL allows much greater resolution in modeling the real world. It is therefore capable of representing differences between the OWL-DL models developed by the individual intelligence agencies. An analogy to this line of reasoning is the use of real mathematics to understand integer arithmetic. The CCM will also need to address the concept of uncertainty [10]. Pragmatically, this approach does not impose a common vocabulary across the Intelligence Community, or force substantial rework to harmonize agency-specific ontologies. Rationalization across ontologies is achieved via mapping to a higher-order logical standard.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2.3. Rationale for Approach Rationale for the viability of this approach is derived from a prototype ontology-based knowledge framework, for which we have been able to demonstrate both fusion and reasoning capabilities of the sort required. Our framework is designed to provide services useful to intelligence analysts by allowing them to draw on the functionality of ontologies without imposing the use of a common vocabulary. At the same time however our approach addresses needs being increasingly felt across the intelligence community to force substantial harmonization of agencyspecific approaches to knowledge representation. As in the biological domain of the OBO Foundry, it exploits the benefits of modularity in building a common upper-level framework to which agency-specific representations can be mapped according to specific needs, in ways which guarantee the interoperability of the modules used. As this research continues, we will continue to extend the range of incorporated modules, again drawing on the example of the OBO Foundry family of ontologies, to achieve broader coverage across the intelligence / security domains.
3. Discussion of Research and Conclusions 3.1. Semantic Image Markup Tool One component in our framework is the GARCON™ tool referenced above, which is receiving attention especially in the GEOINT (Geospatial Intelligence) community as a tool for annotating images. GARCON uses an ontology, written in OWL-DL, as the
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
62
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
basis for an analyst’s semantic mark-up of an image. Semantic mark-up implies that the entities and relationships extracted from an image are able to be placed in the correct context (as specified by the ontology). The value of semantic markup of imagery is illustrated in the following example. If an analyst is viewing an image of an airfield, and detects the presence of aircraft and other entities, semantic mark-up (as opposed to non-semantic markup) enables the following types of inferences to be made automatically, in addition to the mere augmentation of the metadata associated with that image. a) The relative position of the aircraft with respect to other entities in the image (e.g., a fuel truck) may indicate the operational status of the aircraft, b) The type of aircraft itself indicates the range of possible activities that it may be engaged in, and possible insight into the intent of cognitive entity (e.g., the commander) which caused that aircraft to be in that location, c) The possible situation / context which “best explains” the pattern of all of the entities in the image (i.e., abductive logic). The role that semantic image mark-up played in this research is that it demonstrated how one type of intelligence data (e.g., from GEOINT) can be represented in a manner which enabled it to be automatically fused with other types of intelligence data, to allow deeper situational awareness.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3.2. Semantic Multi-INT Data Integration Our envisioned unified, but modular, framework for intelligence-related information will contain three levels of ontological models: general, intel-related, and INT-specific. These levels correspond to the industry terminology of upper, mid-level, and domain ontologies. Upper ontologies describe concepts such as time, location, part-ness, process, event, etc. Mid-level ontologies introduce concepts of interest across the intelligence domain (as well as other large domains), such as person and organizational relationships, threat, capability, etc. These models enable the integration of information described in domain-specific ontologies. For our purposes, domain-specific ontologies contain the concepts unique to a particular intelligence agency. For example, electron transactions. The framework provides a coherent methodology for capturing, expanding, and reusing sets of models, at all levels, developed by master analysts and subject matter experts. The various types of intelligence (e.g., Geospatial (GEOINT), Signals (SIGINT), Human (HUMINT), Open Source (OSINT), etc.) will be integrated as shown in Figure 1. Our hypothesis is that even though different bodies of information are described using different ontologies based on different logical approaches, they can be unified and reasoned over by automated tools given the right sort of representational and computational framework. Crucial to this framework is the use of a more powerful logic – the CL (Common Logic) standard – to integrate information annotated using ontologies formulated using semantically weaker languages, particularly OWL-DL (Description Logic). Following the lesson of the OBO-OWL conversion project referred to above, we have created similar facilities to convert OWL-DL ontologies to the CL format within the framework of our IRAD project. We have converted information from different
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
63
sources into a common format, and been able to reason over the aggregated knowledgebase. Among the benefits of this approach is that first-order logic allows a level of model sophistication that is not easily achieved (or may not be possible) with DL. That is, CL provides significantly enhanced expressivity, as befits a first-order logic (FOL). We believe that pulling OWL-DL content into an FOL reasoner is the only viable approach to providing what is called “all-source” analysis. As far as we know, this project is the only application of this approach. The Vampire project by the OWL working group Manchester translated OWL into FOL [11], as did Stanford University, but neither focused on end-user application. Tradeoffs are of course involved, since with the increased expressivity of CL means that the resultant framework is no longer marked by the feature of decidability. We believe, however, that decidability is of lesser concern, when compared with expressivity. This assertion has been demonstrated in projects such as the Large Knowledge Collider [12], trading off completeness for scalability. This trade-off is justified as seen in an effort supported by IKRIS [13] as well as the results described below. In biology and other sciences, gaps or inconsistencies in data can be addressed via increased experimentation. In the intelligence domain, gaps and inconsistencies, along with a variety of related problems, are not typically resolvable in this way. Specifically, our information comes from multiple uncoordinated sources, many of which will produce either:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
a) Similar information about the same instances, b) Similar information about distinct instances (which may accordingly be confused), c) Differing information about the same instance (this can produce conflicting information, e.g., concerning the spatial or temporal location of an event), d) Differing information about different instances (there may be two separate but related items being tracked in different ways). There is inevitable uncertainty whenever we are attempting to reason about how instance-level data fit together to form a common operating picture. Intelligence reports are noisy and information is incomplete. In addition, there is a conscious attempt by adversaries to both deny access to information and to increase the uncertainty / ambiguity of that information which is observable. This means that all of the mentioned alternatives will generate knowledge problems, for example because we sometimes believe that two instances are identical when they are in fact distinct. In practical terms it means that combining probabilistic reasoning with semantic technology is an important enabling capability for multi-INT fusion. Facilities for probabilistic reasoning are accordingly an essential component of our project. 3.3. The Need for High-Expressivity Ontology Languages Intelligence agencies have developed INT-specific terminologies for describing qualitative data which define terms and relationships in semantically similar but not necessarily identical ways. Many of the most advanced of these models have been encoded in OWL-DL. OWL-DL is a W3C standard with many attractive algorithmic properties. However, OWL-DL does not provide the degree of expressivity needed for
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
64
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
practical models in the intelligence domain. That is, it cannot easily express complex qualitative information, especially in areas where time and change are involved [14]. For this reason our project draws on the resources not only of OWL-DL but also of the more expressive power of Common Logic (CL), an ISO standard language for firstorder logic and related logics. A code fragment written in a Common Logic is shown below. This construct states that a “meeting” is an event (i.e., an instance of type event), in which two people (i.e., agents) are inferred to have met if they are both related to the same event which occured at a particular time. :Prop Meeting :Inst EventType :Sup Event :Name “Meeting” :Lex “?1 is a meeting” ( (and (Meeting ?e) (agent ?one ?e) (agent ?two ?e) (occursAt ?e ?t))
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
(holdsIn ?t (meetWith ?one ?two))
We draw specifically on the resources of CL with Well-Founded Semantics [15], which has desirable computational properties when used to reason over large bodies of data. CL also has the high expressivity we need to represent complex real-world situations, particularly the expressivity needed to describe things that are changing / evolving over time. For example, dynamic entities such as organizations which gain and lose members, change locations, establish differing relationships and associations, etc.[16]. Well-Founded Semantics [17] provides fast and efficient query-answering capabilities even when addressing large data collections containing 10s of millions of assertions. Like OWL-DL, Common Logic is XML compliant. At the same time, CL is marked by a high degree of syntactic flexibility, so that individual CL systems may use a variety of non-XML syntactic frameworks which are always mapable to a fully XML-compliant syntax. Our system will be useful only if it is responsive to the query needs of intelligence analysts in both accuracy and timeliness. It must therefore by computationally tractable. The solutions provided by HighFleet, Inc. (formerly, Ontology Works) addresses this tractability by including proprietary heuristic algorithms in the reasoner which detect and resolve conditions of non-decidability or runaway. Many organizations in the intelligence and security domains have adopted XML for interoperability between systems and organizations; as is evidenced for example by the XML Directive from the Dept of Navy [7]. XML is, or will become, the de facto standard for the DoD and IC. RDF and OWL-DL, and eventually CL, are logical steps in the progression of standardization, in part because they have been developed on top of XML.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
65
3.4. System / Software Architecture
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 2 shows our proposed system architecture for a semantically-enabled system to support intelligence analysis. The primary interface with analysts is through a query engine operating on a combined, multi-INT knowledgebase. This virtual knowledgebase, has access to all of the information extracted from the intelligence products produced by each of the intelligence agencies, subject to security constraints. We assume that the extraction and markup is based on each agency’s native ontological models and representation methods. The transformation of this information to our proposed comprehensive ontological framework will occur on the knowledgebase side of the interface. Therefore, these transformations (if needed) will be transparent to the contributing agencies, after the semantic compatibility issues have been resolved. The ontological model (or set of ontological models) is the proposed comprehensive model for the intelligence domain. This is what we propose to write in Common Logic. It will drive the knowledgbase, as well as the transformations of agency-unique data into the common format. The set of automated tools (e.g., reasoners, etc.) will operate on the knowledgebase itself, and will support the understanding and response to queries received via the query engine. It must be noted that although the intelligence analyst appears to be peripheral to this architecture, the entire system is intended to be a support tool to him. The analysts and other subject matter experts are also integrally involved in the development of the ontologies themselves.
Figure 2. Proposed System Architecture
3.5. Experiment This project itself was designed to validate the concepts discussed above by demonstrating how the utility of a social network model could be improved via the integration of dynamic spatial and temporal relationships. This would in effect illustrate how data from various intelligence agencies could be fused. For example, social relationships generated by one agency could be combined with location information generated by another agency.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
66
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
A fragment of code is shown below. It provides the logic to determine if there is an entity at a certain location during a particular time interval. ( (and
(fnd.entity ?x12 ?x11) (fnd.temporalIndex ?x12 ?x109) (time.intervalContains ?x109 ?x106) (time.intervalStartedBy ?x106 ?x107) (garcon_time.xsdDateTime ?x107 ?start) (or
(georss.where ?x12 ?x39) (fnd.spatialRegion ?x12 ?x39))
(or
(geo.representativePoint ?x39 ?x38) (rcc.invPart ?x39 ?x38))
(or
(gml.pos ?x38 ?coords) (garcon_space.pos ?x38 ?corrds)))
(holdsIn ?start (locatedIn ?x11 ?coords)))
( (and
(fnd.entity ?x12 ?x11) (fnd.temporalIndex ?x12 ?x109) (time.intervalContains ?x109 ?x106) (time.intervalFinishedBy ?x106 ?x107)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
(garcon_time.xsdDateTime ?x107 ?start) (or
(georss.where ?x12 ?x39) (fnd.spatialRegion ?x12 ?x39))
(or
(geo.representativePoint ?x39 ?x38) (rcc.invPart ?x39 ?x38))
(or
(gml.pos ?x38 ?coords) (garcon_space.pos ?x38 ?corrds)))
(holdsIn ?finish (locatedIn ?x11 ?coords))) The resultant Lockheed Martin multi-INT ontology (organized in the framework structure) was created to accommodate data gathered from a variety of intelligence sources. All data is stored in a knowledge server, which is retrieved / processed via a Common Logic query engine. The knowledge server used in this project was provided by HighFleet, Inc. (formerly known as Ontology Works). The ontology defines both a
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
67
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
data model and a conceptual model, effectively tying together the data from multiple sources into a single understanding of the world. To test the viability of our approach, we conducted the following experiment. Three datasets from the National Counterterrorism Center (NCTC), the Mondial project (using the 1996 CIA World Factbook), and the Garcon-F project were incorporated into our pilot framework together with fictional (randomly created) HUMINT and SIGINT data. The ontology was created in Common Logic to integrate the semantics of these datasets. Data ingest programs were written to pull the data into a knowledge server, which was then used to answer queries that spanned multiple domains. In our chosen scenario, we asked for details about a kidnapping (i.e., an event recognized by NCTC), the membership and leadership of groups blamed for the action, communications between group members mentioning particular keywords, and imagery that positioned those communications in a geo-spatial, geo-temporal framework. Figure 3 illustrates how an analyst may write a query, associated with the above problem, in a graphical language provided by HighFleet, Inc. The relationships needed are defined in the ontology, and the instance data is contained in the knowledgebase. Once the query is defined, the knowledge server will respond with whatever data matches the defined patterns. Figure 4 contains the response to this query, also in graphical format, based on representative data developed by this project. One of these rules of inference is used to conclude that two people met at a particular time given an event description of a meeting. These rules of inference provide clarity both for the ontologist and for the query writer, who can concentrate on the relatively simple relationship between two people rather than a complex event. The rules of inference were also used to integrate data from different sources. In one case, the Garcon-F geospatial/imagery ontology was incorporated into the MultiInt ontology. Rules were written that combined facts based on this OWL-DL ontology to conclude the location of objects at particular times.
Figure 3. Illustration of Graphical Query Interface
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
68
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
Figure 4. Illustration of Graphical Query Response
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. Conclusions The execution of this experiment did in fact demonstrate the ability to integrate multiINT data, natively expressed in lesser semantic formats, into a cohesive whole. That is, to provide the capability to reason over a consolidated knowledgebase. The models in the framework provided the information needed to make the implicit semantics of the native models more explicit. The experiment also confirmed, albeit in a non-exhaustive manner, that systems based on Common Logic (with Well-Founded Semantics) could overcome some of the computational difficulties associated with first order logics. Even if not “provable,” such systems could greatly enhance the analytical capabilities provided to intelligence analysts. In the foregoing we have described only the basic outlines of the project. In addition we are realizing a number of additional components, including image annotation, data import and results visualization. Our major focus is to construct the engineering required to take our approach to multi-INT data integration into production, and to bring our pilot testing on artificial data to the level where the approach can be thoroughly tested by information analysts on large bodies of real-world data.
Acknowledgments We are grateful to the Lockheed Martin Corporation, and also to the following for crucial contributions to this work: William Andersen, Ryan Kohl, Mitch Kokar,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Janssen et al. / A Multi-INT Semantic Reasoning Framework for Intelligence Analysis Support
69
Kathryn Blackmond Laskey, Eric Little, Sarah Taylor, Jim Pridgen, Tony Chen, Kirsten Gheen and Jessica Hartnett.
References [1] [2] [3] [4] [5]
[6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[17]
http://www.america.gov/st/washfileenglish/2004/August/20040827173509adynned0.614330.html http://obofoundry.org/. Smith B, Ashburner M, Rosse C, et al. “The OBO Foundry: Coordinated evolution of ontologies to support biomedical data integration”, Nature Biotechnology (in press). Maureen Donnelly, "Relative Places”, Applied Ontology, 2005 (1). Eric Little, Kathryn B. Laskey, Terry Janssen, “Ontologies and Probabilities: Working Together for Effective Multi-INT Fusion”, Ontologies for the Intelligence Community, Columbia MD, November, 2007. GARCON is a product and trademark of BBN Technologies, Inc. (http:www.bbn.com). http://doni.daps.dla.mil/Directives/05000%20General%20Management%20Security%20and%20S afety%20Services/0500%20General%20Admin %20and%20Management%20Support/5000.36A.pdf. http://www.dtic.mil/whs/directives/corres/pdf/832003p.pdf. https://www.icmwg.org/ciss/introduction.asp. Kathryn B. Laskey, Paulo C. G. Costa, Terry Janssen, “Probabilistic Ontologies for Knowledge Fusion”, Proceedings of the 11th International Conference on Information Fusion, July 2008. http://www.cs.man.ac.uk/~horrocks/Publications/download/2004/TRBH04a.pdf. http://www.larkc.eu/overview/. http://nrrc.mitre.org/NRRC/ikris.htm. Leo Obrst, “Ontologies for Semantically Interoperable Systems”, Proceedings of the Twelfth International Conference on Information and Knowledge Management, November 2003, 366-369. Weidong Chen, Terrance Swift and David S Warren, "Efficient Top-Down Computation of Queries Under the Well-Founded Semantics", Journal of Logic Programming, 1993, 12, 1-199. K. Hornsby and S. Cole, “Modeling moving geospatial objects from an event-based perspective”, Transactions in GIS, 11(4): 2007, 225-243. A. Van Gelder, K.A. Ross and J.S. Schlipf, The Well-Founded Semantics for General Logic Programs. Journal of the ACM 38(3) pp. 620--650, 1991.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-71
71
Chapter 5
Ontologies for Rapid Integration of Heterogeneous Data for Command, Control, & Intelligence
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Leo OBRST1, Suzette STOUTENBURG, Dru MCCANDLESS, Deborah NICHOLS, Paul FRANKLIN, Mike PRAUSA, and Richard SWARD The MITRE Corporation
Abstract. Ontologies enable explicit expression of collective concepts and support Machine-to-Machine (M2M) interactions at the semantic level. Ontologies expressed in a standard language, such as the Web Ontology Language (OWL) and exposed on a network offer the potential for unprecedented interoperability solutions since they are semantically rich, computer interpretable and inherently extensible. In this chapter, we describe how we applied ontologies in OWL for rapid enterprise integration of heterogeneous data sources to track objects in a battlespace. We found that once a robust foundational domain ontology is established, it is easy and quick to integrate new data sources and therefore rapidly provide new system capabilities. In particular, we demonstrate how moving tracks can be quickly integrated with intelligence and space events to provide enhanced situational awareness using ontologies. This chapter also describes the overall SEER and SWORIER systems we developed, the latter of which translated OWL ontologies (and RDF instances) and Semantic Web Rule Language (SWRL) rules into Prolog, applied knowledge compilation techniques, and then at runtime, utilized a combined OWL/logic programming reasoned for efficient automated reasoning. We also briefly describe a more recent extension to the prototype and to the ontologies that we made to address more rigorous geospatial rules for unmanned autonomous vehicle (UAV) avoidance. Finally, we consider some issues raised by our work and future lines of research to address these. Keywords: Ontology, Web Ontology Language, OWL, Semantic Web, Enterprise integration, Agile systems, Service-oriented architecture, Logic programming, Automated reasoning, Knowledge compilation, Command and Control, C2, Intelligence, Situational awareness.
Introduction Increasingly Command and Control (C2) systems require the ability to respond to rapidly changing environments and intelligence. C2 systems must be agile, able to integrate new sources of information rapidly for enhanced situational awareness and
1
Corresponding Author: Leo Obrst, The MITRE Corporation, 7515 Colshire Drive, M/S H305, McLean, VA 22102-7508, USA. E-mail: [email protected]. Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
72
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
response to real-time events. Data from varied sources across the world must be integrated and transformed into knowledge that can be leveraged. Machine-to-machine capabilities are also increasingly necessary to accomplish mission goals. To this end, we developed ontologies and rules to address emerging mission needs. We have found that ontologies and rules offer a powerful tool for rapid enterprise integration. With these, we were able to integrate new sources of data within hours, instead of weeks or months as with traditional software development methods. Our work was showcased at the Joint Expeditionary Force Experiment (JEFX) 2008 [11] for its quick integration of data into usable intelligence-fed C2. This chapter describes the use case and its extensions, the ontologies used to model the use case, and how they support rapid, enterprise integration of C2 and intelligence information, and our prototype Semantic Environment for Enterprise Reasoning (SEER) and the ontology translation and reasoning subsystem called Semantic Web Ontologies and Rules for Interoperability with Efficient Reasoning (SWORIER). Also we acknowledge previous related research in using ontologies for situational awareness and information fusion such as [1], [2]. In addition, there has been interesting non-ontological work concerning convoy movement (the basis of our use case, below) [3], [4]. We initially describe the primary use case, a command and control (C2) and military intelligence scenario involving convoy movement across a theater of operations. The scenario addresses automated support for situational awareness (SA) in order to provide machine assistance to the commander in determining better courses of action (COA) for the convoy. We also discuss aspects of the design of the ontologies, then discuss the prototype, including the overall SEER system but also its SWORIER component. SEER integrated an approach to a service-oriented architecture (SOA) using an enterprise service bus (ESB) with the SWORIER ontology translation and runtime automated reasoning system, and attached Google Earth to the bus as visualization support for the convoy movement. All of these components were extended to later support maritime and space events in a military theater, rather than just ground events. Part of this involved our transitioning to the JEFX 08 experiment focused on Global Space Effects Information Services (GSEIS) to support the United States Strategic command (USSTRATCOM). We also briefly describe a more recent extension to the prototype and to the ontologies that we made to address more rigorous geospatial rules for unmanned autonomous vehicle (UAV) avoidance. Finally, we discuss how our approach can provide value to data integration and interoperability in other venues and for other purposes and consider some issues raised by our work and future lines of research to address these.
1. Use Case and Extensions Initially our research focused on a military C2 domain with a supply convoy moving through an unsecured area. Figure 1 depicts a convoy moving north along a primary route, approaching the location where intelligence has reported an enemy
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
73
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 1. Convoy movement using theater, routes, regions of interest (shown as green circle), etc.
sniper is stationed. Because we decided to focus on the integration of real-time tracks with intelligence information, we modeled Cursor On Target (CoT) Ground Moving Target Indicators (GMTI) and US Message Text Format (USMTF) Intelligence Summary (INTSUM) messages. New information can become available at any time, such as the discovery of a new enemy object in theater, change in weather, etc., either via immediate convoy recognition or through various intelligence information communicated to the convoy by way of intelligence summaries (INTSUM) and visual and ground moving target indications (VMTI and GMTI). We also wanted to show how an Unmanned Aerial Vehicle (UAV) could be re-tasked to collect additional intelligence information, given a particular set of circumstances. Both sources of military intelligence, INTSUM provides a summary of the most current enemy situation covering a period of time designated by the commander whereas GMTI/VMTI provides real time information on ground movers. Both are the result of human reported and sensor based intelligence. Through the ontologies and associated rules, the system provides alerts and recommendations to the convoy commander. The alerts and recommendations enhance situational awareness by fusing events; that is, multiple events from different intelligence sources are combined to form battlefield conditions, which trigger alerts and recommendations. In Figure 1, a convoy has moved so that now its region of interest (the circle surrounding the convoy) has encompassed an enemy unit. In this situation, the system might generate an alert based on an intelligence report of enemy sniper in the vicinity and recommend that the convoy takes an alternate route [5]. Our first extension to the base use case was to focus more on dynamically modifying a service behavior based on real-time events. We had to select a dynamic event to add to the scenario to trigger a change in rule sets. We selected a change in visibility as the real-time trigger event. Therefore, the most important extension to the mission use case was the expansion of the scenario to cover dynamic visibility conditions. Changes in visibility conditions may arise both routinely (day/night) and because of variable environmental conditions (such as sandstorms, etc.) From an operational perspective, such conditions require adjustments in operations. From a technical perspective, it is necessary to identify requirements for applying semantic rules to changing situations that arise in the real-time battlespace. This extension to the
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
74
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
scenario required the addition of rules that could infer different alerts and recommendations under different visibility conditions. So we constructed a second set of rules for low visibility, replicating the first set, then making changes in just two areas. First, Region of Interest (ROI) sizes were enlarged under the low visibility rules. Second, the treatment of moving objects with unknown intent was changed. Instead of deploying a UAV to take a closer look at unknown objects (which wouldn’t make sense in low visibility), we instead treat the unknown mover as a potential threat. So the rules are designed to alert the Convoy Commander that the unknown mover is a potential threat that should be avoided.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 2. Demonstration Scenario to Highlight Dynamic Change in Rule Sets
We also designed our demonstration scenario to highlight the dynamic change in rule sets under the dynamic event that a sandstorm occurs. Our approach is shown in Figure 2. Our demonstration consisted of 2 main steps. First, we use the same demonstration from year 1, in which a Convoy nears a sniper position and changes route. Next, the Convoy Commander is alerted on an unknown mover is now approaching, and the Commander sends out his UAV to take a closer look. The UAV confirms that the unknown mover is in fact, an enemy, and the appropriate alerts are generated. The second step of the scenario is designed to demonstrate the dynamic rule swap. We use the same scenario as before except that we inject a simulated real-time sandstorm event. When the event is injected, the ROIs immediately are enlarged. Also, when the unknown mover appears, no UAV is dispatched, and instead, the unknown mover is treated as a threat and the Convoy Commander is advised to avoid the potential threat. After showing how ground position and intelligence data could be integrated using ontologies, we extended our prototype by adding event types, including space events,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
75
live satellite positions and ship movement, as reported by additional intelligence sources. We added these events in just hours. The events included: (1) Missile Events added to Convoy Scenario (2) Space Events in support of a Global Strike scenario (3) Live Satellite positions (4) Maritime Events in support of a Homeland defense scenario. As an example, Figure 3 shows a pilot entering into an area in which satellite communication is degraded. Figure 4 shows a constellation of satellite positions.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 3. A pilot enters an area of degraded satellite communication. The ROI in red shows the projection of the satellite coverage area onto the Earth.
Figure 4. Google Earth view showing constellation of satellites in real time (satellite positions obtained from WWW).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
76
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
We implemented the case in which satellite element sets and propagated satellite positions are made available via web services, and reasoned about. Rules were slightly modified to detect satellite proximity. That is, we used location data in X,Y,Z coordinates to detect satellite proximity. Rules can now alert when two real satellites are too close together. Finally, we built a maritime demonstration for United States Northern Command (USNORTHCOM). This scenario involved two ships passing one another just off the coast of the mid Atlantic area. Events are simulated to indicate that the two ships are maneuvering, that one ship is performing evasive maneuvers, and that some of the crew have links to extremist organizations. These events are combined in different ways to form alerts and recommendations, ultimately leading to the recommendation to intercept the ship.
2. Ontology Design To model the objects and events described in Section 2, we constructed five ontologies:
• • • • •
TheaterObject – battlefield objects and reports about them. RegionOfInterest – battlefield regions of interest. Convoy – the convoy, its mission, components, etc. Convoy Routes – routes the convoy might take. ConditionsAndAlerts – how the knowledge base aggregates events, resulting in conditions and alerts that affect the convoy.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 5 shows the high level relationships between each original ontology and its major concepts (in black and dark gray; subsequent modifications are in light gray). TheaterObjects are MilitaryUnit, Sniper, RoadObstacle, and Facility. TheaterObjects have a location, and may have a speed, heading, and combatIntent (hostile, friendly, etc.)
Figure 5. Original ontologies, with modifications (in light gray).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
77
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
To distinguish the entity in theater from reports about it, we specified the class of ObservationArtifacts, intelligence reports about objects in theater. ObservationArtifacts have properties such as timeOfObservation, locationOfObservation, speedObservation, etc. The distinction between theater object and observation is important, allowing inference over multiple reports about the same object in theater. The RegionOfInterest (ROI) ontology models the geospatial areas of special interest surrounding TheaterObjects. An ROI is centered on the position of its focal object, and has shape, dimensions and area -- the dimension and area dependent on the type of threat or interest. ROIs are used to define a “safety zone” around a convoy which must not be violated by hostile or suspicious objects. An ROI also models the area around a reported hostile track that defines the potential strike area of the threat. The Convoy ontology, using [6], models the organized blue (US & allied) forces moving on the ground and includes the Convoy's mission, components and personnel. The ConvoyRoute ontology represents the paths of a convoy, including critical points (CPs) for primary and alternate routes. Recommended routes can change based on application of rules. The ConditionsAndAlerts ontology models situations on the battlefield based on aggregations of events and actions of theater objects. Conditions based on events result in alerts and recommendations to blue forces. The ontologies were applied in an overall framework for situational awareness, described in detail in Section 4. We built rules that operated over ontologies as part of an integrated knowledge base that could be queried dynamically. For example, we built queries such as, “Provide all the TheaterObjects within the region of interest of a particular convoy” or “Provide all threats within the convoy’s ROI.” The dynamic nature of semantic queries offers agile capabilities since any portion of the knowledge base can be queried. With SEER we are able to provide new capabilities very quickly. For example, by adding satellite positions and maritime events (displayed in yellow in the figure) to the TheaterObject ontology, instances of those classes are automatically retrieved. We are thus able to integrate new sources of data in hours.
3. SEER Prototype Design We integrated the ontologies and rules that model C2 scenarios and battlefield intelligence into a loosely coupled service-oriented architecture that uses XML-based messages. The high-level design of the application is shown in Figure 6. The components of the system include the following.
• • •
Enterprise Service Bus (ESB) Google Earth2 Client SWORIER (Semantic Web Ontologies and Rules for Interoperability with Efficient Reasoning) [7]: o Reasoner, implemented in AMZI! Prolog Logic Server3
2 3
http://earth.google.com/ http://www.amzi.com/
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
78
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
o Knowledge Base (KB), composed of ontologies in OWL with instances, • • • •
rules in SWRL Situational Awareness Service (SAS) Event Mediation Services (EMS) Adaptors Message Simulator (MS)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 6. SEER architecture with SWORIER.
We use Mule4 as the ESB abstraction layer over disparate messaging technologies, allowing interaction between components with minimal code development. Mule supports transport and transformation of publisher/subscriber pairs, applying the XSLTs of the Adaptors when appropriate. Mule also detects events, including trigger events that cause the swapping of knowledge bases, enabling us to integrate sources for satellite information and other events. We use Google Earth since it offers seamless integration of multiple data sources via its Keyhole Markup Language (KML), and also provides excellent maps and zoom capabilities. AMZI! is the platform on which we host the integrated ontologies and rule base to perform efficient runtime reasoning. The KB consists of integrated ontologies, rules and instances. OWL ontologies and SWRL rules were translated to Prolog, then optimized [7]. Together with the reasoner, these constitute SWORIER, which we will describe in more detail in the next section. SAS detects events (message exchanges over the ESB), consults the knowledge base, and delivers appropriate alerts and recommendations to the convoy commander via Google Earth clients. Events can be object movement, changes in weather, changes in alert conditions, etc. These events constitute reception of simulated INTSUM, GMTI, VMTI, and other intelligence reports. The service can dynamically query the KB.
4
http://mule.codehaus.org/
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
79
EMS handles different types of service communication including SOAP synchronous request/response, SOAP pub/sub, polling and REST. SEER uses EMS to interact with outside message sources. The Adaptors are a set of XSLTs that are invoked by the ESB to translate messages to the appropriate format as they move between components. Events are in an XML format that contains the AMZI! command format, and are asserted to the KBs and translated to KML for display on Google Earth. The active KB generates alarms and recommendations (when queried by the MS) and these messages are translated to KML for display. The MS sends messages over the ESB to simulate events on the battlefield. The SEER application works as follows. First, messages are received on the ESB, either from network sources or by the MS. The ESB applies the appropriate XLSTs of the Adaptor and commits the new information to the KB and sends KML to Google Earth.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. SWORIER Design5 We developed SWORIER (Semantic Web Ontologies and Rules for Interoperability with Efficient Reasoning), which is a system that uses Logic Programming to reason about and answer queries about ontologies and rules (Grosof et al. 2003; Hitzler et al. 2005; Volz et al. 2003). OWL (Web Ontology Language) ontologies (Bechhofer et al. 2004), along with rules in the Semantic Web Rule Language (SWRL) (Horrocks et al. 2004) or the Rule Markup Language (RuleML) (Hirtle et al. 2004) are all translated into Prolog using XSLTs (Extensible Stylesheet Language Transformations). In addition, we have written a set of General Rules in Prolog in order to enforce the semantics of OWL primitives. To do this, it was necessary to address a number of issues related to negation, the open world assumption, complementary and disjoint classes, disjunctive conclusions, enumerated classes, equivalent individuals, error messages, existential quantification, cardinality constraints, duplicate facts, cyclical hierarchies, and anonymous classes. We have imposed strong efficiency requirements on SWORIER, demanding that queries are answered in a matter of seconds. And, unlike previous work, SWORIER can assimilate dynamic changes that are provided at run time, including adding new facts, removing facts, and swapping rule sets, which also must be done in seconds. To achieve this level of efficiency, we established three techniques: extensionalization, avoiding reanalysis, and code minimization. Most knowledge representation languages and knowledge-based systems utilize a restricted version of First Order Logic (FOL). FOL, however, is semi-decidable. It is decidable in that if a theorem is logically entailed by a FOL theory, a proof will eventually be found, but it is undecidable in that if a theorem is not logically entailed, a proof of that may never be found. But decidability here does not mean tractability, and in general even inference in the simpler propositional calculus is NP-complete (Cadoli et al. 1999), i.e., usually unable to be processed in less than exponential time. To make inference tractable, various approaches in the field of knowledge compilation, which involves converting a knowledge base into a more concise or tractable representation, have been devised (Cadoli & Donini 1997; Darwiche & 5
This section is adapted from Samuel et al (2008).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
80
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Marquis 2001). One approach is to syntactically restrict the knowledge representation language, sacrificing expressiveness for tractability and efficiency (de Bruijn et al. 2004). Logic programming (LP), description logic (DL), and description logic programming (an emerging field that weds DL and LP) take this approach (Ait-Kaci 1991; Grosof et al. 2003; Hitzler et al. 2005; Van Roy 1990; Volz et al. 2003). For example, OWL is a DL that defines a tractable subset of First-Order Logic (Bechhofer et al. 2004; Daconta et al. 2003). An alternative is to employ theory approximation (Kautz & Selman 1994; Kautz & Selman 1991), in which the queries that are logically entailed by a knowledge base (a “theory”) can be correctly answered, while, for the rest of the queries, the response is “unknown”. Some researchers preprocess the knowledge in various ways to relax either the completeness or the soundness requirements, perhaps by generating certain default conclusions (Cadoli & Donini 1997). Another possible optimization is to extensionalize the rule base, as we have done.
Figure 7. System Design
Figure 7 shows the system design of SWORIER. A developer creates ontologies, knowledge bases, and/or rules in the formalism(s) of OWL, RuleML, and/or SWRL. Examples of OWL, RuleML, and SWRL are in Table 1a, 1b, and.1c, respectively. This information is translated into Prolog code using XSLTs, resulting in the code shown in Table 2a, 2b, and 2c. (We include words like "is" and "of" in our predicate names to avoid ambiguity. Otherwise, there is the danger of misinterpreting the roles of the arguments. For example, member(X, Y) could be interpreted as "X is a member of Y" or "X has a member, Y", while ismemberof(X, Y) is more clear.) Finally, a set of General Rules (defined in Section 4.2) is appended to the XSLT output to form a complete Prolog program, which can be queried by the user.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
81
Table 1. Examples of OWL, RuleML, and SWRL
a.
b.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
c.
redForceTheaterObject X redForceTheaterObject X
T G
G S
T S
4.1. Translating Facts SWORIER uses a syntax different from that typically found in previous work. For example, Volz et al. (2003) would produce the translation of Table 2d, instead of the translation in Table 2a. But we note that the syntax used by Volz et al. (2003) cannot represent "every class that smith is a member of" with X(smith), because most Prolog implementations disallow predicate variables. In contrast, by making the class names
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
82
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
and property names be arguments instead of predicates, SWORIER has the flexibility to generalize on them with, for example, ismemberof(smith, X). Table 2. Translations
ismemberof(smith, sniper). haspropertywith(smith,hasCombatIntent, friendlyIntent). ismemberof(X, redForceTheaterObject) :isMemberOf(X, redForceTheaterObject). ismemberof(smith, sniper). haspropertywith(T, hasSpeed, S) :hasPropertyWith(T, isDescribedBy, G), hasPropertyWith(G, hasSpeedObservation, S). sniper(smith). hasCombatIntent(smith, friendlyIntent).
a. b.
c.
d.
4.2. General Rules The General Rules are meant to capture the semantics of the primitives in OWL. For example, the rules in Table 3a enforce the transitivity of subclass. Note that there are two different predicates: issubclassof and isSubClassOf. One predicate would be insufficient, because Table 3b has left recursion, resulting in an infinite loop. Table 3. The Transitive Closure of Subclass
a.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
b.
isSubClassOf(C, D) :- issubclassof(C, D). isSubClassOf(C, E) :- issubclassof(C, D), isSubClassOf(D, E). isSubClassOf(C, E) :- isSubClassOf(C, D), isSubClassOf(D, E).
