Weaving the Semantic Web: Contributions and Insights

Home , Calculus ratiocinator, ENQUIRE, SAPPHIRE (Health care), Semantic integration, Semantic wiki, Upper ontology

Anne Cregan

Doctoral Thesis submitted in partial requirement of the award of PhD from the University of New South Wales

September 2008

Research reported in this thesis has been partially financed by NICTA (http://www.nicta.com.au). NICTA is funded by the Australian Government’s Department of Communications, Information Technology and the Arts, and the Australian Research Council through Backing Australia’s Ability and the ICT Centre of Excellence program. It is supported by its members the Australian National University, University of NSW, ACT Government, NSW Government and affiliate partner University of Sydney. Abstract

The semantic web aims to make the meaning of data on the web explicit and machine processable. Harking back to Leibniz in its vision, it imagines a world of interlinked information that computers ‘understand’ and ‘know’ how to process based on its meaning. Spearheaded by the World Wide Web Consortium, ontology languages OWL and RDF form the core of the current technical offerings. RDF has successfully enabled the construction of virtually unlimited webs of data, whilst OWL gives the ability to express complex relationships between RDF data triples. However, the formal semantics of these languages limit themselves to that aspect of meaning that can be captured by mechanical inference rules, leaving many open questions as to other aspects of meaning and how they might be made machine processable. The Semantic Web has faced a number of problems that are addressed by the included publications. Its germination within academia, and logical semantics has seen it struggle to become familiar, accessible and implementable for the general IT population, so an overview of semantic technologies is provided. Faced with competing ‘semantic’ languages, such as the ISO’s Topic Map standards, a method for building ISO-compliant Topic Maps in the OWL DL language has been provided, enabling them to take advantage of the more mature OWL language and tools. Supplementation with rules is needed to deal with many real-world scenarios and this is explored as a practical exercise. The available syntaxes for OWL have hindered domain experts in ontology building, so a natural language syntax for OWL designed for use by non-logicians is offered and compared with similar offerings. In recent years, proliferation of ontologies has resulted in far more than are needed in any given domain space, so a mechanism is proposed to facilitate the reuse of existing ontologies by giving contextual information and leveraging social factors to encourage wider adoption of common ontologies and achieve interoperability. Lastly, the question of meaning is addressed in relation to the need to define one’s terms and to ground one’s symbols by anchoring them effectively, ultimately providing the foundation for evolving a ‘Pragmatic Web’ of action.

i Oﬃcial Statements

I hereby grant the University of New South Wales or its agents the right to archive and to make available my thesis or dissertation in whole or part in the University libraries in all forms of media, now or here after known, subject to the provisions of the Copyright Act 1968. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertation. I also authorise University Microﬁlms to use the 350 word abstract of my thesis in Disser- tation Abstracts International (this is applicable to doctoral thesis only). I have either used no substantial portions of copyright material in my thesis or I have obtained permission to use copyright material; where permission has not been granted I have applied/will apply for a partial restriction of the digital copy of my thesis or dissertation.

Signature: Date:

Authenticity Statement

I certify that the Library deposit digital copy is a direct equivalent of the ﬁnal oﬃcially approved version of my thesis. No emendation of content has occurred and if there are any minor variations in formatting, they are the result of the conversion to digital format.

Signature: Date:

Originality Statement

I hereby declare that this submission is my own work and to the best of my knowledge it contains no materials previously published or written by another person, or substantial proportions of material which have been accepted for the award of any other degree or diploma at UNSW or any other educational institution, except where due acknowledgment is made in the thesis. Any contribution made to the research by others, with whom I have worked at UNSW or elsewhere, is explicitly acknowledged in the thesis. I also declare that the intellectual content of this thesis is the product of my own work, except to the extent that assistance from others in the projects design and conception or in style, presentation and linguistic expression is acknowledged.

Signature: Date:

iii Preface

The Semantic Web is a fascinating enterprise with the aim of inter-linking data on a massive scale and machine processing it based on its semantics. It is a mammoth effort, now involving thousands of people, of which I am merely one. Whilst much has been achieved, some of the questions regarding meaning and making it machine processable are tantalizingly unresolved. Our generation has a great opportunity to address this question and to decide how it will be tackled. This thesis details some of the contributions and insights I as one of those thousands of people have offered to the overall process I like to refer to as ‘Weaving the Semantic Web’, an expression that tips its hat to world wide web and semantic web inventor Sir Tim Berners-Lee’s famous book titled ‘Weaving the Web’. As a UNSW student sponsored by NICTA, Australia’s Centre of Excellence for Infor- mation and Communications Technology, I have embraced NICTA’s values of use-inspired research and close engagement with real-world users of technology. The publications included in this thesis thus reflect a focus on delivering what has been or is needed to make the Semantic Web as usable as possible, as well as attempting to deliver on its vision of machine processable meaning. Some of the included publications have helped the semantic web along its evolutionary path to get where it is now, some are at the forefront of its current evolution, whilst others are very long-sighted and examine what is needed to reach the ultimate goal of machines that ‘understand’ information and ‘know’ how to process it according to its meaning. Such an effort involves many activities, of which only some result are visible as the publications included in this thesis: others result in community building, application building, education and outreach and material such as WG reports and recommendations. I have been involved in many of these activities and would like to thank all those who have ac- companied me on this exciting journey, particularly my colleagues at NICTA, and fellow members of the W3C’s OWL Working Group and Reasoning on the Web Incubator Group (UR-W3). I would also like to thank my supervisors Emeritus Professor Norman Foo, Dr Thomas Meyer and Dr Maurice Pagnucco, my family and particularly James for their invaluable support throughout the time I have been engaged in this work.

Anne Cregan September 2008.

v Structure of the Thesis

As the Semantic Web is a very dynamic field of activity and my work has focused on active areas of research and development, it is very difficult to offer a complete doctoral thesis that is current at the time of submission. Therefore, in accordance with the regulations of UNSW, I have opted to submit this thesis as a series of publications. Each of these has been researched and published or accepted for publication within the period of my PhD candidature at UNSW, and has been subjected to a peer review process prior to publication. The publications are grouped into chapters according to commonality of content. Each publication is preceded by details of where and when it was published and the author’s personal contribution to each publication, acknowledging the contribution of others where the candidate is not the sole author. The personal contribution percentage was arrived at by asking each co-author to estimate my personal contribution to the publication, and averaging their estimates. The copyright permission of each co-author to reproduce the publication was also obtained. The thesis contains an introductory chapter, Chapter 1, that introduces the Semantic Web, places the published works in the context of problems facing the Semantic Web, and summarizes each publication’s content and contribution. The publications themselves are presented in Chapters 2 through 7 in their published formats, overlaid with pagination within this thesis. In summary: • Chapter 2 Overview of Semantic Technologies is an overview of semantic technologies, which explains the field to mainstream information technologists in order to make it more accessible and to encourage wider adoption of the technologies. In the context of this thesis, it provides background information that is relevant for setting the scene for the following chapters, which are essentially independent of each other. • Chapter 3 Integrating Topic Maps into the Semantic Web details work conducted to map the ISO’s Topic Map standards into the W3C’s Semantic Web stack of technologies, so that it might take advantage of the formal semantics and tools associated with the latter. • Chapter 4 Adding Rules to OWL Ontologies describes joint work conducted to explore the need for, and use of rules in conjunction with OWL ontologies, editors and reasoners. • Chapter 5 Controlled Natural Language Syntaxes for OWL describes joint work on designing a controlled natural language syntax for OWL 2. The initial work on Sydney OWL Syntax was presented at OWLED in 2007, where a CNL task force formed to explore various alternatives, reported in the second paper of this chapter. • Chapter 6 Encouraging Ontology Reuse details joint work on the issue of ontology reuse, its importance in semantic interoperability, and plans for a tool which would encourage reuse by leveraging social factors and networking tools. • Chapter 7 Foundational Issues in Meaning examines the issue of meaning and definition at a fundamental level, and its importance for Semantic Web technologies. Following these chapters, a concluding chapter, Chapter 8, summarises the overall conclusions of the published works, articulates the contributions made to the Semantic Web, and examines the extent to which the publications addressed the problems raised in the introductory chapter. It also outlines the remaining issues and makes recommendations for future work.

vii Contents

1 Introduction 1 1.1 Leibniz’s Dream in the 21st Century ...... 1 1.2 The Semantic Web Vision ...... 3 1.3 The Semantic Web Eﬀort ...... 5 1.3.1 W3C Semantic Web technology stack ...... 5 1.3.2 Additional Items ...... 8 1.3.3 Semantic Web Tools and Applications ...... 9 1.4 A Web of Data ...... 10 1.4.1 RDF Graphs ...... 10 1.4.2 Interlinking for Semantic Interoperability ...... 11 1.4.3 The Linking Open Data Project ...... 12 1.5 Describing Data using Ontology Languages ...... 12 1.5.1 What is an Ontology? ...... 13 1.5.2 RDF/OWL Ontologies ...... 13 1.5.3 RDF-Schema ...... 13 1.5.4 RDF-Schema Vocabulary ...... 14 1.5.5 RDF-Schema Design Choices ...... 15 1.5.6 Web Ontology Language (OWL) ...... 15 1.5.7 OWL Vocabulary ...... 17 1.5.8 Diﬀerences between Sublanguages of OWL ...... 19 1.5.9 OWL 2 Extensions ...... 20 1.5.10 OWL Design Choices ...... 21 1.6 Notations ...... 21 1.6.1 RDF Abstract Syntax ...... 21 1.6.2 RDF Transfer Syntax and OWL Exchange Syntax ...... 22 1.6.3 OWL Abstract Syntax ...... 22 1.6.4 Non-Normative Syntaxes ...... 23 1.7 Formal Semantics ...... 25 1.7.1 Reasoning Services ...... 26 1.7.2 Model Theoretic Semantics ...... 27 1.7.3 Meaning and Formal Semantics ...... 28 1.8 Beyond Model Theory ...... 29 1.9 Problems Facing the Semantic Web ...... 30 1.9.1 Adoption and Accessibility ...... 30 1.9.2 Competing Standards for Building Semantic Structures ...... 31 1.9.3 Rules and the limits of Ontology Languages ...... 32 1.9.4 Readable Syntax ...... 32 1.9.5 Interoperability of Ontologies ...... 32 1.9.6 Making Meaning Machine Processable ...... 33 1.10 Included Publications ...... 34 1.10.1 Chapter 2: Overview of Semantic Technologies ...... 34 1.10.2 Chapter 3: Integrating Topic Maps into the Semantic Web ...... 37 1.10.3 Chapter 4: Adding Rules to OWL DL Ontologies ...... 40 1.10.4 Chapter 5: Natural Language Syntax for OWL ...... 42

ix CONTENTS

1.10.5 Chapter 6: Encouraging Ontology Reuse ...... 45 1.10.6 Chapter 7: Foundational Issues in Meaning ...... 48

2 Overview of Semantic Technologies 55 2.1 Overview of Semantic Technologies ...... 55

3 Integrating Topic Maps into the Semantic Web 77 3.1 Building Topic Maps in OWL-DL ...... 77 3.2 An OWL DL Construction for the ISO Topic Map Data Model ...... 113