With two different subclass predicates, some questions must be answered. Should the user submit queries with issubclassof or isSubClassOf? Also, which form should the input from the XSLTs be? If the input used isSubClassOf, then neither of the rules in Table 3a would ever succeed, thus the input must use issubclassof. On the other hand, queries should use isSubClassOf because the issubclassof set of facts is incomplete ── none of the subclass relationships that are derived by transitivity are captured by issubclassof. Note that the issubclassof set of facts is a subset of the isSubClassOf set of facts, because of the first rule in Table 3a, which is called the conversion rule for subclass. For consistency, we created two cases of each predicate, all-lowercase and camelcase.6 Also, each predicate has a conversion rule. The XSLT facts always use the all-lowercase forms of predicates, while the user queries are always in camelcase. (However, the developer decides how to spell the names of constants, such as hasSpeed in Table 2c.) And any rules, other than recursive rules and conversion rules, follow the convention: 6 Any predicates that are not used for input or output are written in an underscore case, such as is_sub_class_of_but_not_equal_to. Also, for some predicates, there are two sources of recursion, requiring three cases of the predicate. An example of this is the member relation, for which the three cases are ismemberof, is_member_of, and isMemberOf.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
83
All predicates in the body of the rule are camelcase, and the predicate in its head is all-lowercase. As an example, see the second rule in Table 2c. Using camel-case predicates in the rule's body guarantees that the rule is triggered by everything that can be derived for that predicate in either case. And using the all-lowercase predicate for the rule's head insures that any facts generated by the rule will hold for both cases of the predicate. 4.3. Translating Rules Some of the inputs provided to SWORIER are RuleML or SWRL rules that were created by the developer. It is not difficult to translate these rules into Prolog, because they are written in Horn Clause form. However, we cannot control which rules are provided nor how they are written. Problems can emerge, such as the infinite loop in Table 2b, which was generated by the RuleML rule in Table 1b. In addition, the order in which rules are listed and the order of the terms in the rules' bodies can have significant effects, much like order of join evaluation in database languages such as SQL. Another concern is that the input could include rules that produce duplicate copies of a fact.7 And the rules might be inefficient. There are ways to correct or at least mitigate some of these problems. For example, we could apply transformations as for logic queries in the form of rewrite rules such as “magic sets” optimization (Cadoli et al. 1999; Sippu & Soisalon-Soininen 1996). But currently, we must impose strong requirements on the developer, who may need to be very familiar with Prolog programming techniques, the logical consequences of the facts in the ontologies, and the General Rules.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4.4. Challenges We followed the groundbreaking work of Volz et al. (2003), who were among the first researchers to investigate OWL-to-Prolog translation. They discussed a number of problems that they encountered in the course of their work. These included issues such as logical negation in the first-order logic and description logic worlds vs. negation as finite failure (NAFF) in the logic programming and database worlds. Another issue is that between the open world assumption and the closed world assumption. The Web Ontology Language OWL has an open world assumption, meaning that a term is false only if it can be proven false. In Prolog, the closed world assumption holds, which means that anything that cannot be proven true must be false. Our solution to these two issues was to allow for both kinds of negation by permitting normal NAFF negation and additionally creating a form of constructive negation for true logical negation (cf. Barták & Roman, 1998), in the form of an asserted logicNot construction, i.e., a specifically constructed negation such as: logicNot(isMemberOf(jones, sniper)). In addition, one can attach pragmas, i.e., instructions to the interpreter/compiler of General Rules, as to how one wants to behave with respect to open vs. closed world assumptions. For example, one could use the open world assumption and thereby make equivalences between objects validly concluded as being in a model of the ontology; or using the closed world assumption, one could flag a 7
We direct the interested reader to the discussion in Samuel et al (2008), section 4.9.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
84
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
warning that two objects validly concluded violate a cardinality constraint and thus not assert an equivalence relation between them. Other issues included allowing multiple terms in the head of the rule. This is disallowed in logic programming, which is based on a Horn Logic syntactic restriction of first-order logic which disallows it. For conjunction of terms in the head of the rule, the solution is easy: create multiple rules. For disjunction of terms in the head -- which more generally requires disjunctive logic programming as in Maedche & Volz (2003), Minker & Seipel (2002) – we developed a kind of constructive disjunction. Although there were other challenges (see Samuel et al (2008)), some additional primary challenges was dealing with existential quantification, cardinality restrictions, cyclic hierarchies, and anonymous classes – all of which the OWL standard permits.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4.5. Dynamic Changes Another useful capability is to change the knowledge base at run time. For example, in our convoy task, intelligence reports can come in at any time during a scenario, and we want SWORIER to be able to incorporate the new information into the knowledge base. This must be done within a few seconds. So, we have enabled SWORIER to accommodate dynamic changes of facts, adding or removing facts at run time. (We have not yet tried dynamically changing classes, properties, or rules.) Unfortunately, dynamic assertions significantly decrease efficiency, because the Prolog compiler can no longer be used. It is not possible to assert or retract any facts with predicates that are compiled. This issue had to be addressed by our knowledge compilation techniques (section 4.6). SWORIER is also capable of dynamically changing rules, but only in a restricted way. We require that all of the desired rule sets are available in advance. This can still be quite useful. For example, under low visibility conditions, different rules might be desired from the rules used with high visibility. Both rule sets can be developed in advance, and then SWORIER can generate a separate program for each case. These rules might be considered different policies or rules invoked by different contexts. At run time, when visibility is high, the user queries are submitted to the first program. But whenever visibility is lost, such as at the onset of a sandstorm, the two programs are swapped, and the user queries are sent to the second program. Note that whenever facts are added or removed from the knowledge base, both programs must be modified appropriately. 4.6. Efficiency Initially, the SWORIER system was too slow. In order to make the system tractable at run time, we implemented an offline technique to speed up the program. We modified SWORIER to extensionalize all of the facts that can be derived from the input (that a user might want to query on), converting the program from an intensional form to an extensional form. Figure 8 shows the modified system design.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
85
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 8. Extensionalization
At compilation time, the extensionalization algorithm runs generalized queries in order to derive all of the desired facts and save them in the extensional program. (Note that it is easy to keep duplicate facts out of the extensional code, because they always look identical.) Then, at run time, the extensional program uses the derived facts, in effect utilizing table lookup. This preprocessing technique enabled the system to work much faster. However, because the offline extensionalization step still required too much time, we introduced a couple of other techniques to speed up the compilation: 1) avoiding reanalysis, and 2) minimizing code. The former took a declarative approach to minimizing the reevaluation of code by establishing sentinels so that the evaluation process could see where it had already been – in addition to Prolog’s usual checkpoint evaluation. The latter, minimizing code, is an efficiency improvement that can be implemented if certain knowledge is available prior to run time. Given a list of all of the predicates that are used in 1) the ontology, 2) the dynamic changes, and 3) the queries, it may be possible to eliminate some of the rules, thus improving efficiency of the extensionalization process. In addition, the same technique can be used to eliminate rules in the run time program. The idea is to figure out which of the rules are actually necessary, because the unnecessary rules can be dropped, improving efficiency. If a rule can never successfully fire, then it is unnecessary. Also, if no query will ever result in testing a rule, then that rule is unnecessary. More precisely, a rule is a necessary rule only if it is both satisfiable and testable. To determine which rules are satisfiable and testable, it is necessary to figure out which predicates are satisfiable and testable, respectively.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
86
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
5. Applying Geospatial Rules for UAV Deconfliction In a later extension to our original work, we applied our architecture and results to another scenario, thus demonstrating the flexibility of our design and the extensibility of our ontology. We chose a scenario involving predicting and avoiding mid-air collisions between Unmanned Aerial Vehicles (UAVs). We also integrated other processing tools and data sources that use and build on the ontology. For this effort, we selected a UAV simulator that provides realistic UAV performance and used its inputs as our primary data source. 5.1. Use Case
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In the figure, UAV 642 is heading west to east and UAV 716 is heading east to west on a collision course with UAV 642. The controller system is receiving telemetry from both UAVs. The controller also has the ability to send a new flight plan up to UAV 716 and automatically redirect it to this new flight plan. By using the UAV’s telemetry and a mathematical formula, the controller system calculates when the two UAVs will reach their closest point of approach along their flight paths. Using a set of rules, the controller is able to detect if the UAVs will come within a pre-determined distance that is less than a safe distance resulting in a collision. The controller can also calculate at what time this distance will be reached and detect if the amount of time remaining before the collision is enough time to allow for evasive maneuvers. If the controller detects either of these conditions, it sends a new flight plan to UAV 716 and redirects the UAV to the right. In the use case, we chose to simply make UAV 716 “break right” because that is the normal evasive maneuver done by a private pilot in order to avoid a collision. In the use case, we did not consider any other factors from the situation. In a more sophisticated use case, we could consider such things as airspace restrictions, terrain, physical object obstructions, UAV mission priority, etc.
Figure 9. Collision Avoidance Use Case
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
87
5.2. Ontology Extension The addition of UAVs fit neatly into the ontology. Reports of UAV position and velocity data was organized and treated as a subclass of ObservationArtifact. Because more detailed information on the UAV’s velocity was needed, a specific subclass of ObservationArtifact called UAVObservation was created that included a vertical velocity component and explicitly represented the heading and speed as Vnorth and Veast, which is a common attribute of aircraft. The additions to the ontology are shown in the diagram in Figure 10.
Figure 10. Additional Ontology Class, shown with new Data Properties
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5.3. Rule Extensions The most significant addition was a set of rules to perform vector-based mathematical calculations of an object’s future position relative to another object or some fixed location. Several rule additions were needed to accomplish this. First, the geodetic coordinates of latitude, longitude, and altitude needed to be converted into a set of Earth Centered Earth Fixed (ECEF) metric-based XYZ Cartesian coordinates. This step was necessary because performing distance and speed calculations in geodetic coordinates is very difficult and prone to inaccuracy. This was done using a set of mathematical formulas readily obtainable from several sources and easily implemented in Prolog as a rule [13]. The implementation was kept very general in order to preserve its usefulness to a wide range of problems. The need to reason over time and space is common to many applications, and this approach was designed to be easily reusable. Second, the velocities of TheaterObjects also had to be converted from heading/speed or Vnorth, Veast, and Vdown into their XYZ Cartesian coordinates (Vx, Vy, and Vz). This involved inverting the usual conversion matrix, and again implementing the conversion as a rule.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
88
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
The third rule addition was designed to express the conditions that could lead to a collision. This was done using a Time of Closest Approach algorithm. Once the system could express positions and velocities in an XYZ coordinate frame, it is fairly easy to compute the time of closest approach. The algorithm works as follows (see [14] for a complete mathematical derivation). Given two objects P and Q, with initial positions P0 and Q0, and each moving in a straight line with velocities u and v, (in meters per second) respectively, if their initial distance apart is |w0| = |P0 – Q0| then the time tc at which the two objects will be closest to each other is given by:
tc =
− w 0 ⋅ (u − v ) u −v
2
Where tc is the number of seconds in the future from the current time. If t c is negative, then the two objects have already been as their closest point, and are now moving away from each other. Also, if |u – v| = 0 then the two objects are traveling along the same or parallel lines, and will remain a constant distance apart. If two UAVs were determined to have a tc that was less than some threshold value (we initially used 30 seconds, but later reduced it to 20 and still had good results) then the distance dc between them at time tc was found by:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
d c = (P0 + tc ⋅ u ) − (Q 0 + tc ⋅ v ) If dc was found to be less than another threshold (300 meters turned out to be a useful value) then the rule returned an alert that a potential collision situation had arisen and one of the UAVs was directed to immediately turn toward a location 45 degrees to the right and ½ a kilometer away (i.e., break right). This approach is a departure from our original methodology based on convoy movement in that the calculations are no longer based on intersecting ROIs, but rather on direct numerical calculations based on actual position and velocity. Furthermore, the calculations are memory-less; that is, the knowledge base neither keeps track of previous positions nor performs any type of predictive tracking – it simply uses the current position and velocity observations when they are reported. We are not entirely happy with this last accommodation, made for efficiency purposes, since we want to investigate issues involving correlation and fusion of multiple tracks in the future. Additionally, by keeping historical tracking information, predictive tracking may be facilitated – something we would like to also address.
6. Conclusions and Next Steps Ontologies can be applied for rapid enterprise integration, allowing delivery of new capabilities for example in C2-Intelligence applications in hours instead of weeks or months. In addition, we maintain a clear distinction between reception of intelligence information and actual theater object representation and behavior, so that we might address more complicated issues such as information fusion, wherein multiple intelligence sources might report on similar or dissimilar objects and be correlated, fused, or de-correlated or un-fused based on eventual further information and reasoning.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies for Rapid Integration of Heterogeneous Data
89
By simply adding new classes to an existing, robust ontology, we were able to leverage the power of the situational awareness framework and deliver new information and capabilities rapidly. Of course, the construction of the original set of ontologies was a manual process, and therefore expensive and time-consuming. Therefore, we are proposing additional research at MITRE to investigate, enhance and/or develop approaches for automatic ontology generation and mapping. In addition, we believe that eventually semantic interoperability requirements across multiply developed ontologies and their supporting data will require embedding these ontologies within common super domain, and upper/middle/foundational ontologies [8], [9], [10].
References [1]
[2] [3] [4] [5]
[6]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[7]
[8]
[9]
[10] [11]
[12]
[13] [14]
Matheus, C. J.; K. P. Baclawski; and Kokar, M. M.. 2003. Derivation of Ontological Relations Using Formal Methods in a Situation Awareness Scenario. Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2003, pages 298–309. SPIE. Matheus, C. J.; Kokar, M. M.; and Baclawski, K. 2003. A Core Ontology for Situation Awareness. Proceedings of FUSION’03, Cairns, Queensland, Australia. Chardaire, P.; McKeown, G. P.; Harrison, S. A.; and Richardson, S. B. 2005. Solving a time-space network formulation for the convoy movement problem. Operations Research 53(2): 219–230. Chardaire, P.; McKeown, G.P.; Harrison, S. A.; and Richardson, S. B. 2001. Convoy planning in a digitized battlespace. Journal of Defence Science, 6(2):168–175. Stoutenburg, S.; Obrst, L.; Nichols D.; Samuel, K.; and Franklin, P. 2006. Applying Semantic Rules to Achieve Dynamic Service Oriented Architectures. Second International Conference on Rules and Rule Markup Languages for the Semantic Web, at ISWC 2006, Athens, GA, November 10-11, 2006, LNCS 4294, 2006, p. 581-590. Heidelberg: Springer. Convoy Operations Battle Book. 2005. Tactics, Techniques, and Procedures for Training, Planning and Executing Convoy Operations. Tactical Training and Exercise Control Group. Marine Aviation Weapons and Tactics Squadron One. Samuel, Ken; Leo Obrst; Suzette Stoutenberg; Karen Fox; Paul Franklin; Adrian Johnson; Ken Laskey; Deborah Nichols; Steve Lopez; and Jason Peterson. 2008. Applying Prolog to Semantic Web Ontologies & Rules: Moving Toward Description Logic Programs. The Journal of the Theory and Practice of Logic Programming (TPLP), Massimo Marchiori, ed., Cambridge University Press, Volume 8, Issue 03, May 2008, pp 301-322. http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=1853440. Semy, S. K.; Pulvermacher, M. K.; and Obrst, L. J. 2004. Toward the Use of an Upper Ontology for U.S. Government and U.S. Military Domains: An Evaluation. http://www.mitre.org/work/tech_papers/tech_papers_05/04_1175/04_1175.pdf. Pulvermacher, M.K.; Stoutenburg, S.; and Semy, S. K. 2004. Netcentric Semantic Linking: An Approach for Enterprise Semantic Interoperability. http://www.mitre.org/work/tech_papers/tech_papers_04/04_1174/04_1174.pdf. Upper Ontology Summit Joint Communiqué. 2006. http://ontolog.cim3.net/cgibin/wiki.pl?UpperOntologySummit/UosJointCommunique. van der Oord, Capt. Larry. 2007. Joint Expeditionary Force Experiment 2008 (JEFX 08) Underway. Air Force Print News Today, 11/15/2007. Air Force Global Cyberspace Integration Center (GCIC) Public Affairs. http://www.globalsecurity.org/military/library/news/2007/11/mil-071115-afpn05.htm. Keen, M.; Susan Bishop; Alan Hopkins; Sven Milinski; Chris Nott; Rick Robinson; Jonathan Adams; Paul Verschueren; Amit Acharya. 2004. Patterns: Implementing an SOA Using an Enterprise Service Bus (1st ed.): International Business Machines (IBM) Redbooks. July 25, 2004. http://www.redbooks.ibm.com/abstracts/sg246346.html. Dana, Peter H. Department of Geography, University of Texas at Austin, 1995; Available at http://www.colorado.edu/geography/gcraft/notes/datum/datum.html. http://geometryalgorithms.com/Archive/algorithm_0106/algorithm_0106.htm.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-91
91
Chapter 6
Ontology-Driven Imagery Analysis Troy Self, Dave Kolas, Mike Dean BBN Technologies 1300 N. 17th Street, STE 400 Arlington, VA 22209 [email protected], [email protected], [email protected] Abstract. This paper presents a new paradigm for imagery analysis where imagery is annotated using terms defined in ontologies, enabling more powerful querying and exploitation of the analysis results. The ontology terms represent the concepts and relationships necessary to effectively describe the objects and activities within a domain of interest. A platform for viewing and editing imagery annotations is described along with a specialized semantic knowledge base capable of efficiently querying the information using semantic, spatial, and temporal qualifiers. The ontologies used for representing the annotations and domain of interest are also described.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction Imagery analysis is the process of examining overhead imagery, identifying the objects and activities present in the image, and correlating this data with information not available in the image to derive new knowledge. Current practices for capturing imagery analysis results as narrative text or in relational databases become a burden when an analysts needs to search past reports, correlate facts across multiple reports, and search for specific examples of general scenarios. The purpose of this chapter is to describe an imagery analysis environment where observations are recorded as structured annotations using descriptive semantic concepts defined in an ontology, enabling more powerful search and exploitation of the annotations than can be achieved using traditional methods. The first section provides background information about imagery analysis including the goals of the analysis and current methods of recording analysis results. Section 2 provides an overview of the imagery annotation approach proposed in this chapter. Implementation details of the approach are described in Section 3 including the ontologies used in the representation, the user environment, and the specialized knowledge base developed to enable efficient storage and retrieval of the annotations.
1. Background The goal of imagery analysis is to determine what is happening on the ground by examining overhead imagery. This begins by identifying the objects seen in the image, Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
92
T. Self et al. / Ontology-Driven Imagery Analysis
such as buildings and vehicles. From there, the analyst may try to determine the activity occurring in the image. Are there construction vehicles suggesting construction of a new facility? Are particular vehicles present to suggest a military exercise? Is cargo being unloaded? What type of cargo is it? These are simple examples of the questions that imagery analysts try to answer. Understanding what is happening occurs at multiple levels of detail. At the lowest level, the analyst determines what is occurring in a single image. The next level may involve combining this information with what is observed in other images to determine if both sets of observations are related. Is the military exercise observed here related to the government limousines observed at a nearby airport? How do the observations in this image relate to the observations of the same area in previous weeks? Is there a trend to the activity? To perform this second level of analysis requires that either the same analyst is working on all of the related images and can combine the information in his head or the multiple analysts looking over the imagery are communicating and collaborating to determine these conclusions. Given the sheer magnitude of new imagery generated daily, neither of these cases is likely. Analysts are too busy keeping up with the incoming imagery to perform broader analysis that spans multiple images, regions, or countries. For this reason, it becomes important that the method of recording imagery observations enables computers to help analysts find trends, identify related activities, and understand the big picture. Unfortunately, current techniques for recording imagery observations drastically limit both the analysts’ and computers’ ability to search and use this information. Analysts currently record their observations using narrative reports. The report may include the objects that the analysts see along with any attributes they feel are worth reporting. For example, a report about a military airfield may include, “Two MiG-23s are parked on the ramp with no visible munitions.” The analyst is assuming that anyone reading this report understands that a MiG-23 is a fighter jet with certain capabilities and that the ramp refers to a particular area of the airfield. This is true for the people that need an updated report about this specific airfield. However, what if another analyst studying a nearby area needs to know if there are any airfields with 100 miles with more than one fighter aircraft? That analyst is limited to a keyword search and therefore needs to know that she is searching for the word ‘MiG-23’ since ‘fighter aircraft’ is not part of the report. Once she expands her search to include ‘MiG-23’, she may get back thousands of reports that include ‘MiG-23’ from airfields around the world. This example question illustrates two key limitations with the current method of recording imagery observations. The first limitation is content representation. A MiG23 is represented with the string, “MiG-23”, and therefore cannot be found by searching for aircraft or military asset or other abstract terms that “MiG-23” implies. The second limitation is the lack of structured geospatial data in the report. Modern Geospatial Information Systems (GIS) enable complex geospatial analysis of structured geospatial data. The analysis methods include performing geospatial queries (e.g. Find all armed tanks within 100 yards of a particular road) and combining various types of information into a single, common picture. The lack of structured geospatial information on imagery observations prevents analysts from using these modern geospatial processing techniques. The result of these two key limitations is the lack of interoperability between imagery analysts of different groups and organizations or between imagery analysts and non-imagery analysts.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
93
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2. System Overview The system described in this chapter eliminates the limitations mentioned above by allowing analysts to capture the where, what, and when of imagery observations using terms and relationships described in ontology. This provides the semantics necessary to enable automated implications and the structure necessary for the complex geospatial queries and analysis common in GIS. This system eliminates the separation between the imagery viewing application and the annotation capture application by allowing analysts to capture annotations by drawing graphics directly on the imagery. Satellite imagery files contain additional data that enable a system to determine a geospatial location (i.e. latitude and longitude) for each pixel in the image. This extra data can be used to determine a geospatial representation for any graphics drawn by the analyst. Thus, any point, line, or polygon drawn by the analyst is immediately de-referenced to the geospatial geometry it represents on Earth. This is commonly referred to as vectorizing the data since it converts raster data (graphics) into geospatial vectors. The imagery file also includes a timestamp for when the image was captured by the satellite. This means that simply by drawing on the image, the analyst has capture the where and when of a particular observation. The analyst then provides the what by selecting terms and relationships from the ontology. Using a term from the ontology means that the annotation automatically includes any logical implications defined in the ontology. While a reference to the imagery file is maintained, the particular image being studied is no longer required for understanding the observation. Annotations are attached to their geospatial location and temporal information. This enables annotations to be projected onto other imagery files or maps for further geospatial analysis. Analysts can query and view annotations in a GIS environment by applying spatial, temporal, and ontology-based filters. The spatial filter is implied by the current extent of the GIS. In other words, if the GIS application is zoomed-in on New York City, then only annotations in New York City will be displayed. The temporal filter is applied by selecting a temporal interval. The temporal interval could focus on a specific second or multiple years. The ontology-based filters enable an analyst to state what he is looking for. Setting the filter to the Airplane class will query for annotations of anything known to be an Airplane. The approach described here provides mechanisms for capturing imagery analysis results as geospatial vectors with appropriate semantic information to allow users or computers to reason and query annotations based on geospatial, temporal, and logical axioms. Exposing these annotations using standards-based representations enables interoperability between various imagery analysis groups and other intelligence analysts. The following section describes the ontologies developed to implement this system.
3. Ontology Design The implemented ontology is designed to formalize a conceptual model of the world, enable a dynamic, context relevant user interface, and meet the data requirements of the overall system. As Figure 1 shows, the ontology is structured into three separate, but interrelated component ontologies. The foundational ontology formalizes the conceptual model used by the system and exists independent of the analytical domain.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
94
T. Self et al. / Ontology-Driven Imagery Analysis
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 1. Overview of the component ontologies designed for this system.
The domain ontology captures concepts of unique relevance to a domain (e.g. Air Defense). The application ontology meets the particular information requirements of the imagery analysis system. Each of these ontology components is discussed in more detail in the sections below. Partitioning the ontologies in this manner enables a loosely-coupled knowledge representation system that provides two key benefits. First, it allows augmentation of any partition without disruption of the others. For example, domain-specific implications and capabilities can be added into the domain ontology without requiring updates to the application or foundational ontologies. The second benefit is that each ontology partition can be used by itself or in combination with other ontologies for a separate application. The ontologies are defined using the OWL Web Ontology Language (OWL) [1]. OWL provides the interoperability requirements necessary to allow the imagery annotations captured with this system to be integrated with other sources of information to form a common, logical picture. 3.1. Foundational Ontology The foundational ontology is designed as an application independent, domain agnostic, conceptual model of the world. It contains formalizations of basic notions of time and space, mereology (part/whole relationships), and event participation. It is a suitable model for information systems that maintain information about objects in the physical world over time. The foundational ontology is an integration and augmentation of bestof-breed, publicly available ontologies. 3.1.1. Temporal Representation The temporal representation used is OWL-Time [2], a product of the W3 Semantic Web Best Practices and Deployment Working Group. It includes interval and instant
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
95
based time representations and is aligned with XML Schema built-in data types. This alignment eases application of existing RDF and XML software tools. Figure 2 provides an example of a time interval attributed to an entity using terms from the OWL-Time ontology.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 2. Example of a temporal interval using the OWL-Time ontology components within the Foundational Ontology.
The representation includes multiple levels of indirection, which may appear to add unnecessary complexity for a simple time interval. However, this representation enables intervals to be defined in terms of other intervals and for multiple intervals to share DateTime objects. This allows representation of temporal intervals that are known to exist even if the specific details are unknown. This is critical for imagery analysis when associating mobile objects to specific locations. The analyst cannot determine from an image when an object arrived in a location; only when it was first observed. Likewise, the representation must support when an analyst first notes that an object is no longer where it was before. Figure 3 illustrates how the OWL-Time concepts can represent interval data based on what is known without making any assumptions that can lead to analysis errors. 3.1.2. Geospatial Representation The concrete geospatial representation is an adaptation of GeoRSS [3], which includes a profile of the Geography Markup Language (GML) [4]. Use of GML makes exchange of geospatial data with external tools feasible. OWL-Time and GeoRSS have both been integrated with the Basic Formal Ontology (BFO) [5]. BFO is a widely studied and published formal ontology that enumerates concepts at the highest levels of abstraction. In particular, all entities in the BFO formulation of the world are either
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
96
T. Self et al. / Ontology-Driven Imagery Analysis
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 3. Representation of a time interval based on known time data and temporal relationship to other known intervals. The exact interval is unknown, so it is represented that it contains a known interval and occurs before another known interval.
continuants or occurents. Continuants are those entities that have a continuous existence and endure through time. Examples include a piece of rock or the planet Earth. Occurents are those entities that are bound in time and include processes and events. Examples of occurents include the process of walking the dog and the lifecycle of a frog. Reusing a common and well documented formal ontology makes explicit the conceptual model being used by the information system and adopted in domain specific extensions. Figure 4 provides a simple example of an imagery observation using concepts from geoRSS, GML, and BFO. It is notable that the publicly available OWL formulation of BFO contains only a class hierarchy; it does not contain any representation of intra-ontological relations. The formalization of such relations in OWL is a challenge because they are minimally ternary relations and OWL naturally models only binary relations. The foundational ontology thus contains OWL classes which model the n-ary relations necessary to capture a relevant (but incomplete) subset of the BFO intra-ontological relations. 3.1.3. Mereology Representation One of the fundamental concepts necessary to describe object interaction is mereology. Mereology is the study of objects as parts of other objects. This parthood could refer to a physical relationship (e.g. My arm as part of me). It could also refer to a logical membership (e.g. A college as part of a larger university). The goal of this representation of mereology is to describe the relationships between objects and their
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
97
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 4. Example of a point imagery observation using terms from geoRSS and GML.
parts in a way that is both fundamentally sound and also practical from an imagery analysis point of view. This representation combines the practical intentions of the SimplePartWhole ontology [6] with the logical principles of Minimal Mereology [7]. The SimplePartWhole ontology was created by the W3C Semantic Web Best Practices and Deployment Working Group. It describes parthood and direct parthood, the latter of which is useful for easily being able to determine the proper hierarchy of items in a parthood tree. Minimal Mereology provides the underlying fundamental concepts of parthood, proper parthood, Atoms (indivisible objects), and Composites (objects composed of Atoms or other Composites). No OWL representation of Minimal Mereology existed and was therefore created as part of the foundational ontology and linked to the SimplePartWhole ontology using the appropriate OWL axioms. 3.1.4. Process Representation Appropriate foundations were necessary to describe processes and events. The representation used in the foundational ontology was developed based on the Participation relations described in [8], which extends BFO. These relations are used to describe how a continuant, such as a car, participates in an occurent process, such as refueling. Figure 5 shows a class hierarchy for the various types of participation. Table 1 lists the 9 types of process participation. This participation hierarchy was added to the foundational ontology with Event as a common superclass. Each Event has a single participant and refers to a single process. The process object is how multiple participants are tied into a common process. Figure 6 provides an example representation of a repair event using the ontology terms described above.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
98
T. Self et al. / Ontology-Driven Imagery Analysis
Figure 5. Class hierarchy noting the 9 subtypes of participation in a process.
Table 1. Participation Types Participation Type Perpetration
Initiation Perpetuation Termination
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Influence
Facilitation
Hindrance
Mediation
Patiency
Description Relation indicates a substance perpetrates an action (direct and agentive participation in a process). Examples include: a referee fires the startingpistol; the captain gives the order. Relation indicates a substance initiates a process. Examples include: referee starts the race; the attorney initiates the appeal process Relation indicates a substance sustains a process. Examples include: the singer sings a song; the charged filament perpetuates the emission of light Relation indicates a substance terminates a process. Examples include: the operator terminates the projection of the film; the judge terminates the imprisonment of the pardoned convict Relation indicates a substance has an effect on a process. Examples include: the hilly slopes affect the movement of the troops; the politicians influence the course of the war Relation indicates a substance plays a secondary role in a process (for example by participating in a part or layer of the process). Examples include: the catalyst provides the chemical conditions for the reaction; the traffic-police facilitate our rapid progress to the airport Relation indicates a substance has a negative effect on the unfolding of a process (by participating in other processes). Examples include: the drug hinders the progression of the disease; the strikers prevent the airplane from departing Relation indicates a substance plays an indirect role in the unfolding of a process relating other participants. Examples include: the Norwegians mediate the discussion between the warring parties; the mailman brings Mary's letter to John Relation indicates a substance is being acted on by a process. Examples include: the pistol is being fired; the song is being sung. This is the dual of agentive participation (perpetration)
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
99
Figure 6. Example representation of a repair process using the foundational ontology. The event is a patience process that points to a Repair Process, which implies that the participant is the thing being repaired.
The foundational ontology integrates BFO, OWL-Time, GeoRSS, Minimal Mereology, and Smith and Grennon’s Process Participation hierarchy in a manner that makes it useful in applications which have a requirement to model and capture entities in the physical and temporal world.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3.2. Domain Ontology The domain ontology is designed as a pluggable ontology used by the user interface to enable imagery annotation within a particular domain of interest. It defines the terms and relations that the analyst can use for annotation and the system can use to process logical implications. It can be as simple or as complex as the domain requires for appropriate annotation so long as the domain can be appropriately described using OWL-DL. The ontology is loosely coupled to the rest of the ontologies so that it can be swapped for a different domain with minimal disruption. Air defense was used as the domain of interest for the purpose of this prototype. The domain ontology described here formalizes air defense concepts. Many of these air defense concepts are adapted from publicly available sources of information on air defense topics, such as the Federation of American Scientists. The ontology is also aligned with National System for Geospatial-intelligence (NSG) feature catalog to promote reuse. This catalog provides a list of features and some relationships among the features. Names of features from the catalog are consistent with names used in the ontology. The NSG feature catalog does include subsumption (subclass/superclass relationships). These relationships are added, where appropriate, when NSG features are added to the domain ontology.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
100
T. Self et al. / Ontology-Driven Imagery Analysis
The domain ontology is aligned with the foundational ontology in order to determine which concepts are available to the analyst for annotation purposes. Specifically, some classes are subclasses of IndependentContinuant to express that they are standalone entities which an analyst can use to annotate an image (e.g. MiG-21). Other classes are subclasses of Qualities to indicate that these concepts can only be used as temporally changing attributes of an IndependentContinuant (e.g. the operational status of a MiG-21). Another example of alignment with the foundational ontology is that some classes are subclasses of Process. This indicates that these classes are to be used to indicate that some process or event is taking place (ex. fueling a MiG21). The air defense ontology leverages the Description Logic characteristics of OWL to appropriately describe relationships between terms and relations. For example, equivalence between classes is used to allow specification of aircraft using the official name or the NATO designation (e.g. MiG-23 versus Flogger). Cardinality and owl:hasValue restrictions are used to provide default attributes for physical objects represented in the ontology. This enables the system to find annotations based on attributes that the analyst may not have actually annotated (e.g. Finding an aircraft based on its cargo capacity). 3.3. Application Ontology The application ontology represents data that is specific to the function of this application. For this system, the application ontology contains terms related to imagery analysis. The imagery analysis application ontology includes image metadata such as the date and time an image was taken and the name of the file. The imagery analysis application described in the next section is appropriately coupled to this ontology. The domain ontology and foundational ontology do not rely on the application ontology, allowing them to be used for a different application area, such as document annotation.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. Imagery Analysis Environment The imagery analysis application is implemented as a plugin for the ESRI ArcMap GIS [9], a popular mapping and geospatial analysis tool. The plugin includes a custom layer for viewing geo-registered imagery and marking annotations as well as custom user interface controls for creating new and searching existing annotations. The user interface controls’ content is generated dynamically based on the ontology terms and relationships in the knowledge base. This allows users to immediately leverage modifications and enhancements to the domain ontology without having to wait for deployment of a new version of the application. The user creates new annotations by using the custom controls to describe the observation using the semantic terms defined in the domain ontology. The application automatically captures the timestamp and geospatial details of the annotation. By capturing the temporal and spatial extent of the observation, annotations can be linked and searched using time, space, and description regardless of whether they originated from one or more images. When viewing imagery, the user can use the custom query controls to filter the visible annotations based on spatial, temporal, or semantic qualifiers. For example, when viewing an airport, the user can choose to only show observations of support
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
101
vehicles within the hangar area within the past 7 days. This search relies on the ability of the knowledge base to understand what qualifies as a support vehicle and to efficiently eliminate observations that occur outside the specified spatial and temporal extent. The user environment also includes an advanced query interface that allows the user to write custom SPARQL queries that cannot be defined using the UI controls. As an example, this interface allows the user to query for all cases where aircraft maintenance was observed twice within the same week, within the same airport.
5. Spatiotemporal Semantic Knowledge Base The knowledge base (KB) is the repository for all data in the system. This includes data created by the analyst along with any inference from the ontology. The knowledge base therefore must support fast access using spatial extents, temporal extents, and combinations thereof. The knowledge base uses the Jena Semantic Web Framework [10] for query and graph processing, BBN’s Asio Parliament KB [11] as an underlying RDF [12] storage mechanism, and libraries from BBN’s Openmap GIS [13] application for spatial indexing. Custom Jena Graph interfaces were developed to integrate Asio Parliament KB and the spatial and temporal indexes into the knowledge base. The custom interfaces encapsulate the implementation details, allowing transparent use by the query interface. The custom graph interface for the indexes facilitates ordering and splitting the queries between the semantic and spatiotemporal processing components.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5.1. Approach The knowledge base storage approach is modeled after the well-established method used in relational databases that support spatial operations. The primary storage mechanism is attached to an external index which is capable of efficiently processing certain portions of the query. Here, this means that a RDF/OWL storage mechanism is connected to indices for spatial and temporal information. The design deviates from the object-relational model in that no new concepts are introduced into the data manipulation language. This will be discussed in more detail in the following sections. 5.1.1. Design and Implementation At a high level, the knowledge base has the same components that any data processing system has. There is a network interface layer, a query processor, and physical storage for the data. In designing these components, there were several primary goals: • • •
Provide efficient processing of SPARQL queries that contain spatial and temporal components Refrain from modifying the SPARQL query language, so that SPARQL clients would not have to be modified Reuse as much existing Semantic Web architecture as possible
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
102
T. Self et al. / Ontology-Driven Imagery Analysis
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Efficient processing of spatiotemporal SPARQL queries is required to achieve the types of analysis with ontological data that have been discussed in this chapter. By not modifying the SPARQL query language, any client software that can create SPARQL queries can create queries for our knowledgebase. A full description of why this is desirable can be found in [14]. Finally, by reusing as much Semantic Web existing code as possible, maximum interoperability is ensured. The implementation of the knowledge base is rooted in the Jena Semantic Web framework and the Joseki HTTP query interface. Jena and Joseki together can provide end-to-end persistent RDF storage, but it was preferable to modify this solution to meet the goals of this implementation. The HTTP query interface was left unchanged. The following diagram shows the default Jena stack versus the modified version for spatiotemporal indexing:
Figure 7. Knowledgebase query processing stack. The stack on the left represents the default stack when using Jena within the Joseki interface. The stack on the right illustrates the changes necessary to accommodate the spatial and temporal indices as well as the high-performance RDF storage component, Parliament.
This change in architecture required changes to three primary parts of the Jena stack. First, the Jena Memory graph is removed, and is replaced with the Indexing Graph built for this purpose. The goal of the Indexing Graph is to handle the splitting of both inserted data and requested queries between the primary storage and the indices. These processes will be described in later sections. Underneath the Indexing Graph, persistent storage is provided by Asio Parliament. This was done to increase efficiency of RDF/OWL queries [15]. Finally, the spatial and temporal index processors are
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
103
attached to the indexing graph to provide efficient processing of the spatial and temporal portions of queries. 5.2. Assumed Ontologies
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In order to process data insertion and queries without altering SPARQL, the knowledge base must be able to understand certain ontologies of time and space. By assuming that these ontologies are used, the knowledge base can know which data must be inserted into the index and which data must come from which index in a query. As suggested in [14], the knowledge base uses GeoRSS to describe the spatial portions of entities, and an ontology based on RCC8 for expressing spatial relationships. Analogously, for temporal entity descriptions, OWL-Time is used [2], and an ontology based on the Allen time intervals [16] is used for temporal queries. Figure 8 is an RDF representation of a spatiotemporal annotation.
[] a foundational:SpatialLocation ; foundational:entity [ a airdef:SA-19 ]; georss:where [ a snap:SpatialRegion ; geo:representativeRegion [ a kb-space:Polygon ; kb-space:exterior [ a kb-space:LinearRing ; kb-space:posList "31.8149 -106.3090 31.8161 -106.3102 31.8169 -106.3085 31.8162 -106.3071 31.8157 -106.3075 31.8149 -106.3090" ] ] ]; foundational:temporalIndex [ a time:ProperInterval ; time:intervalContains [ a kb-time:ProperInterval ; time:intervalStartedBy [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime "2006-12-13T16:35:36-05:00"^^xsd:dateTime ]; time:intervalFinishedBy [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime "2006-12-14T16:35:36-05:00"^^xsd:dateTime ] ]; time:intervalBefore [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime "2006-12-15T16:35:36-05:00"^^xsd:dateTime ] ]. Figure 8. RDF representation of an annotation using OWL-Time and GeoRSS.
Similarly, spatiotemporal ontological queries can now be expressed in SPARQL as shown in Figure 9.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
104
T. Self et al. / Ontology-Driven Imagery Analysis
SELECT ?location ?entity ?startTime ?finishTime WHERE { ?location a garcon:SpatialLocation ; garcon:entity ?entity ; garcon:temporalIndex ?validInterval . ?validInterval a time:ProperInterval ; time:intervalContains [ a kb-time:ProperInterval ; time:intervalStartedBy [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime "2000-12-15T16:35:36-05:00"^^xsd:dateTime ]; time:intervalFinishedBy [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime ?endTime ] ]. ?validInterval time:intervalBefore ?finish . ?finish a kb-time:DateTimeInterval ; kb-time:intervalBefore [ a kb-time:DateTimeInterval ; kb-time:xsdDateTime ?finishTime ] } Figure 9. Example SPARQL query that is selecting SpatialLocations (observations) whose interval starts on December 15, 2000 at 4:35pm.
5.3. Data Insertion
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In order to populate the indices and thus improve query efficiency, appropriate items must be detected for insertion into the index. First, consider the addition of a point location, which consists of the objects, its semantic type, and its geo-location. Example (1) represents an SA-19 Surface-to-Air Missile launcher and its geo-location. [] a airdef:SA-19; georss:where [ a gml:Point; gml:pos “38 -77” ].