4 Adding Rules to OWL DL Ontologies 127 4.1 Pushing the limits of OWL, Rules & Prot´eg´e ...... 127 4.2 Exploring OWL & Rules ...... 139

5 Controlled Natural Language Syntaxes for OWL 161 5.1 Sydney OWL Syntax ...... 161 5.2 A Comparison of three CNLs for OWL 1.1 ...... 173

6 Encouraging Ontology Reuse 185 6.1 n2mate ...... 185

7 Foundational Issues in Meaning 193 7.1 Towards a Science of Deﬁnition ...... 193 7.2 Symbol Grounding for the Semantic Web ...... 203

8 Conclusion 219 8.1 Overview of Semantic Technologies ...... 219 8.2 Competing Standards ...... 220 8.3 Rules and the Limits of Ontology Languages ...... 222 8.4 Readable Syntax ...... 224 8.5 Interoperability of Ontologies ...... 226 8.6 Making Meaning Machine Processable ...... 227 8.6.1 Further Analysis ...... 228 8.6.2 Building Situation Awareness from Sensor Input ...... 229 8.6.3 A Strategy for Symbol Grounding ...... 231 8.7 Summation ...... 234 List of Figures

1.1 Berners-Lee and Swick’s current Semantic Web Technology Stack or ‘layer cake’...... 6 1.2 Example of an RDF Triple ...... 10 1.3 RDF Graphs may be interlinked to form a semantic structure interlinking disparate datasets. Image courtesy of Ivan Herman, Semantic Web Actibity Lead...... 11 1.4 The Linking Open Data dataset cloud, representing an RDF graph of over two billion RDF triples interlinked by around 3 million RDF links, as at October 2007 ...... 12 1.5 RDF/XML syntax shown in the SWOOP ontology editor ...... 22 1.6 OWL DL syntax shown in the Protégéontology editor ...... 23 1.7 OWL Abstract Syntax shown in the SWOOP ontology editor ...... 24 1.8 OWL Turtle syntax shown in the SWOOP ontology editor ...... 25 1.9 OWL Manchester Syntax shown in the Protégéontology editor ...... 26 1.10 Lars Marius Garshol’s diagram of ISO Topic Map standards vs W3C standards, 2005 ...... 31 1.11 Vocabularies exist within social and business contexts ...... 47 1.12 Sample n2mate interface screen ...... 48 1.13 Semiotic Triangle - Odgen and Richard’s 1923 version ...... 51

8.1 SAIL Prototype System Architecture ...... 230

xi 1 Introduction

1.1 Leibniz’s Dream in the 21st Century

The dream of representing all human knowledge and reasoning with it may be traced back at least as far as the great German polymath Gottfried Leibniz (1646-1716), who dreamed of a universal language which could represent any concept, combined with a theoretical logical calculation framework for the purpose of reasoning. In 1677 in his Preface to the General Science [82], Leibniz observed

“It is obvious that if we could ﬁnd characters or signs suited for expressing all our thoughts as clearly and as exactly as arithmetic expresses numbers or geometry expresses lines, we could do in all matters insofar as they are subject to reasoning all that we can do in arithmetic and geometry. For all investigations which depend on reasoning would be carried out by transposing these characters and by a species of calculus.”

Leibniz identified two components as necessary for the realization of this dream (as recorded by Hintikka [69]): 1. a characteristica universalis or lingua characteristica, which was to be a universal language of human thought, and would have a symbolic structure directly reflecting the structure of the world of human concepts. This ‘alphabet of human thought’ would be a universal and formal language able to express mathematical, scientific, and metaphysical concepts. 2. a calculus ratiocinator, which was a method of symbolic calculation which would mirror the processes of human reasoning. This was to provide a framework for universal logical calculation, where any difference of opinion could be resolved by recourse to calculation. Using the characteristica universalis and calculus ratiocinator as instruments, Leibniz believed it would be possible to construct an encyclopedia that would be the key to all the sciences and a compendium of all human knowledge, and hoped to complete such a project. However, whilst his lifetime achievements were nothing short of monumental, he unfortunately never achieved this dream. In a March 1706 letter to the Electress Sophia of Hanover, he wrote:

“It is true that I once planned a new method of calculation proper to subjects having nothing in common with mathematics, and if this manner of Logic were put into practice, all reasoning, even analogical ones, would be carried out in a

1 Chapter 1 Introduction

mathematical way. Then modest intellects could, with diligence and good will, not accompany but at least follow greater ones. For one could always say ”let us calculate” and judge properly, insofar as reason and the data can furnish us the means to do so. But I do not know whether I will ever be able to execute such a project, one requiring more than one hand, and it would even seem that humanity is not yet suﬃciently mature to pretend to the advantages to which this method could lead.” Translation of a passage reproduced in [26] :p118, 1706.

Whilst unachieved in his lifetime, Leibniz’s grand vision has continued to inspire others through the centuries. Hintikka argues that Leibniz’s vision inspired two strands of research tradition, corresponding to the two components identified [69]. He argues that the ‘algebraic school’, represented by Boole, Peirce, and Schröder,were inspired by the calculus ratiocinator and sought to develop mathematical techniques for capturing forms of reasoning, whilst the characteristica universalis was the primary aim of Frege’s Begriffsschrift [43]. However, Frege’s approach was later criticized by Husserl and Tarski who argued that a lingua characteristica cannot be purely formal. This criticism seems appropriate in light of the Characteristic Universalis being intended to capture the world of human thought: unless it can be shown that human thought operates as a formal system, the approach does indeed seem to require some means to bridge between whatever the language of human thought is on the one hand, and the logical formalism on the other. Leibniz himself appreciated the difficulty of constructing a suitable characteristica universalis, but remarked “I think that some selected men could finish the matter in five years” [98] :p224. Unfortunately however, Leibniz never described any operational details or methods for attacking the project. The very idea has even been derided by some philosophers as an absurd fantasy: “Leibniz’s views about the systematic character of all knowledge are linked with his plans for a universal symbolism, a Characteristica Universalis. This was to be a calculus which would cover all thought, and replace controversy by calculation. The ideal now seems absurdly optimistic...” [119] :p. ix ) The challenge of creating a language which can be used to capture any human concept whilst simultaneously supporting formal reasoning is indeed a particularly challenging one. It must answer deep questions about the nature of the world and how the mind represents it, as well as the nature of (correct) reasoning, and then grapple with the question of how to connect the two so that appropriate and accurate logical calculations may be made. Today, the semantic web [16] is facing this problem. Using ontologies both as a means of capturing conceptualizations, and as a framework for reasoning by virtue of their description logic formal semantics, it echoes Leibniz’s notions of being able to both capture any concept and to apply symbolic calculation to provide sound reasoning and the calculation of truth. Assisted by the Semantic Web’s interlinking techniques, the digital age now provides the means to implement the formal methods Leibniz inspired on an unprecedented scale across vast distributed arrays of data, and potentially deliver the results of reasoning processes virtually instantaneously. Where Leibniz imagined creating an encyclopedia that would be the work of a few learned men, the Semantic Web can take advantage of the conceptual knowledge of multitudes of men and women. And today, we may hope that humanity, or at least a significant portion of it, is now mature enough to appreciate the advantages such an advance can bring. However, the unresolved questions raised by the lingua characteristica continue to loom. Whilst the formal semantics of ontologies and their associated inferencing procedures are firmly rooted in the algebraic tradition, the ontologies themselves usually seek to represent 3 some domain of interest that ventures outside the realm of formal logics and into the conceptual and physical worlds: that is, they seek to describe concepts in the minds of human beings, as well as concrete things in the world that exist independently not only of formal systems, but also of language and human concepts. The question of bridging between the world of formal methods and the worlds of cognitive processes and physical things seems to be one that the Semantic Web must, by its very nature, address. The underlying challenges are addressed in depth in the later sections on formal semantics at Sections 1.7 and 1.8 and foundational issues in meaning at Section 1.10.6 and Chapter 7.

1.2 The Semantic Web Vision

In inventing the World Wide Web, Sir Tim Berners-Lee was seeking to implement a mechanism that would be both a collaborative framework for sharing knowledge, and would also be machine-understandable so that “the networks, operating systems and commands.. become invisible, and leave us with an intuitive interface as directly as possible to the information” [13]. The World Wide Web which took oﬀ so dramatically in the 1990s provided the ability to share electronic information on an unprecedented scale, using the simple mechanisms of hypertext and unique resource identiﬁers on top of the existing Internet Transmission Control Protocol TCP/IP. Notably, its main achievement was the ability to display information to humans, rather than the ability for machines to process that information. However, Berners-Lee’s original plan had also envisaged that the Web should have semantics that would make the information shared on it machine processable. His original ‘Enquire’ program created at CERN, on which his plan for the web was based, had

“assumed that every page was about something. When you created a new page it made you say what sort of thing it was: a person, a piece of machinery, a group, a program, a concept, etc. Not only that, when you created a link between two nodes, it would prompt you to ﬁll in the relationship between the two things or people. For example, the relationships were deﬁned as ‘A is part of B’ or ‘A made B’.” [13].

In order to make the World Wide Web a web of well-deﬁned, machine-processable information, Berners-Lee proposed a need for what he called a ‘Semantic Web’ in the late 1990s.

“I have a dream for the Web [in which] machines become capable of analyzing all the data on the Web - the content, links, and transactions between people and computers. A ‘Semantic Web’, which should make this possible, has yet to emerge, but when it does, the day-to-day mechanisms of trade, bureaucracy and our daily lives will be handled by machines talking to machines.... The ‘intelligent agents’ people have touted for ages will ﬁnally materialize.” Tim Berners-Lee, ‘Weaving the Web’ p: 169, 1999 [15].

The Scientific American article of 2001 [16], which launched the Semantic Web vision on a wider public, described a utopian future in which unified data would be utilized by intelligent agents and accessed via any desktop or handheld device to perform tasks like finding medical practitioners that met certain criteria and automatically booking medical appointments at the most appropriate locations and times. In this vision, machines had the ability to process information according to its meaning, and the article even referred to Chapter 1 Introduction computers being able to ‘understand’ information and ‘know’ how to process it according to its meaning. In this article, the idea of defining meaning and making it machine processable was paramount, with Berners-Lee, Hendler and Lassila stating that “the Semantic Web is an extension of the current web in which information is given well-defined meaning, better enabling computers and people to work in co-operation” [16]. The World Wide Web Consortium (W3C) leads the co-ordination of Semantic Web activities worldwide, and its current Semantic Web Activity Statement [154] explains the Semantic Web as a vision as follows:

“The Web can reach its full potential only if it becomes a place where data can be shared and processed by automated tools as well as by people. For the Web to scale, tomorrow’s programs must be able to share and process data even when these programs have been designed totally independently. The Semantic Web is a vision: the idea of having data on the web deﬁned and linked in a way that it can be used by machines not just for display purposes, but for automation, integration and reuse of data across various applications.”

As the Semantic Web technologies have developed, the Semantic Web is no longer only a vision, but also refers to the technical recommendations published by the W3C and the various applications that implement those recommendations. The December 2007 Scientiﬁc American article [39] which revisited and updated the progress on the 2001 vision [16], emphasised this aspect of the Semantic Web. Keeping the notion of understanding and meaning in the picture, it described the Semantic Web as:

“A set of formats and languages that ﬁnd and analyze data on the World Wide Web, allowing consumers and businesses to understand all kinds of useful online information.”

Today in 2008, the World Wide Web Consortium (W3C) describes the Semantic Web on its home page as follows [157]:

“The Semantic Web is about two things. It is about common formats for integration and combination of data drawn from diverse sources, where on the original Web mainly concentrated on the interchange of documents. It is also about language for recording how the data relates to real world objects. That allows a person, or a machine, to start oﬀ in one database, and then move through an unending set of databases which are connected not by wires but by being about the same thing.”

Clearly the idea of creating a coherent view of the world that machines can traverse and use is an important part of the Semantic Web vision. The theme of making the meaning of information explicit and machine processable, so that machines ‘understand’ and ‘know’ what to do with it, is also a recurrent one in the Semantic Web vision. It is reasonable then to assume that the published W3C recommendations and the applications that adhere to them are, at least to some extent, intending to deliver this or at least to move towards it as a goal. The following sections present the Semantic Web technologies that have been delivered to date, as well as some of the W3C’s future plans, and then consider to what extent these are actually able to deliver this envisaged ‘web of meaning’. 5

1.3 The Semantic Web Eﬀort

The W3C is leading the Semantic Web implementation by providing language and technology recommendations that build on existing W3C-endorsed World Wide Web standards. Other organizations, both academic and corporate, are providing tools such as editors and reasoners that implement the W3C recommendations. Semantic Web applications use ontologies and reasoners in conjunction with traditional web programming technologies and databases.

1.3.1 W3C Semantic Web technology stack As the W3C’s Semantic Web Activity statement [154] describes, the principal technologies of the Semantic Web fit into a set of layered specifications, with core components being RDF and OWL. As shown in Berners-Lee and Swick’s current Semantic Web ‘layer cake’ diagram [17] shown in Figure 1.1, the W3C has sought to implement a stack of technologies and standards, each building on and extending the achievements of the previous layer. Broadly speaking, the lower layers (up to RDF) are completed and stable, the middle layers (OWL and SPARQL) have finalized initial, implementable offerings but are still undergoing evolution and extension, (RIF is currently under active development) whilst the higher and rightmost layers (Unifying Logic, Proof, Trust and Crypto) are in an exploratory phase where suitable approaches are being identified. A description of each element of the stack and its current status follows.

• Unicode: a standardized character set and encoding for computing purposes, allowing most characters of the world’s writing systems to be represented. Status: Unicode is developed by the Unicode Consortium in conjunction with the International Organization for Standardization and its character set is captured in ISO/IEC 10646: the Universal Character Set. It is commonly used across many computing platforms and is not speciﬁc to the Semantic Web.

• URI: Uniform Resource Identiﬁer. A string of characters used to uniquely identify or name a resource. The resource may be anything at all: an information resource, a real world thing like a person, an idea or anything else. If it is an information resource, the URI can usually be dereferenced to obtain the actual resource through a protocol such as http or ftp. Status: URIs are one of the basic mechanisms underpinning the World Wide Web. The current generic URI syntax speciﬁcation is RFC 3986 / STD 66 (2005).

• XML: Extensible Markup Language. A general-purpose specification for creating custom markup languages by allowing users to define and structure their own elements. XML may be used to impose both a syntax and structure to which conforming documents must adhere. It facilitates the sharing of structured data across information systems. Status: XML was formalized via the W3C in the late 1990s and the current W3C XML Recommendation is dated August 2006 [20]. It is used across many computing applications and is not specific to the Semantic Web.

• RDF: Resource Description Framework. Allows the user to make statements about resources in the form of subject-predicate-object expressions, called ’triples’ in RDF terminology. These directed triples combine to form an RDF graph, described Chapter 1 Introduction

Figure 1.1 Berners-Lee and Swick’s current Semantic Web Technology Stack or ‘layer cake’

in detail in Section 1.4.1. The standard exchange syntax is XML. Status: RDF ﬁrst became a W3C recommendation in 1999 [94]; the current revision is dated 10 February 2004 [34].

• RDF-S: RDF-Schema. RDF’s Vocabulary Description Language, RDF-Schema is an extensible knowledge representation language, providing basic elements such as classes and properties intended to structure RDF resources. It is described in detail at Sections 1.5.3 and 1.5.4. Status: The ﬁrst version was published by W3C in April 1998, and the ﬁnal W3C recommendation, introducing class and subclass relations and domain and range on properties was released in February 2004 [21].

• SPARQL: SPARQL Protocol and RDF Query Language. SPARQL is a language for querying and updating RDF graphs. It is designed to query collections of RDF triples either natively or via middleware, and has similar querying functionality for an RDF knowledge base as SQL has for relational databases. Output may be 7

represented either as RDF graphs or as result sets. A SPARQL endpoint enables users to query a knowledge base via the SPARQL language, and may be used either by humans, software or Web Services. Status: SPARQL became a W3C Recommendation on 15th January 2008 [132].

• OWL: Web Ontology Language. An ontology language for the Web which extends the core RDF-S components. OWL has three ﬂavours: Lite, DL (Description Logic) and Full, each of which adds additional expressivity. OWL is described in detail at Section 1.5.6. Status: The original OWL speciﬁcations became a W3C Recommendation on 10 February 2004 [31, 102, 122, 144]. A W3C OWL Working group is currently working on further extensions to the OWL language originally referred to as OWL1.1 [121] and currently as OWL 2 [108].

• RIF: rule interchange format. The W3C’s RIF Working Group [156] is chartered to produce a core rule language plus extensions which together allow rules to be translated between rule languages and thus transferred between rule systems. Rather than creating a new rule language, the focus is on creating a standard interchange format for existing rule languages such as the Semantic Web Rule Language (SWRL) [76]. Status: This work is currently in active development - six RIF working documents were issued for comment by the W3C in August 2008.

• SWRL: Semantic Web Rule Language. SWRL [76] is not shown explicitly in the Semantic Web stack but is included in this list as it is the most commonly used Semantic Web rule language. It utilises Horn clauses - essentially the intended meaning of a SWRL rule is that whenever all the antecedent conditions in the clause are true, the consequent condition must also hold. SWRL’s abstract syntax extends that of OWL and the formal meaning of OWL ontologies with SWRL rules is given as an extension of OWL’s model-theoretic semantics. For compatibility with RDF and OWL, SWRL also uses URIs for resource identiﬁcation. SWRL contains builtins for comparisons, mathematical operations, Boolean values, strings, date and time, URIs and lists. Status: SWRL became a W3C member submission in May 2004.

Whilst the lower layers are well-deﬁned, little has been formally written about the upper layers and ‘Crypto’. Thus these are the author’s interpretations based on her own W3C involvement.

• Unifying Logic The logical underpinning or formal semantics of RDF, RDF-S, OWL, SPARQL and RIF that allows them to be used in conjunction with inference engines. This semantics allows them to interoperate on shared data and produce consistent results. As expert John Sowa put it in a 2007 email to Pat Hayes which copied in the ontolog-forum [147], ”Unifying Logic is the framework that includes the others as subsets: RDF, RDF-S, Rule RIF, OWL, and SPARQL. Each of these subsets is tailored for a specific kind of inference engine and/or a specific range of uses. What unifies them is the common model-theoretic semantics. That semantics enables all of them to interoperate on shared data and produce consistent results.”