(1)
This example demonstrates that three RDF statements are necessary to represent an object at a location. Therefore, the trigger for indexing must be based on all three statements rather than a single property. The rule mechanism built into Jena allows specification of multi-statement triggers as rules. The indexing graph contains an inference graph layered above the storage graph. This inference graph does not contain any more statements than the underlying graph because the trigger rules do not result in the assertion of any new statements. Instead, the rules used by the spatial and temporal index processors perform operations when matched. These operations call back to the index processors, allowing items to be added to the indices. Example (2) is the Jena rule that triggers on the data shown in Example (1).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
105
T. Self et al. / Ontology-Driven Imagery Analysis
[point: (?x rdf:type gml:Point) (?x gml:pos ?pos) -> point(?x, ?pos) ]
(2)
Similar rules are provided for adding polygons, temporal instants, and temporal intervals. This method of adding items to the index has both positive and negative attributes. First, it allows triggering of index insertions from multiple statements without requiring any particular triggering mechanism from the underlying storage. Second, it would allow new indices to be added dynamically, since the new index processors merely register their insertion triggers. Also, non-persistent indices will be automatically regenerated when the system is started. However, rules and triggers are not perfectly aligned. There is no guarantee that a given rule will not fire multiple times with the same input data, so the indices must actively avoid multiple copies of the same data being inserted.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5.4. Query Processing The job of query processing for the indexing graph has several steps. First, the query must be split into the parts that are processed by each index processor and the primary storage mechanism. Next, these query parts must be ordered based on an estimation of the selectivity of each portion. Finally, the result rows are spooled from the various portions of the query together, returning overall result rows that match the entire query. The query splitting portion of the process is quite straightforward. For each triple in the query, the namespace of the predicate (if the triple is not a rdf:type statement) or the namespace of the object (if the triple is an rdf:type statement, and the object is bound) is inspected. This namespace is compared against the set of namespaces registered by each of the index processors. Once the query is divided into partitions based on in which storage mechanism the answer will be found, the partitions are divided again by shared variables. This results in a set of sub-queries that will be used to answer the overall query. The next step is the ordering of the sub-queries. For each sub-query, the index processor associated with it, or the primary storage mechanism otherwise, is given an opportunity to estimate the selectivity of the sub-query. In this case, selectivity refers to the ratio of number of results to total number of items in the index. For example, querying a phone book for people with the name “Adam” is not as selective as querying for people with a specific phone number. This sub-query estimation step is exactly like the query planning stage in a spatial RDBMS. These estimates of selectivity are then used to order the sub-queries, with the more selective sub-queries going earlier. In this implementation, the estimation of the sub-query selectivity is actually rather naïve; however, it is expected that this will be an important area of future work. Finally, the results of the query are spooled to the user. The sub-query with the highest selectivity estimation is started, and the variable bindings coming from that sub-query are passed into the other sub-queries. This is modeled after the behavior of query portions in the ARQ query processor, the SPARQL query engine for Jena.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
106
T. Self et al. / Ontology-Driven Imagery Analysis
5.5. Indices For this implementation, to achieve rapid development, simple indices were used that could easily be integrated into the system. This meant that any index to be used must be fairly lightweight and implemented in Java. For the spatial index, the in-memory gridfile index used by BBN’s Openmap software was used. This was convenient because it was accessible as a Java library, and already contained methods that could process geospatial functions such as intersection, buffer calculation, and distance. However, since indexing of large datasets is not the actual purpose of this library, there are certainly shortcomings. As the data sets grow larger, the amount of memory required to maintain the index becomes quite large. This is at least in part the cause of slowdowns in query processing of data sets beyond a certain size [17], though the rule-based index insertion previously discussed likely plays a role as well. The temporal index, requiring only one dimension of storage, is much simpler. For this implementation, a sorted TreeSet is used to keep instants ordered. Intervals are stored as their two endpoints. Again, this implementation suffers from all of the advantages and disadvantages of an in-memory index; faster times for small datasets, but slower times once a certain threshold of memory is reached. One of the primary areas of future work on this system will be to replace the inmemory indices with persistent ones, and increase the effectiveness of the selectivity estimations on them. In doing this, the system should become stable and robust enough to handle the needs of analysis on real world data sets.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6. Results The resulting environment provides an application through which a user can examine and annotate geo-registered imagery using air defense concepts and relationships described in the domain ontology. The inference capabilities provided by the ontology enable the system to automatically enrich each annotation and draw further conclusions. This allows users to search for annotations using abstractions and characteristics that were never specifically captured by the analyst. The spatiotemporal capabilities of the knowledge base combined with the semantics of the ontology enable analysts to efficiently query for observations that occur within a spatiotemporal extent or are related spatially or temporally. Finally, the representation of the ontology and data allows the annotations to be easily linked to annotations from other intelligence sources.
7. Conclusion This chapter presents an imagery analysis environment that allows imagery to be annotated using highly descriptive semantic concepts and relationships defined in an ontology. By combining efficient semantic storage and retrieval techniques with efficient spatial and temporal indexing, these annotations can be queried and exploited in more powerful ways than can be achieved using traditional keyword search or relational database techniques.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
T. Self et al. / Ontology-Driven Imagery Analysis
107
References [1] [2]
[3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[17]
OWL Web Ontology Language (OWL). [Online] http://www.w3.org/2004/OWL/. Hobbs, Jerry. and Pan, F. 2004. An Ontology of Time for the Semantic Web. ACM Transactions on Asian Language Processing (TALIP): Special issue on Temporal Information Processing, Vol. 3, No. 1, March 2004, pp. 66-85. GeoRSS version 1.0. GeoRSS::Geographically Encoded Objects for RSS feeds. [Online] http://georss.org/1. Cox, Simon, et al. OpenGIS Geography Markup Language (GML) Implementation Specification, Version: 3.00. s.l. : OpenGIS, 2003. Specification. SNAP and SPAN: Towards Dynamic Spatial Ontology. Grenon, Pierre and Smith, Barry. 2004, Spatial Cognition & Computation, pp. 69-104. Simple part-whole relations in OWL ontologies. [Online]. August 11, 2005. http://www.w3.org/2001/sw/BestPractices/OEP/SimplePartWhole/index.html. Mereology. Stanford Encyclopedia of Philosophy. [Online] http://plato.stanford.edu/entries/mereology/. The Cornucopia of Formal Ontological Relations. Grenon, Pierre and Smith, Barry. 2003, Dialectica, Vol. 58, No 3, pp. 279-296. ESRI ArcGIS Desktop. ESRI. [Online] http://www.esri.com/software/arcgis/about/desktop_gis.html. Jena - A Semantic Web Framework for Java. [Online] http://jena.sourceforge.net. Asio. [Online] BBN Technologies. http://www.asio.bbn.com. Resource Description Framework (RDF). [Online] http://www.w3.org/RDF/. OpenMap – Open Systems Mapping Technology. [Online] http://openmap.bbn.com. Kolas, Dave. Supporting Spatial Semantics with SPARQL. Manuscript submitted for publication. Rohloff, Kurt, Dean, Mike, Emmons, Ian, Ryder, Dorene and Sumner, John. An Evaluation of TripleStore Technologies for Large Data Stores. OTM Workshops (2), 2007, pp. 1105-1114. Allen, James. Maintaining Knowledge about Temporal Intervals. Communications of the ACM, Vol. 26, Issue 11, pp 832-843. Kolas, Dave. A Benchmark for Spatial Semantic Web Systems. Manuscript submitted for publication.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-109
109
Chapter 7
Provability-Based Semantic Interoperability for Information Sharing and Joint Reasoning Andrew SHILLIDAY and Joshua TAYLOR and Micah CLARK and Selmer BRINGSJORD Rensselaer Artificial Intelligence and Reasoning (RAIR) Laboratory Departments of Cognitive and Computer Science Rensselaer Polytechnic Institute, Troy NY 12180, USA
Abstract. We describe provability-based semantic interoperability (PBSI), a framework transcending syntactic translation that enables robust, meaningful, knowledge exchange across diverse information systems. PBSI is achieved through translation graphs that capture complex ontological relationships, and through provability-based queries. We work through an example of automating an unmanned aerial vehicle by reasoning over information from a number of sources. Keywords. Provability-based semantic interoperability, translation graphs, logic, ontologies, unmanned aerial vehicles
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction Provability-based semantic interoperability (PBSI) is a logic-based approach to information sharing and joint reasoning among knowledge-driven systems that employ disparate representation schemes. The PBSI approach achieves a high degree of semantic interoperability without requiring a privileged interlingua ontology, and it preserves semantic consequences among ontologies even when bidirectional translation is impossible. In this chapter, we first describe semantic interoperability (§1), and the need for it within certain areas, e.g., in the Semantic Web and in the defense and intelligence communities (§2). Next, we review two relevant prior approaches to semantic interoperability, their merits, and their shortcomings (§3). Then, we formally present PBSI, including the treatment of ontologies, incremental ontology modification, translation graphs, and provability-based queries (§4). Afterward, we present a concrete example of PBSI-enabled interoperability (§5), and summarize several other illustrative applications of the PBSI approach (§6). As a capstone, we detail an extended example wherein PBSI is used to enable automation of an unmanned aerial vehicle (UAV) so that it acts in accordance with rules of engagement (§7). Finally, we consider future research directions (§8), and remark on PBSI’s role in system interoperability (§9).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
110
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
1. What is Semantic Interoperability?
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Prolification of knowledge-driven systems has spawned great diversity in reasoning capabilities. Ideally, these systems ought to apply their reasoning capabilities to knowledge exchanged freely and transparently with their peers—and do so despite differences in underlying representation. Unfortunately, lack of either a common representation scheme or an effective method for information exchange has greatly hindered meaningful interoperability among knowledge-driven systems. Meaningful interoperability requires not only shared protocols for communication and information exchange (what we call syntactic interoperability) but also a shared meaning (what we call semantic interoperability) so that the understanding, and consequences, of information are preserved faithfully across systems. The former, syntactic interoperability, is well-addressed by standardized interchange formats such as XML [1], RDF [2], IKL [3], KIF [4], and Common Logic [5,6]. The latter, semantic interoperability, has, as of yet, no general solution; it continues to be an ever-present difficulty in a number of areas, for example, in database integration. Existing approaches for addressing the problem of semantic interoperability (e.g., ontology mappings [7], schema matchings [8], and schema morphisms [9]) do achieve a modicum of success, and may be sufficient in certain limited circumstances. However, none of these approaches achieves semantic interoperability in the broad sense that we envision—a sense that requires seamless information sharing and joint reasoning. Of course, the sufficiency of any approach to semantic interoperability must be measured against the approach’s interoperability goals and a given situation’s technical requirements. The interoperability goals of our approach, PBSI, are meaningful information sharing, and joint reasoning, among disparate knowledge-driven systems. This pair can be explicated as follows. Information Sharing references the capability to take information expressed in the language of one ontology, and re-express it in the language of another. Note, the ability to re-express, or translate, knowledge from one system to another does not guarantee the ability to do the reverse, that is, to translate knowledge in the other direction [10]. Asymmetries of translation can arise, for instance, from a difference in subject coverage or in the expressivity of representation formalisms. Meaningful information sharing requires meta-knowledge about the ontology translations, that is, knowing what types of information can be translated from source ontology to target ontology, and what types of information cannot be translated. Joint Reasoning references the capability to reason over the knowledge contained in a number of systems while preserving semantic consequences between systems. Note that such reasoning is possible, and profitable, even among systems that cannot share information—that is, where translation between systems is impossible. For instance, a sentence in a knowledge-base may have semantic consequences in another knowledge-base, even if it cannot be translated to the other knowledge-base. Meaningful joint reasoning requires meta-knowledge about ontologies and understanding of the relationships between ontologies. These interoperability goals dictate that we adopt an expressive formalism, and rigorous process, for capturing the complex relationships of ontology vocabularies and between multiple ontologies.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
111
Given that knowledge-driven systems play increasingly critical roles in our “technologized” world, and the subtle ways in which semantic errors may arise, the trustworthiness of interoperating systems can only be established through formal analysis and verification. This analysis and verification speaks not only to the interoperability approach, but also to output products, for example, query responses. Mission-critical interoperating systems ought to produce inspectable justifications for query results. We evaluate PBSI and other approaches to semantic interoperability using provabilitybased queries—queries wherein the response is accompanied by a justification.1 This evaluation methodology reveals the limitations of prior approaches (e.g., inability to capture asymmetry in translation, or to ensure translatability of semantically relevant information), and yet is general enough to subsume more common query-correctness evaluation methods.
2. The Need for Semantic Interoperability
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In the Semantic Web. The Semantic Web requires that data be published in a structured and meaningful way. When it is, information from many sources can be used and combined easily. Currently, this requirement is met at the syntactic level by generating, storing, and presenting information in RDF [19] using, e.g., RDFS [2] or OWL [20]. At the semantic level, this requirement is only partially met, and even then, only because developers possess detailed knowledge about all of the ontologies on which their applications depend. The complete vision of the Semantic Web will be realized only when web-oriented systems share information without a priori knowledge about their peers [21]. In the Defense and Intelligence Communities. The defense and intelligence communities recognize the need for information sharing, and actively sponsor research on interoperability. Results of this sponsorship include the DARPA Agent Markup Language (DAML) [22] for the markup of information, DAML+OIL [23] for describing ontologies, and the IKRIS Knowledge Language (IKL) [3] for describing relationships between, and exchanging information among, ontologies. Furthermore, past, present, and future development of knowledge-driven systems within the defense and intelligence communities exposes the need not just for static information sharing, but for dynamic collaboration as well. Only full semantic interoperability can ensure the trustworthiness of such knowledge-driven systems.
3. Relevant Prior Approaches to Semantic Interoperability We review two prior approaches to the problem of semantic interoperability, namely, schema matchings, and schema morphisms. In the review, we consider: (i) whether an approach is logically based; (ii) whether asymmetry in translation is preserved; and (iii) whether semantic consequences can be obtained even when translation between ontologies is impossible.
1 Our evaluation methodology accords with the traditions of logic-based artificial intelligence and cognitive science [11,12,13,14,15,16] and logic-based semantic interoperability [17,18].
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
112
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
Both schema matchings and schema morphisms have achieved a degree of semantic interoperability in real-world applications. However, neither approach possesses the necessary qualities for full semantic interoperability, nor produces the inspectable justifications needed to ensure trustworthiness. By building on the foundation laid by these prior approaches, our method of PBSI takes a further step toward the ideal.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Schema Matchings. Schema matchings are applicable when the subject domains and vocabularies of two ontologies are similar, and the information expressed within them is not overly complex. In a schema matching, two corresponding terms from the vocabularies are presumed synonymous; this allows certain expressions from either ontology to be syntactically recast into the other.2 Schema matchings are non-logical. Vocabulary terms are matched based on lexical considerations. For example, a matching might be based on the fact that one ontology’s ‘lname’ field and the other ontology’s ‘surname’ field share the lexical component ‘name.’ Schema matchings are one-to-one between vocabulary terms; therefore, any asymmetry is discarded. Schema matchings do not allow semantic consequence to be transferred when translation is impossible: when there is no match for a vocabulary term, the consequences of information expressed using that term are restricted to the native ontology. Schema matchings are meta-level artifacts expressed in a formalism that differs from the representational formalism of ontologies. Schema Morphisms. A schema morphism maps a subset of expressions (as opposed to simply vocabulary terms) drawn from one ontology into expressions of another. Like schema matchings, schema morphisms are non-logical, syntactic manipulations, but schema morphisms have some key advantages over schema matchings. Firstly, schema morphisms allow the correctness of information translations to be formally verified [25]. Secondly, schema morphisms allow some semantic considerations to be realized. For example, a schema morphism might formally state that one ontology’s ‘full-name’ field corresponds to a function (e.g., concatenation) applied to the other ontology’s ‘first-name’ and ‘lastname’ fields. Thirdly, schema morphisms are able to capture some asymmetry in translation. For example, concatenating one ontology’s ‘firstname’ and ‘lastname’ fields may always yield a meaningful ‘full-name’ in the other ontology, but it is not necessarily the case that every ‘full-name’ can be decomposed into a meaningful ‘first-name’ and ‘last-name.’ Like schema matchings, schema morphisms do not allow semantic consequence to be transferred when translation is impossible. When an expression cannot be translated using a morphism, its semantic consequences are restricted to its native ontology. Schema morphisms, like schema matchings, are meta-level artifacts expressed in a formalism that differs from the representational formalism of ontologies.
4. Provability-Based Semantic Interoperability We now lay out the components of provability-based semantic interoperability. Specifically, we describe: (i) how ontologies are represented; (ii) the incremental modifications used to construct ontologies; (iii) how the modifications are recorded and associated with axioms in a translation graph; and finally, (iv) how provability-based queries are used to evaluate PBSI. 2
The process of identifying and making schema matchings has been partially automated [8,24].
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
113
4.1. Ontologies We treat ontologies as pairs of the form ۃȭǡ Ȱ ۄwhere ȭis a signature in a many-sorted logic, and Ȱis a set of sentences in the language of ȭ.3 Real-world interoperability does not always require the sophistication of many-sorted or first-order logics, but in cases where expressivity is needed, one has it, and the less demanding scenarios are easily handled in all cases as well. A signature in many-sorted logic defines the set of well-formed expressions of a particular language. Formally, a signature ȭis a pair ߪۃǡ ߶ۄ. ߪ is a partition of the language’s universe of discourse, and its cells are the sorts belonging to ȭ. The set of all sorts of all signatures is denoted ܵ כ, and thus ߪ כ ܵ ك. ߶ψis a set of functors. A functor ݂ψis a function ݏ ൈ ǥ ൈ ݏିଵ ՜ ݏ where each ݏ כ ܵ א. If, for a signature ߪۃǡ ߶ۄ, ݂ ߶ א, then each ݏ ;ߪ אthat is to say, a signature contains every sort referenced by its functors.4 4.2. Incremental Ontology Modifications
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Equations (1)–(4) define four atomic operations corresponding to primitive modifications that can be performed on ontology signatures.5 Using just these four atomic operations, signatures can be constructed that specify the vocabulary of a language. ሺݏǡ ߪۃǡ ߶ۄሻ ൌ ߪۃሼݏሽǡ ߶ۄ
(1)
ሺݏǡ ߪۃǡ ߶ۄሻ ൌ ̳ߪۃሼݏሽǡ ߶ۄ
(2)
ሺ݂ǡ ߪۃǡ ߶ۄሻ ൌ ߪۃǡ ߶ ሼ݂ሽۄ
(3)
ሺ݂ǡ ߪۃǡ ߶ۄሻ ൌ ߪۃǡ ߶̳ሼ݂ሽۄ
(4)
However, ontologies require not only a vocabulary, but also vocabulary semantics. To address this, we attach axioms to applications of atomic operations during signature construction (§4.3), and these axioms are available when evaluating provability-based queries (§4.4). Existing ontology description languages include constructs for defining various facets of vocabulary semantics. Semantic Web languages, such as RDFS [2] and OWL [27], include constructs for defining class hierarchies and specifying properties of relations (e.g., transitivity and reflexivity). In other systems, semantics are defined via arbitrary axioms; e.g., Athena and KIF provide define-symbol [28] and deffunction [4], respectively, for introducing such axioms. 4.3. Translation Graphs A translation graph is a visual trace of the incremental construction and interrelation of ontology signatures (an example translation graph is shown in Figure 1). The process of interrelating ontologies reduces to interrelating ontology signatures. The process 3 Many-sorted logic is reducible to standard first-order logic [26]. Our use of many-sorted logic is a matter of convenience—it facilitates parsimonious description of ontological constructs. 4 A signature is actually a triple wherein the third element is a mapping from function symbols to functors, but to ease exposition, and because this mapping is not relevant here, we ignore this third element. 5 There is one exception: ܴ݁݉ݐݎܵ݁ݒሺݏǡ ߪۃǡ ߶ۄሻ is undefined if ݏis used by any functors of ߪۃǡ ߶ۄ.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
114
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
starts with an empty signature (i.e., one with no sorts or functors), and then the signatures of individual ontologies are created (or recreated) by adding sorts, relations, and functions via the atomic operations defined above.6 The ontology signatures are then interrelated by the gradual transformation of one signature into another. This transformation is achieved by the successive application of atomic operations, and the associating of an axiom, called a bridging axiom, with each application. Formally, a translation graph is a directed graph whose vertices are signatures, and whose edges are relationships between signatures. Associated with each edge ۃȭ ǡ ȭ ۄ is an axiom that captures the relationship between ȭ and ȭ . Note that these relationships are associative and transitive, thus a ȭ ǡ ȭ̴݆-path in a translation graph captures the relationship between ȭ and ȭ , and between the ontologies to which they belong.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4.4. Provability-Based Queries A query is a yes/no question posed to a set of knowledge-driven systems that have been interrelated by some interoperability approach (e.g., schema matchings, schema morphisms, and translation graphs). To answer the query, the information stored in the interrelated systems must be reasoned over, along with any ancillary information required by the interoperability approach. (We use the word ‘reason’ here in a very broad sense, and construe it to include, e.g., table transformations, algorithmic search, model finding, and theorem proving.) A provability-based query is a query whose yes/no response must be accompanied by an explicit justification. Formally, a query is a tuple ۃȞǡ ߰ ۄwhere Ȟ is a set of formulae, and ߰ is a single formula of interest, the question. In the simplest case, Ȟ consists of all the information in the interrelated ontologies along with the additional information generated by the interoperability approach. In real-world applications, Ȟ can be restricted for various reasons, e.g., for efficiency, or to understand how different ontologies contributes to the answers. In our approach, justifications are formal proofs and models within many-sorted logic. This results from the treatment of ontology vocabularies as signatures in manysorted logic, and of ontology knowledge and bridging axioms as formulae in manysorted logic. (These justifications can be generated and inspected by existing tools; we inherit this benefit from prior research in formal foundations.)
5. A Simple Genealogical Example We present a simple genealogical example to illustrate the use of translation graphs. We construct a small translation graph that relates two genealogical ontologies that describe similar domains, but have no terms in common; and then we detail the evaluation of a provability-based query over the constructed translation graph. Though our genealogical example is simple, it demands capabilities beyond the reach of the aforementioned prior approaches (§3). Specifically, the example’s genealogical ontologies have no semantically equivalent terms, thus schema matchings are ineffective. Furthermore, schema morphisms would provide only an incomplete solution because there are some sentences that cannot be directly translated between the 6 Creating, or recreating, one or more ontologies from the empty signature is largely a formality; we elide doing so when there is no danger of confusion.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
115
Figure 1. This translation graph interrelates ࣩଵ and ࣩଶ , two genealogical ontologies. Primitive modifications and corresponding bridging axioms appear as labels on the edges of the graph.
genealogical ontologies. PBSI overcomes these difficulties by relating the ontologies by axioms extracted from a translation graph. 5.1. Constructing the Translation Graph The first ontology, ࣩଵ , has a single sort, ܖܗܛܚ܍۾, and the binary relations ۱܌ܔܑܐ, ܚ܍ܜܛܑ܁, and ۰ܚ܍ܐܜܗܚ. In addition, ۱܌ܔܑܐሺݔǡ ݕሻholds provided that ݔψis a child of ;ݕ ۰ܚ܍ܐܜܗܚሺݔǡ ݕሻand ܚ܍ܜܛܑ܁ሺݔǡ ݕሻhold just in case ݔψis a brother or sister, respectively, of ݕ. ࣩଵ also includes axioms defining the semantics of its vocabulary. For instance, one such axiom is Axiom (5), which states that no individual is its own child. ௫ ۱܌ܔܑܐሺݔǡ ݔሻ
(5)
The second ontology, ࣩଶ , has the same sort, ܖܗܛܚ܍۾, and two binary predicates ܜܖ܍ܚ܉۾, and ܖܑܔ܊ܑ܁. ܜܖ܍ܚ܉۾ሺݔǡ ݕሻis true if and only if ݔψis a parent of ݕ, and ܖܑܔ܊ܑ܁ሺݔǡ ݕሻholds if and only if ݔψand ݕψare siblings. ࣩଶ includes, for instance, Axiom (6) which defines ܖܑܔ܊ܑ܁: ݔψand ݕψare siblings exactly when ݔψand ݕψare distinct and share at least one parent.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
௫ǡ௬ ሾሺ ר ݕ ് ݔ௭ ሾܜܖ܍ܚ܉۾ሺݖǡ ݔሻ ܜܖ܍ܚ܉۾ רሺݖǡ ݕሻሿሻ ֞ ܖܑܔ܊ܑ܁ሺݔǡ ݕሻሿ
(6)
Initially, ࣩଵ is the only ontology in the translation graph. The predicate ܜܖ܍ܚ܉۾is added to ࣩଵ , along with Axiom (7) that defines ܜܖ܍ܚ܉۾in terms of ۱܌ܔܑܐ. This operation, of course, yields a new ontology, ࣩଵǤଵ , containing the terms ۱ ܌ܔܑܐ, ܚ܍ܜܛܑ܁, ۰ܚ܍ܐܜܗܚ, and ܜܖ܍ܚ܉۾. To move toward ࣩଶ , ܖܑܔ܊ܑ܁and Axiom (8) are added to ࣩଵǤଵ , yielding ontology ࣩଵǤଶ . (One common approach in interoperability is to use an intertheory to translate between ontologies. Here, ࣩଵǤଶ would be the intertheory for ࣩଵ and ࣩଶ because its vocabulary is the union of the vocabularies of ࣩଵ and ࣩଶ .) To complete our translation graph, the predicates ܚ܍ܜܛܑ܁, ۰ܚ܍ܐܜܗܚ, and ۱ ܌ܔܑܐare removed from ࣩଵǤଶ in a three-step process that results in ࣩଶ . This transformative process is traced in the translation graph shown in Figure 1, and the bridging axioms produced by this process are shown below. ௫ǡ௬ ሾܜܖ܍ܚ܉۾ሺݔǡ ݕሻ ֞ ۱܌ܔܑܐሺݕǡ ݔሻሿ
(7)
௫௬ ൣ൫۰ܚ܍ܐܜܗܚሺݔǡ ݕሻ ܚ܍ܜܛܑ܁ שሺݔǡ ݕሻ൯ ֞ ܖܑܔ܊ܑ܁ሺݔǡ ݕሻ൧
(8)
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
116
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
5.2. Evaluating a Provability-based Query We now evaluate a query ۃȞǡ ߰ ۄwith respect to the translation graph and the ontologies that the graph interrelates. Ȟ includes all axioms that define vocabulary semantics, any other declarative sentences in the ontologies, and the bridging axioms that interrelate the ontologies. In the present case, Ȟincludes Axioms (5)–(8), and any other contents of the knowledge-bases of ࣩଵ and ࣩଶ . The question, ߰ , posed in the query is: ௫ ሾܚ܍ܜܛܑ܁ሺݔǡ ݔሻሿ? That is to say, ߰ asks whether any individual is its own sister. Evaluating this query generates a negative response: no known individual is its own sister, and, in fact, no such individual can ever exist within the given ontologies. An explanatory proof of, roughly, the following form accompanies the response: If an individual were its own sister, then, by Axiom (8), it would be its own sibling. An individual that is its own sibling, by Axiom (6), must not be the same object as itself, but this is a contradiction. Therefore, no individual can be its own sister. The purpose of this example has been to illustrate the construction and use translation graphs and provability-based queries. We recognize that in the example, some of the difficulties typically encountered in interoperability have been glossed over. For instance, it is particularly convenient that the two ontologies happened to use the same ܖܗܛܚ܍۾sort. Realistically, this is often not the case, and additional axioms are required to formalize the relationships between similar sorts.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6. Illustrative Applications of PBSI The PBSI approach has been applied to a number of IC-relevant systems. For illustration, we present several applications of PBSI in §6.1–4. For each illustrative application, we describe the specific type of interoperability achieved, the systems involved, and then sketch some of details of the effort. The illustrative applications given below all center on our Slate [29] system. Slate is a domain- and ontology-agnostic argument analysis and visualization environment designed for use in both formal (e.g., engineering, mathematics, logic) and informal reasoning (e.g., business and intelligence analysis). Information in Slate is presented visually and through a natural language interface, Solomon. Internally, however, Slate represents knowledge using many-sorted logic. The use of many-sorted logic enables automated reasoning tools such as theorem provers and model finders to be integrated easily with Slate, and generated proofs and models can be presented to users in a rudimentary graphical format (e.g., see Figures 3 and 4). 6.1. Between Systems in the IKRIS Workshop The PBSI approach was developed, in part, to enable meaningful interoperability between systems participating in the IKRIS workshop, a Disruptive Technology Office (DTO) sponsored workshop on Interoperable Knowledge Representation for Intelligence Support (IKRIS). The IKRIS workshop addressed the problems of enabling “interoperability of knowledge representation and reasoning (KR&R) technology developed by the multiple DTO programs and designed to perform different tasks, and how to practically represent knowledge that is relevant to intelligence analysis tasks in a form that enhances support for analysts.” [30]
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
117
As part of the IKRIS workshop, three IC systems exchanged information as needed to solve collaboratively Case Study 4: Sign of the Crescent [31], an intelligenceanalysis training scenario used at the Joint Military Intelligence College. Specifically, the three systems were KANI [32], Noöscape [33], and Slate [29], and each system used a different knowledge representation scheme: KANI used Common Logic [5,6], a predecessor to IKL, Noöscape used CycL [34], and Slate used many-sorted logic. The IKRIS evaluation included three types of interoperability tests (viz., round-trip, systemto-system, and multi-system), and culminated in a capstone demonstration [35]. Round-Trip Tests. In the round-trip tests, information was translated from a system’s native knowledge representation format into IKL [3], and then back into the system’s native format. The round-trip was considered successful if the valid inferences were preserved through the translation. Because each of the three systems employed different representation schemes, the round-trip tests involved not only changes in ontology, but also in language semantics and syntax.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
System-to-System Tests. After systems completed round-trip tests, they progressed to system-to-system tests. The system-to-system tests determined whether the translators from IKL to participant systems would properly handle formulae that were not handcrafted in IKL but products of other translators. For instance, a KANI knowledgebase would be translated into an IKL knowledge-base, which was then translated into a Noöscape knowledge-base. The test was considered successful if consequences of the original knowledgebase were also consequences of the final knowledge-base. Multi-System Tests. Finally, after successful round-trip and system-to-system tests, the multi-system tests were performed. The multi-system tests were designed to imitate the flow of information between specialized analysts working collaboratively on a task, yet using different analytical software. The knowledge-bases for the multi-system tests, and ultimately the capstone demo, were drawn from Case Study 4 [31]. To mimic the workflow of multiple analysts, no system contained all of the relevant information; instead, information in the case study was partitioned according to analytic specialties, for example, finance, and sociology. The multi-system tests were qualitative; evaluation considered whether the interoperability of systems aided case study resolution. The multi-system tests also revealed a need to exchange analytic artifacts such as arguments and possible models of the world—first-class results in the analytic process [36,37,38]. In the workshop, a point solution met this need; we return to the topic of a general solution in §8. 6.2. Between Slate and Oculus’ GeoTime Interoperability between Slate and Oculus’ GeoTime [39,40] focused on automatic translations between low-level descriptions and high-level abstractions. The GeoTime tool provides analysts with sophisticated visualizations of interleaved geospatial and temporal information. The benefits of Slate and GeoTime interoperability included allowing analysts using Slate to access GeoTime’s visualization tools, and allowing analysts using GeoTime access to Slate’s argumentation and reasoning capabilities. Interoperability focused on how high-level abstractions in Slate, e.g., “John flew from Albany to Chicago,” could be translated to and from low-level descriptions of events in GeoTime, e.g., “John is a participant in a transportation event, ࣟ. A Boeing 747 participates in ࣟ . ࣟ began at 3:00 at 42°44̵54̵̵N 73°48̵06̵̵W. ࣟ ended at 5:00 at
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
118
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
41°58̵43̵̵N 87°54̵17̵̵W:” More generally, corresponding high-level abstractions were inferred from the existence of events and assertions about events, e.g., their types and participants. 6.3. Between Slate and SUNY Albany’s HITIQA Interoperability between Slate and the University at Albany’s High-Quality Interactive Question Answering (HITIQA) System [41] transformed structured document representations into propositional content then used in Slate-based argumentation analysis. HITIQA is a QA system that retrieves documents from a corpus based on an interactive question-and-answer dialogue with users. Internally, HITIQA employs natural language processing to produce structured representations of documents, and documents are retrieved based on HITIQA’s assessment of their structured representation’s relevancy to user questions. HITIQA’s formal representation of documents is used for retrieval, but contains much of the information that an analyst interested in the document would use in constructing related arguments and reports. Interoperability focused on translating HITIQA’s document-centered representations into propositions expressed by the documents. For instance, given HITIQA knowledge that “document ݀ψdescribes an attack,” “the event described by ݀ψwas performed by ݔ,” and “ݕψis the object of the event described by ݀,” it can be inferred that “ݔψattacked ݕ.”
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.4. Between Slate and Discourse Representation Theory-Based Systems Interoperability between Slate and systems based on Discourse Representation Theory [42] enabled translation between controlled natural language and many-sorted logic by transforming ontology-agnostic reified representations to ontology-specific propositions. Slate’s natural language interface (Solomon) is based on Attempto Controlled English (ACE) [43,44], a restricted subset of English grammar designed to be easily read and written by humans, and unambiguously parsed by machines. ACE text is parsed into discourse representation structures (DRSs) containing reified linguistic content. In turn, the DRSs are translated into many-sorted logic. Using a userspecified signature, the translation from DRSs to many-sorted logic maps the reified linguistic content to proposition expressions. As a result, domain-specific linguistic content is translated to ontology-specific propositional content via an intermediary, ontology-agnostic, controlled natural language.
7. Automating an Unmanned Aerial Vehicle: An Extended Example As a capstone, we present an example wherein the process of vetting targets for an unmanned aerial vehicle (UAV) is automated. Specifically, four data-source ontologies are made available, by PBSI, to a fifth ontology that contains a targeting policy. This example illustrates several of PBSI’s advantages, namely, (i) only construction of the translation graph requires knowledge of multiple ontologies; (ii) ontologies may be models of systems, that is, interoperating systems need not be implemented as knowledge representation systems; (iii) though not proscribed, an intertheory is not required; and (iv) interoperability relationships are not limited to simple paths: they may be complex networks that integrate knowledge across multiple ontologies.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
119
Table 1. The Geospatial ontology signature. Sorts Predicates
Functions
۰ܖܑ܌ܔܑܝ, ܜܖܑܗ۾ ك ܜܖ܍܋܉ܒ܌ۯ۰ ܖܑ܌ܔܑܝൈ ۰ܖܑ܌ܔܑܝ ك ܔܗܗܐ܋܁۰ܖܑ܌ܔܑܝ ۶ ك ܔ܉ܜܑܘܛܗ۰ܖܑ܌ܔܑܝ ۳ ك ܡܛܛ܉܊ܕ۰ܖܑ܌ܔܑܝ ܜܖܑܗ۾ܖܑ܌ܔܑܝ܊۰ ܖܑ܌ܔܑܝ՜ ܜܖܑܗ۾ ܜܖܑܗܘ۰ ܜܖܑܗ۾ ܖܑ܌ܔܑܝ՜ ۰ܖܑ܌ܔܑܝ
7.1. Ontologies In this UAV example, declarative information is distributed across five ontologies (described in §7.1.1–5). Each ontology’s vocabulary is limited to the terms needed to express information within a particular domain. This conceptual compartmentalization allows knowledge engineers to use the simplest terminology possible, and removes the need to know the vocabulary of other external ontologies. To illustrate, the Geospatial ontology given in §7.1.1 has vocabulary terms for mapping geospatial coordinates (e.g., latitude and longitude) to buildings, and terms for classifying buildings (e.g., as academic, medical, or residential).7 In contrast, the ontology given in §7.1.4 has vocabulary terms for capturing various capabilities of UAVs (e.g., flight range, maximum altitude, and stall speed). It is appropriate for these two ontologies to be distinct because there is little, if any, semantic overlap between them. However, a UAV controller—human or automated—must be aware of, and reason over the information collectively expressed in, these ontologies; hence the need for PBSI.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.1.1. Geospatial The Geospatial ontology (shown in Table 1) includes two sorts, ۰ ܖܑ܌ܔܑܝand ܜܖܑܗ۾, denoting buildings and geospatial coordinates (points). Furthermore, the ontology includes two functions, ܜܖܑܗܘ۰ ܖܑ܌ܔܑܝand ܜܖܑܗ۾ܖܑ܌ܔܑܝ܊map between buildings and points. Given a point , ܜܖܑܗܘ۰ܖܑ܌ܔܑܝሺሻdenotes the building closest to point , and given a building ܾ, ܜܖܑܗ۾ܖܑ܌ܔܑܝ܊ሺܾሻdenotes the point at the center of the building. The Geospatial ontology also has a binary relation, ܜܖ܍܋܉ܒ܌ۯ, and a number of monadic predicates that describe various types of buildings, e.g., ܔܗܗܐ܋܁. For example, buildings ܾଵ and ܾଶ are declared adjacent by ܜܖ܍܋܉ܒ܌ۯሺܾଵ ǡ ܾଶ ሻ and ܔܗܗܐ܋܁ሺܾଵ ሻdeclares that ܾଵ is a school. 7.1.2. UAV Data The UAV is responsible for knowing its own location and what points are in range. The UAV’s ontology (shown in Table 2) includes the sorts ܜܖܑܗ۾and ܍ܕܑ܂, and the predicates ܑ܍ܖ܉܀ܖ, and ܜ܍ܚ܉܂. That a point ψis in the range of the UAV at time ݐψis denoted ܑ܍ܖ܉܀ܖሺǡ ݐሻ. While the UAV might not contain a declarative knowledge-base, in the tradition of logic-based subsumption architectures [46,47], data from the UAV is considered a virtual, continually updated, knowledge-base. Commanding the UAV is similarly recast as asserting formulae in the UAV’s virtual knowledge-
7
The geospatial ontology presented here is simplistic. For realistic geospatial ontologies, see [45].
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
120
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
Table 2. The UAV Data ontology signature. Sorts Predicates
ܜܖܑܗ۾, ܍ܕܑ܂ ܜܖܑܗ۾ ك ܜ܍ܚ܉܂ ܑ ܜܖܑܗ۾ ك ܍ܖ܉܀ܖൈ ܍ܕܑ܂ Table 3. The Known Insurgent Locations ontology signature.