• Proof Formal Semantics gives the ability to provide proofs for the inferences made on the semantic web. In addition, it is desirable not only to give a formal proof, but Chapter 1 Introduction

also to provide a human understandable justification for the inference, such as that given by SWOOP [86]. • Trust Providing proof and justification for inferences made gives users the confidence to trust the results of machine processing. Trust is essentially a social and subjective human phenomena, simply supported by the technologies. However, the work of the W3C Protocol for Web Description Resources (POWDER) Working Group described at Section 1.3.2 is relevant to both Proof and Trust for its ability to provide verifiable assertions about Web resources. • Crypto The inclusion of this element in the Semantic Web stack is simply an acknowledgement of the importance of cryptography for ensuring the privacy and security of data shared and distributed via the web. • User Interface and Applications This recent (late 2006) inclusion to the Semantic Web technology stack is an acknowledgement of the importance of the user interface. Whilst the end user will see the user interface and applications, the semantic technologies that power these and make them appear to ‘understand’ the user and provide machine processing of information are ‘under the hood’ and thus should essentially be transparent to the user.

1.3.2 Additional Items The W3C Semantic Web Activity statement [154] includes some additional activities and specifications of interest outside those shown explicitly in the Semantic Web layer cake diagram, including those of the GRDDL, POWDER and SKOS working groups, summarized below. • GRDDL: (Gleaning Resource Descriptions from Dialects of Languages) is a mechanism for embedding extractable RDF in XML and XHTML documents. Using the GRDDL specification [33], one may use markup based on existing standards to declare that an XML document includes Resource Description Framework (RDF) data and link to algorithms (typically represented in XSLT) for extracting this data from the document. It thus provides a useful bridge between RDF and existing web standards. GRDDL became a W3C Recommendation on 11 September 2007. • POWDER: (Protocol for Web Description Resources) As its charter explains [151], this W3C working group’s mission is to develop a mechanism through which structured metadata (“Description Resources”) can be authenticated and applied to groups of Web resources. This mechanism will allow retrieval of the description resources without retrieval of the resources they describe, thus providing the ability to make verifiable assertions about groups of web resources. Being in RDF, it can be joined to other RDF data and thus utilized by RDF-based searches and other services. • SKOS: (Simple Knowledge Organization System) [152] is developing standards for representing the basic structure and content of concept schemes such as thesauri, classification schemes, subject heading lists, and other types of controlled vocabulary in RDF. Many of the knowledge organization systems mentioned share a similar structure, and this commonality is described by the SKOS data model, which describes and documents the concepts used in such schemes, making them explicit and share- able via the Web. SKOS provides language for capturing concepts, definitions and (linguistic)semantic relations between terms such as broader or narrower. The SKOS data model is formally defined as an OWL Full ontology. 9

1.3.3 Semantic Web Tools and Applications Semantic Web tools provide functionality for building, using and mapping ontologies. Some of the key areas are:

• Editors to assist in ontology building. These provide a graphical user interface for building and displaying ontologies. Usually they work in conjunction with reasoners for consistency checking. Some leading examples include the free, open source Prot´eg´e [41] editor developed by Stanford Center for Biomedical Informatics Research, and the commercial product TopBraid Composer [149], amongst many others.

• Programming Environments for building Semantic Web applications include Jena [130] and Sesame [1] for Java development and many others across a wide range of programming languages including Python, C, C++, C#, Javascript, PHP, Lisp, Prolog, Perl, Ruby and Haskell.

• Reasoners to implement the formal semantics of the ontology language and provide various reasoning services including consistency checking, inferencing, etc. Leading examples include RacerPro [60], FaCT++ [114], Pellet [96], and KAON2 [42].

• Mapping and Alignment Ontologies often need to be mapped together: tools can assist in identifying the classes and individuals to be mapped and to do so via a graphical interface. Leading examples include MAFRA [143].

• Querying tools provide support for query building and execution over ontologies, including provision of SPARQL endpoints e.g. the Talis Platform [56].

• Browsers traverse RDF graphs and retrieve data directly by dereferencing URIs. Leading examples include Disco [11] (primarily server side) and Tabulator [37] (client side).

• RDF Data Stores As Semantic Web applications may involve many thousands or millions of RDF triples, database technology is needed to store and access these on demand. Sesame [1] provides an open source RDF database with support for RDF Schema inferencing and querying, whilst Oracle has started to oﬀer RDF data store support from version 10gR2 onwards [25].

Early in its development, Semantic Web applications tended to be small prototypes built in academic environments. The yearly Semantic Web challenge [52] has charted the evolution of Semantic Web applications into large scale real-life software. The 2007 Scientific American article [39] which reported on the overall progress of the Semantic Web, mentions a number of leading real world examples of applied Semantic Web technology. It covers Semantic Web tools provided on Science Commons for sharing of scientific data, which are used by Pfizer and others for drug discovery, and public health problem detection via a system called SAPPHIRE, which integrates and uses data from local health care provider, hospitals, environmental protection agencies and scientific literature using semantic web technologies. Currently Semantic Web technologies are at the exploratory and early adopter phases, rather than being in mainstream use. However, it is likely that there may soon be a shift towards mainstream adoption, as more traditional technologists become aware of these technologies and start to integrate them into corporate and government frameworks, and researchers join new commercial enterprises to apply semantic technologies in commercial and public domains. Chapter 1 Introduction

Figure 1.2 Example of an RDF Triple

1.4 A Web of Data

As discussed above, the semantic web is envisaged to be a web of data that has well-deﬁned meaning. In order to become a web it requires some kind of mechanism for interlinking data. RDF [101] provides this mechanism, enabling data in the form of URIs, data values and blank nodes to be structured into a directed graph of RDF triples, referred to as an ‘RDF Graph’.

1.4.1 RDF Graphs The key idea of RDF [101] is that it structures data into directed triples that may be combined to form a directed graph. The triples correspond to Subject-Predicate-Object assertions, and the resulting graph may be traversed by RDF Browsers such as Tabulator [37] or Disco [11], and may be queried using the SPARQL [132] Query Language. An RDF graph may also be serialized as XML [34]. An RDF triple contains three components: 1. the subject, which is an RDF URI reference or a blank node 2. the predicate, which points from subject to object, and is an RDF URI reference 3. the object, which is an RDF URI reference, a literal or a blank node RDF triples correspond to assertions that connect subjects to objects via predicates, also referred to as properties. As shown in the example at Figure 1.2, an RDF triple may assert that the resource identiﬁed by www.nicta.com.au/people/cregana (Subject) has foaf:ﬁrstname(Predicate) of “Anne” (Object). An RDF triple is conventionally written in the order subject, predicate, object. As each element of the triple may be a URI, RDF enables distributed information to be woven together. An RDF graph is simply a set of RDF triples. The set of nodes of an RDF graph is the set of subjects and objects of triples in the graph, and its edges are the predicates of those triples.

Literals A literal in the object position is simply a data value. XML Schema (xsd) Datatypes [18] may be used if a datatype for the literal is required. 11

Figure 1.3 RDF Graphs may be interlinked to form a semantic structure interlinking disparate datasets. Image courtesy of Ivan Herman, Semantic Web Actibity Lead.

Blank Nodes

Blank nodes are used in the object position when an extra node is needed for the purpose of structuring the data model but there is no need for the node to be externally accessible, and thus it does not need to be named and its identity may be left blank.

Reiﬁcation

RDF also has vocabulary for reiﬁcation which permits an RDF triple to be reiﬁed as a resource in its own right, enabling other RDF triples to make assertions about it.

1.4.2 Interlinking for Semantic Interoperability

RDF graphs may be interlinked on common nodes to form a larger RDF Graph, as shown at Figure 1.3. Such a structure provides semantic interoperability between the datasets intermapped, as the structure created may then be queried and processed as a whole. Exposing data held in distinct data stores through the use of RDF provides the ability to visualize, query and reason across the combined datasets regardless of origin. Chapter 1 Introduction

Figure 1.4 The Linking Open Data dataset cloud, representing an RDF graph of over two billion RDF triples interlinked by around 3 million RDF links, as at October 2007

1.4.3 The Linking Open Data Project

Recently a W3C Semantic Web Education and Outreach (SWEO) Community Project called ‘Linking Open Data’ [36] has created what is essentially an open data commons in the form of an RDF graph. By publishing open data as RDF triples and linking on common URIs, an RDF graph has been created that consists of over two billion RDF triples interlinked by around 3 million RDF links (as at October 2007). The dataset cloud, shown at Figure 1.4, represents the largest RDF graph created to date. The construction of this shared, interlinked RDF graph may be taken as incontrovertible evidence that the ‘Semantic Web’ as a web of data on a massive scale is no longer just an idea, but is an actual artefact with a real and tangible existence.

1.5 Describing Data using Ontology Languages

Whilst RDF [101] provides the ability to construct directed RDF graphs, the ability to describe data is provided by ontology languages which include RDF-Schema [21] and the various ﬂavours of the Web Ontology Language OWL [144]. 13

1.5.1 What is an Ontology? The word ‘ontology’ is borrowed from philosophy, where it refers to the study of what exists. Gruber’s 1993 definition [58] that “an ontology is a specification of a conceptualization”, is the one most often cited in IT and AI contexts. According to Noy and McGuinness [113], who have been closely involved in both the evolution of OWL [102, 144] and of the ontology building tool Protégé[41], an ontology defines a common vocabulary for researchers who need to share information in a domain, and includes machine-interpretable definitions of basic concepts in the domain and relations among them. In detail, Noy and McGuinness state “an ontology is a formal explicit description of concepts in a domain of discourse (classes (sometimes called concepts)), properties of each concept describing various features and attributes of the concept (slots (sometimes called roles or properties)), and restrictions on slots (facets (sometimes called role restrictions)). An ontology together with a set of individual instances of classes constitutes a knowledge base.”

1.5.2 RDF/OWL Ontologies Embracing Noy and McGuinness’s notion of ontology in the context of RDF/RDF-S and OWL, an ontology may be considered to be a structure that describes individuals, classes, properties and the relationships between them. Intuitively, the basic entities of an ontology may be introduced as follows:

1. Individuals. These are the basic entities of the universe of discourse. Typically, these exist independently of the knowledge structure, and are named rather than deﬁned.

2. Classes. These are intended to group individuals who have something in common. Classes are often also referred to as Concepts, particularly within the Description Logic tradition.

3. Properties. These may be thought of as attributes of individuals, that connect them to other individuals or to data values. At the class level, these may be restricted in various ways to reﬂect the entities being described.

1.5.3 RDF-Schema As the W3C’s RDF Schema Recommendation [21] explains, whilst RDF properties represent relationships between resources, they provide no mechanisms for describing these properties, or the relationships between these properties and other resources . This is the role of the RDF vocabulary description language, RDF Schema [21], which is a semantic extension of RDF that provides mechanisms for describing groups of related resources and the relationships between these resources. Rather than providing a vocabulary of specific descriptive properties such as ‘author’, RDF Schema specifies mechanisms that may be used to name and describe properties and the classes of resource they describe in general. RDF Schema thus provides vocabulary descriptions written in RDF that may be used to determine the characteristics of other resources. For instance, RDF Schema permits the specification of domains and ranges of RDF properties, by virtue of creating the constructs rdfs:domain and rdfs:range for use in specifying the domains and ranges of specific user-defined RDF properties (instances of rdf:Property). Chapter 1 Introduction

1.5.4 RDF-Schema Vocabulary In the following discussion, the prefix rdf: indicates use of the namespace http://www.w3. org/1999/02/22-rdf-syntax-ns# and the prefix rdfs: indicates use of the namespace http: //www.w3.org/2000/01/rdf-schema#. Following the RDF Schema specification [21], RDF Schema is considered to include constructs such as rdf:type that were originally introduced by the 1999 RDF Model and Syntax specification and are in the rdf: namespace. ‘Appendix A: RDF Schema as RDF/XML’ of the W3C’s February 2004 RDF-Schema Recommendation [21] is taken as indicative of the vocabulary included in RDF-Schema.

Resources The most inclusive construct of RDF Schema is rdfs:Resource. All things described by RDF are considered to be resources, and are instances of the class rdfs:Resource, which may be considered to be the class that includes everything. Note that resources are intended to encompass anything at all that users may wish to refer to, and should be considered to encompass real world objects and human conceptualizations as well as information resources. URIs provide the mechanism for indicating which resource is being referred to.

Classes, SubClasses and Instances RDF Schema introduces rdfs:Class, the class of classes. Using this construct in combination with rdf:type it becomes possible to assert that a given RDF resource is an instance of an RDF class. The RDF property rdfs:subClassOf then permits RDF classes to be arranged in a subsumption hierarchy. Note that all other classes are a subclass of the class rdfs:Resource, whilst rdfs:Resource is itself an instance of rdfs:Class. The formal semantics of such assertions has ramiﬁcations that are discussed in the following Section 1.5.5.

Properties and SubProperties rdf:Property is the class of RDF properties, a subclass of rdfs:Resource. rdfs:subPropertyOf enables one to assert that the pairs connected by a particular RDF Property are a subset of the extension of pairs connected by the parent Property.

Domains and Ranges The RDF-S domain and range mechanisms rdfs:domain and rdfs:range enable speciﬁcation of the classes that are the domain and range of an RDF property. These may be considered to constrain the property: the property may only connect pairs from these classes.

Other RDF Schema vocabulary RDF Schema also provides vocabulary for:

• specifying lists and containers. This includes rdfs:List and rdfs:Container and its associated subclasses rdf:Bag, rdf:Seq and rdf:Alt.

• reifying RDF statements so they may themselves be treated as RDF resources. This is enabled by the RDF class rdf:Statement, and RDF properties rdf:subject, rdf: predicate, rdf:object. 15

• annotation purposes, that is, for humans to refer to. This includes the RDF properties rdfs:label, rdfs:comment, rdfs:seeAlso and rdfs:isDeﬁnedBy.

1.5.5 RDF-Schema Design Choices

Whilst the RDF vocabulary description language class and property system is similar to the type systems of object-oriented programming languages such as Java [54], it differs from many such systems in that instead of defining a class in terms of the properties its instances may have, it describes properties in terms of the classes of resource to which they apply. For example, the rdfs:subClassOf property described above is characterized by having both its domain and range set to rdfs:class, whilst rdfs:class is not defined in terms of the properties it may have. This design choice was made in order to enable adding to the definition of existing resources after their creation. One of the architectural principles of both the Web and the Semantic Web is openness: to different viewpoints and to the addition of more material. This design choice embraces the ‘Open World’ principle: when an RDF-Schema class is created, there is no assumption that all its properties are specified, so anyone may add additional properties to it.

1.5.6 Web Ontology Language (OWL)

The Web Ontology Language OWL [31, 102, 122, 144] provides additional vocabulary for describing data. It facilitates greater machine interpretability of Web content than that supported by XML, RDF, and RDF Schema (RDF-S), by providing a formal semantics for this additional vocabulary [122].

OWL Ontologies

An OWL ontology contains a sequence of axioms, annotations and facts. Annotations on OWL ontologies may be used to record human-readable information associated with an ontology, including authorship and imports references to other ontologies. However, the main logical content of an OWL ontology is carried by its axioms and facts. These provide information about the classes, properties, and individuals of the ontology. As will be described below, classes, properties, individuals and the relationships between them make up the core content of an ontology.

OWL Ontologies and Meaning

OWL Ontologies are also intended to be a way of representing meaning explicitly. Quoting from the OWL Web Ontology Language Overview [102]:

“OWL can be used to explicitly represent the meaning of terms in vocabularies and the relationships between those terms. This representation of terms and their interrelationships is called an ontology. OWL has more facilities for expressing meaning and semantics than XML, RDF, and RDF-S, and thus OWL goes beyond these languages in its ability to represent machine interpretable content on the Web. OWL is a revision of the DAML+OIL web ontology language incorporating lessons learned from the design and application of DAML+OIL.” Chapter 1 Introduction

OWL Ontologies as Virtual Structures OWL ontologies may import other ontologies. owl:imports is an annotation property that has the additional eﬀect of importing the target ontology. Additionally, every aspect of an ontology other than literals and blank nodes is a URI, and may therefore reside anywhere. Whilst an ontology usually has its own namespace, it may include classes, properties and individuals from anywhere on the web. Therefore, an ontology should be considered to be a virtual structure that connects various pre-existing imported ontologies and referenced URIs with newly created material.

OWL Sublanguages The 2004 OWL Specifications [31,102,122,144] encompassed three sub-languages or ‘flavours’ of OWL that provided increasing levels of expressivity: OWL Lite, OWL DL, and OWL Full. Each flavour of OWL fully includes the previous one: every OWL-Lite ontology is an OWL-DL ontology, and every OWL-DL ontology is an OWL-Full ontology, whilst the inverses do not hold. However, whilst OWL Lite and OWL DL are logically decidable, OWL Full is not.

• OWL Lite was designed to support users needing only a classification hierarchy and simple constraints, providing a quick migration path for existing artefacts such as thesauri. Whilst supporting cardinality constraints, these are limited to values of 0 or 1. • OWL DL was designed to be in the ‘sweet spot’, maximizing expressiveness whilst retaining decidability. DL stands for Description Logic, which provides the formal semantics for both OWL DL and OWL Lite. Description Logics [5] are a decidable fragment of First Order Logics, with their own specific notation. OWL DL includes all the OWL language constructs but limits their use: for example, cardinality restrictions may not be placed upon properties which are declared to be transitive. • OWL Full is based on a different semantics from OWL Lite or OWL DL, and was designed to preserve compatibility with RDF Schema by including all the RDF Schema vocabulary as well as all the OWL vocabulary. OWL Full is thus far more expressive than OWL DL: for example, in OWL Full a class can be treated simultaneously as a collection of individuals and as an individual in its own right; whilst this is not permitted in OWL DL. Due to its greater expressiveness, reasoning in OWL Full is undecidable.

Relationship between OWL and RDF/RDF-Schema Whilst OWL goes beyond RDF and RDF-Schema in its expressivity, it should be noted that only OWL Full includes the entire RDF Schema vocabulary. Thus whilst OWL Full can be viewed as an extension of RDF, OWL Lite and OWL DL must be viewed as extensions of a restricted view of RDF. Therefore, whilst every valid RDF document is a valid OWL Full ontology, it is not necessarily a valid OWL Lite or OWL DL ontology. RDF/RDF-Schema allow higher-order and non-extensional classes, whilst these are not allowed by either OWL Lite or OWL DL. For instance, RDF/RDF-Schema allows a resource to be both an individual and a class, and to be an instance of itself, whilst OWL DL and OWL Lite require a strict separation of classes, properties and individuals, and thus do not utilize all the RDF/RDF-Schema vocabulary. Speciﬁc details of the restrictions and vocabulary for each sublanguage are given at Section 1.5.8. 17

Reasoning over OWL Ontologies The reason for these restrictions is to guarantee tractable reasoning over OWL Lite and OWL DL ontologies. OWL DL has a Description Logic formal semantics that gives it tractable reasoning properties that OWL Full lacks. OWL Lite also has a Description Logic formal semantics with a lower computational complexity than OWL DL and is thus faster to run. Roughly speaking, OWL Lite and OWL DL correspond respectively to the description logics known as SHIF(D) and SHOIN(D), with some limitation on how datatypes are treated [74]. Reasoning support for OWL Full is very limited and generally speaking, the use of OWL Full is not advisable if reasoning services are required.

1.5.7 OWL Vocabulary Compared to RDF/RDF-Schema, the OWL vocabulary provides additional capability for constructing class deﬁnitions and for restricting properties, as well as stating equality and inequality between classes, properties and individuals respectively. The discussion of vocabulary below includes all the OWL DL vocabulary. OWL DL and OWL Full both include all the OWL vocabulary, (but as explained above, OWL DL has certain restrictions relative to OWL Full and does not include the full RDF/RDF-S vocabulary), whilst OWL-Lite is a strict subset of OWL-DL. Speciﬁc details of the restrictions and vocabulary for each sublanguage of OWL are given at Section 1.5.8.