Sorts Predicates
ܖܗܛܚ܍۾, ۰ܖܑ܌ܔܑܝ, ܍ܕܑ܂ ܑܛ۷ܖܗܛܚ܍۾ ك ܜܖ܍ܚܝܛܖ ܑܖ۰ ܖܗܛܚ܍۾ ك ܖܑ܌ܔܑܝൈ ۰ ܖܑ܌ܔܑܝൈ ܍ܕܑ܂
Sorts
۰ܖܑ܌ܔܑܝ ۲ ك ܍ܔ܊܉ܡܗܚܜܛ܍۰ܖܑ܌ܔܑܝ
Table 4. The UAV Capabilities ontology signature.
base. The UAV is commanded to target a point ψby asserting ܜ܍ܚ܉܂ሺሻin the UAV’s virtual knowledge-base. 7.1.3. Known Insurgent Locations To decide whether a building is a potential target, the UAV controller needs to consider whether there might be known insurgents in the building. (While there certainly are other relevant criteria that ought to be factored in deciding whether a building is a potential target, in this example we consider only whether a building has insurgents in it.) The ontology of known insurgent locations (shown in Table 3) includes three sorts, ۰ܖܑ܌ܔܑܝ, ܖܗܛܚ܍۾, and ܍ܕܑ܂, and two predicates, ܑܖ۰ ܖܑ܌ܔܑܝand ܑܛ۷ܜܖ܍ܚܝܛܖ. That a person ψis in a building ܾψat time ݐψis denoted by ܑܖ۰ܖܑ܌ܔܑܝሺǡ ܾǡ ݐሻ, and that ψis a known insurgent is denoted by ܑܛ۷ܜܖ܍ܚܝܛܖሺሻ.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.1.4. UAV Capabilities The particular capabilities of the UAV are also captured in an ontology (shown in Table 4), though in this example, the capabilities are simplified and abstracted into a single unary predicate, ۲܍ܔ܊܉ܡܗܚܜܛ܍, that applies to elements of the sort ۰ܖܑ܌ܔܑܝ. 8 Thus, ۲܍ܔ܊܉ܡܗܚܜܛ܍ሺܾሻasserts that the UAV is capable of destroying building ܾ. 7.1.5. An MMOI-Based Targeting Policy Conceptually, we wish to automate realization of the UAV controller’s intent to have the UAV destroy certain targets, and to leave certain others unscathed. In this example, though, we detail only the automation of a simple target-selection policy. Policies for target selection are captured in a modified MMOI ontology (shown in Table 5), where ‘MMOI’ stands for the Wigmorean elements of motive, means, opportunity, and intent. In MMOI frameworks, an agent intends to perform an action for which it has motive, means, and opportunity. In our modified MMOI ontology, intent is implicit: the automated UAV system implicitly intends to destroy any permissible target for which there is motive, means, and opportunity; furthermore, the system will act autonomously to 8 We are unaware of any actual UAV capability ontology, but the practice of codifying device capabilities—e.g., those of mobile computing devices [48]—is widespread in other domains.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
121
Table 5. The MMOI-Based Targeting Policy ontology signature. Sorts Predicates
Functions
۰ܖܑ܌ܔܑܝ, ܍ܕܑ܂, ܖܗܑܜ܋ۯ ܖܗܑܜ܋ۯ ك ܍ܞܑܜܗۻ ܖܗܑܜ܋ۯ ك ܛܖ܉܍ۻ ܖܗܑܜ܋ۯ ك ܡܜܑܖܝܜܚܗܘܘ۽ ܖܗܑܜ܋ۯ ك ܍ܔ܊ܑܛܛܑܕܚ܍۾ ۲ ܡܗܚܜܛ܍۰ ܖܑ܌ܔܑܝൈ ܍ܕܑ܂՜ ܖܗܑܜ܋ۯ
achieve its intent. The MMOI-based ontology includes the sorts ۰ܖܑ܌ܔܑܝ, ܍ܕܑ܂, and ܖܗܑܜ܋ۯ, and the predicates ܍ܞܑܜܗۻ, ܛܖ܉܍ۻ, ܡܜܑܖܝܜܚܗܘܘ۽, and ܍ܔ܊ܑܛܛܑܕܚ܍۾. The ontology also includes a function ۲ܡܗܚܜܛ܍, which is used to denote the action of destroying a building. Specifically, ۲ܡܗܚܜܛ܍ሺܾǡ ݐሻ denotes the action of destroying building ܾ ψat time ݐ. The meanings of ܍ܞܑܜܗۻ, ܛܖ܉܍ۻ, ܡܜܑܖܝܜܚܗܘܘ۽, and ܍ܔ܊ܑܛܛܑܕܚ܍۾are explained as follows. Motive. The ܍ܞܑܜܗۻpredicate holds on an action if there are reasons that the action is, in some way, desirable. In the present example, the UAV has motive to destroy a building if an insurgent is in the building, and this is formally stated in Axiom (9). Of course, one may have motive for an action as well as a reason not to act. For instance, there would be motive to target a building containing known insurgents, but if the building were a school full of children then there is a compelling reason not to fire. Reasons for not acting are subsumed by the predicate ܍ܔ܊ܑܛܛܑܕܚ܍۾, which is explained shortly.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
ǡǡ௧ ൣ൫ܑܛ۷ܜܖ܍ܚܝܛܖሺሻ ܖܑ ר۰ܖܑ܌ܔܑܝሺǡ ܾǡ ݐሻ൯ ֜ ܍ܞܑܜܗۻ൫۲ܡܗܚܜܛ܍ሺܾǡ ݐሻ൯൧ (9) Means. The UAV is physically capable of destroying some buildings, but not others. For instance, the UAV may be able to destroy residential and office buildings, but it might be less effective against fortified buildings such as concrete bunkers. The UAV Capabilities ontology (§7.1.4) used ۲܍ܔ܊܉ܡܗܚܜܛ܍ሺܾሻto indicate that the UAV is capable of destroying building ܾ. Thus, Axiom (10) formally states that when a building ܾψis ۲܍ܔ܊܉ܡܗܚܜܛ܍, the UAV has ܛܖ܉܍ۻto ۲ܾ ܡܗܚܜܛ܍. ǡ௧ ൣ۲܍ܔ܊܉ܡܗܚܜܛ܍ሺܾሻ ֜ ܛܖ܉܍ۻ൫۲ܡܗܚܜܛ܍ሺܾǡ ݐሻ൯൧
(10)
Opportunity. The UAV has ܡܜܑܖܝܜܚܗܘܘ۽to destroy a building ܾψwhen the building is within the range of the UAV; this is formally stated in Axiom (11). ǡ௧ ሾܑ܍ܖ܉܀ܖሺǡ ݐሻ ֜ ܡܜܑܖܝܜܚܗܘܘ۽ሺ۲ܡܗܚܜܛ܍ሺܜܖܑܗܘ۰ܖܑ܌ܔܑܝሺሻǡ ݐሻሻሿ
(11)
Permissibility. Some actions are impermissible (e.g., destroying occupied schools and hospitals) even when ܍ܞܑܜܗۻ, ܛܖ܉܍ۻ, and ܡܜܑܖܝܜܚܗܘܘ۽are at hand. The ܍ܔ܊ܑܛܛܑܕܚ܍۾predicate is used to modulate what, if any, actions are actually taken. For instance, schools, hospitals, and embassies are considered protected buildings, thus it is impermissible to target them. Axiom (12) defines a new predicate, ك ܌܍ܜ܋܍ܜܗܚ۾ ۰ܖܑ܌ܔܑܝ, in terms of ܔܗܗܐ܋܁, ۶ܔ܉ܜܑܘܛܗ, and ۳ܡܛܛ܉܊ܕ. (Needless to say, in a real targeting policy, the semantics of ܌܍ܜ܋܍ܜܗܚ۾would be much more complex.)
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
122
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
ൣ൫ܔܗܗܐ܋܁ሺܾሻ ש۶ܔ܉ܜܑܘܛܗሺܾሻ ש۳ܡܛܛ܉܊ܕሺܾሻ൯ ֞ ܌܍ܜ܋܍ܜܗܚ۾ሺܾሻ൧
(12)
Notice that an unprotected building is not necessarily a permissible target. Targeting mechanisms and other physical systems have a degree of inherent imprecision. This imprecision is addressed in Axiom (13), which formally states that it is permissible to destroy a building if the building is not protected and it is not adjacent to a protected building. This definition of permissibility is obviously simplified: typical considerations in rules of engagement include, for example, whether buildings are occupied by lawful combatants, unlawful combatants, or non-combatants, and the extent of resulting collateral damage. Adequate representation of such considerations requires fullfledged deontic logic [49,50,51].
ሺ܌܍ܜ܋܍ܜܗܚ۾ሺܾሻ רᇲ ሾܜܖ܍܋܉ܒ܌ۯሺܾǡ ܾ ᇱ ሻ ܌܍ܜ܋܍ܜܗܚ۾ רሺܾ ᇱ ሻሿሻ ൨ ֜ ܍ܔ܊ܑܛܛܑܕܚ܍۾ሺ۲ܡܗܚܜܛ܍ሺܾǡ ݐሻሻ
(13)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Targeting Criterion. The four elements of the modified MMOI framework can be combined into a single target selection policy dictating when the UAV can, and will, target the location of a given building ܾψat a specified time ݐ. This targeting policy is formalized in Axiom (14), and it states that the UAV will target (i.e., attack) the location of a building if the building is a permissible target, and the UAV has motive, means, and opportunity to destroy the building. When the autonomous UAV system determines whether or not to attack a given building, it seeks justification for a formula of the form ܜ܍ܚ܉܂ሺܜܖܑܗ۾ܖܑ܌ܔܑܝ܊ሺܾሻǡ ݐሻ. In other words, before taking action, the system proves that the building is a valid target. Any proof will be based on information from all five ontologies, so the way in which data is incorporated is particularly important to the correct realization of the targeting policy. ܽ ൌ ۲ܡܗܚܜܛ܍ሺܾǡ ݐሻ ר ۍ ې ܍ܞܑܜܗۻሺܽሻ ܛܖ܉܍ۻ רሺܽሻ ר ൲ ֜ ܜ܍ܚ܉܂ሺܜܖܑܗ۾ܖܑ܌ܔܑܝ܊ሺܾሻǡ ݐሻۑ ǡǡ௧ ێ൮ ܡܜܑܖܝܜܚܗܘܘ۽ሺܽሻ ר ێ ۑ ے ۏ ܍ܔ܊ܑܛܛܑܕܚ܍۾ሺܽሻ
(14)
7.2. The Translation Graph The translation graph shown in Figure 2 relates the five ontologies described in §7.1. The five signatures shown with solid borders in the translation graph are those of the five ontologies, while ones shown with dotted borders are intermediate signatures. The incremental modifications of the translation graph are depicted with directed edges. Though the operations are reversible (e.g., adding a functor in one direction is the same as removing it in the other direction), the directed paths in this graph illustrate the likely flow of information between the five ontologies. The construction of the translation graph is relatively straightforward, and we will not detail each individual edge. However, we note below several important characteristics of the translation graph.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
123
Figure 2. The translation graph relating the five ontologies.
Recording Inexpressible Axioms. None of Axioms (10)–(14) is expressible in any of the five ontologies; each uses vocabulary terms from two or more of the ontologies. For instance, Axiom (11) uses ܑܛ۷ ܜܖ܍ܚܝܛܖfrom the Known Insurgent Locations ontology, ܡܜܑܖܝܜܚܗܘܘ۽from the MMOI ontology, and ܜܖܑܗܘ۰ ܖܑ܌ܔܑܝfrom the Geospatial ontology. This would present a problem if these axioms had to be stored as declarative knowledge in a single ontology, but the PBSI approach allows each of these axioms to be a bridging axiom expressible in an intermediate signature.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
124
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
Table 6. Sample geospatial data. Label ଵ ଶ ଷ ସ ହ
Classification Factory Office Building Office Building School Apartment
Notes Adjacent to ସ Adjacent to ଷ Adjacent to ଶ Adjacent to ଵ
Multiple and Indirect Connections. Although the translation graph is connected (i.e., between any two of the five ontology signatures there is some path), there are not direct paths connecting every pair of the five main ontologies. For example, there is no direct path between the UAV Capabilities ontology and the Known Insurgent Locations ontology; they are connected only by a path through the MMOI ontology. Furthermore, there are multiple paths between some of the signatures. For example, there is a direct path between the Geospatial ontology and UAV Data ontology, and another path through the MMOI ontology. Relations Using Auxiliary Information. Some of the ontology relationships are expressed using auxiliary information drawn from other ontologies. For instance, there is a path from the Geospatial ontology to an intermediate signature in the (directed) path from the UAV Data ontology to the MMOI ontology. The influence from the Geospatial ontology allows intermediaries between the UAV Data ontology and the MMOI ontology to use vocabulary from the Geospatial ontology (viz., ܜܖܑܗܘ۰ )ܖܑ܌ܔܑܝto convert from the UAV Data ontology’s ܜܖܑܗ۾to the MMOI ontology’s ۰ܖܑ܌ܔܑܝ. In general, this technique enables various kinds of data transformations, such as mappings between datatypes (e.g., times, coordinate systems), and between individuals and their characteristics (e.g., a person and his/her phone number).
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.3. Provability-Based Queries Suppose the Geospatial ontology is populated with the information presented in Table 6. Specifically, there are five buildings in the database, labeled ܤଵ through ܤହ , consisting of two office buildings, a factory, a school, and an apartment building. The table also describes which pairs of buildings are adjacent. Further, suppose that a UAV is within range of these buildings and has sufficient firepower to destroy any building except ܤହ (this data is present in the UAV Capabilities ontology). The UAV’s response to an insurgent’s occupation of the factory ܤଵ —i.e., whether the UAV will target ܤଵ —can be determined using a provability-based query. The consequent of the query is ܜ܍ܚ܉܂ሺܜܖܑܗ۾ܖܑ܌ܔܑܝ܊ሺܤଵ ሻǡ ݐ௪ ሻand, as described in §7.1.5, the response should be affirmative if and only if there are motive, means, opportunity to destroy ܤଵ and it is permissible to do so. The data contained in the UAV Capabilities ontology, augmented with knowledge from the Geospatial ontology, entail the UAV’s means and opportunity. Knowledge of the insurgent’s location, viz., that the insurgent is in ܤଵ , yields motive. However, ܤଵ is adjacent to ܤଷ , and since ܤଷ is a school, it is not permissible to target ܤଵ . Thus, the autonomous UAV system will fail to prove that it ought to target ܤଵ , and so the UAV will not be commanded to target ܤଵ . Slate [29]—the reasoning system used for this example—answers the query with a counter-model demonstrating that the consequent does not follow from the current knowledge (see Figure 3).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
125
Figure 3. A counter-model has been found in Slate, thus denying that the UAV should target ܤଵ .
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Were it the case that the insurgent occupies instead the office building ܤଷ , which is adjacent only to another office building, the autonomous system would conclude that it is permissible to destroy the occupied building. The consequent of this second query would be ܜ܍ܚ܉܂ሺܜܖܑܗ۾ܖܑ܌ܔܑܝ܊ሺܤଷ ሻݐ௪ ሻ, and it is logically entailed, as Slate demonstrates with a deductive proof (see Figure 4).
8. Future Work We now turn to three topics for future work: applying PBSI-related techniques to metalevel knowledge representation artifacts; automating the construction of translation graphs; and addressing scalability issues that may arise in larger applications. Complex Knowledge Artifacts. Provability-based queries require accompanying justifications, and in our approach, we use formal proofs and models. However, just as propositional content in one ontology can have semantic consequences in another, so too proofs and models in one formalism can have consequences in another. Thus, there is a need for research toward information sharing and joint reasoning over these metalevel constructs. Results in this area would have practical as well as theoretic value; for instance, they would address the need for sharing arguments and possible models of the world, which was exposed by the IKRIS multi-system tests (§6.1).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
126
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 4. A proof has been found in Slate, thus confirming that the UAV should target ܤଷ .
Automation. We have demonstrated that PBSI can achieve information sharing and joint reasoning, specifically by using translation graphs and bridging axioms. The process of extracting bridging axioms from translation graphs has been automated, but there remains a pressing need to automate fully the construction of translation graphs. Because translation graphs are ultimately software artifacts, the task of automatic construction is, in effect, the challenge of automatic programming [52]. Scalability. The PBSI approach works in the smaller examples that we have presented here, but guaranteeing scalability will require further research into distributed reasoning, parallel reasoning, and into ways to compartmentalize reasoning whenever possible. Issues of scalability also include treating other sophisticated aspects of the systems involved, e.g., allowing inconsistency within individual systems, and accurately representing partial information [53].
9. Conclusion We have described an approach toward semantic interoperability, particularly toward information sharing and joint reasoning, that is based on translation graphs and bridging axioms. Through discussion of illustrative uses (§6) and a capstone example (§7) we have shown some of the key features of this approach, viz.: it is logically based, preserves asymmetry in translation, captures semantic consequences even when translation is impossible, it can capture semantic statements that could not be expressed in any
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
127
single ontology to be captured as bridging axioms, relationships between ontologies may involve any number of ontologies, and ontologies may be related in multiple ways. Future work on PBSI includes: extending the approach to share and reason about complex knowledge and reasoning artifacts, such as proofs and models; automating the most labor intensive part of the process, i.e., constructing translation graphs; and addressing scalability issues by distributing the reasoning process, and accurately representing nuances of the systems involved in that distribution.
References [1] [2] [3] [4] [5] [6]
[7] [8]
[9] [10]
[11]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[12] [13] [14] [15] [16] [17]
[18] [19] [20] [21] [22] [23]
Bray, T., Paoli, J., Sperberg-McQueen, C.M., Maler, E., Yergeau, F.: Extensible Markup Language (XML) 1.0 (Fifth Edition). Recommendation, W3C, November 26 (2008) Brickley, D., Guha, R.V., McBride, B.: RDF Vocabulary Recommendation, W3C (2004) Hayes, P., Menzel, C.: IKL Specification Document. Specification, IKRIS Interoperability Group (2006) IKL specification document. Genesereth, M.R., Fikes, R.E.: Knowledge Interchange Format Version 3 Reference Manual. (1997) Menzel, C.: Common Logic Standard (2003) Presentation at the 2003 Metadata Forum Symposium on Ontologies. International Organization for Standardization: Information technology – Common Logic (CL): a framework for a family of logic-based languages. International Standard ISO/IEC 24707:2007, International Organization for Standardization (2007) Choi, N., Song, I.Y., Han, H.: A Survey on Ontology Mapping. ACM SIGMOD Record 35(3) (2006) 34–41 Wang, G., Goguen, J.A., Nam, Y.K., Lin, K.: Critical Points for Interactive Schema Matching. In Yu, J.X., Lin, X., Lu, H., Zhang, Y., eds.: Advanced Web Technologies and Applications: 6th Asia-Pacific Web Conference, APWeb, Hangzhou, China, Springer (2004) 654–664 Goguen, J.A., Burstall, R.M.: Introducing Institutions. In Clarke, E.M., Kozen, D., eds.: Logics of Programs. Volume 164 of Lecture Notes in Computer Science., Springer (1984) 221–256 Dou, D., McDermott, D.: Deriving Axioms Across Ontologies. In: AAMAS ’06: Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems, New York, NY, ACM Press (2006) 952–954 McCarthy, J.: Programs with Common Sense. In: Proceedings of the Teddington Conference on the Mechanization of Thought Processes. (1959) Nilsson, N.: Logic and Artificial Intelligence. Artificial Intelligence 47 (1991) 31–56 Bringsjord, S.: The logicist manifesto: At long last let logic-based AI become a field unto itself. Journal of Applied Logic 6(4) (2008) 502–525 Bringsjord, S., Ferrucci, D.: Logic and Artificial Intelligence: Divorced, Still Married, Separated…? Minds and Machines 8 (1998) 273–308 Bringsjord, S., Yang, Y.: Representations Using Formal Logics. In Nadel, L., ed.: Encyclopedia of Cognitive Science Vol 3. Nature Publishing Group, London, UK (2003) 940–950 Bringsjord, S.: Declarative/Logic-Based Cognitive Modeling. In Sun, R., ed.: The Handbook of Computational Psychology. Cambridge University Press, Cambridge, UK (2008) 127–169 Buvaþ, S., Fikes, R.: A Declarative Formalization of Knowledge Translation. In: CIKM ’95: Proceedings of the Fourth International Conference on Information and Knowledge Management, New York, NY, ACM Press (1995) 340–347 Dou, D., McDermott, D., Qi, P.: Ontology Translation by Ontology Merging and Automated Reasoning. In: Ontologies for Agents: Theory and Experiences. Birkhäuser (2005) 73–94 Klyne, G., Carroll, J.J.: Resource Description Framework (RDF): Concepts and Abstract Syntax. Recommendation, W3C (2004) Patel-Schneider, P.F., Hayes, P., Horrocks, I.: OWL Web Ontology Language Semantics and Abstract Syntax. Recommendation, W3C (2004) Hendler, J.: Agents and the Semantic Web. IEEE Intelligent Systems (2001) 30–37 DARPA Agent Markup Program: DARPA Agent Markup Language Homepage (2003) http://www.daml.org/. Connolly, D., van Harmelen, F., Horrocks, I., McGuinness, D.L., Patel-Schneider, P.F., Stein, L.A.: DAML+OIL (March 2001) Reference Description. Note, W3C (2001)
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
128
A. Shilliday et al. / PBSI for Information Sharing and Joint Reasoning
[24] Hu, W., Qu, Y.: Discovering Simple Mappings Between Relational Database Schemas and Ontologies. In: ISWC/ASWC 2007. Volume 4825 of Lecture Notes in Computer Science., Springer (2007) [25] Goguen, J.A.: Data, Schema, Ontology and Logic Integration. Logic Journal IGPL 13(6) (2005) 685– 715 [26] Manzano, M.: Extensions of First Order Logic. Cambridge University Press (1996) [27] Bechofter, S., van Harmelen, F., Hendler, J., Horrocks, I., McGuinness, D.L., Patel-Schneider, P.F., Stein, L.A.: OWL Web Ontology Language Reference. Recommendation, W3C (2004) [28] Arkoudas, K.: Athena (2005) http://www.cag.csail.mit.edu/~kostas/dpls/athena/. [29] Bringsjord, S., Taylor, J., Shilliday, A., Clark, M., Arkoudas, K.: Slate: An Argument-Centered Intelligent Assistant to Human Reasoners. In: Proceedings of the 8th International Workshop on Computational Models of Natural Argument (CMNA 08), Patras, Greece (2008) 1–10 [30] MITRE: IKRIS, Workshop site: http://nrrc.mitre.org/NRRC/ikris.htm (2007) [31] Hughes, F.J.: The Art and Science of The Process of Intelligence Analysis: Case Study #4 (The Sign of the Crescent. Technical report, Joint Military Intelligence College, Washington, DC (May 2003) [32] Fikes, R.E., Ferrucci, D., Thurman, D.A.: Knowledge Associates for Novel Intelligence. In: Proceedings of the 2005 International Conference on Intelligence Analysis (IA 2005), McLean, VA (2005) [33] Siegel, N., Shepard, B., Cabral, J., Witbrock, M.: Hypothesis Generation and Evidence Assembly for Intelligence Analysis: Cycorp’s Noöscape Application. In: Proceedings of the 2005 International Conference on Intelligence Analysis (IA 2005), McLean, VA (2005) [34] Lenat, D.B., Guha, R.V.: The evolution of CycL, the Cyc representation language. ACM SIGART Bulletin 2(3) (1991) 84–87 [35] Cheikes, B.A.: MITRE Support to IKRIS. Final Report MTR060158, MITRE (2006) [36] Johnston, R.: Analytic Culture in the US Intelligence Community: An Ethnographic Study. Imaging and Publishing Support, CIA (2005) [37] Sherman Kent School: A Tradecraft Primer: Structured Analytic Techniques for Improving Intelligence Analysis. Tradecraft Review 2(2) (2005) [38] Heuer, R.J.: Psychology of intelligence analysis. Center for the Study of Intelligence, Central Intelligence Agency (1999) [39] Kapler, T., Harper, R., Wright, W.: Correlating Events with Tracked Movements in Time and Space: A GeoTime Case Study. In: Proceedings of the 2005 International Conference on Intelligence Analysis (IA 2005), McLean, VA (2005) [40] Chappell, A., Bringsjord, S., Shilliday, A., Taylor, J., Wright, W.: Integration Experiment with GeoTime, Slate, and VIKRS. ARIVA Principal Investigator Meeting Handout (2007) [41] Strzalkowski, T., Small, S., Hardy, H., Yamrom, B., Liu, T., Kantor, P., Ng, K.B., Wacholder, N.: HITIQA: A Question Answering Analytical Tool. In: Conference Proceedings of the 2005 International Conference on Intelligence Analysis , McLean, VA, MITRE, Sponsored by the Office of the Assistant Director of Central Intelligence f or Analysis and Production (2005) [42] Kamp, H., Reyle, U.: From Discourse to Logic: Introduction to Modeltheoretic Semantics of Natural Language, Formal Logic and Discourse Representation Theory. Kluwer Academic, Dordrecht; Boston (1993) [43] Fuchs, N.E., Kaljurand, K.: Attempto Controlled English: Language, Tools and Applications. Presentation (2006) [44] Bringsjord, S., Arkoudas, K., Clark, M., Shilliday, A., Taylor, J., Schimanski, B., Yang, Y.: Reporting on Some Logic-Based Machine Reading Research. In: Proceedings of the 2007 AAAI Spring Symposium on Machine Reading. (2007) [45] Lieberman, J., Singh, R., Goad, C.: W3C Geospatial Ontologies. Incubator group report, W3C (2007) [46] Amir, E., II, P.M.R.: Logic-Based Subsumption Architecture. In Dean, T., ed.: Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence, IJCAI ’99, Morgan Kaufmann (1999) 147–152 [47] Brooks, R.A.: How to Build Complete Creatures Rather than Isolated Cognitive Simulators. In VanLehn, K., ed.: Architectures for Intelligence: The 22nd Carnegie Mellon Symposium on Cognition, Lawrence Erlbaum Associates (1991) 225–239 [48] Klyne, G., Reynolds, F., Woodrow, C., Ohto, H., Hjelm, J., Butler, M.H., Tran, L.: Composite Capability/Preference Profiles (CC/PP): Structure and Vocabularies 1.0. Recommendation, W3C (2004) [49] von Wright, G.H.: Deontic Logic. Mind 60(237) (1951) 1–15 [50] Bringsjord, S., Arkoudas, K., Bello, P.: Toward a General Logicist Methodology for Engineering Ethically Correct Robots. IEEE Intelligent Systems 21(4) (2006) 38–44 [51] Horty, J.: Agency and Deontic Logic. Oxford University Press (2001) [52] Rich, C., Waters, R.C.: Automatic Programming: Myths and Prospects. Computer 21(8) (1988) 40–51 [53] Ghidini, C., Serafini, L.: Distributed First Order Logics. In Gabbay, D.M., de Rijke, M., eds.: Frontiers of Combining Systems 2. Studies in Logic and Computation. Research Studies Press (2000) 121–139
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-129
129
Chapter 8
The Use of Ontologies to Support Intelligence Analysis Richard Lee Booz Allen Hamilton Abstract. For some years now, the Intelligence Community has been using XML "tagging" of documents in an effort to make the documents more usable for data discovery, sharing, and other processing. In implementing a system (METS) which automates the identification of relevant data in documents, we noted several limitations of that XML tagging approach, and therefore chose to also provide an OWL ontology-based representation of that data. Here, we discuss the goals of METS, those XML limitations, and the OWL approach, showing how the latter should support better analysis. (As we discuss, clients have thus far not made use of the OWL results, so the benefits are still hypothetical.) We also discuss issues we encountered in developing the ontologies, outline the design and use of the operational METS for processing message feeds and other documents, and conclude with future plans which include greater ontology coordination and sharing, and assisting with the incorporation of tools for benefiting from the semantic information provided by the OWL.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Keywords. Ontologies, OWL, information extraction, METS
Introduction In this paper we describe the genesis and evolution of the Metadata Extraction and Tagging Service (METS) system in use in the Intelligence Community (IC). We describe the rationale for such a system, and its intended fusion of Information Extraction (IE) with "XML tagging". We explore the limitations we saw with using a traditional XML approach to do this tagging. We describe the system's use of OWL and ontologies to represent the data it produces. We outline the various advantages and disadvantages we see with such an approach. We describe the current and upcoming METS systems, with an emphasis on the integration of IE technology with OWL production. We briefly describe an experiment we conducted on using ontologies for multi-INT data fusion. We conclude with a discussion of work – especially ontology and data coordination -- we are doing or hope to do in the future. The work described here was performed by the author and a varying team of colleagues from various companies.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
130
R. Lee / The Use of Ontologies to Support Intelligence Analysis
1. Background and Goals Over 5 years ago, the notion was expressed in the Intelligence Community (IC) that the decades of message traffic they were storing would be of greater use if it were all "tagged". All that unstructured text was of very limited usefulness to automated tools, and the hope was that the information they contained could be automatically converted into a more structured – and therefore more usable – representation. To examine this idea, we were tasked with evaluating the accuracy and usability of commercial Information Extraction (IE) tools and with determining the benefits of using them to do such tagging. IE tools process free-text documents and extract from them items of interest. These items can cover a wide range of entity types:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
• • • • • • • • • • • • •
Persons Organizations Locations Facilities Equipment (computers, radios, centrifuges, etc) Weapons Vehicles Documents (books, newspapers, passports, license plates, etc) Narcotics Money Events Dates/Times etc
It is important to note that the tools do far more than simply identify the presence of a reference to such an item in the document – they extract the provided information about an item. For a person, this information could include name(s), title, occupation, age, hair color, etc. It could also include information about relationships between the person and other entities and events – associates and relations, membership in a group, ownership of things, instigation of or participation in an event, etc. Similarly, the information about an event would typically include relationships indicating its dates, locations, participants, involved objects, etc. As a rule, the "out-of-the-box" extractor configurations do not handle all the entity types listed. However, many of the extractors allow the specification of new types, and the creation of new rules, actions, etc which enable the extraction of information about instances of those types. This "knowledge engineering" can also be used to make the extraction process more accurate by tailoring the rules to the peculiarities of the actual documents, changes in terminology, etc. We should also note that the above entity type list is just to provide an overview. The extraction tools – either "out-of-the-box" or after some knowledge-engineering – assign types to the items of interest more precisely than most of the types listed above. Weapons, for example, would be identified more precisely as Firearms, Missiles, IEDs, Chemical Weapons, etc, and perhaps even more precisely than that. In deciding the appropriate technologies and representations to use in this project it was important to consider the uses to which the data could and should be put. In
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
131
particular, we were mindful of the Presidential Directive EO 13355 on information sharing, which said (in part): "(g)(1) Establish ... standards ..., with special emphasis on facilitating: (A) the fullest and most prompt sharing of information practicable, assigning the highest priority to detecting, preventing, preempting, and disrupting terrorist threats against our homeland, our people, our allies, and our interests; and (B) the establishment of interface standards for an interoperable information sharing enterprise that facilitates the automated sharing of intelligence information among agencies within the Intelligence Community."
Accordingly, we wished to use a representation of the extraction results which followed the appropriate standards, and was amenable to automated processing for discovering, sharing, and integrating data across projects and agencies.
2. XML Tagging and its Limitations The usual notion of "XML tagging" of documents called for inserting tags into the document, turning each item reference into an XML element. The Intelligence Community Metadata Standard for Publication (IC-MSP) [1] defines a small set of "inline" tags for this purpose. Although the IC-MSP standard does allow for a modest number of attributes, including the xlink set, it was apparent that it – or indeed any representation based on such in-line tags – would be hard-pressed to capture all the useful information produced by IE. Consider the sentence from a sample document shown in Figure 1.
South of Baghdad near the town of Hillah, a suicide bomber blew up his car outside the house of Police chief Maj. Ahmed Suleiman, killing himself andwounding seven, officials said.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 1. Sample Input Text
The tagging allowed by IC-MSP 1.1 (the version in place at the time our work began) would look something like Figure 2. In the 5 years since, the standard has improved, based on requests from METS and other programs. The current version – 4.1 – would allow tagging as shown in Figure 3.
South of Baghdad near the town of Hillah, a suicide bomber blew up his car outside the house of Police chief Maj. Ahmed Suleiman, killing himself and wounding seven, officials said. Figure 2. Text Tagged in IC-MSP 1.1
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
132
R. Lee / The Use of Ontologies to Support Intelligence Analysis
South of Baghdad near the town of Hillah, a suicide bomber blew up his car outside the house of Police chief Maj. Ahmed Suleiman, killing himself and wounding seven, officials said. Figure 3. Text Tagged in IC-MSP 4.1
While the text indicating specific entities and events can be tagged -- although the tags still indicate the types a bit imprecisely -- even the 4.1 version still fails to capture all their relationships:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
• • • • • •
owner of car owner of house agent, location, instrument, intended victims, etc of the bombing event victims of the killing and injuring events causal relationships between the events spatial relations amongst the locations
Many of these concerns could be addressed by abandoning the inline-tag representation in favor of a more item-centric representation. This allows for a cleaner and more complete representation of the information, which facilitates discovery and linking of information across data sources. Using something other than IC-MSP also allows us to indicate the types more precisely. We therefore chose to supplement the IC-MSP representation with such an item-description representation. However, simply switching to some other XML representation would not resolve a more fundamental problem: the lack of a semantic underpinning for XML. DTDs and XSDs specify syntax, not semantics, so they provide no formal clue as to how to interpret the data. As coverpages.org (http://xml.coverpages.org/xmlAndSemantics.html) put it, bluntly, many years ago: "… we must reckon with the cold fact that XML does not of itself enable blind interchange or information reuse. XML may help humans predict what information might lie "between the tags" in the case of , but XML can only help. For an XML processor, and and are all equally (and totally) meaningless."
In other words, there are no clues to tell a person or system how to interpret, say, . Which of the following is it tagging (to quote from dictionary.reference.com)? • • • •
a large receptacle, container, or structure for holding a liquid or gas a natural or artificial pool, pond, or lake an armored, self-propelled combat vehicle … moving on a caterpillar tread a prison cell or enclosure … as for prisoners awaiting a hearing
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
.
R. Lee / The Use of Ontologies to Support Intelligence Analysis
133
Conversely, there are no clues that would facilitate the realization that one representation's is equivalent to another representation's . These limitations led us to look for a representation better suited to automated processing and integration.
3. Ontologies, Original Accordingly, we elected to use an RDF-based semantic representation. Initially, we used DAML (DARPA Agent Markup Language), and then the new W3C standard OWL (Web Ontology Language). At this time, METS uses a set of three inter-related OWL ontologies which were developed on the program. As shown in Figure 4, each consists today of a relatively small number of classes and properties. ontology # of classes # of properties
core ct 400 29 295 10
icmsp 32 60
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 4. Class and Property Counts
The core ontology was designed to arrange a broad set of domain-independent concepts into a class hierarchy. A set of properties, both for simple text values (hasName, hasColor, etc) and for relations (memberOf, uses, eventParticipant, etc) was also arranged into a hierarchy. The properties are also identified where appropriate as transitive, inverses of each other, etc, to further facilitate inferencing. This ontology borrowed the physical geography concepts (Island, River, Forest, etc) from SUMO's Geography sub-ontology, and adapted concepts from relevant database schemas and discussions with analysts, but was mostly the author's own creation. The ct ontology was designed, in like fashion, to cover classes and related properties that were deemed to be specific to the Counter-Terrorism (CT) domain; these were tied into the core hierarchies at various points (via subClassOf and subPropertyOf declarations). The icmsp ontology was designed to mirror the PublicationMetadata portion of the IC-MSP specification, following its XML structure as closely as possible within the added constraints of RDF. It follows that specification's use of the IC-ISM specification for representing security markings. It deviates a bit from the MSP specification to use key items (Person, Organization, Date, etc) out of the core ontology. This ontology is used, as the name implies, to capture document metadata much as it is captured in IC-MSP XML. A portion of the core and ct ontologies is shown in Figure 5. Fragments of the class and property (relationship) hierarchies are shown graphically in Figure 6. These hierarchies are one obvious way in which ontologies provide semantics for the concepts. They also enable the same data to serve different clients that want or need differing degrees of type precision. METS makes a point of identifying the type (ontology class) of each extracted item as precisely as the ontologies allow and as precisely as the extractors are capable of distinguishing, based on the input text. For clients that care about the distinction between, say, a missile and a firearm, the
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
134
R. Lee / The Use of Ontologies to Support Intelligence Analysis
Terrorist Person
Figure 5. Ontology Excerpts
distinction is made in the data wherever possible and their needs are met. For clients that only wish to view either one as a weapon, that need is met as well thanks to the class hierarchy. In other words, if a data item is declared to be a Missile or a Firearm, then the ontology supports the trivial inference that it is also a Weapon. Similarly, both clients that care about the distinction between the relationships employee, member, leader, etc. and clients that don't care, are accommodated by the property hierarchy. As noted earlier, properties can be defined as transitive, symmetric, and the inverse of another property, which provides additional avenues for inferencing. For example, relative is declared in the core ontology as being transitive, associate is declared
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
R. Lee / The Use of Ontologies to Support Intelligence Analysis
135
Figure 6. Class and Property Hierarchies
symmetric, and isEmployeeOf is declared the inverse of hasEmployee. These declarations allow the inferences that, for any three Persons A, B, C: A relative B A associate B
&
B relative C
A isEmployeeOf B
=>
A relative C
B associate A
B hasEmployee A
In addition to using the inferences directly sanctioned by OWL ontologies, systems often specify additional rules – using rule languages such as SWRL – to allow additional inferences. Typical examples of this might be (where E is an (appropriate kind of) Event, and L and M are Locations):
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
136
R. Lee / The Use of Ontologies to Support Intelligence Analysis
A hasParent B A locatedIn L
& &
A eventParticipant E
B hasBrother C L subLocationOf M
A hasUncle C
=> =>
& B eventParticipant E
A locatedIn M =>
A associate B
Inferences such as these are typically used in conjunction with an OWL data store in one of two ways: •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
When new data is added to the store, it triggers the application of relevant rules, which try to infer additional data to add to the store (forward chaining) When a request for certain data is received, it triggers the application of relevant rules in reverse, which try to see whether the desired data can be inferred from the existing data (backward chaining)
Note that these processes are called "chaining", because each successful application of a rule results in additional fodder for further rule application, each step thus forming a link in the chain between the known data and the conclusion(s). The key thing to note is that, under either approach, the use of such ontology-supported inferencing enables the discovery of useful data that was not directly stated. Another benefit of the OWL approach was just hinted at earlier. Just as subClassOf and subPropertyOf were used to tie concepts in ct and icmsp to concepts in core, they, along with equivalentClass and equivalentProperty, can be used to tie concepts in two ontologies that may have been developed completely independently. Such declarations, perhaps augmented with rules similar to those discussed above, can be used to at least partially automate the conversion of data between the two ontologies, thus facilitating data sharing and fusion. It should be noted that, along with the benefits of using OWL ontologies, there are a few drawbacks. One is the lack – or perhaps just the perceived lack -- of available tools to exploit OWL and RDF data. This is less of a problem now than when we began this work 5 years ago. The second is the resistance to the whole idea of OWL and ontologies on the part of the client community. Part of this is no doubt due to the first drawback, but part appears to be due to our failure to convince clients of the benefits; and even clients that see the value of the representation are still slow to implement tools to use it. The third drawback is a crucial technical one. In the IC, it is critical that the data representations support attaching various kinds of metadata directly to specific statements in the data. This metadata is similar to the metadata required on a document as a whole: • • • •
Security markings Sources Dates Confidence
In the IC's standard XML representations, it is a relatively simple matter to add attributes to the appropriate elements to implement this, and the use of those attributes is easy to mandate where desired in the corresponding XML schemas. Not so with
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
137
OWL/RDF. Instead, it is necessary to "reify" a statement, which in the XML rendering of OWL/RDF means creating an element for the statement which contains its 3 components and the metadata. For example, associating a classification with our statement 'A hasParent B' would look like Figure 7. This is, needless to say, a bit cumbersome; more importantly, it does not appear to be possible for an ontology developer to mandate the use of such a construct where desired.
Figure 7. Example of Reification for Security
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. Ontologies, New As noted earlier, the ontologies currently in use with METS are almost entirely homegrown. They originated as DAML and were trivially converted to OWL when that became a standard; no effort was made to adopt the emerging “best practices” for OWL. For METS 3.0, we are developing a new set of ontologies which make better use of the industry's ontology development and of the IC's XML standards. And while implementing this, we have tried to follow the guidance of the ontology development tutorials of Kendall and McGuiness [2], and the ontology development book by Allemang and Hendler [3]. Specifically, we are following the lead of the IC's XML standards group in replacing the document metadata portion of IC-MSP with DDMS [4] (and using ICISM [5] for security, of course); we are using the SUMO family [6] (SUMO, MILO, Geography, etc) of ontologies as the centerpiece of the new set, augmented by most of the concepts from the TWPDES [7] XML standard. For all of these, we are constructing OWL representations from the original XML or KIF. We are also constructing OWL versions of all the controlled vocabulary lists from FIPS, ISO, etc used by DDMS and TWPDES, using the W3C’s time ontologies [8], and are augmenting the W3C's small GML-based OWL ontologies [9] with the additional bits used by DDMS. We will also add equivalent… and sub… declarations to relate concepts between these ontologies, and from them to other popular ontologies such as Dublin Core, FOAF, etc.
5. METS (Current) Description The initial evaluation confirmed that the idea of using IE to tag documents was sound, and we were directed to further develop the evaluation system into an operational one. METS (Metadata Extraction and Tagging Service) is the result of that work, and has been operational on the classified network for a few years, providing its services to the IC. The architecture of the current release (2.5) is shown in Figure 8.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
138
R. Lee / The Use of Ontologies to Support Intelligence Analysis
Figure 8. METS Architecture
METS is fronted by 4 operational (and 1 experimental) web services: •
Persistent service ties to the agency's feed of messages and newswire articles (WISE) and processes them into a set of data stores
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
• • •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
139
On-Demand service accepts arbitrary documents and processes them back to the submitter Query service retrieves processing results from the data stores matching the query Bulk Transfer service retrieves all results produced and stored in the specified time interval (Thingfinder service provides a stand-alone experimental interface for trying that extraction tool)
METS accepts documents in a variety of forms (text, HTML, XML, Word, PDF) and first normalizes each into a standard IC-MSP XML form. As part of this process, it identifies all the document metadata (security markings, dates, sources, document IDs, title, etc). This metadata is converted to IC-MSP and OWL. METS then hands the normalized text to a set of commercial IE tools to extract information as described earlier. It converts – using custom code – the results from each extractor into OWL conforming to the ontologies. This conversion is a mostlystraightforward mapping from each of the Java objects provided by each extractor's API to an XML element (rdf:Description) for that object; in some cases, a single object gets converted into more than one Description, and of course any objects representing relationships need to be handled differently. The conversion code was written to use mapping rules specified in static files, so that changes to the ontologies or to extraction capabilities could (usually) be accommodated without changing the code. Note that the extraction tools know nothing about the ontologies; it is a common misconception that changing (or replacing) ontologies will automatically cause the extractors to behave differently. METS then merges the OWL results from the extractors, and runs them through a series of deconfliction steps. These steps use hand-crafted heuristic rules to decide how to handle each disagreement. The rules answer questions such as: "If tool A says a piece of text refers to a Person, and tool B says the same text refers to an Organization, what do we put in the result?" The answer could be "neither", "both", "Person", "Organization", or something else, like "Agent". In reality, the problem is even more complicated, as the two pieces of text could be nested or overlapping rather than identical, and any tool may have itself assigned multiple types to the text. And, of course, adding a 3rd extractor adds a whole new level of complexity. METS then runs the merged and deconflicted result through a series of name clean-up steps, looking for and correcting assorted known problems with the various extractors. These include stripping off leading articles and prepositions, trailing verbs, etc. METS then uses the OWL representation of the extraction results to insert the appropriate in-line tags into the IC-MSP. Although not shown in the examples provided here, the OWL includes offset information indicating where in the normalized document each reference was found. Since the version of IC-MSP – 3.1 – which was current when METS 2.5 was released still lacked many of the needed tags, METS offered a choice of MSP outputs. The 'standard' choice mapped each METS tag (OWL class in core or ct) to the best MSP tag, often with considerable loss of precision; the 'extended' (default) choice used MSP's ExtensionElements to wrap each METS element, allowing the result to stay compliant with the IC-MSP specification, with a considerable increase in bulk.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
140
R. Lee / The Use of Ontologies to Support Intelligence Analysis
Finally, METS combines the metadata with the in-line tagged text for the IC-MSP, and the metadata with the extraction results for the OWL, and passes the resulting pair of documents either into the data store or back to the submitter as requested. Figures 9 and 10 show the IC-MSP and OWL (respectively) for our sample sentence. The IC-MSP is the 'extended' version; we've added a few carriage returns for readability. Also, we have omitted most of the document metadata from both Figures. The attentive reader will note that the outputs shown do not capture all the location relationships and event relationships that were pointed out in the earlier discussion. There is no denying that IE tools have not reached the level of human proficiency, particularly where events and relationships are concerned.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5A8CAA84-A482-4743-BE33-47DBCC635F19
http://example.com/20072004_TURKMENISTAN.xml
South of Baghdad near the town of Hillah, a suicide bomber blew up
his car outside the house of Police chief Maj. Ahmed Suleiman, killing himself and wounding seven , officials
said . … ...
Figure 9. Sample IC-MSP Output
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
5A8CAA84-A482-4743-BE33-47DBCC635F19
URL http://example.com/20072004_TURKMENISTAN.xml
USA
Baghdad
Hillah the town
a suicide bomber his himself
blew up
his car
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
141
142
R. Lee / The Use of Ontologies to Support Intelligence Analysis
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
the house of Police chief Maj. Ahmed Suleiman
Maj. Police chief Ahmed Suleiman
killing
wounding
seven 7
officials
said
...
Figure 10. Sample OWL Output
6. A Multi-INT Experiment The data processed by METS for storage is message traffic (largely HUMINT) and newswire articles from WISE. We undertook an experiment to see whether we could effectively correlate the data produced by METS with data from an IMINT system based on location overlap. First, we extended the ontologies with more geographic and geometric concepts to support this. We converted a very large table of NGA location records into OWL and pre-loaded that OWL into our RDF store. This data included a coordinate centerpoint for each location but not, unfortunately, any boundary coordinates, so we estimated radii based on location type. We added code which:
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
143
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
-85.74 +120.45
-85.74 +146.13
-88.62 +110.37
-88.62 +136.05
Figure 11. Sample OWL Conversion of IMINT Data
• • •
attempted to match each extracted location with a location in the RDF store converted the IMINT (imagery coverage) data into OWL looked for location overlaps (using coordinates) between the data sets
The IMINT data specified a series of "image blocks", with the coverage for each specified by the four corner coordinates. The conversion of this data, using the same (enhanced) ontologies as the normal METS processing, was straightforward. A sample image block definition is shown in Figure 11.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
144
R. Lee / The Use of Ontologies to Support Intelligence Analysis
The results of the experiment were encouraging, but suffered from: • •
the inability to disambiguate (and therefore know the coordinates for) location references in many cases a rather naïve and slow region-overlap detection algorithm
The lessons learned from this are that we should have: • • •
based our ontology additions on existing conversions of GML pieces used geotaggers, which use document context to help disambiguate locations consulted with geo experts to devise a more efficient region-overlap detection algorithm
7. METS (Upcoming) Description We have been directed to reduce the complexity and cost of METS by stripping it down to a single IE tool, which also removes the need for the merge and deconfliction steps, and by replacing other non-free COTS products with free or custom-coded alternatives. We are in the process of implementing the next version (3.0) accordingly. We have elected to replace the ontologies for this new version, as described in the “Ontologies, New” section above. We are also upgrading the IC-MSP XML support from version 3.1 to 4.1. We will be implementing tailored extraction rule sets (KBs) for 3 kinds of perclient requirements:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
• • •
Specific entity, event, property and relation types Precision / Recall balance Peculiarities of the input documents
8. Future Work We are coordinating with efforts such as the DNI’s Catalyst entity resolution program, in the hopes of achieving ontology compatibility with other projects working in the IC. We are also trying to work with our commercial developers and our customer to foster the development and adoption of OWL-enabled tools. These steps will, we hope, enable the OWL data which METS is generating to finally demonstrate the full potential benefits we have described here.
References [1] [2] [3]
IC-MSP (Intelligence Community Metadata Standard for Publication). http://www.dnidata.org. E. Kendall, D. McGuiness. Advanced Topics in the Web Ontology Language (OWL). 2008 Semantic Technology Conference. D. Allemang, J. Hendler. Semantic Web for the Working Ontologist. Morgan Kaufmann Publishing 2008.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
R. Lee / The Use of Ontologies to Support Intelligence Analysis
DDMS (Defense Discovery Metadata Standard). http://metadata.dod..mil. IC-ISM (Intelligence Community Information Security Markings). http://www.dnidata.org. SUMO (Suggested Upper Merged Ontology). http://www.ontologyportal.org/. TWPDES (Terrorist Watchlist Person Data Exchange Standard). http://www.dnidata.org. Time Ontologies. http://www.w3.org/TR/owl-time/. GML Ontologies. http://www.w3.org/2003/01/geo/.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[4] [5] [6] [7] [8] [9]
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
145
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-147
147
Chapter 9
Probabilistic Ontologies for Multi-INT Fusion Kathryn Blackmond LASKEYa1, Paulo C. G. COSTAa, and Terry Janssenb a C4I Center, George Mason University, USA b Lockheed Martin Corporation, USA
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Abstract. Systems are increasingly required to fuse data from geographically dispersed, heterogeneous information sources to produce up-to-date, missionrelevant results. These products focus not only on traditional military forces and systems, but to an increasing degree also on non-traditional combatants and their social networks. Successful multi-INT fusion requires that the constituent systems interoperate not just at the level of syntax and formats, but also at the level of semantics. Ontologies are vital enablers for semantic interoperability. Because uncertainty is a fundamental aspect of multi-INT fusion, lack of support for uncertainty is a major limitation of current-generation ontology formalisms. Probabilistic OWL (PR-OWL) extends the OWL Web Ontology Language to enable the construction of probabilistic ontologies. Ontologies constructed in PROWL can represent complex patterns of evidential relationships among uncertain hypotheses. Recently, a system for specifying and reasoning with PR-OWL ontologies has been released in beta version. This paper describes the PR-OWL ontology language, the probabilistic logic on which it is based, and the reasoning system implementation. A hypothetical case study in the counterterrorism domain illustrates the capabilities of PR-OWL.
Introduction Multi-INT fusion is a critical technology for the next generation of military and intelligence systems. As connectivity and bandwidth increases, commanders and analysts are deluged with ever-greater volumes of data from geographically dispersed, heterogeneous information sources. In today’s military engagements, fusion products must focus not only on traditional military forces and systems, but also on nontraditional combatants and their social networks. Successful multi-INT fusion requires that the constituent systems interoperate not just at the level of syntax and formats, but also at the level of semantics. That is, interoperating systems should interpret terminology in a consistent way; or if not, appropriate translations must be established between vocabularies used by different systems. Techniques for making semantic information explicit and computationally accessible are key to effective exploitation of data from diverse sources. Shared formal semantics enables systems with different
1
Corresponding Author: Kathryn Blackmond Laskey, C4I Center, George Mason University, Fairfax, VA, 22030, USA; E-mail: [email protected]. Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
148
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
internal representations to exchange information, and provides a means to enforce business rules such as access controls for security. When heterogeneous systems are required to interoperate in an open world, vocabularies that were developed for individual stand-alone applications break down. Ontologies provide shared representations of the entities and relationships characterizing a domain, into which vocabularies of legacy systems can be mapped. However, a major limitation of traditional ontology formalisms is the lack of consistent support for uncertainty. Because uncertainty is a fundamental aspect of multi-INT fusion, this is a serious deficiency. Current ontology formalisms provide no principled means to ensure semantic consistency with respect to issues of uncertainty or data quality. Probabilistic ontologies [1] augment standard ontologies with probabilistic information about the domain. Probabilistic OWL (PR-OWL) extends the OWL Web Ontology Language to enable the construction of probabilistic ontologies. PR-OWL is based on (MEBN), a first-order probabilistic logic that combines the representational power of first-order logic (FOL) and Bayesian Networks (BN) [2]. Ontologies constructed in PR-OWL can represent complex patterns of evidential relationships among uncertain hypotheses. Recently, a system for specifying and reasoning with PROWL ontologies has been released in beta version [3, 4]. This system, called UnBBayes-MEBN, provides a graphical user interface for defining entities, attributes, and probabilistic relationships, defining instances, entering evidence, and entering queries. It also includes a reasoning system for performing Bayesian inference to calculate responses to probabilistic queries. The following section describes the PR-OWL ontology language and the MEBN logic on which it is based. Section 2 describes the UnBBayes system for entering and reasoning with PR-OWL probabilistic ontologies. Section 3 illustrates the capabilities of PR-OWL with a hypothetical case study in the counterterrorism domain.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1. Probabilistic Ontologies Initial attempts to represent uncertainty in ontology languages tend to begin with constructs for attaching probabilities as attributes of entities. This approach is clearly inadequate, in that it fails to account for structural features such as conditional dependence (or independence), double counting of influence on multiply connected graphs, and context-specific independence. Many researchers have pointed out the importance of structural information in probabilistic models (e.g. [5, 6, 7]), and it is well known that some questions about evidence can be answered entirely in structural terms (e.g., [6], page 271). For instance, Shafer ([8], pages 5-9) stated that probability is more about structure than it is about numbers. This is particularly true in domains such as intelligence analysis and Human Intelligence (HUMINT), which rely on complex chains of argument with many interacting uncertain hypotheses, in which subtle features of an argument may augment or diminish its force [6]. Structural information also plays a major role in the way evidence collected from multiple sensors with different degrees of reliability and trust is evaluated. In many cases, different aspects of the same piece of information have to be analyzed and weighed based on incomplete knowledge about the source. Structural information is a key asset to provide an in-depth analysis of what each piece of knowledge means in the overall context of an evidential chain. Special-purpose stand-
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
149
alone systems may not explicitly represent many of these subtle structural features, leaving them as implicit assumptions underlying the algorithmic processing performed by the system. However, when systems interoperate, it is essential to represent explicitly the assumptions underlying the processing, and to share information about the context of reasoning, to enable the consuming system to properly assess credibility of the information and its import within the overall context of reasoning. That is, systems must share not only conclusions, but also semantic information about how those conclusions were reached and the conditions under which the conclusions are valid. This requires semantic interoperability. State-of-the-art systems are increasingly adopting ontologies as a means to ensure formal semantic support for knowledge sharing. Uncertainty is becoming recognized as an important aspect to be represented and used in reasoning. A common mistake is to provide support for uncertainty representation by simply annotating ontologies with numerical probabilities. This is a weak approach that leads to fragile intelligence systems, as too much information is lost due to the lack of a representational scheme that can capture structural nuances of the probabilistic information. Clearly, more than mere annotation is needed. Indeed, there is a need for a new category of ontologies. Definition 1 (from [1]): A probabilistic ontology is an explicit, formal knowledge representation that expresses knowledge about a domain of application. This includes:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1. Types of entities that exist in the domain; 2. Properties of those entities; 3. Relationships among entities; 4. Processes and events that happen with those entities; 5. Statistical regularities that characterize the domain; 6. Inconclusive, ambiguous, incomplete, unreliable, and dissonant knowledge related to entities of the domain; and 7. Uncertainty about all the above forms of knowledge; where the term entity refers to any concept (real or fictitious, concrete or abstract) that can be described and reasoned about within the domain of application. Probabilistic ontologies are used for the purpose of comprehensively describing knowledge about a domain, along with its associated uncertainty, in a principled, structured and sharable way. Ideally, this knowledge should be represented in a format that can be read and processed by a computer. Probabilistic ontologies also expand the possibilities of standard ontologies by introducing the requirement of a proper representation of the statistical regularities in a domain, and uncertain evidence about entities in a domain of application. Another aspect that must be emphasized when devising data sharing schemes for intelligence systems is the level of expressivity of a representation formalism. In other words, any representational scheme that attempts to convey all the details and idiosyncrasies of a complex domain must be highly expressive. Although tractability requirements often motivate restrictions on the ability of reasoning engines to process highly expressive representations, if ontologies are to be general-purpose repositories of shared knowledge, then restrictions on reasoners should not dictate what it is possible to say about a domain. PR-OWL, which is used in this paper as the language for building probabilistic ontologies, can achieve the required level of expressivity
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
150
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
because it is based on a First-Order Bayesian Logic that represents probability distributions over interpretations of arbitrary first-order domain theories [2]. MEBN is a first-order Bayesian logic that integrates classical first-order logic with probability theory. Classical first-order logic (FOL) is by far the most commonly used, studied and implemented logical system, serving as the logical basis for most currentgeneration AI systems and ontology languages. MEBN logic provides a logical foundation for extending the capability of ontology languages to include a logically coherent representation for uncertainty. Because a MEBN theory represents a coherent probability distribution, Bayes Theorem provides a mathematical foundation for learning and inference that reduces to classical logic in the case of certain knowledge (i.e., all probabilities are zero or one). MEBN represents the world as comprised of entities that have attributes and are related to other entities. Knowledge about the attributes of entities and their relationships to each other is represented as a collection of MEBN fragments (MFrags) organized into MEBN Theories (MTheories). An MFrag represents a small, repeatable piece of knowledge about the probabilistic relationships among a set of interrelated hypotheses about attributes of or relationships among entities of given types. The generic knowledge represented by the MFrag can be instantiated repeatedly on different entities of the allowable types, thus composing complex argument structures from repeated sub-structures. An example of this is shown below in the case study. Specifically, an MFrag contains context, input, and resident random variables (RVs), a fragment graph and local distributions. The RVs represent uncertain hypotheses; the fragment graph represents dependency relationships among the RVs; and the local distributions provide quantitative information about the strength of the relationships encoded by the fragment graph. Together, the fragment graph and the local distributions define conditional probability distributions for instances of the resident random variables (RVs), conditional on the values of instances of their parents in the fragment graphs, and given the context constraints. Distributions for the input and context RVs are defined in other MFrags. Context nodes represent conditions assumed for definition of the local distributions. A collection of MFrags that satisfies certain consistency constraints implicitly defines a joint probability distribution on instances of its random variables. Such a collection of MFrags is called an MTheory. MEBN semantics integrates the standard model-theoretic semantics of classical first-order logic with random variables as formalized in mathematical statistics. Specifically, a theory in first-order logic defines a set of possible worlds; and any world in which all the axioms of the theory are satisfied is called a model of the axioms. Beyond ruling out worlds inconsistent with the axioms, classical logic cannot say anything about relative plausibility of the possible worlds. A first-order Bayesian logic such as MEBN can grade the possible worlds according to plausibility. Thus, from a given set of axioms, first-order logic can do no more than assert that an assertion is proven, disproven, or neither proven nor disproven. As with FOL, MEBN logic assigns probability zero to assertions that can be disproven from the axioms of an MTheory, and probability one to assertions that can be proven. However, MEBN logic can assign probabilities between zero and one to hypotheses that can be neither proven nor disproven.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
151
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 1. Main Elements of PR-OWL
As a full integration of first-order logic and probability, MEBN provides: (1) a means of expressing a globally consistent joint distribution over models of any consistent, finitely axiomatizable FOL theory; (2) a proof theory capable of identifying inconsistent theories in finitely many steps and converging to correct responses to probabilistic queries; and (3) a built in mechanism for adding sequences of new axioms and refining theories in the light of observations. Thus, even the most complex situations can be represented in MEBN, provided they can be represented in FOL. Furthermore, because MEBN is a first order Bayesian logic, its use as the underlying semantics of PR-OWL not only guarantees a formal mathematical foundation for a probabilistic extension to the OWL language (PR-OWL), but also ensures that the advantages of Bayesian Inference (e.g. natural “Occam’s Razor”, support for learning from data, etc.) will accrue to PR-OWL probabilistic ontologies. A comprehensive explanation of MEBN logic is not on the scope of this paper, but the interested reader is directed to [2]. PR-OWL was developed as an extension enabling OWL ontologies to represent complex Bayesian models in a way that is flexible enough to be used by diverse Bayesian probabilistic tools (e.g. Netica, Hugin, Quiddity*Suite, JavaBayes, etc.) based on different probabilistic technologies (e.g. PRMs, BNs, etc.). More specifically, PROWL is an upper ontology for probabilistic systems that can be used as a framework for developing probabilistic ontologies (as defined above) that are expressive enough to represent even the most complex probabilistic models. DaConta et al. define an upper ontology as a set of integrated ontologies that characterizes a set of basic commonsense knowledge notions ([9], page 230). In PR-OWL, these basic commonsense notions are related to representing uncertainty in a principled way using OWL syntax (itself a specialization of XML syntax), providing a set of constructs that can be employed to build probabilistic ontologies. Figure 1 shows the main concepts involved in defining an MTheory in PR-OWL. In the diagram, ellipses represent general classes while arrows represent the main relationships between these classes. A probabilistic ontology (PO) has to have at least one individual of class MTheory, which is basically a label linking a group of MFrags that collectively form a valid MTheory. In actual PR-OWL syntax, that link is expressed via the object property hasMFrag (which is the inverse of object property isMFragIn). Individuals of class MFrag are comprised of nodes, which can be resident, input, or context nodes (not shown in the picture). Each individual of class Node is a random variable RV and thus has a mutually exclusive, collectively exhaustive set of possible states. In PR-OWL, the object property hasPossibleValues links each node with its possible states, which are individuals of class Entity. Finally, random variables (represented by the class Nodes in PR-OWL) have unconditional or conditional probability distributions, which are represented by class ProbabilityDistribution and linked to its respective nodes via the object property hasProbDist. Figure 2 depicts the
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
152
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
main elements of the PR-OWL language, its subclasses, and the secondary elements necessary for representing an MTheory. The relations necessary to express the complex structure of MEBN probabilistic models using the OWL syntax are also depicted. In addition to [1] the prospective reader will find more information on the PR-OWL language at http://www.pr-owl.org.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 2. PR-OWL Elements
The first step to build a probabilistic ontology in compliance with Definition 1 is to import into any OWL editor an OWL file containing the PR-OWL classes, subclasses, and properties (one is available at http://www.pr-owl.org/pr-owl.owl). After importing the PR-OWL definitions, the next step in ontology design is to construct domainspecific concepts, using the PR-OWL definitions to represent uncertainty about their attributes and relationships. Using this procedure, an ontology engineer is not only able to build a coherent generative MTheory and other probabilistic ontology elements, but also make it compatible with other ontologies that use PR-OWL concepts.
2. A Reasoner for Bayesian Ontologies At its current stage of development, PR-OWL contains only the basic representation elements that provide a means of representing any MEBN theory. Such a representation could be used by a Bayesian tool (acting as a probabilistic ontology reasoner) to perform inferences to answer queries and/or to learn from newly incoming evidence via Bayesian learning. However, building MFrags in a traditional ontology editor is a manual, error-prone, and tedious process. Avoiding errors or inconsistencies requires deep knowledge of the logic and of the data structures of PR-OWL. The user would have to know many technical terms such as hasPossibleValues, isNodeFrom, isResidentNodeIn, etc. Furthermore, reasoning with a PR-OWL ontology requires creating instances of the random variables needed to respond to a given query, assembling them into a Bayesian network, and entering that Bayesian network into a software application that can
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
153
perform the desired inference. This too is a tedious, manual, error-prone process. Ideally, much of this work could be automated by a software application designed to enforce the consistency of a MEBN model and to respond correctly to queries.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 3. The UnBBayes-MEBN GUI
The development of UnBBayes-MEBN, an open source, Java-based application,2 is an important step towards this objective, as it provides both a GUI for building probabilistic ontologies and a reasoner based on the PR-OWL/MEBN framework. UnBBayes-MEBN was designed to allow building probabilistic ontologies in an intuitive way without having to rely on a deep knowledge of the PR-OWL specification. Figure 3 shows a snapshot of the UnBBayes-MEBN user interface. In the figure, a click on the “R” icon and another click anywhere in the editing panel will create a resident node, for which a description can be inserted in the text area at the lower left part of the screen. Clicking on the arrow icon would allow one to graphically define the probabilistic relations of that resident node with other nodes, as much as it would be done in current Bayesian packages such as Hugin™. All those actions would result in the software creating the respective PR-OWL tags (syntactic elements that denote particular parts of a PR-OWL ontology) in the background. Probabilistic Ontologies in UnBBayes-MEBN are saved in PR-OWL format (*.owl file), while application-specific data is stored in a text file with the *.ubf extension (UnBBayes file format). Support for MEBN input/output operations is provided via the Protégé-OWL API3, which is based on the class JenaOWLModel. By using a common API, UnBBayes-MEBN ensures that MTheories created using its GUI can be opened and edited in popular ontology editor Protégé4 (and vice-versa). This compatibility is important because it ensures that files created in UnBBayes-MEBN can be opened and edited not only in Protégé, but also in any OWL-compliant application (although these applications will not be able to understand the ontology’s probabilistic characteristics). In addition, ontologies that have already been defined using an OWL-compliant editor can be extended to the PR-OWL format in a quick and 2
At this writing, UnBBayes-MEBN is in beta phase (public release January 09). http://protege.stanford.edu/plugins/owl/api/index.html 4 http://protege.stanford.edu 3
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
154
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
direct way. All that is needed is to open the OWL file in UnBBayes-MEBN, create an MTheory for the ontology, and save the result. UnBBayes-MEBN provides not only a GUI for building probabilistic ontologies, but also a probabilistic reasoner that allows for plausible inferences using Bayes Theorem as evidence accrues. Currently, only a restricted class of queries has been implemented. Specifically, a query asks for the distribution of a single random variable instance, as well as some additional limitations in the Beta version on the kinds of queries that can be processed. Future releases will include the ability to perform queries on several random variable at the same time, as well as queries involving the application of logical operations on random variables. When a query is submitted, the knowledge base is searched for information to answer the query. If the available information does not suffice, then the KB and the generative MTheory are used to construct a BN to answer the query. This process is called Situation Specific Bayesian Network (SSBN) construction. In the current implementation, a query consists of a single random variable (RV) instance. The following procedure takes a node name and a list of entity instances as arguments. It is called initially with the query node and its arguments. PROCEDURE SSBN-CNSTR(NODE,ENTITY-LIST) 1. 2.
3.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. 5.
6.
7. 8.
For the RV instance NODE(ENTITY-LIST), search for evidence in the KB. If there is a finding for this given entry, finish. Search for the resident node that has the name NODE and get its MFrag. Once NODE(OV-LIST) is found, verify if the type of ENTITY-LIST is the same as OVLIST (where OV-LIST is the list of ordinary variable arguments for NODE in its home MFrag). Verify in the MFrag which context nodes refer to the OVs in OV-LIST. Then replace each OV by the appropriate instance in ENTITY-LIST and evaluate it according to the KB. If any context variable is false, mark the MFrag to use the default distribution. Check whether the truth-value of the context node in (3) can be determined from the evidence. If not, make that context node a parent of NODE. For each child of NODE in the same MFrag, identify any instance of the child that can be constructed by replacing the OVs by the known entities (contained in the query or KB), and has not yet been added to the SSBN. For each such child instance, call procedure SSBN-CNSTR for the child node and its arguments. Search for MFrags where NODE is an input node. For each child of NODE (in all MFrags found), identify any instance of the child that can be constructed by replacing the OVs by the known entities (contained in the query or KB), and has not yet been added to the SSBN. For each such child instance, call procedure SSBN-CNSTR for the child node and its arguments. If NODE is a query or finding node, then mark as permanent. If NODE is not dseparated by the finding nodes from the query node, then mark as permanent.5 If NODE is permanent, then for each parent of NODE, identify any instance of the parent that can be constructed by replacing the OVs by the known entities (contained in the query or KB), and has not yet been added to the SSBN. For each
5 The d-separation check is still under development; thus, the beta version cannot handle some kinds of queries. A fully capable SSBN algorithm for single node queries is expected in summer 2009.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
155
such parent instance, call procedure SSBN-CNSTR for the parent node and its arguments. 9. Create the NODE's CPT. 10. Finish.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This algorithm is easily enhanced to allow multiple query nodes. This enhancement is currently under development. A few performance issues had to be considered in implementing UnBBayesMEBN. Depending on the complexity of the domain, the algorithm may reach a context node that cannot be immediately evaluated. This happens when all ordinary variables in the parent set of a resident random variable term do not appear in the resident term itself. In this case, there may be an arbitrary, possibly infinite number of instances of a parent for any given instance of the child. In this case, the local distribution for a random variable must specify how to combine influences from all relevant instances of its parents. However, especially in complex formulas this may have a strong impact in the performance of the algorithm, so the current algorithm uses the default distribution instead. A planned future enhancement is to give the user options for how to handle this case. Another design option was to restrict memory usage in a way that a possible memory overload triggers a warning to the user and stops the algorithm. In step (3), a design optimization over the general SSBN algorithm in [2], only the necessary context nodes for a given MFrag are evaluated, in contrast with the original solution of evaluating all the context nodes for that MFrag. Although the implementation addressed other optimization issues, for the sake of conciseness only the most relevant are listed here. UNBBayes-MEBN is a work in progress that is still in beta status, but it already provides a major contribution to the development of probabilistic ontologies. Its basic functionality was enough to support our work in designing a case study employing POs as a knowledge sharing enabler.
3. Case Study: Attack in Lahore To illustrate the capabilities of PR-OWL to represent the kinds of multi-INT fusion problems faced by today’s net-centric systems, we consider a hypothetical counterterrorism case study. Our simple illustrative scenario concerns an attempted attack on a high-profile meeting in Pakistan that is detected and prevented through collaboration between two intelligence analysts and interoperation of diverse fusion systems. Although the scenario is hypothetical, it illustrates the role of semantic technology and probabilistic reasoning in enabling a successful intervention to prevent a terrorist plot from succeeding. The analysts. Intelligence analyst IA1 has been assigned the task of compiling and maintaining social networks of persons-of-interest in Pakistan. Over time, he has developed a social network that includes a known arms dealer (AD) in Islamabad and his associates. Meanwhile, intelligence analyst IA2 has been tasked with compiling and maintaining an intelligence profile of the city of Lahore. In this role, IA2 has access to all intelligence reports associated with people, events, communications, etc within his area of responsibility (AOR).
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
156
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 4. MFrags for First SSBN
The meeting. At present, IA2 is aware of, and is monitoring, a conference occurring in Lahore. Attendees include six Tribal Leaders (TL1 – TL6). This is a high-profile meeting that is receiving heavy coverage by news agencies all over the world, and is therefore of concern as a potential terrorist target. The arrest. At the Lahore airport, a canine unit has detected explosive residue on a Lahore resident (P) attempting to leave the city. Upon receiving this report, IA2 declares P a person-of-interest. This declaration initiates an automatic action to add P to the scope of IA1’s social network, and to alert IA1 to report any significant results concerning P coming from the social network analysis. IA1’s analysis uncovers a thirdorder relationship between P and AD: P’s brother, BP, has the same religious advisor, C, as AD. Figure 4 shows a set of MFrags that could be used to support the above analysis. These MFrags are shown as screenshots from the UNBBayes-MEBN system. The MFrags involve reasoning about entities of different types and the relationships among them. In an operational analyst support system, the PR-OWL ontology that represented the uncertain aspects of this problem would import existing upper ontologies and domain ontologies. For this illustration, we constructed a simple, stand-alone PR-OWL ontology. Plan Agent and Target MFrag. This MFrag represents basic information about attacks using explosives. The context random variables, drawn as pentagons at the top of the MFrag, represent logical conditions assumed to hold when the probability distributions are assigned. In this case, the context random variables state that pln
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
157
represents an attack plan, agt (the agent about whom we are reasoning) and v (the potential victim) represent persons-of-interest, and tgt represents a venue (the potential target of the attack). In our simple example, we take a venue to denote a localized space-time region that might be the focus of an attack, or might otherwise be of relevance to the situation. The MFrag contains two input random variables, whose distributions are defined in other MFrags. These are shown as trapezoids in the figure. They represent whether agt is a weapons supplier and whether the potential victim v and the subject of our inquiry agt are rivals in the social network. Root nodes in the MFrag are random variables representing whether the plan is active, the political importance of the target, and whether a rival of agt is expected to be present at the venue. Whether the venue is targeted depends on whether the plan is active (if the plan is not active, then no venue is targeted by the plan), and the political importance of the venue (important venues are more likely to be targeted). Whether agt is an agent of the plan, i.e., is actively involved in bringing it about, depends on whether the plan is active, whether a rival of agt is expected to be at the venue (agents may try to target their rivals), and whether the agent is a weapons supplier (weapons suppliers are more likely to be agents in attacks using explosives). Finally, whether agt plays the role of supplying weapons depends on whether agt is an agent of the plan and whether agt is a weapons supplier. Social Network MFrag. This MFrag represents the actors and their relationships. Its context variables state that agt1 and agt2 are persons-of-interest and pln is an attack plan. It represents the knowledge that two agents of the same plan are likely to be related in the social network. It also represents probabilities that two persons-of-interest are rivals and that a person-of-interest is a weapons dealer. Plan Execution MFrag. This MFrag represents the knowledge that an agent of a plan may execute the plan, and one of the activities a plan executor might perform is to plant explosives at the targeted venue. Forensic Report MFrag. This MFrag represents the possibility that an individual who plants explosives may be apprehended and explosive residues detected. Of course, the model described here is highly simplified – its purpose is to illustrate the capabilities of the language and not to provide a sophisticated representation of terrorist attacks. Our ability to represent the problem is limited by the inability of the beta implementation of UNBBayes-MEBN to represent subtypes. We expect this limitation to be removed in future versions. We could not use UNBBayes-MEBN to construct a situation-specific Bayesian network (SSBN) for IA2’s analysis problem because of limitations in the kinds of query that can be handled by the beta version. Nevertheless, we did construct a SSBN by hand for this problem. To do this, we first defined instances of the relevant entities: P, C, and AD (persons of interest), Conf (a venue), and ConfAtk (an attack plan). The query of concern is whether the conference is targeted.. This is represented by the RV instance IsTarget(Conf, ConfAttack). This is an instance of the generic RV IsTarget(tgt, pln) in the Plan Agent and Target MFrag, in which Conf has been substituted for the ordinary variable tgt and ConfAtk has been substituted for the ordinary variable pln. Evidence random variable instances ExplosiveResidueReport(P), SNRelated(P, AD), SNRelated(C, AD), IsWeaponSupplier(AD), and PoliticalImportance(Conf) are also created to represent the information that explosive residues were found on P, that P and AD are related through the social network analysis, that C and AD are related through the social network analysis, that AD is an arms dealer, and that the conference has high political importance. SSBN construction begins with these RV instances, identifies any
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
158
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
Figure 5. Additional MFrags for Second SSBN
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
additional RV instances needed to compute a query response, instantiates them, and uses the fragment graphs to compose them into a Bayesian network. After declaring the evidence, a standard belief propagation algorithm is used to compute a query response. The prior probability of an arbitrary venue being targeted for an attack was set at 0.02%. For an event of high political importance such as the conference in question, the probability is 0.3%. After incorporating the information that AD is a weapons dealer who is related in the social network to both C and P, and that explosive residues were detected on P, the probability that the conference has been targeted for attack becomes about 10%. As part of his continuing analysis, IA2 has been monitoring the system for current intelligence information related to the conference. A query for the current locations of TL1 through TL6 reveals a HUMINT report that TL6 was seen in Karachi five hours ago. A query for IMINT change detection indicates that a vehicle that was present during the conference is now missing from the conference location. A further analysis of the HUMINT report reveals that TL6 was seen entering the residence of C. Finally, a query to the social network system reveals that TL6 and TL5 are bitter rivals. Figure 5 shows a set of MFrags that can be instantiated to incorporate this new information. Agent Location MFrag. This MFrag represents the knowledge that an individual who is expected at a venue is likely to be at the venue unless the individual is an agent of a plan that targets the venue.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
159
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 6. Situation-Specific Bayesian Network for Conference Bombing
Meeting Venue MFrag. This MFrag represents the residence of one of the agents involved in a plan is a candidate for the venue at which a meeting will be held to organize the plan. Location Report MFrag. This MFrag represents the information that the vehicle of an agent expected to be at a given location is likely to be at the location, and that the reported location of a HUMINT location report is likely to be the place where the agent is located. Location Constraint MFrag. This MFrag represents the constraint that an agent cannot be two places at once. Whenever there are two instances of the random variables AgentAt(agt, loc) with the same agent and two different venues, such that the venues are different physical locations at the same time, an instance of this RV is created and instantiated to the value possible. The value impossible has probability 1 if both parents are true. Thus, this RV enforces the constraint that an agent cannot be two places at once. After constructing the situation-specific Bayesian network and adding the evidence that TL6 was reported to be entering C’s residence, that C’s residence was a location other than the conference, that TL6 and C are related in the social network (inferred by logical reasoning from the visit to C’s residence), that TL6’s car was missing from the conference, and that TL6 and TL5 are rivals, the probability has increased to about 71% that the conference has been targeted for an attack. Figure 6 shows the SSBN constructed by hand using the Netica® Bayesian network software package. Comparing this SSBN with the MFrags, we see that its random variables are instances of the random variables from the MFrags, obtained by substituting problem-specific entity instances for the ordinary variables of the MFrags. We are currently extending the SSBN construction algorithm in UnBBayes-MEBN to be capable of constructing the SSBN for this problem.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
160
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
This problem requires bringing knowledge to bear about events in space and time, how agents use objects such cars, social interactions among agents, and other sophisticated kinds of reasoning. Many of these reasoning patterns are reusable across a wide variety of problems. Examples include the knowledge that individuals may meet with each other to coordinate joint activities, and that they use cars for transportation. In an operational system, these kinds of reasoning would make use of existing ontologies. PR-OWL allows the user of such an ontology to add probabilistic information to represent relationships that fall short of certainty. To conclude our case study, after using PR-OWL and Bayesian reasoning to explore the implications of the evidence, IA2 appreciates the significance of the combined Multi-INT data, and issues an Actionable Intelligence Report to interdict the possible terrorist attack.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4. Conclusion Exactly how ontologies should work with probabilities is still an open research issue. The Intelligence Analysis of the knowledge sharing use case presented in this work has shown how probabilistic ontologies can be used to address that issue. UnBBayesMEBN, which was used to support the use case, is still in beta phase and should see various improvements in the near future. This system is a promising environment for building probabilistic ontologies to support knowledge sharing in open world environments. Ontologies provide the “semantic glue” to enable knowledge sharing among diverse systems cooperating in data rich domains such as Intelligence Analysis, but fail to provide adequate support for uncertainty, an ubiquitous characteristic of open world environments. Effective multi-INT fusion requires uncertainty management to be effective, and recent advances in research on probabilistic ontologies have the potential to integrate uncertainty management smoothly with semantic technology. The case study presented in this work has shown that such research, albeit in its infancy, can help to support interoperability among Intelligence systems in an open environment, addressing issues of fusing multiple sources of noisy information into a coherent overall situation picture.
Acknowledgements Kathryn Laskey gratefully acknowledges partial support from Lockheed Martin Corporation for the research reported in this paper. The authors thank Rommel Carvalho for assistance in using UnBBayes-MEBN.