OWL Classes and Individuals According to the OWL Web Ontology language overview [102], an OWL class is intended to deﬁne a group of individuals that belong together because they share some properties. An OWL class is created and named by using owl:Class and setting rdf:ID to a text string corresponding to the chosen class name e.g. < owl:Class rdf:ID =“Person”\ > creates an OWL class called “Person”. An individual is asserted to be an instance of an OWL class using rdf:type. Every OWL individual must belong to at least one class, but may belong to more than one class (note: in OWL Full exceptions apply). OWL introduces a most general class named owl:Thing, that is the class of all individuals and a superclass of all OWL classes, and a class named owl:Nothing, that is the class that has no instances and is a subclass of all OWL classes. Thus every OWL class is a subclass of owl:Thing and a superclass of owl:Nothing. OWL uses rdfs:subClassOf to deﬁne class hierarchies. An OWL class may be a subclass of multiple OWL classes. As OWL individuals may be instances of more than one class, and classes may be subclasses of multiple classes, an OWL ontology is a graph rather than a strict tree hierarchy.

OWL Class Descriptions OWL provides the ability to deﬁne classes using what is referred to as a ‘Class Description’. The mechanisms available for deﬁning classes in OWL using Class Descriptions are the following:

1. Enumeration. A class may be deﬁned to be exactly a set of enumerated individuals, using owl:oneOf. For example, a class named Month may be deﬁned as exactly the twelve individuals January, February, March,...., December. Chapter 1 Introduction

2. Boolean combinations of other classes. A class may be deﬁned to be a Boolean union, intersection or complement of other classes, using owl:unionOf, owl:intersectionOf and owl:complementOf.

3. Restriction. Restrictions may be used to deﬁne a class by deﬁning local constraints on properties. Restrictions use owl:allValuesFrom and owl:someValuesFrom to limit permissible values. This may be done over either data values via owl:DataProperty, or individuals via owl:ObjectProperty. Restrictions may also specify how many values are allowed, using owl:Cardinality, owl:minCardinality and owl:maxCardinality.

4. A combination of the above.

OWL Class Axioms Class Descriptions may form part of class axioms that specify whether the description is complete or partial. A complete modality indicates that the class is exactly equivalent to the given description, whereas a partial modality indicates the class is a subclass of the given description. Other class axioms include:

1. owl:EnumeratedClass, which makes a class consist of exactly a set of enumerated individuals.

2. owl:DisjointClasses which lists classes required to be pairwise disjoint.

3. owl:subClassOf which asserts one class to be a subClass of another class.

4. owl:equivalentClasses which states that two or more classes are equivalent. This is generally used in mapping classes from diﬀerent source ontologies.

OWL Properties In terms of logical axioms, the two key kinds of OWL properties are Data Properties (owl:dataProperty), which connect individuals to data values, and Object Properties (owl: ObjectProperty) which connect individuals to other individuals. The OWL vocabulary also includes ontology properties (owl:ontologyProperty), which are used to describe OWL ontologies, and annotation properties (owl:annotationProperty), which enable any aspect of the ontology to be annotated with information intended for humans.

OWL Property Axioms Property restrictions may be used within class descriptions to deﬁne OWL classes as described above. There are also a number of other axioms which allow OWL properties (either object or data properties) to:

1. be arranged in subsumption hierarchies, using owl:subPropertyOf.

2. be declared to be equivalent, using owl:equivalentProperty. As with owl:equivalentClass, this is generally used in ontology mapping.

3. be declared to be functional i.e. to have no more than one target value or individual.

4. be declared to have a speciﬁc class or class description as an rdfs:domain. 19

5. be declared to have a speciﬁc class (in the case of object properties) or data range (in the case of data properties) as an rdfs:range.

Additionally, there are a number of axioms that are applicable only to object properties, that enable one to assert that an OWL object property is:

1. the inverse of another object property, using owl:inverseOf.

2. a symmetric property, that is, if it holds between individuals A and B, it also holds between individuals B and A.

3. an inverse Functional property, that is, there is no more than one domain individual for which it holds.

4. a transitive property, that is, if it holds between individuals A and B, and between individuals B and C, then it also holds between individuals A and C.

OWL Facts Facts give the ability to assert information about OWL individuals. There are two kinds of OWL facts:

1. those that assert information about a particular individual, including the classes that the individual belongs to and the properties/values of that individual. These use the vocabulary constructs rdf:type, owl:objectProperty and owl:dataProperty described previously.

2. those that make individuals the same or distinct. Since OWL makes no unique name assumption, two identifiers may possibly refer to the same individual. This may be asserted explicitly using owl:sameAs, whilst owl:differentFrom may be used to assert that two individual identifiers refer to distinct individuals. Additionally, owl: allDifferent allows a list of any number of individual identifiers to be asserted to be all distinct individuals.

1.5.8 Diﬀerences between Sublanguages of OWL Both OWL DL and OWL Full use the complete OWL vocabulary described above, but OWL DL is subject to some restrictions that do not apply in OWL Full, due to the exclusion of some RDF/RDF-Schema features. OWL Lite has additional restrictions compared to OWL DL and uses only a subset of the OWL vocabulary.

Incremental Diﬀerences between OWL Full and OWL DL Unlike OWL Full, OWL DL requires that:

• Every URI used as a class name must be explicitly asserted to be of type owl:Class, not rdfs:class.

• Similarly, properties must be either owl:DataProperty or owl:ObjectProperty rather than rdf:property. (Note: OWL also allows for ontology and annotation properties).

• Every OWL-DL individual must be asserted to belong to at least one class (even if only owl:Thing). Chapter 1 Introduction

• Strict separation of types is required, thus the URIs used for classes, properties and individuals must be mutually disjoint: an OWL class may not be a property or instance; an OWL property may not be a class or individual, and an OWL individual may not be a class or property. OWL Full thus allows classes to be used as individuals, whereas OWL DL does not. This also implies that restrictions cannot be applied to the language elements of OWL itself in OWL DL, whereas this is allowed in OWL Full.

Incremental Diﬀerences between OWL DL and OWL Lite OWL Lite obeys all the restrictions of OWL DL laid out above, and additionally restricts OWL DL in the following ways:

1. Whereas OWL DL permits class to be described as an enumeration of individuals, OWL Lite does not.

2. In OWL DL, a property can be required to have a speciﬁc individual as a value using hasValue; OWL Lite does not permit this.

3. OWL DL permits classes to be stated to be disjoint; OWL Lite does not.

4. OWL DL permits classes to be described as arbitrary Boolean combination and restrictions whereas OWL Lite does not.

5. Whilst OWL Lite still permits the use of cardinality vocabulary, the cardinalities themselves are restricted to O or 1, whilst OWL DL permits the use of arbitrary non-negative integers.

6. In many constructs, OWL Lite restricts the syntax to use single class names rather than the arbitrarily complex class descriptions allowed by OWL DL.

1.5.9 OWL 2 Extensions A W3C OWL Working Group is currently developing extensions to the OWL language, to be released as OWL 2. OWL 2 [108] is based on an OWL 1.1 member submission dated 19 December 2006 [121] and, with the exception of syntax, is planned to be fully backwards compatible with the 2004 OWL Recommendations, oﬀering a small but useful set of additional features that have been requested by users and for which which eﬀective reasoning algorithms are now available. It should be noted that what was being referred to as OWL 1.1 in 2006 and 2007 has undergone a name change and is now referred to as OWL 2. New features planned for inclusion in OWL 2 [108, 121] are:

• extra syntactic sugar to make some commonly used patterns easier to state. This includes DisjointUnion, to deﬁne a description as the disjoint union of a set of descriptions, and NegativeObjectPropertyAssertion and NegativeDataProp- ertyAssertion, for use in stating negative property membership assertions.

• additional property and qualiﬁed cardinality constructors. OWL 2 moves from the SHOIN Description Logic underlying OWL DL to the SROIQ Description Logic, for which reasoning algorithms are now available. Consequently, OWL 2 provides extra Description Logic expressive power for qualiﬁed cardinality restrictions, property chain inclusion axioms (provided there are no cyclic inclusions), and (for non-complex 21

properties) local reflexivity restrictions, reflexive, irreflexive, symmetric, asymmetric and disjoint properties.

• extended datatype support, to allow user-deﬁned datatypes.

• simple metamodelling, referred to as ‘punning’, which allows a name to be used simultaneously as any or all of an individual, class or property. Whilst they may share a name, the underlying individual, class and/or property are carefully preserved as distinct entities that do not impinge on each other in order to avoid the computational problems encountered when a strict separation of types is not enforced, as is the case in RDF/RDF-Schema and OWL Full.

• extended annotations.

• a functional-style syntax, which is easier to read and linearize. Whilst similar to the OWL 1.0 Abstract Syntax, it is not fully backwards compatible with it.

1.5.10 OWL Design Choices Unlike previous knowledge representation systems, OWL has been designed for knowledge representation over the distributed environment of the World Wide Web. Two key design choices made by OWL that assist it to operate in this environment are its open world assumption and its no unique name assumption.

Open World Assumption OWL assumes that we are operating in an open world; that is, that we never have complete knowledge and it is always possible that more material may be added. The implications of this assumption are explored in two of the included papers, as explained in Section 1.10.3.

No Unique Name Assumption This assumption relates to identity between individuals in an ontology. OWL does not assume that because two names of individuals are different that they necessarily refer to different individuals. Under certain circumstances, reasoning may infer that they are in fact the same individual. OWL provides the owl:sameAs, owl:differentFrom and owl:allDifferent constructs to make explicit assertions about individual equality and inequality of identity.

1.6 Notations

1.6.1 RDF Abstract Syntax RDF has a normative abstract syntax [91] which is the set of RDF triples, also known as the RDF graph. As described previously, each RDF triple consists of a Subject, which may be an RDF URI reference or blank node, a Predicate, which may only be an RDF URI reference, and an Object, which may be an RDF URI reference, literal or blank node. An RDF URI reference is simply a Unicode string that does not contain any control characters, and produces a valid URI under the speciﬁed Unicode encoding [91]. Two RDF URI references may then be determined to be equal if and only if their Unicode strings match character by character. Chapter 1 Introduction

Figure 1.5 RDF/XML syntax shown in the SWOOP ontology editor

1.6.2 RDF Transfer Syntax and OWL Exchange Syntax Additionally, RDF has a transfer syntax called RDF/XML [34], which is an XML format. In order to be encoded in XML, the nodes and arcs of the RDF graph need to be represented in XML terms as element names, attribute names, element contents and attribute values. The normative exchange syntax for OWL is also RDF/XML. The OWL Language Reference [31] describes all OWL’s modelling primitives in the RDF/XML exchange syntax, and these primitives were detailed in Section 1.5.7 on the OWL vocabulary. An example of RDF/XML syntax is shown at Figure 1.5.