References [1] [2]
Costa, P. C. G. 2005. Bayesian Semantics for the Semantic Web. PhD Diss. Department of Systems Engineering and Operations Research, George Mason University. 315p, July 2005, Fairfax, VA, USA. Laskey, K.B., MEBN: A Language for First-Order Bayesian Knowledge Bases, Artificial Intelligence, 172(2-3), 2007.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
K.B. Laskey et al. / Probabilistic Ontologies for Multi-INT Fusion
[3]
[4]
[5] [6] [7] [8]
Carvalho, R. N., Santos, L. L., Ladeira, M., and Costa, P. C. G. 2007. A GUI Tool for Plausible Reasoning in the Semantic Web using MEBN. In Proceedings of the Seventh International Conference on Intelligent Systems Design and Applications, 381-386. IEEE Press. Costa, P. C. G.; Ladeira, M.; Carvalho, R. N.; Laskey, K. B.; Santos, L. L.; and Matsumoto, S. 2008. A First-Order Bayesian Tool for Probabilistic Ontologies. In Proceedings of the 21st Florida FLAIRS UNBBayes-MEBN Pearl, J. 1988. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. San Mateo, CA, USA: Morgan Kaufmann Publishers. Schum, D.A. 1994. Evidential Foundations of Probabilistic Reasoning, John Wiley & Sons, Inc., New York, NY. Kadane, J. B.; and Schum, D. A. 1996. A Probabilistic Analysis of the Sacco and Vanzetti Evidence. New York, NY, USA: John Wiley & Sons. Shafer, G. 1988. Combining AI and OR. University of Kansas School of Business, Working Paper No. 195. DaConta, M. C.; Obrst, L. O.; and Smith, K. T. 2003. The Sematic Web: A Guide to the Future of Xml, Web Services, and Knowledge Management. Indianapolis, IN, USA: Wiley Publishing, Inc.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[9]
161
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-163
163
Chapter 10
Design Principles for Ontological Support of Bayesian Evidence Management Michael N. HUHNS, Marco G. VALTORTA, and Jingsong WANG University of South Carolina, Columbia, SC, USA
Abstract. This chapter describes work on an integrated system that can assist analysts in exploring hypotheses using Bayesian analysis of evidence from a variety of sources. The hypothesis exploration is aided by an ontology that represents domain knowledge, events, and causality for Bayesian reasoning, as well as models of information sources for evidential reasoning. We are validating the approach via a tool, Magellan, that uses both Bayesian models and logical models for an analyst’s prior knowledge about how evidence can be used to evaluate hypotheses. The ontology makes it possible and practical for complex situations of interest to be modeled and then analyzed formally. Keywords. causal ontology, Bayesian reasoning, evidence management
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction Much of the extensive work on ontologies to date has focused on modeling and representing the world of objects. The ontologies needed for our research supporting the management of hypotheses and evidence for analysts, however, must additionally model events and causality. Less work has been done on this aspect of ontologies. In this paper we show how concepts from a causal ontology can be used directly as variables in Bayesian networks and how the attributes of the causal concepts can be used in matching evidence to the variables. Moreover, subclass relationships in the ontology enable the extension of Bayesian reasoning over types.
1. Bayesian Reasoning for Evidence Management There are numerous real-world situations about which an analyst might wish to hypothesize and investigate, but it would be impractical to encode all of them explicitly in a support system for analysts. Instead, our approach is to represent fragments of situations and provide a mechanism for combining them into a wide variety of more complete ones [1,2]. The combination occurs dynamically as evidence about a situation becomes available or as an analyst revises or enters new hypotheses. A situation fragment is represented as a Bayesian network with nodes for hypotheses, events, and evidence, and links for relating them. Our ability to combine the fragments into more complete situation models is dependent on having a consistent terminology
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
164 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
in which the fragments are described. The focus of our work has been on (1) defining and representing the terminology, including terms of a domain and terms for evidence in that domain, (2) capturing new fragments from a variety of sources, and (3) incorporating the terminology and BN fragments into an integrated end-to-end tool, Magellan. 1.1. Recognizing and Representing Situations
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Our objective is to be able to model and reason probabilistically about a wide variety of situations that might be of interest to analysts. Unfortunately, there are too many situations for system developers to encode a priori, which even if possible would make the resultant system too complex for analysts to use, and it is unrealistic to expect most analysts to be able to use the requisite formal mechanism to encode situations a posteriori. Instead, our approach is to represent small, common aspects of situations generically, and then provide a means to combine them dynamically into representations for real-world situations. We term the small generic situation aspect a fragment, and choose a first-order representation for it. An example situation aspect that we might represent as a fragment would be a “suspicious transfer of money,” with variables corresponding to banks, organizations, deposits, withdrawals, and the transferring agent. The fragment would be instantiated when evidence matched the variables, e.g., “a church attended by Syrians in Detroit deposited funds into a Michigan bank and the funds were transferred to a bank in Cairo.” More precisely, each variable (node) in a fragment has a set of identifying attributes and their collective instantiated values specify a particular instance of a random variable. Because the evidence might be uncertain, there would be probabilities associated with the instantiated fragment, and we would treat the instantiated fragment as a Bayesian network. This is shown in Figure 1. Note that the probability distribution described in the Bayesian network is a joint distribution on the nodes only, not on the nodes and the attributes.
Figure 1. A commonly occurring part of a situation for a suspicious bank transfer of money, represented as an uninstantiated Bayesian network. Notice that the nodes (variables) have attributes, making them equivalent to concepts or classes in an ontology. If one or more items of evidence matched the nodes, then details of the evidence would be used to instantiate the attributes of the variables.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 165
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
An advantage of using fragments of situations instead of more complete situations is that many more situations can be represented efficiently. More precisely, N fragments can potentially be combined in N! ways to represent N! situations. The combining is guided by available evidence. For example, three other situations that we might represent as fragments are “purchases of weapons,” “influencing an election,” and “bribing a politician.” If evidence matched one of these, and the resulting instantiated fragment had one or more variables in common with the money fragment, then we would merge the fragments at the point of the common variables to produce a representation of a more complete situation, such as “transferring money to influence an election.” Note that fragments can be merged only if the attributes of their common variables unify. Also note that it is not necessary for the fragments to have any variables in common in order to merge them and represent larger situations. As a result, the fragments could represent situations such as “bribing a politician to influence an election” and “purchasing weapons to influence an election.” Further, because each fragment could be instantiated multiple times, we could represent several different money transfers being used to purchase weapons. Our system, Magellan, considers all of the possible situations that are consistent with available evidence. Magellan then performs Bayesian reasoning on whichever complex situation representation resulted from instantiating fragments with the available evidence and integrating those fragments. The overall process for merging instantiated fragments and reasoning over them is shown in Figure 2.
Figure 2. Fragments (left side)—in this case about people being on a train, having made reservations on a train, and being at the same location as the train—are merged based on the evidence (center) that instantiates them.
1.2. Capturing the Terminology and Prior Knowledge for a New Domain A key activity of an intelligence analyst is to distinguish among competing hypotheses, determine the likelihood of their occurrence, and reduce the uncertainty in the outcomes of the hypotheses, upon which decision makers will then base their decisions.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
166 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
Hypothesis outcomes1 are related to observable evidence via direct or indirect causal relations, and therefore ontological support for analysts should involve cause-and-effect. This is best supported by an ontology emphasizing events and their causal relationships, along with a hypothetical world of possible events, actions, and causes. However, causal relationships must be interpreted in the context of the state of the real world— primarily consisting of objects and their physical properties—which can be represented in a conventional ontology, such as those that are part of SUMO. The evidence for reasoning about hypotheses can come from a variety of sources, and the acquisition of evidence and events from these sources must also be represented, constituting a third kind of ontological representation describing the information sources. Figure 3 depicts the three ontological models we use for (1) modeling situations and relating them to (2) background knowledge about the state of the world, and (3) acquiring evidence, all of which enables an assessment of the likelihood of the situations using Bayesian reasoning.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 3. An ontology for intelligence analysts has three related parts, corresponding to (1) the world of causality and hypothetical events needed for Bayesian reasoning, (2) the real world of things needed to model situations, and (3) the world of information and information sources needed for evidence management.
A situation might represent an analyst’s query or, more generally, provide context and support for a hypothesis. A situation would be comprised of one or more items of interest and each such item of interest has information provided by several information sources. An item of interest may be specialized to Person, Organization, Event, or Place, and of particular interest would be items relating events involving people at significant places. Information sources can be maps, images, reports video, audio, email, websites, and database records. Typically, an item of interest would have many information sources describing aspects of that item, for example a meeting held by members of a suspected terrorist organization might be described by audio, video, and email surveillance or reports by insiders. Our tool, Magellan, uses Protégé [3] (see Figure 7) for capturing the ontologies, RDF (Resource Description Framework [4]) for representing the terminology, XMLBIF (eXtensible Markup Language Bayesian Interchange Format [5]) for representing the causal relationships, and RDF and SPARQL (a query language for RDF [6]) for requesting evidence from information sources. It also makes use of logical, non-probabilistic models, as shown in Figure 4 and described next. 1 In our ontology, an outcome is thus an important and necessary property (“slot” in Protégé) for hypotheses and, indeed, for any concept that may be in a causal relationship. The relationship is a link in a Bayesian network.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 167
1.3. Situation Fragments Represented by Logical Models Our objective is to produce models of systems and situations that will be sufficiently accurate that they can be used—where appropriate—to predict future states, to understand operations, to illuminate the factors relevant to decisions, and to control behaviors. We have realized that some knowledge is more easily and naturally represented in the form of statements in a logic language and some is more naturally represented in a Bayesian-network formalism. For example, logic is best for expressing: • • • • •
Class-subclass statements, such as “C4 is an explosive” Part-whole statements, such as “triggers are part of IEDs” Definitional statements, such as “triangles have three sides” Temporal statements, such as “3:00 p.m. occurs before 4:00 p.m.” Spatial statements, such as “Irbil is located in Kurdistan”
Other knowledge is probabilistic, such as: • •
“Terrorist cell X planned the bombing” “Suspect Y met with cell leader Z in Syria last March”
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Our resultant reasoner takes advantage of the strengths of each formalism, while integrating them into a single coherent system.
Figure 4. The BALER framework for integrating logical models with probabilistic models, with an ontology developed in Protégé providing a consistent vocabulary for all domain concepts
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
168 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
An example of the situations that can be represented by such an integrated system is shown in Figure 5. This system would help analysts confront problems of credibility, relevance, contradictory evidence, and pervasive uncertainty, using: • • • • • •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
A unique combination of the power of logical and probabilistic reasoning Numerical analysis of competing hypotheses Automated linking of relevant evidence Automated propagation of uncertainty values: good arguments from uncertain data still add strength to a conclusion Robust reasoning over contradictory information allows analysts to exploit maximal amounts of information A provision for analysts to enter their own knowledge directly, allowing the system to learn from its users The use of probabilities to quantify belief in hypotheses to support optimal decision making according to the principle of maximum expected utility.
Figure 5. An example illustrating the need for both Bayesian and logical reasoning
Formal logical tools are able to provide some amount of reasoning support for information analysis, but are unable to represent uncertainty. Bayesian network tools represent probabilistic and causal information, but in the worst case they scale as Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 169
poorly as some formal logical systems and require specialized expertise to use effectively [7]. The framework (BALER) we have developed for intelligence reasoning incorporates the advantages of both Bayesian and logical systems [8]. The framework includes a formal mechanism for the conversion of automatically generated natural deduction proof trees into Bayesian networks. This is indicated by the information flow shown in Figure 6. We have proven that the merging of such networks with domainspecific causal models forms a consistent Bayesian network with correct values for the formulas derived in the proof. In particular, we show that hard evidential updates (see Section 1.5) in which the premises of a proof are found to be true force the conclusions of the proof to be true with probability one, regardless of any dependencies and prior probability values assumed for the causal model. Information Logic Knowledge
Finding Knowledge
Probabilistic Knowledge
Logic Rules
Findings
Probabilistic Model Instantiated BNs
Scenario Filter Logic Goals, Rules SILK Theorem Prover
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Proofs Convertor
Proof BNs
Composer
Composite Model Figure 6. The BALER software process flow, which is supported by the tripartite ontology of real world concepts, events, and information sources
1.4. Causality Causality is a special relationship among events for which certain properties hold probabilistically. For example, causality is logically irreflexive and asymmetric, but probabilistically transitive. Causality, like the relation subevents, generates a strict partial order among events. Causal models are very useful, because they allow prediction of the effect of interventions [9,10]. Our interest is in a causal Bayesian network.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
170 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 7. Protégé is used to enter the ontology concepts that form the basis for representing situations and evidence.
A causal Bayesian network consists of a causal graph, a directed acyclic graph (DAG) expressing causal relationships, and a probability distribution respecting the independence relation encoded by the graph [8]. A link between two nodes in a Bayesian network is often interpreted as a causal link. However, this is not necessarily the case. When each link in a Bayesian network is causal, then the Bayesian network is called a causal Bayesian network or Markovian model. A Markovian model is a popular graphical model for encoding distributional and causal relationships. To summarize, a Markovian model consists of a DAG G over a set of variables V = {V1; . . . ; Vn}, called a causal graph and a probability distribution over V that has some constraints on it. The interpretation of such a model consists of two parts: the association of the variables to events and the assignment of probability distributions to the links. For causality, variable assignment must satisfy the obvious constraint that (Event A causes Event B) (timeA < timeB) The probability distributions must satisfy two constraints. The first constraint is that each variable in the graph is independent of all its non-descendants given its direct parents. The second constraint is that the directed edges in G represent causal influences between the corresponding variables. A Markovian model for which only the first constraint holds is called a Bayesian network, and its DAG is called a Bayesian
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 171
network structure. This explains why Markovian models are also called causal Bayesian networks. As far as the second condition is concerned, causality requires that, when a variable is set, the parents of that variable be disconnected from it: this is called the excision model of causality. In our prototype tool, Magellan, new variables are added to the causal and event portion of an analyst’s ontology using Protégé, so that all of the nodes in a Bayesian network fragment are represented in a standard and consistent terminology. We extend SUMO with this terminology, so that we can take advantage of SUMO’s existing description of general knowledge of the world. Each variable has a set of identifying attributes, which are used to combine fragments (fragments can be combined only if their attributes unify) [1,2]. Probabilities are assigned to events in the fragment by performing experiments, estimating beliefs, or counting outcomes. Once assigned, they are updated by conditioning on evidence using Bayes rule and the laws of probability. The fragments are stored in a repository, where they can be matched with evidence and combined with other fragments to produce models of situations that are as complete, accurate, and specific as possible. 1.5. Evidence Fragments are instantiated by evidence, which we define informally as information (perhaps wrong, perhaps incomplete) about what happened (events). For example, a bank clerk might be uncertain whether a money transfer was to a Cairo bank or a Boston bank. We represent in the information source ontology the level of credibility of items of evidence, and provide a Bayesian interpretation of credibility. Formally, we define evidence to be a collection of findings, each of which describes the state of a Bayesian network variable, and distinguish three kinds [7]:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
11. A hard finding specifies that the variable has a particular value. For example, whether or not a money transfer occurred or whether or not a suspect is a terrorist (Male_TerroristSuspect = true) 12. A soft finding is a distribution on the states of a variable, usually corresponding to an “objective” statistical distribution that is not expected to change within a scenario [11]. For example, there might be an observation that 95% of terrorists are male (and 5% are not), i.e., Q(Male_TerroristSuspect)=(0.95, 0.05) 13. A virtual finding is a likelihood ratio corresponding to the credibility associated to an evidence source, such as a witness. For example, witness Bill might have observed a suspect entering a men's-only area of a mosque, which would be interpreted as 4-to-1 that the suspect is a male L(Male_TerroristSuspect)=(0.8, 0.2) Unlike soft findings, virtual findings allow for an update of the posterior probability of the evidence variable. Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
172 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
The relationships among the evidence types are shown in Figure 8.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 8. Evidence consists of a set of findings, which can be of three different types, hard, soft, and virtual.
Figure 9. Magellan’s extended ACH interface is integrated with the ontology of events through pull-down menus, i.e., each hypothesis (such as “TerroristGroupAttack”) and each type of relevant evidence (such as “DetectedChemical”) is a concept from the domain ontology.
Our modified version of the tool ACH2 [12] is used by an analyst to enter the appropriate hypotheses and any initial evidence that might be available. The terminology available to the analyst is provided via drop-down menus as shown in Figure 9, where the menu entries are the ontology terms from our ontology developed
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 173
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
in Protégé [3]. The use of terms from an ontology is essential for (1) enabling logical proofs to be constructed out of both new knowledge and prior knowledge, (2) taking advantage of known generalizations and specializations for reasoning and fragment matching, (3) guiding analysts in the kinds of concepts that can be used to represent hypotheses and evidence, and (4) enabling new fragments to be composed with existing fragments to represent situations more comprehensively. The resultant Analysis of Competing Hypotheses (ACH) [13] matrix is converted automatically into a bipartite Bayesian network, with initial probabilities assigned based on the relevance factors assigned to cells of the matrix. An example of the network is shown in Figure 10. The network is saved into a repository of fragments, from where it can be retrieved for matching to evidence and then composed with other fragments.
Figure 10. A Bayesian network fragment constructed automatically from an ACH matrix. The conditional probabilities needed for Bayesian reasoning are derived from the user-entered values in the matrix indicating whether or not a finding is consistent with an analyst’s hypothesis.
2. Use of Tripartite Ontology for Intelligence Analysis Figure 11 shows an end-to-end architecture for Bayesian reasoning, which would be used as follows. The process might be triggered by the arrival of evidence in the form of a message, such as the following: FBI Report Date: 10 April 2003. FBI: Abdul Ramazi is the owner of the Select Gourmet Foods shop in Springfield Mall. Springfield, VA. (Phone number 703-659-2317). First Union National Bank lists Select Gourmet Foods
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
174 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
as holding account number 1070173749003. Six checks totaling $35,000 have been deposited in this account in the past four months and are recorded as having been drawn on accounts at the Pyramid Bank of Cairo, Egypt and the Central Bank of Dubai, United Arab Emirates. Both of these banks have just been listed as possible conduits in money laundering schemes.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 11. Magellan architecture for Bayesian Reasoning used to explore an analyst’s hypotheses, indicating how the ontology makes it possible for evidence to be combined with generic situation fragments to produce models that can be reasoned over probabilistically to explain the evidence.
Based on such a message, or based on a hypothesized situation that an analyst would like to investigate, an appropriate scenario represented as a Bayesian model is chosen by the analyst and a corresponding form is displayed listing initial evidence and the domain variables for the scenario. The evidence values for the variables can be supplied automatically from the triggering messages, by matching message terms with ontology concepts as shown in Figure 12, or can be entered by the analyst. Because the probabilities of the variables represented in a situation are updated to be consistent with the evidence at hand, the situation tracks the variables of interest to an analyst. When the probability of a particular value of a variable of interest becomes sufficiently high, an alert could be issued to the analyst. Then the Bayesian reasoning component, using a value-of-information calculation, identifies the variables that have the most potential impact on the probability profile of a variable of interest. (Algorithm 1 contains the algorithm that we use to calculate the value of information for a chosen variable.) That is, it determines which pieces of evidence would be most useful in confirming or denying the analyst’s hypothesis. Such especially informative variables can then become the subject of focused queries. A request for this evidence is sent to the analyst, who returns the result to the Bayesian reasoner for incorporation into the situation, and the likelihood of the analyst’s hypothesis is reassessed. The process is repeated until the analyst decides to stop or there is no more evidence available that changes the plausible outcomes.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 175
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Figure 12. A small portion of the tripartite ontology indicating how an item of evidence would be classified and used to instantiate one or more fragments. Shown here is the outcome of representing the FBI evidence message above using the ontology that we defined in Protégé.
3. Evaluation An early anecdotal evaluation of Magellan was conducted at NIST. The evaluators (three naval reservists with a background in intelligence analysis) tested the hypothesis generation aspect of the system for four hours. In this test, the analysts were presented with several items of evidence (similar to the FBI Report in of section 3) and asked to generate hypotheses, using an interface such as is shown in Figure 13. After they had finished, they were shown hypotheses generated by Magellan and were asked to rate these hypotheses in comparison to the ones they had generated. The NIST summary of the evaluation indicated that the analysts generated more hypotheses than Magellan and that Magellan’s hypotheses did not take into account all the possible variables. However, analysts’ ratings for Magellan-generated hypotheses are equal to the ratings for the analyst-generated hypotheses in 1/3 of the cases. In 7/9 cases the ratings for the Magellan-generated hypotheses were given mid-level ratings or higher.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
176 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
Algorithm 1. Value-of-Information Calculation • • •
Let V be a variable whose value affects the actions to be taken by an analyst. For example, V indicates whether a bomb is placed on a particular airliner. Let p(v) be the probability that variable V has value v. The entropy of V is:
H(V ) = − ∑ p(V = v) log( p(V = v)) v∈V
• •
Let T be a variable whose value we may acquire (by expending resources). For example, T indicates whether a passenger is a known terrorist. The entropy of V given that T has value t is:
H(V | t ) = − ∑ p(V = v | T = t ) log( p(V = v | T = t )) v∈V
•
The expected entropy of V given T is:
E[H(V | t )] = ∑ p(T = t )H(V | t )) t∈T
•
The value of information is then:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
VOI (V ) = −(E[H(V | t )] − H(V ))
Figure 13. The Magellan interface showing an evidence message (upper right), the ontology concepts it contains (lower right), the fragments that it instantiates, composed into a situation (lower left), and the posterior probability for an hypothesis about the situation
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management 177
4. Discussion The key features of our approach to reasoning about evidence are the ability to model fragments of abstract situations, to base the models on concepts from a causal ontology, to use a combination of both logic and probability for reasoning about the models, to ground situation models by instantiating the ontological concepts in the fragments with evidence, to compose the instantiations of situation fragments into complete situation models based on evidence, and to analyze the resultant situation models for sensitivity and surprise. The heart of our approach is Bayesian reasoning. However, there are alternative approaches for reasoning over uncertain evidence about ontological concepts, notably Pronto [14,15], a probabilistic extension to OWL [16], and P-Classic [17]. Pronto provides reasoning services for knowledge bases containing uncertain knowledge. It extends the Pellet reasoner by enabling probabilistic knowledge representation and reasoning in OWL ontologies. Pronto represents uncertainty by probability intervals, instead of point probabilities and tables of conditional probabilities, as in Bayesian networks. The advantages of the Bayesian approach are: • •
•
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
Bayesian networks directly support causality, which to do the equivalent in Pronto would require an additional logical theory. Both approaches can handle logical conflicts, but Pronto relies on a mechanism of model ordering via the use of preferences, whereas Bayesian networks make use of explicit models that describe the conflicts, so that they can be reasoned about in the same way as non-conflicting evidence. As evidence about an uncertain variable accumulates, the variable’s probabilistic interval becomes wider and it becomes more difficult to base a decision on the variable. Probabilistic interval updating as done in Pronto is more complex than the updating of point probabilities in Bayesian reasoning.
In P-Classic, which supports conditional probabilities as in Bayesian networks, links represent subclasses, as opposed to representing causality. P-Classic is most useful for problems of identification, i.e., given some uncertain features about an unknown concept x, it can conclude that x is most likely an instance of class Y. The work on probabilistic extensions to OWL by Ding and Peng [16] improves on PClassic by focusing on formal rules for translating OWL ontologies into Bayesian networks. Note that our ontology is not itself probabilistic and we do not translate it into a Bayesian network—we just use concepts from it in Bayesian networks and ensure that the Bayesian networks are consistent with the causality knowledge in the ontology.
5. Conclusion Our work is predicated on the observation that ontologies make it easier for tools to interoperate. We have found that our ontologies need to describe both the physical world and the on-line information world, because our reasoning system relies on the relationships and links between both kinds of domains. The reasoner, BALER, enables
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
178 M.N. Huhns et al. / Design Principles for Ontological Support of Bayesian Evidence Management
first-order logic sentences to be combined with Bayesian networks by generating Bayesian networks for any first-order natural deduction proof (that uses the ReevesClarke inference rules). This exploits the complementary powers of both logical and Bayesian representations.
6. Acknowledgements This work was funded in part by the Disruptive Technology Office Collaboration and Analyst System Effectiveness (CASE) Program, contract FA8750-06-C-0194 issued by Air Force Research Laboratory (AFRL). The views and conclusions are those of the authors, not of the US Government or its agencies.
References [1]
[2]
[3] [4] [5] [6] [7] [8]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[9] [10] [11] [12] [13] [14] [15] [16] [17]
John Cheng, Ray Emami, Larry Kerschberg, Eugene Santos, Jr., Qunhua Zhao, Hien Nguyen, Hua Wang, Michael Huhns, Marco Valtorta, Jiangbo Dang, Hrishikesh Goradia, Jingshan Huang, and Sharon Xi, “OmniSeer: A Cognitive Framework for User Modeling, Reuse of Prior and Tacit Knowledge, and Collaborative Knowledge Services,” Proceedings of the 38th Hawaii International Conference on System Sciences, HICSS38, 2005. Katherine Laskey and Suzanne Mahoney, “Network Fragments: Representing Knowledge for Constructing Probabilistic Models,” in Proceeding of the Thirteenth Conference on Uncertainty in Artificial Intelligence, AAAI Press, 1997, 334-341. The Protégé Ontology Editor and Knowledge Acquisition System, http://protege.stanford.edu/. Resource Description Framework (RDF), http://www.w3.org/RDF/. Fabio G. Cozman, “The Interchange Format for Bayesian Networks,” 1998. http://www.cs.cmu.edu/~fgcozman/Research/InterchangeFormat/ SPARQL Query Language for RDF, http://www.w3.org/TR/rdf-sparql-query/. Marco Valtorta and Yimin Huang, “Identifiability in Causal Bayesian Networks: A Gentle Introduction,” Cybernetics and Systems 39:4 (2008) 425-442. Marco Valtorta, John Byrnes, and Michael Huhns, “Logical and Probabilistic Reasoning to Support Information Analysis in Uncertain Domains,” Proceedings of the Third Workshop on Combining Probabilty and Logic (Progic-07), Canterbury, England, September, 2007, 5-7 Yimin Huang and Marco Valtorta, “Pearl’s Calculus of Intervention is Complete,” Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence (UAI-06), 2006, 217-224. Judea Pearl, Causality: Modeling, Reasoning, and Inference, Cambridge, England: Cambridge University Press, 2000. Marco G. Valtorta, Y.-G. Kim, and Jirka Vomlel, “Soft Evidential Update for Multiagent Systems,” International Journal of Approximate Reasoning 29:1 (2002) 71-106. Peter Pirolli and Lance Good, “Evaluation of a Computer Support Tool for Analysis of Competing Hypotheses,” UIR Technical Report, Palo Alto Research Center, 2004. Richards J. Heuer, Jr., Psychology of Intelligence Analysis, Center for the Study of Intelligence (at http://www.cia.gov/csi/books/19104/index.html), 1999. Thomas Lukasiewicz, “Probabilistic Description Logics for the Semantic Web,” INFSYS Research Report 1843-06-05, Institut Für Informationssysteme, Technische Universität Wien, March 2007. Pronto--A Probabilistic Reasoner for OWL DL and Pellet, http://pellet.owldl.com/pronto. Zhongli Ding and Yun Peng, “A Probabilistic Extension to Ontology Language OWL,” Proc. 37th Hawaii International Conference on System Sciences, HICSS37, 2004. Daphne Koller, A. Levy, and Avri Pfeffer, “P-CLASSIC: A Tractable Probabilistic Description Logic,” in Proc. AAAI-97, AAAI Press, 1997, 390-397.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-179
179
Chapter 11
Geospatial Ontology Trade Study James RESSLERa, Mike DEANb, Dave KOLASb a Northrop Grumman Corporation b BBN Technologies
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Abstract. It has been estimated that up to 80% of all information contains some notion of location. This is helping create a greater understanding of the utility of geospatial information as a framework for organizing, portraying and better understanding other information and the relationships of people, places, things and events. Geospatial capabilities are entering the mainstream of information technology and spatial data infrastructures (SDI's) are being implemented to bring together the technologies, policies, standards, and human resources to better utilize geospatial data. SDI's such as the National System for Geospatial-Intelligence (NSG) are using a standards baseline of ISO, Open Geospatial Consortium and other relevant consensus standards and putting service oriented architectures in place to achieve distributed, data-centric, net-centric operations. This stage of development of SDI's is bringing an unprecedented level of interoperability to geospatial data and technology and is setting the stage for an even greater level of future interoperability and data integration. The development of geospatial ontologies and semantic capabilities for integrating well structured geospatial data with unstructured geospatial information existing in other data sets will be the catalyst for this next major step forward. The National Geospatial-Intelligence Agency is in the forefront of examining the current state, future potential and implementation requirements for a semantically enabled geospatial web. This Geospatial Ontology Trade Study is a broad survey of ontologies. An ontology is a formal, explicit, shared conceptualization of a domain and defines the concepts and vocabulary used within a community of interest. The study report outlines the characteristics of the ontologies surveyed and makes recommendations about which are best suited for certain types of uses and identifies further research and work to formalize geospatial ontologies. The report concludes that there are a number of existing standards-based ontologies which provide building blocks for geospatial representations and makes recommendations for strategic actions to incorporate ontologies and semantic knowledge into the growing base of Geospatial Intelligence capabilities.
Executive Summary This trade study is a broad survey of 45 ontologies that apply to the spatial, event, and temporal granularity concepts. The study has reviewed thousands of classes and properties in order to find the characteristics that make these ontologies best suited to the geospatial intelligence community for uses in annotation, qualitative reasoning and interoperability. The study has made several strategic recommendations to further the development of semantic technology and the application of these ontologies. The principal recommendations are:
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
180
J. Ressler et al. / Geospatial Ontology Trade Study
• • • • •
There are a number of acceptable OWL ontologies and related representations available that can be re-used and extended for a domain. Update, formalize and control ontologies in a best practice document aligned with existing standards. Create an ontology library for the ISO Geographic Information technical committee specifications. Use the National System for Geospatial Intelligence (NSG) Application Schema as the basis of a standard feature type ontology. Representatives from geospatial organizations should participate in the ISO Technical Committee 211 project 19150 to promote spatial upper ontologies.
More specific recommendations are presented in Section 7. The study also recommends the following guidelines for reusing geospatial ontologies: • • •
Use OWL for ontology definition. Use the simplest OWL representations that meet application needs. Geospatial Ontologies should be based upon standards consistent with the NSG Architecture and the GEOINT Standards listed in the Defense Information Standards Registry (DISR).
The Spatial Ontology Community of Practice (SOCoP) of the US Federal CIO Council provides a good forum for exposing and coordinating geospatial ontologies. As organizations employ semantic technology to geospatial information, interoperability should be considered from the outset because semantic queries are not inherently interoperable when performed across domains.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
1. Introduction As part of the Geospatial Semantic Web (GSW) project sponsored by the National Geospatial-Intelligence Agency’s InnoVision Basic and Applied Research Directorate (NGA/IB), Northrop Grumman TASC and BBN Technologies were tasked to conduct this trade study of available ontologies in areas of likely interest to NGA and its customers. This included geospatial, event, and granularity ontologies. This paper contains the results of the trade study. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Geospatial-Intelligence Agency. Gruber [25] defines an ontology as “an explicit specification of a conceptualization” that “defines the vocabulary with which queries and assertions are exchanged among agents.” This paper uses an ontology as a formal, explicit, shared conceptualization of a domain. As such, it defines the concepts and vocabulary used within a community of interest. The formal, machine-understandable representation allows use of the ontology to support logical inference. Ontologies can be viewed as an extension of other means of expressing vocabularies and data models, including thesauri, schemas, class hierarchies, entity-relationship-attribute models, and UML diagrams. A profile of UML has been developed [1] for representing ontologies and instances in UML, and is discussed further in Section 3.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
181
Ontologies can be codified in a variety of forms, including First Order Logic and the W3C OWL Web Ontology Language, as discussed in Section 3. This study investigated many ontologies and data models relevant to the domains of interest, from multiple sources and employing multiple representations. More detailed analysis was performed on those ontologies expressed in OWL. The report is structured as follows: • • • • • • • •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
Section 2 describes three use cases that set the context for the study. Section 3 is an overview of languages used to represent knowledge. Section 4 contains an overview of all the ontologies studied. Section 5 describes the characteristics of foundational ontologies. Section 6 describes in detail the characteristics of the spatial ontologies. Section 7 describes the characteristics of the event ontologies. Granularity is addressed in the context of spatial and temporal ontologies, rather than independently. Section 8 is a set of recommendations for adopting and advancing the use of spatial and event ontologies by NGA. Appendix A is a table summarizing the source, format, content, categories, use, and metrics of each ontology studied. Appendix B is a hierarchical listing of the concepts contained in those ontologies that are available for public release.
The appendices are contained in a referenced report and intended to be used as a reference for readers to search, discover, and apply the ontologies to their particular domains and problems. To use a domain ontology to query over data described in another ontology, classes and properties need to be translated from the data source ontology to the domain ontology. Another task in the GSW project developed Snoggle, a graphical tool to assist software developers in constructing Semantic Web Rule Language (SWRL [33]) rules to map data from one ontology to another. Snoggle was developed as open source and is available from http://projects.semwebcentral.org/projects/snoggle. The design and use of Snoggle is addressed in Ressler, 2007 [2] and other documentation on the Snoggle project site.
2. Use Cases To focus discussion and applicability, the study considered three primary use cases motivating the development and selection of geospatial and related ontologies. Use cases are used here to describe a nominal application of ontologies, without the formal UML definition of actors, process flow, input and output conditions. That is, geospatial and related ontologies were judged according to their support of three uses – for annotation of geospatial information, modeling qualitative relationships and as a means to integrate and relate alternative conceptual views of geospatial information. The spatial and temporal ontology use cases are described in the subsequent sections.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
182
J. Ressler et al. / Geospatial Ontology Trade Study
2.1. Annotation The term annotation is used to describe the process of recording geospatial information in a geospatial context. An ontology provides the classes and/or properties used to express spatial information, such as locations and regions, associated with other objects, such as buildings, automobiles, and people. Issues investigated include richness of representation (e.g. points, lines, polygons), support for multiple coordinate systems (latitude/longitude, UTM, alternative datums, etc.), perceived precision (e.g. 47 degrees 31 minutes looks much more precise when converted to a decimal representation), and ease of use. 2.2. Qualitative Reasoning Significant effort in developing spatial and temporal ontologies has gone into modeling qualitative relationships such as containedWithin, connectedTo, or occursBefore and their associated axioms. These are typically based on formal models such as the Region Connection Calculus (RCC8), 9 Intersections, or Allen relations. The utility of such qualitative reasoning, in isolation or as a filter for quantitative reasoning, is still being debated. The study investigated properties and axioms supporting qualitative reasoning and any underlying theory.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2.3. Information Integration In addition to support for annotation, ontologies provide conceptual modeling capabilities beyond those available from other methodologies such as entity/relationship/attribute models or UML. In particular, ontologies take a broader perspective to modeling the world rather than just expressing constraints on data. They also allow concepts to be defined in terms of other more “primitive” or foundational concepts (e.g. as in “a father is a man with one or more children”). Because of this richness, ontologies (particularly large foundational and/or upper ontologies) are often seen as a means to integrate and relate alternative conceptual views. A sufficiently overarching ontology could be used as the hub in a translation network, replacing mappings between each ontology pair, an O(n2) problem, with mappings between each ontology and the hub, an O(n) problem. The study looked for breadth, completeness, and generality in identifying ontologies that are candidates for information integration. A preference was given to ontologies that are complete and (derived) from authoritative sources.
3. Knowledge Representation Languages The study focused primarily on ontologies that are expressed in or can be translated to the OWL Web Ontology Language [3]. The OWL expressions are in themselves considered an ontology within the context of the study. OWL is a W3C Recommendation (industry standard recognized in the Defense Information Standards Registry, with the same status as HTML and XML) developed for the Semantic Web, an international activity that seeks to expose World Wide Web content originally developed for human consumption in a form more amenable to processing by
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
J. Ressler et al. / Geospatial Ontology Trade Study
183
intelligent agents and other computer programs. As a knowledge representation language for such a broad audience, OWL represents a compromise among expressivity, computational complexity, scalability, and ease of use. OWL is heavily influenced by Description Logic (DL), a subset of First Order Logic (FOL) with computational properties that allow for more efficient reasoning implementations. The DL research community uses abbreviations such as SHOIN(D) to indicate the expressiveness of a representation, where each letter corresponds to a specific language feature, as shown in Appendix A. For the interested reader, we include these characterizations for each ontology studied. OWL includes 3 sublanguages: OWL Lite, OWL DL, and OWL Full, with different restrictions on the features used (OWL Lite is a subset of OWL DL, which is essentially a subset of OWL Full). Many tools are targeted at a specific sublanguage or support an even more restricted OWL subset such as the underlying RDF Schema. A more expressive ontology can be used with a less capable tool, but this may cause fewer statements to be inferred (i.e. may not be complete). Several of the ontologies studied were originally expressed in a First Order Logic language such as KIF or CycL and then translated to OWL. This results in some loss of expressivity. Translation to Semantic Web Rule Language [33] (SWRL) rules in addition to OWL, as is done in the Cyc Exporter, reduces this loss. In general, we worked with both the original and translated versions of these ontologies. Unlike static database schemas and similar representations, an OWL ontology can be easily augmented with application-specific classes, properties, or restrictions. Such customization is standard practice in the Semantic Web. While notation for ontology modeling is still emerging, some researchers are recommending UML as a language to graphically represent ontological modeling [4]. Gasevic defined the XSLT transformations to build an ontology from a UML logical model in XMI format [5]. Corresponding UML and OWL features are shown in [1] Table 19, page 213. The Ontology Definition Metamodel (ODM) [1] provides a syntactic representation of ontology in UML that gives a precise representation needed for such a translation. The ODM can be applied to UML data models intended for use as an ontology to define associations and namespaces (packages in UML). Some UML and OWL language elements cannot be represented in the other language and could be problematic (such classes as instances, disjointsWith relation) as shown in [1] Table 20 and 21, pages 213-4. If these elements are required, an alternative representation of these elements needs to be developed.
4. Overview of Ontologies To show the breadth of the 45 ontologies studied in this report, a general categorization and summary of the primary uses of the ontologies is provided. Appendix A contains a detailed description of the characteristics of each ontology, including the categorization, use, metrics and other attributes of the ontology. A concise representation of many of the ontologies is provided in Appendix B. This hierarchical representation gives a quick understanding of the breadth (number of classes) and inheritance depth of each ontology. The use of an ontology varies based upon the level of abstraction and subjectmatter knowledge represented in the ontology. Various ontology classifications have been proposed. Base Spatial ontologies (as discussed in [6]) are a subset of upper
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
184
J. Ressler et al. / Geospatial Ontology Trade Study
ontologies and models that address spatial concepts and relationships upon which spatial domain ontologies are constructed. This study classified ontologies as follows: • • • •
Foundational ontology – a fundamental definition of basic concepts and methods of organization, often focused on avoiding common modeling problems and intended to be specialized Upper ontology – a high-level abstraction of real-world concepts, typically specialized Domain ontology – a specialization unique to a user or community’s field of interest. A domain ontology may be further specialized Data source ontology – while not explicitly studied here, a data source ontology reflects the underlying data model used by a particular data source or common format. For example, a data source ontology may describe feature data represented by a specific geospatial product format in XML schema or in a relational database (derived from the data definition language).
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Folksonomies are a method of tagging web content that is becoming widespread in Web 2.0 applications including blogs and social networking sites. While this study did not address the use of tagging as an ontology, it considered the relation that folksonomies have on formal ontologies On the Semantic Web, it is expected that multiple ontologies will address the same subject domain for different communities of interest (as in this study). Significant research attention has been devoted to the problem of ontology alignment, which has led to the Ontology Alignment Evaluation Initiative (OAEI) 1 [7], which focuses on automated alignment tools. The best automated tools currently identify about 30% of the matches while generating lots of false matches and typically focus on identifying equivalent classes rather than more common subclass relationships. For several of the ontologies studied, a representation of a single feature is illustrated in order to provide concrete examples that can be compared across ontologies. The example feature used is The Pentagon in Washington, D.C., which is in the shape of a five-sided polygon as illustrated in Figure 1. 4.1. Categories The W3C Geospatial Incubator Group defined seven categories of Geospatial Ontologies2: 8. 9. 10. 11. 12. 13. 14.
Geospatial Feature Ontology Feature Type Ontology Spatial Relationship Ontology Toponym (Place name) Ontology Coordinate Reference / Spatial Grid Ontology Geospatial Metadata Ontology (Geospatial) Web Services Ontology
1 The OAEI is an initiative conducted as part of the International Semantic Web Conference. See http://oaei.ontologymatching.org/2006/. 2 http://www.w3.org/2005/Incubator/geo/
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
185
Figure 1. Pentagon Feature (GoogleEarth © Google)
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In order to fully represent the range of geospatial and temporal data considered in this study, these additional categories are recommended:3 15. Geometric Ontology – basic geometry of geospatial entities: points, lines, polygons, surfaces, spheres, etc. 16. Coverage Ontology – a coverage is an intersection between a feature and temporal data, their properties, subclasses and associations are distinct. 17. Geopolitical Ontology – political borders and locations (countries, states, counties, politically determined regions (such as ZIP Code, Area Code, representative jurisdiction) which are distinct from the toponym ontology. 18. Temporal Ontology – time-based entities and events related to geospatial information (new ontology or reference to an existing temporal ontology) A count of the number of ontologies per category is summarized in Table 1. This shows that most categories are fully supported by more than one ontology. Including the Geometric category, Feature is by far the most prevalent ontology category in use at this time. As documented in Appendix A, a comparison of the ontologies mapped to the 11 categories above reveals which categories are most prevalent and which are sparsely covered by existing ontologies. Figure 2 shows the ontologies assessed along with their categories, where “F” indicates the one category to which the ontology most closely applies while “P” indicates any additional categories to which the ontology partially applies. The conclusions from this analysis are discussed in Section 8.
3
These additional categories have been submitted to the W3C Geospatial Incubator group for consideration.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
186
J. Ressler et al. / Geospatial Ontology Trade Study
Table 1. Ontology Categorization Spatial Relationship
Toponym
Coordinate
Metadata
Web Service
Geometric
Coverage
Geopolitical
Temporal
7
6
3
1
4
2
0
4
1
2
6
Ontology Count – Partially Related
6
2
6
4
2
3
2
4
4
3
4
TOTAL
1 3
8
9
5
6
5
2
8
5
5
10
P
Cyc Surface Geometry
P
P
F
Cyc Terrain Cyc Topology
P
F
P
ISO Conceptual Schema Language
P P
ISO Feature Cataloging
F
P
SOUPA geo-measurement
P
SWEET Space P
S-57 - maritime P
P
P
F
F F
P
P
F F
P
P
F
ISO Temporal
F
CycTemporalPredicates
F
F
CycTimeInterval
F
OWL-Time
RDF Calendar F
SOUPA event
F
SWEET Time
P
F F F F F
P
P P
P P
P
SOUPA time
F
P
F
P
F
SUMO Geography SUMO MILO
P P
F
SOUPA rcc SUMO
F P
F
ISO Spatial Refernence System ISO Spatial Schema
KML 2.1
P P
MINDSWAP geoRelations
F
ISO Metadata ISO Spatial Referencing by Geometric Identifier
F
MINDSWAP geoFeatures F
ISO Coverages
F
GML
MINDSWAP geoCoordinateSystems
F
P
F
geoRSS
geonames.org P
FGDC CSDMG ISO Application Schema
P
P
Temporal
F
Geopolitical
Cyc Open Geospatial Consortium
F
Coverage
Cyc Map Projection
Geometric
Basic Geo F
P P
Web Service
F
F F
Metadata
Cyc Geodesy
P
Coordinate
NAS 1.8 (NSG FC 1.7)
Toponym
Enterprise Conceptual Data Model
P
SpatialRelati onship
P
DOLCE
Feat-Type
Feature
Temporal
Geopolitical
Coverage
Geometric
Web Service
Metadata
Coordinate
Toponym
SpatialRelati onship
Feat-Type
Feature
BFO
Cyc Linear Object
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Feat-Type
Feature Ontology Count –Fully Supports
F P
P
Figure 2. Summary of Ontology Categorization
4.2. Use Case Analysis In addition to the 11 categories, each ontology was assessed for relevance to the three use cases in terms of being able to support the use case fully, partially or indirectly. The three levels are subjective terms that provide an overall suitability of the ontology for the use case. “Fully” means that the concepts in the ontology directly apply to the use case and can be applied without modification. “Partially” means that some of the concepts apply to the use case, but it is not is not the primary intent of the ontology or requires modification (either generalization or specialization) to be used in this manner.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
187
J. Ressler et al. / Geospatial Ontology Trade Study
“Indirectly” supporting the use case means that while not being directly applicable, the ontology can contribute to the use case. In general, the classification by use case shows the level of support and number of ontologies that have concepts applicable to the use case. Table 2 shows the summary of use case applicability, which is expanded for each ontology in Appendix A. This analysis shows the annotation use case is fully supported by the most ontologies, many of which also indirectly support the information integration use case. The qualitative reasoning use case has a narrow focus that restricts its support to fewer ontologies. The use case analysis is considered again in the recommended ontologies presented in Section 8. Table 2. Use Case Classification of Ontologies Use Case / Ontology Count
Fully
Partially
Indirectly
Total
Annotation Use Case Support
15
12
5
32
Qualitative Use Case Support
7
8
2
17
Integration Use Case Support
8
7
21
36
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
4.3. Comparative Analysis A more in-depth comparison of two ontologies can be performed by comparing the classes and properties of each. Several such comparisons were performed using the INRIA ontology Alignment API and its default alignment algorithms [8]. These algorithms typically compare the class and property names of two ontologies to find matches based upon the similarity of words used within these names. The Alignment API assigns a level of confidence to the match, with 1.0 being an exact match in name and less than 1.0 being similar in name or meaning. The Alignment API was run selectively on ontologies within the same categories and having some similar concepts. The comparisons were made for the geometric, coordinate reference (geodesy), metadata, temporal and spatial relationship categories. A summary of the Alignment API results for similar ontologies is shown in Table 3. Obvious mis-alignments of name or mis-understood meaning were eliminated from the count. It was revealing to note that when using the Wordnet thesaurus for comparison, INRIA failed to correctly match most entities with < 50% confidence. The most reliable alignment results with Wordnet required a confidence ≥ 60%, which was the threshold for determining the number of classes and properties in common. The results of automated alignment showed while there is a potential for time saving, matching classes and properties between ontologies still requires human reason and decision making. The Snoggle tool was developed to support human based alignment and rule generation.
5. Foundational Ontologies Foundational ontologies provide a reference point for rigorous comparisons among different possible ontological approaches, and a framework for analyzing, harmonizing,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
188
J. Ressler et al. / Geospatial Ontology Trade Study
Table 3. Automated Ontology Alignment Results
Ont 1
Ont 2 # classes Geometry
Ont 2 # properties (data & object)
# Commo n (Inria Word align > 60%)
Cyc OGC
68
47
ISO 19107
109
107
23
14%
Cyc SurfaceGeometry Cyc LinearObject
124 87
61 76
ISO 19107 ISO 19107
109 109
107 107
29 23
14% 12%
GML GML
166 166
121 121
ISO 19107 Cyc SurfaceGeometry
109 124
107 61
163 24
65% 10%
GML
166
121
Cyc LinearObject 87 76 23 10% Coordinate Reference / Map Projection / Geodesy
GML
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Ont 1 # properties Ont 1 # (data & classes object) Ont 2
% Commo n Class & Propert y (Inria Word align > 60%) Conclusions
166
121
Cyc MapProjection
50
25
13
7%
2D geometry and ellipsoid in common 2D geometry and ellipsoid in common Line geometry in common GML and ISO 19107 have much in common; built from the same data model. a few basic shapes in common Cyc is more specialized, a few basic shapes in common Projections are map styling in GML; Point, Surface in common; GML does not use projections, some topological concepts in GML mistakenly mapped; Cyc does not have CRS concept GML has generalized datum types, Cyc has specific datums. Point in common, reference system terms have dissimilar names Terrain is a distinct domain from GML GML & Cyc topology different class names
GML
166
121
Cyc Geodesy
84
42
8
4%
GML
166
121
Cyc Terrain
164
28
2
1%
GML
166
121
Cyc Topology
17
28
2
1%
Cyc Geodesy
84
42
ISO 19111
28
47
11
11%
Cyc MapProjection
50
25
ISO 19111
28 Metadata
47
1
1%
coordinate ref & datum in common Cyc specific projections vs. general in ISO
FGDC CSDGM
165
302
ISO 19115
134 Temporal
309
112
25%
Similar metadata terms, FGDC more detail
OWL-TIME EVENT
12
48
SOUPA TIME
22
1
31
75%
OWL-TIME EVENT
12
48
ISO 19108
45
243
14
8%
OWL-TIME EVENT
12
48
Cyc TimeInterval
211
4
3
2%
OWL-TIME EVENT
12
48
SWEET TIME
69 6 Spatial Relationship
17
25%
basic concepts (begin, end, instants) and some relations (before, after) in common, includes DAML time some basic concepts (begin, end, instants); did not have much common with DAML time Interval in common, Cyc has more specifc time/day/year; only 3 common with DAML time basic concepts (begin, end, instants) and some relations (before, after) in common, includes DAML time
20%
only 2 relations in common (contains, overlaps); some direction clases and property types in common
MINDSWAP georelations
35
46
SOUPA RCC
2
9
9
and integrating existing ontologies and metadata standards. Foundational ontologies can be used to translate across different domain ontologies by providing appropriate (inter-theoretical) semantic content. Conceptual mapping becomes easier and more consistent, if there is a good foundational or upper ontology being used. The foundational ontologies covered in this section are well-known, but are not spatial in nature. The upper spatial ontologies presented in Section 6 are not necessarily built upon a foundational ontology.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
189
5.1. Basic Formal Ontology (BFO) The Basic Formal Ontology (BFO) comprises two primary parts. These are referred to as SNAP and SPAN. SNAP (snapshot) is a set of sub-ontologies that represent pointin-time views of the world. SPAN is a sub-ontology that represents processes over time. In some sense, the temporal modeling aspects of the ontology have been designed into BFO from the start. Within the OWL representation of BFO, temporal instants and intervals are defined, but are not grounded to a calendar. 5.2. DOLCE The Descriptive Ontology for Linguistic and Cognitive Engineering (DOLCE) is the first of several foundational ontologies to be developed under the European Union WonderWeb program. With a “clear cognitive bias”, DOLCE is focused on comparing and elucidating relationships in other ontologies. DOLCE was originally developed using a First Order Logic representation. DOLCE Lite is an OWL representation without modality, temporal indexing, or relation composition. DOLCE Lite Plus adds additional modules including SpatialRelations and TemporalRelations. See http://www.loa-cnr.it/DOLCE.html for more information. Many of the terms used (e.g. endurant vs. perdurant) in DOLCE will likely be unfamiliar to most users. The ontologies use lowercase names with hyphens for both classes and properties rather than the UpperCamelCase class names and lowerCamelCase property names generally found in the Semantic Web – this is a significant distraction. The DOLCE Lite Plus ontologies use multiple namespaces and modules, but these do not seem to be particularly correlated. Latitude and longitude are not mentioned in the ontology. SpatialRelations doesn’t include relations such as RCC8. TemporalRelations includes the Allen relations.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6. Spatial Ontologies & Models The search for spatial ontologies focused on those with the most prevalent use and fullest definition of a spatial vocabulary. Several ontologies addressed multiple categories of spatial concepts. The evaluation included upper ontologies, those developed specifically for the National System for Geospatial-Intelligence (NSG), commercial ontologies and relevant concepts embedded in other ontologies. The study applied a high-level evaluation to each ontology (as summarized in Appendix A), with more depth focused on those ontologies that had unique characteristics. 6.1. Upper Ontologies As stated previously, the group of ontologies classified as “upper” spatial ontologies were selected for the higher level of abstraction they presented when compared to a domain and data source ontology. The upper ontologies evaluated were taken from several well-known sources in the Semantic Web and geospatial standards communities.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
190
J. Ressler et al. / Geospatial Ontology Trade Study
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.1.1. Cyc The Cyc ontology is defined by a research community of contributors and intended to represent the “sum of human knowledge.” 4 Cyc’s knowledge base has been created over the past 20 years by hundreds of people. Cyc stores information about the common knowledge used for reasoning about common entities. It currently contains almost 200,000 terms, with several dozen hand-entered assertions about each term. Two forms of Cyc are available – OpenCyc (non-proprietary) and ResearchCyc 5 . ResearchCyc was examined for this study because OpenCyc contains only a small subset of the knowledge in ResearchCyc [Matuszek, et. al 2006]. Cyc knowledge is represented in CycL, a LISP-based syntax for First Order Logic. The Cycorp OWL Ontology Exporter produces OWL and SWRL corresponding to a subset of the CycL language6. The Cyc ontology contains over 20 predefined “domains” with tens to hundreds of concepts (classes). The Cyc domains are a “clumping of concepts” 7 that pertain to many common fields of knowledge, but are not synonymous with a domain ontology as defined in Section 4. Cyc can be searched for any concept, but was utilized in this study by searching collections of concepts in domains. Searching through the predefined domains of Cyc uncovered seven that pertain to spatial concepts: Geodesy, Linear Object, Map Projection, Open Geospatial Consortium (not an OGC product), Surface Geometry, Terrain, Topology. While there is a small degree of overlap between spatial domains, each domain has a unique set of classes and purpose. Temporal domains in Cyc are covered in section 7.1.1. The Open Geospatial Consortium (OGC) domain in Cyc is the most generalized and useful for geospatial representations. It’s similar in purpose to the simple features profile of GML [10], a subset of the full GML specification for basic geometry types (points, multi-point lines, curves, surfaces) and coordinate reference systems. The specialized applications of geospatial knowledge are addressed in other specific domains. The OGC domain is the most generalized, yet doesn’t share the same classes with most of the other ontologies. The geometry representations are also addressed by Linear Object and Surface Geometry ontologies which have a full range of classes to represent two dimensional lines and polygons and three dimensional geometric projections upon spherical and planar surfaces. These ontologies are also similar to types of feature geometry in the GML simple features profile [10]. These domains shares the basic spatial geometry classes with the Cyc OGC domain but have more complex linear and surface geometrics. The Topology and Terrain domains are the most specialized. Topology classes represent the network of geospatial points, edges and surfaces but these concepts are distinct from the GML topology. The terrain ontology describes types of soil, primarily terrain for military applications and is not concerned with the geometry of terrain. The Geodesy domain has classes for coordinate reference systems, datums and a range of 33 different spheroids. The Geodesy classes are similar to ISO 19111, Spatial Referencing by Coordinates [11], yet the ISO/OGC standard is more generalized to 4
http://www.opencyc.org/ ResearchCyc was licensed for government use on this study. 6 The Cyc OWL Exporter was developed for NGA under contract HM1582-05-C-0020 and used with permission on this study. 7 Taken from Research Cyc Ontology Exporter tutorial http://research.cyc.com/ (licensed) 5
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
191
allow mathematical definition of any datum. The Map Projection ontology is specific to map products using worldwide and local projections and is not similar to any other ontology studied. Specific map projections can be represented using GML, but they are not specified in the breadth of Cyc. The overall assessment of the Cyc ontologies is that they indeed contain some useful concepts and relations to address the core subjects of geospatial and temporal ontologies, though lacking a concise and usable event ontology. The primary difficulty of using Cyc is finding the appropriate ontology subsets, searching for the proverbial “needle in a haystack” while removing the “wheat from the chaff”. This is due in part to the open base of knowledge represented in Cyc, which tends to clutter the ontologies with trivial information and concepts unrelated to a field of interest. Even with the Cyc OWL Exporter, applying ResearchCyc to the geospatial intelligence community is likely to require an expert ontologist who is familiar with CycL. Cyc includes an individual ThePentagonBuilding, with several classifications and assertions. There were 71 classes that applied to “The Pentagon,” such as “Building,” “GovernmentalBuilding,” “LandMark,” “PolyDimensionalThing,” and “GeographicThing”. While geographic location information for ThePentagonBuilding was not found, searching the Cyc Browser discovered assertions for its location. 6.1.2. FGDC Metadata Ontology Drexel University developed an ontology for the Content Standard for Digital Geospatial Metadata (CSDGM) of the Federal Geographic Data Committee (FGDC) using FGDC-STD-001-1998. While differing in structure, the metadata content is similar to ISO 19115 (in 2003, a mapping between the two standards found less than 20 of the 349 FGDC metadata fields had no match in ISO 19115)8.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
6.1.3. ISO TC 211 Geographic Information Ontologies The standards approved by ISO Technical Committee 211 define a model of geographic information that is comprehensive and well organized. In 2004, Drexel University represented the ISO standards as written at the time in separate ontologies. Some of these ISO ontologies are referenced in comparison to other ontologies in this study. The OGC’s GML ontology (ISO 19136) is covered separately in Section 6.3.2. The standard schemas from the ISO Geographic Information series of specifications (191xx) considered for incorporation into ontologies, most of which are in OWL, are as follows. • • • •
8
Conceptual Schema Language [12] – used to build other schemas, not inherently geospatial. Spatial Schema [13] – spatial geometry for points, lines, curves, surfaces, topology (nodes, edges, faces, solids) and variations on these entities. Temporal Schema [14] – representation of time (periods, instants, date & time). This is described in more detail in Section 7.1.2. Rules for Application Schema [15] – schema definition (associations, constraints, feature types, properties) used to develop application schemas using the general feature model (not inherently geospatial).
http://www.fgdc.gov/metadata/documents/FGDC_Sections_v40.xls/view
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
192
J. Ressler et al. / Geospatial Ontology Trade Study
• • • • •
Methodology for Feature Cataloguing [16] – feature catalog (version, registration of features, feature types, attributes, associations). Spatial Referencing by Coordinates [17] – geodesy, coordinate reference systems, datums. Spatial Referencing by Geographic Identifier [18] – location identification such as used in a gazetteer (toponyms). Metadata[19] – metadata elements of a geographic dataset. Coverage [20] – a specialization of features with a range of attribute values that vary in space or time to provide a grid of data for geometric types (point, surface).
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
ISO 19123, Coverage, was not included in the OWL ontologies from Drexel because it was published after Drexel completed the ontologies in 2004. The coverage schema is included in the collection of ontologies because it is crucial to describing geospatial data and should be a part of the foundation of the ISO 19xxx ontologies. In comparing the latest version of the ISO specifications with the versions used by Drexel in 2004, three specifications have been updated since the ontologies were created: ISO 19103, 19109 and 19110 (not including ISO 19123). ISO 19136 (GML), discussed in Section 6.3.2, has also been updated since 2004, though it is not yet an approved ISO standard. The ISO ontologies are well-defined and based upon an authoritative set of geospatial and temporal concepts. They are well suited for use as upper ontologies to be specialized for a domain. Some models such as Temporal, Spatial, and Coordinate Referencing can be imported and directly used in domain ontologies. One shortcoming of the ISO ontologies is the inter-dependence between the ontologies. The Temporal, Spatial, Feature Catalog and Metadata import most of the other ISO ontologies, as shown by the “Dependency” column in Appendix A. The import of multiple ontologies increases the time to load and search the ontology in a browser or rules engine, as well as increasing the impact of changing an ontology. The ISO ontologies would be easier to use if decomposed into modular files with less inter-dependency. A representation of the Pentagon feature using ISO 19107 is as follows:
2
This example reveals some limitations of using the current ISO 19107. First, there are no instances of SC_CRS (CoordinateReferenceSystem from ISO 19111) in the
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
193
standard. Secondly, statements in OWL are inherently unordered, which causes ambiguity regarding which coordinate is latitude and which is longitude. The coordinate ordering limitations can be addressed in a specialization of ISO 19107 that provides sub-properties of coordinate, such as latitude and longitude or x and y. The reference systems can be instantiated from ISO 19111 or by mapping 19111 to another ontology which provides specific coordinate reference systems. 6.1.4. SUMO The Suggested Upper Merged Ontology (SUMO), developed as one of several candidates for an IEEE Standard Upper Ontology, contains 1,000 classes of general interest and utility across a wide range of applications. It is accompanied by the MILO Mid-Level Ontology, which adds 1835 classes with additional detail, and a number of domain ontologies (including Geography and Countries and Regions) focused on specific application areas. These ontologies were originally developed in KIF and then translated to OWL. Some of the pertinent characteristics of SUMO are: • • •
•
Relevant geospatial classes in SUMO include Region and TimePosition. Latitude and longitude are not defined as properties in either SUMO or MILO. They are defined in Geography.kif. Loading sumo.owl and Mid-level-ontology.owl consistently caused SWOOP [32] 2.2.1 to become unresponsive and consume all available CPU resources, even after replacing rdfs:Class with owl:Class. This may be related to SUMO’s use of metaclasses (classes of classes). The OWL translations are incomplete and may not be current with respect to the latest KIF versions. We’re unaware of any OWL translations of the domain ontologies.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
The representation of the example Pentagon feature in SUMO is as follows:
SWEET contains a temporal ontology described in Section 7.1.6.
7. Event Ontologies An Event (such as a birth, meeting, or commercial transaction) in the context of this study is generally associated with one or more locations, a time instance or interval,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
203
and one or more actors (persons, organizations, and/or objects). Locations were discussed in the previous section. We now discuss temporal and actor representations before addressing events. 7.1. Temporal Representations Significant research has been devoted to the representation of time. XML Schema includes a basic set of temporal datatypes derived from ISO 8601. These include date, time, dateTime, duration16, gMonth, gDay, gYear, gYearMonth, and gMonthDay. These accommodate local times, time zone offsets, and Coordinated Universal Time (UTC/GMT). These datatypes are supported by OWL and sufficient by themselves for many applications such as timestamps. The XML Schema datatypes don’t address time intervals other than by loosely associating start and end times or start times and durations. 7.1.1. Cyc Temporal
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
In addition to the upper spatial ontologies described in Section 6.1.1, two temporal representations, CycTimeInterval and CycTemporalPredicate, are discussed here. The ResearchCyc OWL ontologies available through the OWL Exporter lack a succinct set of concepts for events and time, even though they are common entities. The “Event” concept has over 13,000 descendants, many of which are trivial events of no use. The first level of Event specialization has 120 concepts, of which no more than ten would be useful to geospatial events. The Open Cyc glossary contains the definitions for basic time, date and event concepts shown in Figure 5. The ontology for portions of this model was exported from ResearchCyc by writing a CycL query.
Figure 5. Research Cyc Temporal Concepts (©Cycorp)
The “TimeInterval” concept has some useful descendants for creating a temporal schema. Of particular use is “ContinuousTimeInterval” which has calendar date, time 16 XQery 1.0 and XPath 2.0 Functions and Operators further divides duration into two totally ordered subtypes, yearMonthDuration and dayTimeDuration, alleviating some of the concerns expressed in the OWL specification.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
204
J. Ressler et al. / Geospatial Ontology Trade Study
and periods of time concepts. Related to temporal, a collection of properties in Cyc called “ComplexTemporalPredicates” are useful for temporal relations between concepts. These properties contain some trivial relations that need to be filtered. In this collection, the properties inherited from “temporallyRelated” have useful relationships for time intervals “endsDuring,” “startsDuring,” “temporallyIntersects,” and “temporallyCoexist.” The relations encompass most of 13 basic temporal relations defined by Allen [26]. It was difficult to locate the basic Allen relations because the temporal predicates were mingled with more complex relations such as “startsAfterEndingOf”, a combination of two of Allen’s relations [31]. 7.1.2. ISO 19108 The ISO 19108 standard was designed to represent the temporal characteristics of geographic information. While not created natively in OWL, it has been partially made into an OWL ontology by Akm Saiful Islam and others at Drexel University. Its description suggests that more work is yet to be done, however the ontology has not been updated since 2004 and thus can probably be considered abandoned. However, the ontology does support several useful temporal concepts. It defines both instants and intervals, as well as spatial topological relationships. It also gives the capacity to relate the instants and intervals to clock and calendar times and dates. It does not give definitions for events, however, nor does it provide definitions for recurrence.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.1.3. OWL-Time Representing and reasoning about time intervals was one of the motivators for the OWL-Time ontology [27]. OWL-Time contains the basic concepts for time intervals and instants for both clock and calendar time. OWL-Time builds upon the W3C XML Schema for time and time zones. Combined with the W3C XML Schema for time, OWL-Time is the most comprehensive of the temporal ontologies within this study, and as its name suggests, is primarily represented in OWL. It comprises three major temporal areas. The first is temporal topological relations. This includes both definitions of the fundamental concepts of instants and intervals, as well as descriptions of the Allen relations between them. These are a straightforward OWL interpretation of the Allen intervals. The next fundamental area is the measurement of durations. This is accomplished by a conceptualization of durations as sets of smaller-length durations, e.g., 60 consecutive second durations compose a minute. These durations are grounded to a non-relative time conceptualization by adding the third fundamental area, clock and calendar. This portion defines the time and date, going to great length do deal appropriately with considerations like regions being in different time zones at different times of the year. OWL-Time by design is not an event ontology, however, and requires predicates of an event ontology to link its temporal representations to the relevant people, places, and things. Though older versions of the ontology did not have support for recurrence (“The first Monday of every April”), current versions of the ontology account for this as well. 7.1.4. RDF Calendar Recurring time instants and intervals (e.g. the second Tuesday of every month) motivates much of the work in scheduling, such as the IETF iCalendar standard and its Semantic Web representation RDF Calendar. The focus of this representation is on
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
205
communicating the scheduling of events, and thus recurrence is a primary feature. Like SOUPA Time and unlike OWL-Time, RDF Calendar uses XML Schema datatypes for representing dates and times. Its notion of time is essentially driven by calendar events, and thus it has no need to define instants or topological temporal relationships. 7.1.5. SOUPA Time SOUPA Time is one of the set of the SOUPA (Standard Ontology for Ubiquitous and Pervasive Applications) ontologies. Unlike OWL-Time, it makes use of XML Schema datatypes for dateTime, thus drastically reducing the amount of temporal information that is represented in OWL. This could be viewed as either a strength or a weakness, as it is simpler but less rich. SOUPA Time contains the definitions of the basic temporal instants and intervals, as well as definitions for the Allen relations. No representation of recurrence is included within SOUPA Time or the other SOUPA ontologies for Schedule and Event. 7.1.6. SWEET Time The Time ontology included in NASA SWEET includes seasons, times of day, and an assortment of time eras and periods (e.g. Jurassic). It does not contain sufficient object and datatype properties to be used in temporal instances or relations.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.1.7. Temporal Cross Products Though not a temporal ontology per se, the work of Kathleen Hornsby on crossing spatial and temporal ontologies is relevant to the area. This work is taken from conversations and [28], developed as part of NGA grant HM1582-05-BAA-002 to the University of Maine. The method conceived takes geospatial classes from one ontololgy, temporal classes from another ontology, and stores the classes and relations into a relational database. A Cartesian cross product is performed in SQL to create the resultant spatio-temporal combination of classes. For example, one can take a spatial ontology with descriptions of places on a college campus and a temporal ontology of varying granularity such as days, weeks, and seasons to create the cross-product, resulting in concepts like “ParkingLot_Weekend” and “Library_Evening”. The method creates a pair for every combination of spatial and temporal classes, which can become a large ontology. The cross product uses relations in the ontology such as “is-A”, “contained-In” and “component-Of” to infer all terms that related to a query. For example, the concepts related to the pair “Building_Morning” would include the following shown in Table 5. Table 5. Example Temporal Cross Product17
17
Spatial term
Temporal Term
Relation
Building
Morning
cross-product of Building and Morning
Academic Building
Morning
is-A Building
Building
EarlyMorningHour
component-Of Morning
DiningCommons
Morning
is-A Building
Based upon Figure 11 in [28].
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
206
J. Ressler et al. / Geospatial Ontology Trade Study
One benefit of using temporal cross products is it allows a knowledge engineer to review these cross-product concepts to find relevant ones for their domain. No data is yet available to indicate how many useful concepts are derived from a “typical” application of this technique. The resultant set of terms can be deterministically inferred using the reasoning queries in the approach. 7.2. Actor Representations The concept Actor is used to generalize the people, organizations, and/or other objects that participate in an event. Friend-of-a-Friend (FOAF) is a widely used Semantic Web ontology for representing information about people. It includes the classes Agent (their term for Actor), Person, Organization, and Group, but provides the most detailed definitions for Person. As such, it provides a good basis for tying person actor information to definitions of events. It primarily describes peoples’ names and contact information, affiliations, and links to other people. Other ontologies provide more complete definitions of Organization, including typical subclasses such as Company and EducationalInstitution, often with many subclasses. One needs only to choose an ontology that is suited for describing the types of actors that take part in the relevant events. 7.3. Event Representations An event ontology should define concepts for linking spatial, temporal, and actor representations to a description of an event. This linkage can be of varying complexity, ranging from simple declaration of an event class and the properties to link space, time, and actors, to descriptions of actor roles, preconditions, causes, effects, etc. The ontologies within this study, including foundational ontologies, each treat events somewhat differently.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
7.3.1. RDF Calendar RDF Calendar is essentially built around the concept of events; however, this is event in the sense of “meeting” instead of the more general “occurrence at a time and a place.” As such, while it may be adaptable for use cases in which scheduled events with people are the primary focus, it is less suitable for domains dealing with scientific processes, socio-political happenings, or business transactions. 7.3.2. SOUPA Event SOUPA contains a sub-ontology for describing events. It defines SpatialTemporalThing as the intersection of SpatialThing and TemporalThing, and defines SpatialTemporalEvent as the intersection of SpatialTemporalThing and TemporalEvent. However, this is the extent of what is described in this sub-ontology. No predicates for linking actors to these events are provided. 7.3.3. Time as a Sequence of Events Work by Hornsby and Cole [29] uses a sequence of events to model the passage of a continuant object (one which endures over time) through time. The definition of an Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
207
event comprises the standard parts that have been previously discussed: a reference to the continuant object involved, a time for the event, a location, and a description of the event. This work provides an interesting abstraction of the qualitative motion of an object. Rather than attempting to track an object precisely, an object’s motion can be represented as a series of change of ‘zone’ events. Due to the discrete nature of data in OWL, this may be a good fit for systems that do not require very detailed information about the current location of an object. At the time of this writing, however, an OWL representation of these event concepts does not exist. An ontology of movement would be well-suited to describe events that are associated with motion. Developing such an ontology in concert with the existing temporal ontologies is a good subject for further research.
8. Recommendations and Conclusions The recommendations from this study fall into two areas: 1. The ontologies surveyed that are currently best suited to each use case per category 2. Further research and work to formalize the spatial ontologies. There are a number of acceptable OWL ontologies and related representations available. Where an ontology exists, there is no need to build a new one from scratch or to provide geospatial representations unique to each application. We suggest the following guidelines for reusing geospatial ontologies:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
• • •
Use OWL. Users should use the simplest OWL representations that meet their needs. Geospatial Ontologies should be based upon standards consistent with the NSG Architecture and with the GEOINT Standards listed in the Defense Information Standards Registry (DISR), which are also contained in the NSG Architecture Compliance.
The NSG user has additional requirements for ontologies beyond the business- and consumer-oriented mass market. The ontologies recommended for feature, feature type and geometry are intended for a full representation of feature geometry and attributes for geospatial information systems, such as intended by the NSG. These ontologies can be referred to as the “Full spatiotemporal ontology.” Use of semantic information in commercial practice needs to be simpler and readily available to the mass market of information technology, which is only interested in using geospatial data if it works with other web-based applications. Thus, another category of “light” spatial ontologies and “folksonomies” are emerging on the web, regardless of whether there are standards to formalize the ontology. To this end, research is needed into a “light” spatiotemporal ontology for the minimum spatial and temporal representation, such as are becoming popular with GeoRSS, Basic Geo, KML or eventually the GML Simple Features profile. Alignments and mapping technology will be needed to be continuously applied in order for this “full” spatiotemporal ontology to be interoperable with these “light” ontologies and hence interoperable with the mass market.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
208
J. Ressler et al. / Geospatial Ontology Trade Study
A recommended ontology/set of ontologies for each of the ontology categories and use cases listed in Sections 4.1 and 4.2 are shown in Table 6. The rationale for each category recommendation is summarized below each category. Table 6. Recommended Full Spatiotemporal Ontologies Recommended Ontology per Use Case Ontology Category 1. Geospatial Feature
Annotation
Qualitative Reasoning
Information Integration
GeoRSS (Simple or GML)
Not applicable
GML
Rationale – GeoRSS Simple provides a very concise representation allowing basic geometries using WGS84 decimal degrees to be expressed using a single property, but can be readily converted to GML. GeoRSS GML provides the full power of GML for more complex geometries or alternative coordinate reference systems, and is standardized, expressive, and widely used. 2. Feature Type
NSG FC / NAS 1.8 (in OWL)
Not applicable
NSG FC / NAS 1.8 (in OWL)
Rationale – NAS has the largest representation of geospatial features, though it does not have an OWL representation. ECDM is a good alternative in OWL, but is not being maintained. MINDSWAP geofeatures and SWEET cover a small range of feature types. Geonames uses a large feature set from the NGA GEOnet Names Server, but lacks properties. 3. Spatial Relationship
SOUPA rcc
SOUPA rcc
SOUPA rcc
Rationale – SOUPA RCC is a highly modular ontology that focuses on the Region Connection Calculus. It can be used with other ontologies. MINDSWAP includes a set of directional relationships; these are transitive, but can be problematic when composed over varying distances. Cyc’s and GML’s relationships only apply to topological features and surfaces. 4. Toponym
ISO 19112
Not applicable
ISO 19112
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Rationale – ISO 19112, Geometric Identifier, is a standard for named locations (gazetteer toponyms) and includes coordinate reference, temporal extent, and other attributes. NAS 1.8 has a Named Location feature type with a type and administrative level. 5. Coordinate Reference
ISO 19111,
Not applicable
ISO 19111
Cyc Map Projection Rationale - ISO 19112, Geometric Identifier, is a standard model with all of the elements of a coordinate reference system. Cyc is the only ontology with an enumeration of map projections useful for annotating digital maps. 6. Geospatial Metadata
ISO 19115
Not applicable
ISO 19115
Rationale – ISO 19115 is a standard and 96% of FGDC metadata can be described in 19115. 7. Web Service
Not evaluated
Not evaluated
Not evaluated
Rationale – web service ontologies were not in scope of the spatial ontology study. 8. Geometric
ISO 19107
SOUPA rcc
ISO 19107
Rationale – ISO 19107 is a standard with all of the most widely used spatial geometries. The geometry model of GML complies with ISO 19107. GeoRSS is part of the Feature ontology. Common geometries in Cyc’s ontologies are covered by ISO 19107, unique geometries can be mapped to ISO. SOUPA RCC is the ontology best suited to basic geometric relationships, although MINDSWAP may have some application in directional relationships.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
209
J. Ressler et al. / Geospatial Ontology Trade Study
Recommended Ontology per Use Case Ontology Category 9. Coverage
Annotation
Qualitative Reasoning
None available
Not applicable
Information Integration None available
Rationale – Currently lack an ontology for expressing coverage classes. ISO 19123 is recommended as a basis for this ontology. 10. Geopolitical
None recommended
Not applicable
None recommende d
Rationale – Cyc Terrain is not generalized for other political regions; SUMO Geography only includes global water and land. 11. Temporal
XML dataypes in OWL, OWL-Time
OWL-Time
OWL-Time
Rationale – OWL-Time is based upon DAML-Time and has all widely used temporal representations and the 13 Allen relationships. CycTimeInterval contains time instances, but differs in structure from OWLTime. SOUPA Time is a concise representation which also uses DAML-Time (similar to OWL-Time), but contains only 7 temporal relations.
Further recommendations specifically addressing these ontologies and follow-on activities are as follows. •
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
•
Create an ontology library for the ISO 191xx specifications (ISO Project 19150). Because the ISO-based ontologies created by Drexel are over two years old, these should be updated to the latest version of the ISO specifications, published to a maintained URL in a modular library with minimal inter-dependency. In addition, an ontology for 19123, Coverage, is needed. Perform a metrics evaluation of the quality of the selected ontologies similar to the assessment performed by Burton-Jones on the DAML ontology library [30]. This quality assessment developed measurements of an ontology’s syntax, richness, interpretability, clarity, comprehensibility and relevance. While geospatial web services are likely to use the spatial ontologies recommended, web service ontologies were not in scope of this evaluation. Given the active interest in service oriented architectures, the use of ontologies to describe services (such as OWL-S and SAWSDL) is an active area of research and commercial development. An evaluation of ontologies to represent web services is recommended for a future study.
Semantic knowledge can work well when encapsulated in an application of semantic concepts within a domain. However, unless customized for multiple knowledge bases, semantic queries are not interoperable when performed across domains and between multiple semantic sources of knowledge. The formalization of ontologies will lead to greater interoperability in the future to be leveraged by semantic applications. A lack of standardization will result in the creation of separate silos of semantic data that is difficult and more expensive to re-use, having the undesirable effect of slowing the adoption of semantic applications.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
210
J. Ressler et al. / Geospatial Ontology Trade Study
Acknowledgments The authors would like to acknowledge Dan Adams, Todd Pehle, Mike Smith, and Jess Irwin for their contributions to the study, and appreciate the expert review and input from the leadership of the Spatial Ontology Community of Practice: John Moeller, Kevin Backe, Gary Berg-Cross, and Joshua Lieberman. The authors acknowledge Jane Schultz for editing the finished document. The authors acknowledge the cooperation of Cycorp and the University of Maine in the research of this study. This report has been sponsored by NGA Innovision Basic Research under the leadership of Ed Laikin.