1.6.3 OWL Abstract Syntax OWL was developed as a vocabulary extension of RDF and an OWL ontology is also represented as an RDF graph as above. OWL simply assigns an additional meaning to certain RDF triples, as speciﬁed by the OWL Semantics and Abstract Syntax document [122]. OWL DL’s abstract syntax has a frame-like style and details the facts and axioms allowed by OWL. (Note that OWL 2 has opted for a functional style syntax that is easier to 23

Figure 1.6 OWL DL syntax shown in the Protégéontology editor linearize [108].) The semantics is model-theoretic, and relates the abstract syntax directly to a standard model theory, specifying vocabulary and interpretation. The document also defines a normative mapping from the OWL Abstract Syntax to RDF graphs. An example of OWL Abstract Syntax is shown at Figure 1.7.

1.6.4 Non-Normative Syntaxes

Logical Syntax Additionally, OWL may be expressed in a DL style syntax. Utilising its Description Logic Semantics, it may be expressed as logical axioms. This style may be used for creating axioms in Prot´eg´e[41], for example, as shown at Figure 1.6.

Compact Forms As the normative RDF/XML exchange syntax is rather verbose, a number of more compact alternatives have been devised [9], including Notation 3, also known as N3, which is an assertion and logic language which is a superset of RDF, and N-Triples, a subset of N3 for RDF. More recently, Turtle: Terse RDF Triple Language [10] has deﬁned a textual syntax for RDF that has some compatibility with N-Triples and Notation 3, and allows RDF graphs to be written in a compact and natural text form, with abbreviations for common usage patterns and datatypes. An example of Turtle syntax is shown at Figure 1.8. Chapter 1 Introduction

Figure 1.7 OWL Abstract Syntax shown in the SWOOP ontology editor

Manchester Syntax

Another approach that provides both compactness and readability is the human-friendly Manchester syntax [72], which takes the approach of substituting logical operatives with English words - for example, logical operators ∃ and ∀ are replaced by the words “some” and “only”. (Note that quantiﬁcation is only used in OWL in the context of property restriction, so there is no need to represent statements such as “All men are mortal”. If that were not the case “all” would be a more appropriate translation for ∀ than “only”). This approach enables non-logicians to more easily write and understand OWL axioms, making ontology construction and use more accessible for domain experts. An example of OWL Manchester Syntax is shown at Figure 1.9.

Controlled Natural Language Syntaxes

This approach is being developed further to encompass a syntax that may be written and read entirely in English. The author’s contributions to this eﬀort constitute part of this thesis, and are covered by two included papers in Chapter 5 as explained at Section 1.10.4. 25

Figure 1.8 OWL Turtle syntax shown in the SWOOP ontology editor

Graphical Notations There is also some work on developing a graphical syntax for OWL based on UML. The Object Management Group’s work on an Ontology Deﬁnition MetaModel [55] includes UML proﬁles for RDF and OWL, which provide a standard graphical notation for RDF vocabulary and OWL ontology development using UML tools.

1.7 Formal Semantics

The chief utility of a formal semantic theory is simply to ensure logical inferences are correct. The nature of the things being represented by the language of the formal system has no bearing on the formal semantics, the latter being simply a means to guarantee that true premises produce true conclusions. Only the truth or falsity of a statement and its logical components is considered relevant, whilst what is actually asserted about the world is considered extraneous. Two approaches are commonly used for specifying the formal semantics of a system of logical inference. The model-theoretic approach equates logical consequence with truth preservation in models: that is, every model where the premises are true is a model where the conclusions are true. In this approach, inferential rules are correct if they are truth- preserving over models. In contrast, the proof-theoretic approach takes the inference rules as primary. In this approach, for instance, the Modus Ponens rule of inference (inferring B from A and ‘If A then B’) is taken to constitute the deﬁnition of the ‘if ...then’ expression - it is simply the Chapter 1 Introduction

Figure 1.9 OWL Manchester Syntax shown in the Protégéontology editor expression which permits inferences of this form to be made. The semantics of RDF, RDF-Schema and OWL are given as a standard model theory [35, 122]. Note that the original 1999 RDF specifications [94] had no formal semantics: these were added later by the W3C’s RDF Core Working Group and appear in the 2004 recommendations [35] for compatibility with OWL. As described above in Section 1.5, the OWL language variants OWL Lite, OWL DL and the coming OWL 2 correspond to specific Description Logics. The ability to reason over OWL ontologies written in OWL Lite, OWL DL and OWL 2 is enabled by the reasoning algorithms that have been developed for the respective corresponding description logics.

1.7.1 Reasoning Services The formal semantics of semantic web ontology languages enable automated reasoners to process the web of data captured in RDF graphs. Whilst OWL 2 is under development, OWL DL currently provides the most reasoning capability. Reasoning should not be confused with initial parsing of an ontology to ensure it is syntactically correct, or to determine which ‘species’ of OWL Sublanguage it conforms to. Speciﬁcally, the following reasoning services are currently available for OWL DL ontologies:

1. Consistency checking. This ensures that the ontology does not contain any logical contradictions. Generally speaking, an ontology is considered to be consistent if there is some interpretation that satisﬁes each ontology, fact and axiom in the collection [122].

2. Concept satisﬁability. This reasoning service determines whether it is possible for a class (concept in DL terminology) to have any instances, that is, whether deﬁning 27

an instance of the class would cause the entire ontology to be inconsistent. If so, the class is said to be unsatisﬁable, otherwise it is considered satisﬁable.

3. Classiﬁcation. This service computes the subclass relations between every named class to create a complete class hierarchy. This may be useful to retrieve all subclasses of a class.

4. Realization. This service finds the most specific class(es) that an individual belongs to i.e the lowest in the subsumption hierarchy. Usually realization is performed after classification, since direct types (class memberships) are provided with respect to the class hierarchy.

5. Inferencing on Individuals Compute all the instances of a class, or all the types for a given individual. Compute the role ﬁllers of an individual. If ‘No Unique Name Assumption’ is supported, compute whether a given individual is necessarily the same as a diﬀerently named individual.

6. Rule Processing. Ontologies are commonly used in conjunction with rule languages such as SWRL [76] that have a model-theoretic semantics compatible with OWL, and enable rules to be added to ontologies and processed as a complete logical structure.

7. Querying. Query languages such as SPARQL [132] permit construction of queries over RDF graphs to extract information on request, as well as the ability to insert new information.

These reasoning services are currently provided by Description Logic reasoners such as Pellet [96], RacerPro [60], Fact++ [114] and KAON2 [42].

1.7.2 Model Theoretic Semantics The reasoning services described above are realizable due to the underlying model-theoretic semantics provided for RDF and OWL [35, 122]. Model theory involves a model and an interpretation function that enables the truth or falsity of well-formed statements of a language to be determined. As OWL language variants other than OWL Full are based on decidable logics, the truth or falsity of any well-formed OWL statement may be always be determined. A model for a language is simply an ordered pair where A is a set and I is an interpretation function that maps all terms of some language L to A such that any well formed statement of the language may be determined to be true or false under the interpretation. We then say that an interpretation I is a model for a set of well formed statements S of L iff every such statement is true under I. A is the universe that the language refers to, and the model is both the universe and the interpretation function. Model theory primarily concerns itself with finding valid inference procedures, that is, those inference procedures that preserve truth under all interpretations. Note however, that whilst model theoretic semantics includes a “universe”, this is simply a set of symbols, and does not involve any specification of the nature of the things being described. The interpretation function simply maps one set of symbols to another without any deep consideration of what is being represented. In fact, as the RDF Semantics specification [35] describes, model theory actively tries to be both metaphysically and ontologically neutral: Chapter 1 Introduction

“The chief utility of a formal semantic theory is not to provide any deep analysis of the nature of the things being described by the language or to suggest any particular processing model, but rather to provide a technical way to determine when inference processes are valid, i.e. when they preserve truth.”

Thus the key reason to have a model theoretic semantics is purely to guarantee that inferences are sound; that is, that only true statements may be inferenced from true statements.

1.7.3 Meaning and Formal Semantics

Other than the ability to create a web of data, the key contribution of the semantic web and its ontology languages is purported to be their ability to represent meaning and to make it machine processable, as enabled via the formal semantics of the languages. Whilst the model theory and corresponding reasoning services provided are certainly useful, their relationship to meaning in the sense that human beings understand it, and as the vision of the Semantic Web uses it, is not immediately clear. Certainly meaning is related to inference and the determination of truth, but intuitively there is more to it than that. Whilst a formal system is useful for deducing truth from truth, establishing the initial truth of statements that describe the world must be determined with reference to the world itself, one would assume. As well as inferring consequences, meaning involves the aspect of reference: that of establishing what it is that we are talking about. This aspect comes through clearly when the Semantic Web is described as connecting data based on it being about the same thing, as was quoted in section 1.2 as part of W3C’s Semantic Web vision. Yet the only means provided to determine this is the URIs, blank nodes and literals allowed by RDF statements and it is not entirely clear how these are intended to represent the world at large. Furthermore, many types of reasoning commonly used by human beings are not represented - the kind of reasoning that is supported is primarily syllogistic. The usefulness of syllogistic reasoning for real-world applications has been brought into question by writers such as Clay Shirky [142]. However, unless the Semantic Web is willing to compromise on the truth of its inferences (as humans typically do in their everyday reasoning, by using stereotypes and making assumptions and generalizations), it must necessarily stay within the bounds that formal logics set regarding the state of the art for decidable and sound inferencing techniques. (Extensions for dealing with probabilistic and uncertain knowledge are, however, being explored, as described in Section 8.1.) The RDF Semantics document [35] gives some background on the nature of ‘meaning’ captured by the RDF and OWL speciﬁcations. It states:

“Exactly what is considered to be the ’meaning’ of an assertion in RDF or RDFS in some broad sense may depend on many factors, including social conventions, comments in natural language or links to other content-bearing documents. Much of this meaning will be inaccessible to machine processing and is mentioned here only to emphasize that the formal semantics described in this document is not intended to provide a full analysis of ’meaning’ in this broad sense; that would be a large research topic. The semantics given here restricts itself to a formal notion of meaning which could be characterized as the part that is common to all other accounts of meaning, and can be captured in mechanical inference rules.” 29

This conﬁrms that despite the various notions of meaning at a broader level used in the Semantic Web vision, the actual implementations to date restrict themselves to that part of the notion of meaning that may be captured by mechanical, truth-preserving inference rules. These rules are based on the underlying model theoretic semantics speciﬁed in the RDF and OWL Semantics documents [35,122]. Under this notion of meaning, the meaning of an RDF graph may be considered to be simply the assurance of the truth of the statements made by its triples.