References [1] [2] [3] [4] [5] [6] [7]
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[8] [9]
[10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25]
[26] [27] [28]
Object Management Group, “Ontology Definition Metamodel (ODM),” OMG/RFP ad2003-03-40, 5 June 2006. http://www.omg.org/docs/ad/06-05-01.pdf . Ressler, J., Dean, M., Benson, E., Dorner, E., Morris, C. “Application of Ontology Translation,” submitted to the International Semantic Web Conference, 2007. M. Dean and G. Schreiber, eds. OWL Web Ontology Language Reference. W3C Recommendation 10 February 2004. http://www.w3.org/TR/2004/REC-owl-ref-20040210/. Faklovych, K. et al. “UML for the Semantic Web: Transformation-Based Approaches. http://homepages.cwi.nl/~media/publications/UML_for_SW.pdf. Gasevic, D., Djuric, D, Devedzic, V. and Damjanovic, V. “Converting UML to OWL Ontologies,” http://afrodita.rcub.bg.ac.yu/~gasevic/projects/UMLtoOWL, May, 2004. Kolas, Dave, Dean, Mike and Hebeler, John, “Geospatial Semantic Web: Architecture of Ontologies,” Proc. IEEE Aerospace Conference, Big Sky, MT, March 2006. Euzenat, Jerome, et al. “Results of the Ontology Alignment Evaluation Initiative 2006,”, Ontology Alignment Evaluation Initiative (OAEI), Proceedings at the 1st International Workshop on Ontology Matching at the 5th International Semantic Web Conference, Athens, GA, Nov. 2006. http://www.dit.unitn.it/~p2p/OM-2006/7-oaei2006.pdf. Euzenat, Jerome, “An API for ontology alignment (version 2.1),” INRIA Rhone-Alpes, 2006. Matuszek, Cynthia, Cabral, John, Witbrock, Michael, DeOliveira, John. “An Introduction to the Syntax and Content of Cyc,” Proc. 2006 AAAI Spring Symposium, Association for the Advancement of Artificial Intelligence, March 2006. Geography Markup Language (GML) Simple Features Profile, version 1.0, 2006-04-25. Spatial Referencing by Coordinates, 2005-04-25, submitted as ISO 19111. ISO/CD TS 19103:2005, Conceptual Schema Language ISO 19107:2003, Spatial Schema. ISO 19108:2002, Temporal Schema. ISO/ FDIS 19109:2005, Rules for Application Schema. ISO/ FDIS 19110:2005, Methodology for Feature Cataloguing. ISO 19111:2003, Spatial Referencing by Coordinates. ISO 19112:2003, Spatial Referencing by Geographic Identifier. ISO 19115:2003, Metadata. ISO 19123:2005, Coverages. Local MSD Implementation Profile (GML 3.2.1), 2007-03-25. Geography Markup Language (GML) 3.0, Open GIS Consortium, 2003-01-29. Geography Markup Language (GML) 2.0, Open GIS Consortium, 2001-02-20. Open Geospatial Consortium, “KML 2.2 Reference OGC Best Practice”, 2007-09-14. http://www.opengeospatial.org/standards/kml/. Gruber, T. R., “Toward principles for the design of ontologies used for knowledge sharing”, International Journal of Human-Computer Studies, Vol. 43, Issues 4-5, November 1995, pp. 907-928. http://tomgruber.org/writing/onto-design.pdf. Allen, J.F., “Maintaining knowledge about temporal intervals”, Communications ACM 26, 11, 1983. J. Hobbs and F. Pan. Time Ontology in OWL. W3C Working Draft 27 September 2006. http://www.w3.org/TR/2006/WD-owl-time-20060927/. Stewart Hornsby, K. and K. Joshi, “Combining Ontologies to Automatically Generate Temporal Perspectives of Geospatial Domains,” Draft, Dept. of Spatial Information Science and Engineering, University of Maine, Orono, ME, June 2007.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
J. Ressler et al. / Geospatial Ontology Trade Study
211
[29] Stewart Hornsby, K. and Cole, S. “Modeling Moving Geospatial Objects from an Event-Based Perspective,” to appear in Transactions in GIS, 2007. [30] Burton-Jones, A., Storey, V., Sugumaran, V., Ahluwalia, P. “A semiotic metrics suite for assessing the quality of ontologies.” Data & Knowledge Engineering 55 (2005) 84-102. [31] Allen, J. and Ferguson, G. “Action and Events in Interval Temporal Logic,” University of Rochester, Computer Science Department, Technical Report 521, July 1994. [32] Kalyanpur, A., Parsia, B., Sirin, E., Grau, B., Hendler, J., “Swoop: A Web Ontology Editing Browser”, Web Semantics: Science, Services and Agents on the World Wide Web 4 (2006) 144–153. [33] I. Horrocks, P. Patel-Schneider, H. Boley, S. Tabet, B. Grosof, and M. Dean. SWRL: A Semantic Web Rule Language Combining OWL and RuleML. W3C Member Submission 21 May 2004. http://www.w3.org/Submission/2004/SUBM-SWRL-20040521/. [34] Geography Markup Language (GML), ISO/DIS 19136:2005, Draft Interface Specification.
Appendices Two appendices referenced for additional information are found online at: http://projects.semwebcentral.org/docman/?group_id=84
Appendix A - Ontology Survey Data A survey of the ontologies studied is documented in a spreadsheet formation giving the description, source, language, metrics, and applicability of each ontology.
Appendix B: Ontology Hierarchy
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
For the publicly releasable ontologies used in the preparation of this study, a hierarchy of the classes and properties was produced from OWL using the “dumpont” tool. 18 Each group of ontology hierarchies begin on a separate page and subsection for each ontology.
18
http://projects.semwebcentral.org/projects/dumpont/.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved. doi:10.3233/978-1-60750-581-5-213
213
Chapter 12
Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future Leo OBRSTa,1, Werner CEUSTERSb, Terry JANSSENc a The MITRE Corporation, USA b State University of New York at Buffalo, cLockheed Martin Corporation
Abstract: This chapter looks at the intersection of intelligence and ontologies and semantic technologies, and tries to characterize the impact of these in the future. It provides a view into some emerging technologies such as query languages and rule standards for the Semantic Web. It also provides some guidance from a different domain, the biomedical domain, and tries to show that realist ontologies, ontologies based on common real world characterizations, have an effective impact on applications in those domains. Finally, it looks at the potential impact of these technologies on intelligence collection and analysis in the future, and makes some predictions.
Keywords: Ontology, information-sharing, intelligence community, semantic technologies, healthcare.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Introduction In this chapter we look ahead: at ontology and semantic technologies and standards that are emerging, at the prospective evolution of the intelligence community, at where things could go in the future. The future direction and its success depend on many factors, including the commitment to embracing these technologies and the quickness and sophistication of their adoption. To assist our consideration of issues in technology adoption that could affect the intelligence community, we look at a test case, that of the adoption of these technologies by the healthcare community, and its prospective lessons for the intelligence community. Finally, we describe our projections and hopes.
1. Emerging Ontology and Semantic Technologies and Standards We’ve largely focused on Semantic Web technologies in this book. Why? Because Semantic Web technologies represent n emerging set of global standards that are commercially rooted and also driven by a standards process that tends to be shorter in lifespan than older standards processes. This typically shorter, differently regimented 1
Corresponding Author: Leo Obrst, The MITRE Corporation, 7515 Colshire Drive, McLean, VA 221027508, USA; E-mail: [email protected]. Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
214
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
standards process, usually enacted under the World Wide Web Consortium (W3C), does not guarantee better standards, but standards that typically are more immediately adapted to multiple communities of the Internet, e.g., researchers, Web and service developers, database practictioners, digital librarians, ontologists, etc. This is not to diminish the value of ISO standards and, in particular, ISO Common Logic, which is a valuable standard for representing very expressive logical ontologies. 1.1. Complexity of Applications and Costs
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
However, a general maelstrom of activity and acclamation does not guarantee the success of the technologies touted. The value of the work the technologies accomplish, and in fact, the greater value and lesser cost of the work they accomplish even as prorated over time, must be demonstrated. Value, potential value, cost over time – all of these must be estimated. However, as is usual with technological evolution, there is a spectrum or continuum behind the potential adoption of technologies, because there is a spectrum behind the expressivity of the models and the complexity of the potential applications that those models can provide, as Figure 1 depicts [1].
Figure 1. More Expressive Semantic Models Enable More Complex Applications
In Figure 1 shows that as the expressiveness of the semantic model increases, so does the possibility of solving more complex problems. Note that we distinguish term and concept here, where their definitions are the following (from [1]). Terms (terminology) are natural language words or phrases that act as indices to the underlying meaning, i.e., the concept (or composition of concepts). The term is syntax (e.g., a string) that stands in for or is used to indicate the semantics (meaning). A concept (a universal category for referents) is a unit of semantics (meaning) in the mental or knowledge representation model. In an ontology, a concept is the primary knowledge construct, typically a class, relation, property, or attribute, generally associated with or characterized by logical rules. In an ontology, these classes, relations,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
215
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
properties are called concepts because it is intended that they correspond to the mental concepts that human beings have when they understand a particular body of knowledge (subject matter area or domain) but at the philosophical universal level, i.e., as kinds of entities. In general, a concept can be considered a placeholder for a category (way of characterizing) of specific real world referents. From a realist perspective, as will be discussed in the next section, these concepts as placeholders are dispensable. For simple applications, controlled vocabularies, terminologies, and classificational systems, usually structured in topic taxonomies or thesauri, are sufficient. For more complex applications that require precise semantics, more expressive models, i.e., ontologies, are required. Costs of development and maintenance of models have to be tied to use cases and requirements, initially as they exist but also as they evolve over time. A larger cost initially that engenders an ascending benefit over time may be preferable to a much lower initial cost that generates a plateau or even descending slope of accrued benefit, sometimes within a short period of time. Figure 2 [1] notionally depicts this tradeoff between the cost and complexity of the semantic model developed and the prorated value over time of the benefit of using such a model, including reduced maintenance costs. In recent years, there have also emerged better models to estimate the cost of developing ontologies, such as ONTOCOM [2, 3, 4] which also include estimation software in the form of detailed spreadsheets.2
Figure 2. Approximate Cost/Benefit of Moving up the Ontology Spectrum: From Simpler Taxonomies to Ontologies
1.2. Emerging Technologies for Ontologies and the Semantic Web Among the technologies focused on the Semantic Web, in recent years, a number stand out as potentially very useful, for many kinds of applications, but especially for intelligence analysis. These technologies include query languages, repository structures, and rules for rule-based reasoning and interchange. There are also many more kinds of 2
ONTOCOM tools. http://ontocom.sti-innsbruck.at/tools.htm.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
216
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
inference engines, from description logic based classificational reasoners to first-order logic and logic programming based reasoners, and many new end-user Semantic Web applications. This section discusses the SPARQL Protocol and RDF Query Language (SPARQL) [5], triple-stores, the Rule Interchange Format (RIF) [6], and a range of inference engines. These technologies built on the established and more mature Semantic Web languages of the Resource Description Framework (RDF), the Resource Description Framework Schema (RDFS), and the Web Ontology Language OWL, the latter two of which are ontology description languages and the first is a graphstructured instance language. These languages became W3C standards in 2004. OWL 2, however, is a relatively new proposed standard, and increases the expressivity of OWL [7] by providing more datatype support, support for declarations and annotations on ontologies, and “syntactic sugar” for more succinctly and easily defining certain constructs in OWL. The SPARQL query language [5] is a standard graph-based query language defined by the W3C to work RDF triple stores (i.e., n-tuple stores) which are graphstructured, potentially exist anywhere on the Internet or within an Intranet, and are exposed as so-called SPARQL endpoints. A SPARQL endpoint is a way of indexing a triple store, typically by providing an International Resource Identifier (IRI), so that it is known to a query engine. Currently, SPARQL is defined only over RDF, but many Semantic Web inference engines have extended SPARQL support to include the ontology languages RDFS and OWL. These engines extend the SPARQL support by enabling ontology reasoning methods over the queries, in addition to strict retrieval of graph-based instances. Some representative triple stores are: OWLIM [8], Garlik 4Store [9], AllegroGraph [10], Jena [11], Sesame [12], Oracle 11g [13], Mulgara [14], and OpenLink Virtuoso [15]. Some triple stores advertise high storage sizes and various other high access, query, and load rates, with high-end triple stores reporting the ability to store billions of triples. But these claims are not yet independently confirmed. Also, many of these triple stores also support inference engines or link to existing inference engines, both those based on description logics (OWL is mostly a specific kind of description logic) and those based on more general logic and rules. Description logic reasoners include Pellet [16], RacerPro [17], FaCT/FaCT++ [18], etc. More general logic and rule reasoners include Jena, KAON2 [19], SILK [20], and various logic programming (Prolog) engines such as SWI-Prolog [21], Ciao Prolog [22], XSB Prolog [23], HighFleet’s (formerly Ontology Works) High Performance Knowledge Server [24], Cyc [25], and TopQuadrant’s TopBraid, [26] etc. Rules are IF/THEN constructs that specify constraints on classes, relations, and properties (see [27] for more discussion on rules and related Semantic Web notions, from which this section is adapted). They thereby constrain how new classes, relations, and properties are defined, prevent contradictory information from being added to a knowledge base, and enable discovery of new information without explicitly asserting the information. Examples of common rules are 1) a rule that prevents a “child” from being its own “parent”, and 2) a rule that says a “parent” of a “parent” that has a “child” is a “grandparent.” Rules are very closely associated with proof, i.e., rules require a proof mechanism to realize their value. In fact, a class of rules, inference rules, are directly associated with proof insofar as those inference rules license valid deductions as steps in an automated proof. The de facto standard Semantic Web Rule Language (SWRL) is an example of a language for expressing rules and is based on OWL [28]. There is an emerging W3C
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
217
standard rule language based on XML syntax called the RIF [6], which will probably supersede SWRL. Because RIF tries to accommodate many different kinds of rule engines and existing deployed and used implementations, RIF is structured into multiple versions, called dialects or profiles, including the following: Core: the fundamental RIF language and a common subset of most rule engines (providing a basic Datalog, where Datalog is a simplified logic programming language); BLD (Basic Logic Dialect): this adds to Core, by providing logic functions, equality in the then-part, and named arguments (providing a basic Horn Logic, which is the foundation of Prolog, the primary logic programming language) ; and PRD (Production Rules Dialect): this adds a notion of forward-chaining rules, where a rule fires and then performs some action, such as adding more information to the store or retracting some information (providing an expert system-like capability).
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
2. A Prospective Lesson for Intelligence: Realist Ontologies in Healthcare In the domain of healthcare information technology (HIT) it has been commonly accepted for some years now that both the development and use of clinical terminology should be supported by formal methods. Although this is a thesis that we strongly support, we wish no less strongly to insist that formal methods alone are not enough. The use of a Description Logic-based system appears, for example, not to have provided any guarantee for the absence of errors in SNOMED-CT [29], one of the most popular formal biomedical terminologies today. With the extremely positive response to the creation of the Open Biomedical Ontologies (OBO) Foundry [30] it became clear that a role had to be played by realist ontology in making better biomedical terminologies. Realist ontology helped in detecting errors and in ensuring intuitive principles for the creation and maintenance of systems of a sort that can help to prevent errors in the future. More importantly still, however, it helps in ensuring that terminologies are compatible with each other. Note that we say ‘realist ontology’, in order to distinguish ontology in our understanding from the various related things [31] which go by this term in contexts such as formal knowledge representation. It is a realist conception of ontology which underlies statements such as: The UMLS is an extensive source of biomedical concepts. It also provides a large number of inter-concept relationships and qualifies for a source of semantic spaces in the biomedical domain. However, the organization of knowledge in the UMLS is not principled nor consistent enough for it to qualify as an ontology of the biomedical domain [32] In the tradition of analytical philosophy, ontology is understood by the OBO Foundry community not as a software implementation or as a controlled vocabulary, but rather as ‘the science of what is, of the kinds and structures of objects, properties, events, processes and relations in every area of reality’ [33]. Ontology as it concerns us here is a theory of those higher-level categories which structure the biomedical domain, the representation of which needs to be both unified and fully coherent – and as closely allied as possible to the representations used by clinicians in formulating
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
218
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
patient data – if terminologies and coding systems are to have the requisite degree and type of interoperability. Ontology in this realist sense has successfully been used as a method to find inconsistencies in terminologies and clinical knowledge representations [34] such as the Gene Ontology [35] or the UMLS Semantic Network [36]. The method has also proved useful in drawing attention to certain problematic features of the HL7 RIM [37, 38, 39]. One of the major insights brought about by realist ontology in the healthcare domain is that biomedical terminologies can only be compared amongst each other, or used without loss of information within an electronic healthcare record (EHCR) system, if they share a common framework of top-level ontological categories [40]. Often one talks in this connection merely of a shared or common semantics, meaning thereby the sort of regimentation that can be ensured through the use of enabling technologies such as RDF(S) [41] and OWL [42] that currently enjoy a wide interest through their association with the Semantic Web project, not to forget systems such as Protégé that are able to cope with them in a user-friendly way [43]. On closer inspection, however, one discovers that the ‘semantics’ which comes with languages like RDF(S) and OWL is restricted to that sort of specification of meaning that can be effected using the formal technique of mathematical model theory, which is to say that meanings are specified by associating with the terms and sentences of a language certain abstract settheoretic structures, taking Alfred Tarski’s ‘semantic’ definition of truth for artificial languages as paradigm [44]. But model theory is metaphysically and ontologically almost completely neutral. Merely to formulate statements in a language such as OWL is far from building an ontology in the sense of ontology that is employed by analytical philosophers, and neither would translating a terminology into OWL turn it into an ontology. Such translation would indeed allow consistent reasoning about the ‘world’ – but only in the model-theoretic sense of ‘world’ that signifies not the flesh-and-blood reality with which biomedicine is concerned, but rather merely only some highly simplified set-theoretic surrogate. The task of ensuring that the latter somehow corresponds in broad terms to the real world of what happens and is the case, was in the semantics biomedical literature almost never addressed. Now it has become clear that the whole detour via semantic models is in fact superfluous: the job of ontology is not the construction of simplified models; rather, a biomedical ontology should directly correspond to reality itself in a manner that maximizes descriptive adequacy within the constraints of formal rigour and computational usefulness. Applying realist ontology to terminologies and EHCR architectures means in the first place applying it to those entities in reality to which these artifacts of the human intellect refer, such as concrete patients, diseases and therapies. We do this to serve at least one important goal, namely making terminologies coherent, both internally as well as in their relation to the EHCRs in or for which they are used. Already a very superficial analysis of a coding system such as the International Classification of Diseases [45] reveals that this system is not in fact a classification of diseases as entities in reality. Rather it is a classification of statements about disease phenomena which a physician might attribute to a patient. As an example, the ICD-10 class B83.9: Helminthiasis, unspecified does not refer (for example) to a disease caused by a worm belonging to the species unspecified which would be some sub-species of Acanthocephalia or Metastrongylia. Rather, it refers to a statement (perhaps appearing in some patient record) made by a physician who for whatever reason did not specify the actual type of Helminth the patient was suffering from. Neither OWL nor reasoners using models expressed in OWL would complain about making the class B83.9:
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
219
Helminthiasis, unspecified a subclass of B83: Other helminthiasis; from the point of view of a coherent ontology, however, such a view is nonsense: it rests precisely on a confusion between ontology and epistemology [46]. A similar confusion can be found in EHCR architectures, model specifications, message specifications or data types for EHCR systems. References to a patient’s gender/sex are a typical example. Some specifications refer to it as “administrative sex” (leaving it to the reader of the specification to determine what this might actually mean). The possible specifications of administrative sex are then female, male, unknown, or changed. Unknown, here, does not refer to a new and special type of gender (reflecting some novel scientific discovery); rather it refers to the fact that the actual gender is not documented in the record. An interpretation along these lines does not work in every case, however. Consider those specifications which refer explicitly to “clinical observations”, as is the case for Corbamed-COAS (“Clinical Observations Access Server’), which consists of:
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
any information that has been captured about a single patient’s medical/physical state and relevant context information. This [information] may be derived by instruments such as in the case of images, vital signs, and lab results or it may be derived by a health professional via direct examination of the patient and transcribed [sic]. This term applies to information that has been captured whether or not it has been reviewed by an appropriate authority to confirm its applicability to the patient record. [47] When in a EHCR system that claims to follow the COAS specifications the specification “unknown” would be registered for gender, then that specification has to be interpreted that an observation has been made with respect to the patient’s gender, and that as a result of that, an unknown kind of gender has been observed. Of course, that is not supposed to be the idea. European and international efforts towards standardization of biomedical terminology and electronic healthcare records were focused over the last 15 years primarily on syntax. Semantic standardization was restricted to terminological issues around the semantic triangle paradigm [48] on the one hand and to issues pertaining to knowledge representation (and resting primarily on the application of set-theoretic model theory) on the other hand. Moves in these directions are in indeed required, and the results obtained thus far are of value both for the advance of science and for the concrete use of healthcare telematic applications. We can safely say that the syntactical issues are now resolved and also that the problems relating to biomedical terminology (polysemy, synonymy, cross-mapping of terminologies, …) are well understood – at least in the community of specialized researchers. Now, however, it is time to solve these problems by using the theories and tools that have been developed so far, and that have been tested under laboratory conditions. This means using the right sort of ontology, i.e. an ontology that is able explicitly and unambiguously to relate coding systems, biomedical terminologies and electronic health care records (including their architecture) to the real world. To do this properly will require a huge effort, since the relevant existing standards need to be reviewed by experts who are familiar with the appropriate sort of ontological thinking (and this will require some effort in training and education). Even before that stage is reached, however, there is the problem of making all constituent parties –
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
220
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
including patients (or at least the organizations that stand up for them), healthcare providers, system developers and decision makers – aware of how deep-seated the existing problems are. Having been overwhelmed by the exaggerated claims on behalf of XML and similar silver bullets of recent years, that would solve everything, they must be informed about the fact that XML alone isn’t a silver bullet. And for sure, we must also be careful in not giving realist ontology a similar silver bullet status. The message of realist ontology is that, while there are various different views of the world, this world itself is one and unique. It is our belief that it is only through that world that the various different views can be compared and made compatible. To allow clinical data registered in electronic patient records by means of coding (and/or classification) systems to be used for further automated processing, it should be crystal clear whether entities in the coding system refer to diseases or rather to statements made about diseases, or to procedures and observations, rather than statements about procedures or observations. As such, coding systems used in or for electronic healthcare records should be given a precise and formal semantics that is coherent with the semantics of the record as well as with the real world parts that are described by them.
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
3. Intelligence, Ontologies, and Semantic Technologies The previous section underscored that realist ontologies are important, and they are important for intelligence collection and analysis. Realist ontologies are based on common understandings of the real world, and try to avoid conceptualist pitfalls (where concepts are introduced without direct origin in real world objects, relations, and properties) and epistemological, belief-based, or evidential (we use these terms synonymously) observational knowledge. The latter knowledge or approximations of knowledge are extremely important and are the basis of intelligence analysis and collection, but they largely address instance knowledge of the real world, i.e., individuals or particulars, about whom there may be many sources of data, much of which are contradictory. Why? Because the data being received from human and machine sensors are uncertain, error-prone, and often subject to noise, misinterpretation, and deception. It is very important to capture this incoming data, present it to intelligence analysts, and attempt to characterize it according to realist ontologies, but those realist ontologies describe the best knowledge of the real world we have, and the best knowledge of the general properties of that incoming data. They do not presume to be able to adjudicate which belief or observation is actually correct. That’s what an intelligence analyst does, when he/she stitches together the evidence, generates hypotheses, and then either confirms or assigns a value to those hypotheses according to some strength of belief or evidence. Ontology addresses the real entities, relations, and properties of the world; epistemology is about the perceived and belief-attributed entities, relations, and properties of the world, empirical evidence gleaned that will be described or characterized by ontology (see [49, 50], from which this is adapted, for further discussion of the differences between ontology and epistemology). Epistemology is employed in the use and qualification of data and actual data as stored in databases or tagged or indexed in documents. If ontology states that human beings have exactly one birth date, the data about a specific person is epistemological: in a given set of databases the person instance named John Smith (we assume we can uniquely
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
221
characterize this instance, but we may not) may have two or more attributed birth-dates, not one of which are known to be true. Ontology tells us that everything that lives has only one birth-date. Epistemology helps us understand how we can address which one of seven birth-dates is possibly the most accurate, i.e., true. Without ontology, there is no firm basis for epistemology. Epistemological concerns often distort and push off needed ontological distinctions. Why? Because analysts of information often believe that all is hypothesis and argumentation. They really don't understand the ontological part, i.e., that their knowledge is really based on firm stuff: a human being only has one birth date and one death date, though the evidence for that is multivarious, uncertain, and needs to be hypothesized about like the empirical, epistemological notion it is. Often also the charge that knowledge is just too “dynamic” is unjustified. Instance knowledge is very dynamic, i.e., the particular people, places, things, events we are interested in change all the time. My location is different from minute to minute. My activities change every minute, i.e., the events I participate in are new events that occur in time as time moves forward. I cut my hair or dye it. I marry, have children, divorce, move to another city, change jobs, go back to school, start a new hobby, make new friends, lose money in a new investment, watch and like different television programs, books, music, I eat different food and like different food. I think new thoughts and act on them. But the knowledge behind those instances largely remains the same. I am still a human. Families are still families. Organizations, work relationships, friendships, jobs, locations, kinds of events and activities, interests, etc., are the same. Occasionally this generic, ontological knowledge changes. For example, perhaps I join a new kind of organization where I pay the organization to work there. If this would occur (perhaps it’s unlikely), then my ontology about organizations would have to change, to reflect this new real world situation. Perhaps in the future, a collection of men and women can combine to provide genetic material to create a child – in which case, the ontological notion of parent will have to change. The notion of what a parent and a child is, is ontological; which people are the parents of which child is at least partially epistemological: we need evidence, but it is based on our ontological knowledge. Ontologies and semantic technologies are important and will be increasingly important for the intelligence community in the coming years. We have focused this book primarily on ontologies, representing the high end of semantic models, but semantic technologies more generally include a range of semantic models, some of which such as taxonomies, thesauri, and conceptual models are less expressive than ontologies, but useful for particular kinds of applications. Although predictions are notoriously problematic and often overcome by unanticipated events, we think we can make a number of predictions that will become true over the next ten to twenty years: •
The intelligence community will increasingly use semantic technologies in two forms: in the form of vocabularies that enable diverse sub-communities to use their own terms (words and phrases) to express their knowledge and queries, and in the form of ontologies that represent and model the meanings of those vocabularies, so that common information can be shared despite terminological differences among communities. The community of interest (COI) paradigm embraces these notions, and top-down and mid-level vocabularies such as UCore and Command and Control Common Core begin to address the vocabulary (syntactic) side, though are not yet sufficiently focused on the ontology (semantic) side. Largely this is because practictioners
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
222
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
•
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
•
•
are focused purely on XML technologies, and are knowledgeable primarily about database technologies, i.e., structural methods and local semantics only. This will change over time, is changing now. Ontologies and semantic technologies will increasingly provide a basis for intelligence collection and analysis to go beyond local and structural data models, which are the standard currently for the structured data of the database community, and beyond the primarily statistical models, which are the standard currently for the unstructured data of the natural language and information retrieval communities. Semantic analysis and interoperability will be seen to provide more capability over structural and statistical methods and models for sounder and more effective intelligence capture and analysis. For example, greater precision for responses to queries depends on better semantic representation of the data. As is already apparent, analysts and other kinds of users of documents and information, do not have the time to keep performing syntactic, free text searches ala Google – because they do not have the time to read or even skim the documents returned in the result sets of queries, to see whether those documents are really relevant to their queries. How can an analyst tell whether the real answer or best data exists in document 10,000, since he/she will never get to that document? Organizations will change to accommodate ontologies and semantic technologies. The primary issue with technological change is not technological, but sociological. The people and the organizations must change, for better technological methods to be employed to solve existing problems. Institutionalization of change is very hard. The intelligence community, like governmental and even commercial organizations in general, can accommodate technological change, but their sociological milieus and organizational structures in general cannot. Software acquisition processes are monolithic and even in the era of service oriented architecture (SOA), which tries to decompose the older systems and systems of systems into service atoms and molecules, organizational and process change really depends on heroes, i.e., managers and directors who are technologically aware and advocates for change, but who are in oversight and guidance positions for only two years. When they leave after two years, any progress they may have initiated and supported evaporates, and the institution is again left bereft. Ontologies and semantic technologies cannot solve sociological and organizational problems. Unfortunate events will propel change. This is what we all fear. Correlations will not be made, since the data stores are huge, the sources are immense, noise is rampant, collection and analysis resources are insufficient, and there are no overriding descriptions, models, rules, procedures, processes, nor organizational and sociological support that will enable evidence to be stitched together, described under common ontological, semantic, and epistemological characterizations, and acted upon in time to prevent bad events.
4. Cautious Optimism We remain cautiously optimistic about change for the intelligence community, and the prospects for ontologies and semantic technologies to propel those changes. If any
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
223
technology can be seen to enable a revolutionary leap forward for intelligence collection and analysis, if not for information technology in general, it is that of ontologies and semantic technologies. The authors of these chapters and the editors of this book are primarily technologists, and so, by predisposition, optimists about the use of technology to effectively achieve information-technological goals. But we are also realists, as our predisposition to realist ontologies indicates, and pragmatists: we are interested in these technologies, yes, but to serve a purpose, specifically to increase the effectiveness of intelligence collection and analysis. We are interested in technologies serving a purpose, and from our perspective, the best service for ontologies and semantic technologies is to enable the intelligence community, as it is for other scientific communities, to characterize the real world and thereby find out the truth and the probabilities that surround that truth, and so prevent, correct, and adjust to events that threaten nations and peoples. We wish us all sensible heads, sound technologies, stout hearts, and good luck.
References
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
[1]
Obrst, Leo. 2009. Ontological Architectures. Chapter 2 in Part One: Ontology as Technology in the book: TAO – Theory and Applications of Ontology, Volume 2: The Information-science Stance, Michael Healy, Achilles Kameas, Roberto Poli, eds. Forthcoming, Springer. [2] Simperl, Elena Paslaru Bontas; Christoph Tempich; York Sure. 2006. ONTOCOM: A Cost Estimation Model for Ontology Engineering. Proceedings of the International Semantic Web Conference ISWC 2006. http://ontocom.ag-nbi.de/. [3] Elena, Simperl; Igor O. Popov; Tobias Bürger. 2009. ONTOCOM Revisited: Towards Accurate Cost Predictions for Ontology Development Projects. In: Proceedings of the European Semantic Web Conference 2009 (ESWC '09), Heraklion, Greece, May 20 - Jun, 04, 2009. [4] Imtiaz, Ali; Tobias Bürger; Igor O. Popov; Elena Simperl. 2009. Framework for Value Prediction of Knowledge-based Applications. 1st Workshop on the Economics of Knowledge-based Technologies (ECONOM 2009) in conjunction with 12th International Conference on Business Information Systems (BIS 2009) April 27, 28 or 29, 2009, Poznan, Poland. [5] World Wide Web Consortium (W3C), SPARQL Query Language for RDF, http://www.w3.org/TR/rdfsparql-query/, (2008). [6] World Wide Web Consortium (W3C), Rule Interchange Format (RIF), http://www.w3.org/2005/rules/wiki/RIF_Working_Group, (2009). [7] World Wide Web Consortium (W3C), OWL 2.0, http://www.w3.org/TR/2009/PR-owl2-new-features20090922, (2009). [8] Ontotext, OWLIM Semantic Repository. http://www.ontotext.com/owlim/, (2009). [9] Garlik 4Store. http://4store.org/. [10] Franz Inc., AllegroGraph RDFStore™, http://www.franz.com/agraph/allegrograph/, (2009). [11] Hewlett Packard (HP) Labs, Jena, http://www.hpl.hp.com/semweb/, (2009). [12] Aduna (OpenRDF community), Sesame, http://www.openrdf.org/. [13] Oracle Database 11g. http://www.oracle.com/technology/tech/semantic_technologies/index.html. [14] Mulgara Community, Mulgara Semantic Store, http://www.mulgara.org/, (2009). [15] OpenLink Software, Virtuoso Universal Server Platform, http://virtuoso.openlinksw.com/ , (2009). [16] Pellet. http://clarkparsia.com/pellet/. [17] Renamed ABox and Concept Expression Reasoner (RACER) Pro. http://www.racersystems.com/products/racerpro/index.phtml. [18] FaCT++. http://owl.cs.manchester.ac.uk/fact++/. [19] KAON2. http://kaon2.semanticweb.org/. [20] Semantic Inferencing on Large Knowledge (SILK). http://silk.semwebcentral.org/. [21] SWI-Prolog. http://www.swi-prolog.org/. [22] Ciao Prolog. http://clip.dia.fi.upm.es/Software/Ciao/. [23] XSB Prolog. http://xsb.sourceforge.net/. [24] HighFleet’s High Performance Knowledge Server. http://www.highfleet.com/. [25] Cycorp’s Cyc. http://www.cyc.com/,
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
224
L. Obrst et al. / Ontologies, Semantic Technologies, and Intelligence: Looking Toward the Future
[26] TopQuadrant’s TopBraid. http://www.topquadrant.com/. [27] Parmelee, Mary; Leo Obrst. 2010. Technologies for Metadata, Vocabulary, and Ontology Storage. Chapter in: Miguel-Angel Sicilia, ed. Handbook of Metadata, Semantics and Ontologies. World Scientific Publishing Co. Forthcoming. [28] Horrocks, Ian, Peter F. Patel-Schneider, Harold Boley, Said Tabet, Benjamin Grosof, Mike Dean, SWRL: A Semantic Web Rule Language Combining OWL and RuleML, W3C Member Submission. http://www.w3.org/Submission/SWRL/, (2004). [29] Ceusters W, Smith B, Kumar A, Dhaen C. Mistakes in Medical Ontologies: Where Do They Come From and How Can They Be Detected? in Pisanelli DM (ed) "Ontologies in Medicine. Proceedings of the Workshop on Medical Ontologies, Rome October 2003" IOS Press, Studies in Health Technology and Informatics, vol 102, 2004. [30] Smith B, Ashburner M, Rosse C, Bard J, Bug W, Ceusters W, Goldberg LJ, Eilbeck K, Ireland A,Mungall CJ, the OBI Consortium, Leontis N, Rocca-Serra P, Ruttenberg A, Sansone SA, Shah N, Whetzel PL, Lewis S. The OBO Foundry: Coordinated Evolution of Ontologies to Support Biomedical Data Integration. Nature Biotechnology 2007;25:1251-1255. [31] Guarino, N; P. Giaretta, "Ontologies and Knowledge Bases: Towards a Terminological Clarification". In Towards Very Large Knowledge Bases: Knowledge Building and Knowledge Sharing, N. Mars (ed.), pp 25-32. IOS Press, Amsterdam, 1995. [32] Bodenreider O. Medical Ontology Research: A Report to the Board of Scientific Counselors of the Lister Hill National Center for Biomedical Communications. May 17, 2001 (http://etbsun2.nlm.nih.gov:8000/pubs/pdf/2001-MOR-BoSC.pdf) [33] Smith, B. Ontology, in L. Floridi (ed.), Blackwell Guide to the Philosophy of Computing and Information, Oxford: Blackwell, 2003, 155–166 [34] Smith B, Ceusters W. Towards Industrial-Strength Philosophy; How Analytical Ontology Can Help Medical Informatics. Interdisciplinary Science Reviews, 2003, vol 28, no 2, 106-111. [35] Smith, Barry, Jacob Köhler, Anand Kumar: On the Application of Formal Principles to Life Science Data: a Case Study in the Gene Ontology. In: Erhard Rahm (Ed.): Data Integration in the Life Sciences, First International Workshop, DILS 2004, Leipzig, Germany, March 25-26, 2004, Proceedings. Lecture Notes in Computer Science 2994, Springer 2004, 79-94. [36] Schulze-Kremer S, Smith B, Kumar A. Revising the UMLS Semantic Network Medinfo 2004. Proc. Medinfo 2004. [37] Lowell V: Actions in Health Care Organizations: An Ontological Analysis in: Proceedings of MedInfo 2004, San Francisco. [38] Lowell V, Smith B: Speech Acts and Medical Records: The Ontological Nexus. in: EuroMISE 2004, Prague. [39] Smith B, Ceusters W. HL7 RIM: An Incoherent Standard, Stud Health Technol Inform. 2006;124:133138. (Presented at MIE2006) [40] Smith B, Ceusters W. An Ontology-Based Methodology for the Migration of Biomedical Terminologies to Electronic Health Records. AMIA 2005, October 22-26, Washington DC;:669-673. [41] RDF Semantics. W3C Recommendation 10 February 2004 (http://www.w3.org/TR/rdf-mt/) [42] OWL Web Ontology Language Semantics and Abstract Syntax. W3C Recommendation 10 February 2004 (http://www.w3.org/TR/owl-semantics/) [43] Protégé OWL plug-in (http://protege.stanford.edu/plugins/owl/) [44] Model Theory. Stanford Encyclopedia of Philosophy (http://plato.stanford.edu/entries/model-theory/) [45] World Health Organisation. ICD-10 - The International Statistical Classification of Diseases and Related Health Problems, tenth revision (http://www.who.int/whosis/icd10/). [46] Bodenreider O, Smith B, Burgun A. The Ontology-Epistemology Divide: Case Study in Medical Terminology. Submitted to the Third International Conference on Formal Ontology (FOIS) 2004. [47] 3M, Care Data Systems, Inc., CareFlow/Net, Inc., HBO & Company, Philips Medical Systems, Protocol Systems, Inc. CORBAMED-COAS: clinical observations access server specification. Version 1, April 2001. (http://www.medcom.dk/picnic/deliverables/01-04-06%20coas%20specs.pdf. [48] Ogden, C.K., & Richards, I.A. (1927). Meaning of meaning. New York: Harcourt, Brace & Company. [49] Obrst, Leo. 2010. Ontological Architectures. Chapter 2 in Part One: Ontology as Technology in the book: TAO – Theory and Applications of Ontology, Volume 2: The Information-science Stance, Michael Healy, Achilles Kameas, Roberto Poli, eds. Forthcoming, Springer. [50] Poli, Roberto; Leo Obrst. 2010. The Interplay Between Ontology as Categorial Analysis and Ontology as Technology. Chapter 9 in Part One: Ontology as Technology in the book: TAO – Theory and Applications of Ontology, Volume 2: The Information-science Stance, Michael Healy, Achilles Kameas, Roberto Poli, eds. Forthcoming, Springer.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved.
225
Subject Index 71 71 163 37 71 163 71 57 71 163 213 37 129 1, 213 71 57 1, 213 71
logic 109 logic programming 71 Malaysia 37 METS 129 ontology(ies) 1, 71, 109, 129, 213 provability-based semantic interoperability 109 reasoning 57 referent tracking 13 semantic technology(ies) 1, 57, 213 semantic web 71 service-oriented architecture 71 situation awareness 37 situational awareness 71 translation graphs 109 unmanned aerial vehicles 109 web ontology language (OWL) 71, 129
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
agile systems automated reasoning Bayesian reasoning blog mining C2 causal ontology command and control common logic enterprise integration evidence management healthcare human terrain information extraction information-sharing intelligence intelligence analysis intelligence community knowledge compilation
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Ontologies and Semantic Technologies for Intelligence L. Obrst et al. (Eds.) IOS Press, 2010 © 2010 The authors and IOS Press. All rights reserved.
227
Author Index 57 109 v, 1, 13, 213 109 147 57, 91, 179 71 163 v, 1, 57, 147, 213 37 91, 179 147 129 13 37
McCandless, D. Nichols, D. Obrst, L. Prausa, M. Ressler, J. Self, T. Shilliday, A. Smith, B. Stoutenburg, S. Sward, R. Taylor, J. Ulicny, B. Valtorta, M.G. Wang, J.
71 71 v, 1, 71, 213 71 179 91 109 57 71 71 109 37 163 163
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
Basik, H. Bringsjord, S. Ceusters, W. Clark, M. Costa, P.C.G. Dean, M. Franklin, P. Huhns, M.N. Janssen, T. Kokar, M.M. Kolas, D. Laskey, K.B. Lee, R. Manzoor, S. Matheus, C.J.
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook
Copyright © 2010. IOS Press, Incorporated. All rights reserved.
This page intentionally left blank
Ontologies and Semantic Technologies for Intelligence, edited by L. Obrst, et al., IOS Press, Incorporated, 2010. ProQuest Ebook