1.8 Beyond Model Theory

An investigation of the history of the semantic web specifications indicates that the intended specification of the Semantic Web’s semantics was not always so limited in scope. A 2004 paper entitled ‘Meaning and the Semantic Web’ [120] by Parsia and Patel-Schneider, two of the leading Semantic Web designers and theoreticians, explains that in the process of preparing the 2004 RDF specifications there was originally more consideration of machine accessible meaning than the model-theoretic semantics reflected in the current documents. The initial view was that this machine accessible meaning was to be based on the meaning of names (URI references in the semantic web), and the meaning of a name/URI was to be determined by the owner of the name/URI, if one existed, and provided by documents supplied by the owner that would provide a definition of the name/URI. However, during the process of finalizing the RDF specifications, pressure was applied against this view of meaning, as it did not allow for divergence of meaning and thus did not allow for differences of opinion, thus violating one of the basic principles of the web. The alternative proposal put forward in the paper suggested that the meaning of an occurrence of a URI should be determined in a local sense, by its context, that is, the (semantic web) “document in which it appears, plus other documents explicitly mentioned in constructs like the OWL importing mechanism”. However, a brief reflection indicates that neither suggestion is adequate on its own to make meaning machine processable. If the owner were to provide a definition of the name, then presumably it would either be in natural language text, in which case it would assist humans to disambiguate but would not be machine processable, or it would be provided in some machine processable way, but clearly outside of the specifications and formal semantics offered to date, in which case we must consider how and on what basis this is to be done. In addition, the question of determining who has the authority to specify the meaning of a name goes to the very heart of what meaning is taken to be: is it the intention of the person who creates the name or owns the name that determines its meaning, or is it the way the name is used by the community at large? The alternative proposal is designed primarily to accommodate differences of opinion as to what a name means by considering the context in which the name is used. Recognizing that a name / URI may have multiple senses, this gives humans a means to identify which of its senses they intend by indicating the context in which it occurs, usually pinpointing its documented use by some community of practice. Presumably this binds the user to using the name / URI as specified by those documents, their background theory and external intuitions. Whilst this method may be adequate to indicate which sense of a name / URI is being committed to and whether this agrees or disagrees with another user’s use of the term, it still has definite shortcomings. As the authors freely admit, this method is insufficient to capture either the meaning intended by the document writers, or to adequately support the processing of meaning for the purpose of software agents, thus this notion of meaning may Chapter 1 Introduction need to be augmented. The question of making meaning machine processable appears to still be open. This question is a vital one, as it provides the foundation for a ‘Pragmatic Web’ where machines can take practical real-world action based on the outcomes of mechanical inference, without risk of error due to ambiguity and misinterpretation of the underlying semantics.

1.9 Problems Facing the Semantic Web

Having presented the Semantic Web vision, technologies and approach, I now present a number of problems that are or have been obstructing its path that I have sought to address in the included publications. These are not necessarily the only problems faced by the Semantic Web, but are certainly of considerable significance. In keeping with NICTA’s focus on use-inspired research, these problems were identified in 2004-2007, largely through interactions with potential users, and have motivated and guided the work effort presented in the included publications.

1.9.1 Adoption and Accessibility To date, the semantic web has not enjoyed a wide level of adoption, despite the fact that the ontology languages are mature enough for it to now do so. The evidence for this is the lack of widespread use of RDF and OWL in mainstream web applications. From engaging with industry and government through my NICTA sponsorship, I suggest there are a number of reasons for this. The first is the issue of awareness and understanding within the general IT community. The Semantic Web standards have been developed largely in an academic environment, and the specifications are somewhat inaccessible to the usual technologist who does not have a background in AI or formal logics. Mainstream technologists have not really understood what the standards are for, what ontologies are, what a formal semantics is, and so on. The second reason is the issue of implementation. From a mainstream technologist’s point of view, it has not been at all clear how to go about implementing a semantic web application. Whilst the semantic web is described as an extension of the existing web, in practice the W3C’s recommendations have not been well integrated with the existing web. For instance, it was not clear how to embed semantic information in an existing web page. Whilst GRDDL [33] now provides this ability, it is not yet well understood or adopted, and the tools to implement semantic web applications on a large scale have lacked maturity. The third issue is the perceived gap between the vision of the Semantic Web and the reality of what can be done with what has been delivered to date. Applying the existing Semantic Web technologies will not give one an agent that can schedule your medical appointments, as was touted in the original 2001 Scientific American article [16]. To do such a thing requires semantic interoperability on a large scale, and integration with scheduling tools and agent technology. A common criticism of the semantic web is that it is just a pipe dream, and cannot be delivered with the existing technologies. Unfortunately, this view tends to undermine the reality of what can actually be achieved with what has been delivered. Clearly, there has been a need to make the technologies more accessible to the IT community at large, starting with providing an understanding of what they are, how they work, how to implement them, and what they can actually do. This should be couched in the language of business technology, and include strong arguments in business terms as to why it is worth doing. 31

Figure 1.10 Lars Marius Garshol’s diagram of ISO Topic Map standards vs W3C standards, 2005

1.9.2 Competing Standards for Building Semantic Structures The W3C standards are not the only set of international standards seeking to build semantic structures connecting information resources: there is also a set of standards under the ISO umbrella, which may be referred to as the ‘Topic Map Standards’ [100], created independently and in parallel with the W3C approach. When the two development camps started to become aware of each other around 1999, there was some confusion about the relationship and potential compatibility between the two approaches. The Topic Map approach was developed with a slightly different focus than the W3C standards: Topic Maps were intended to be akin to an electronic version of the index typically found at the back of a book connecting topics to their occurrences using associations, using URIs instead of page numbers. Topics could also be connected to published subject identifiers, (analogous to library Subject Headings), in order for the user to clearly identify what the topic was about by citing an external index. Topic Maps were therefore primarily concerned with allowing users to make connections that would assist in finding resources about specific topics or subjects, and in viewing maps showing topics and subjects covered by resources of interest. There were a number of tools to display and search Topic Maps, including Ontopia’s Omnigator [117]. However, Topic Maps had an XML syntax [126] and a plan to create a constraint and query language, as shown in Figure 1.10. Considered as a whole, the Topic Map approach appeared to be paralleling and competing with the W3C approach. The W3C set up an RDF/Topic Map task group in 2004 to address interoperability issues but the way forward was unclear, and the task group did not include OWL within its considerations [125]. For mainstream technologists, the question of whether to adopt Topic Maps or OWL was confusing and for some resulted in a ‘wait and see’ attitude with an understandable reluctance to back the wrong contender and potentially having to backtrack. Chapter 1 Introduction

1.9.3 Rules and the limits of Ontology Languages

A third problem faced by the Semantic Web has to do with the issue of ontology languages and their ﬁt with rule languages. In practice, many classiﬁcation tasks require ontologies to be supplemented with rules. Whilst this was technically understood at logical level, there has been a need for a practical illustration of where this limit was and how rules could be used in tandem with OWL ontologies in practice. It was also important to illustrate the impact of OWL design features, particularly the Open World Assumption. Additionally, it was important to explore the extent to which tools could adequately support ontologies supplemented with rules, and to encourage tool development where necessary.

1.9.4 Readable Syntax

Up until 2005, the syntaxes for RDF/OWL were XML-based [31, 34] and logic-based, as explained in Section 1.6, and whilst compact forms such as N3 [9] and Turtle [10] existed, they were still difficult for non-technicians to read. When developing domain ontologies, it is important to capture domain knowledge accurately, and typically this knowledge is sourced from domain experts who are not necessarily well-versed in formal logics. Without a readable syntax, the domain experts needed to work closely with logicians and technicians to build a domain ontology, and were poorly placed to check its accuracy unless they were trained as logicians. The advent of Manchester syntax [72] in late 2006, supported by Protégé[41], introduced English keywords as alternatives to logical symbols. This made it far easier for domain experts to author ontologies themselves, and be confident that the axioms being created were as they intended. A syntax for OWL entirely in English would extend this readability further. However, it was important to ensure that the logical precision was not compromised by such a syntax. Natural language comes with expectations and understandings inherent from its use as a natural language, and these may compete with the needs of a logical language, whereas OWL’s meaning relates purely to logical inferencing rules.

1.9.5 Interoperability of Ontologies

In order for applications to use data from diﬀerent sources seamlessly, there needs to be semantic interoperability between the ontologies / vocabularies used to mark up that data. Eﬀectively, this means either using the same ontology, or mapping the ontologies together. However, the task of intermapping n ontologies has complexity of n2 for each ontology to be mapped to every other ontology. Clearly this is rather onerous, as whilst ontology mapping can be assisted by tools, it cannot currently be fully automated (current precision results plateau at around 60% and recall about 80% [38]) and therefore requires human mediation. For an n2 problem where n is large, this is obviously impracticable. One of the approaches to tackling this problem has been the use of upper or top-level ontologies such as WonderWeb’s DOLCE: a Descriptive Ontology for Linguistic and Cogni- tive Engineering [47], the Suggested Upper Merged Ontology (SUMO) [111] or Onto-Med’s General Formal Ontology (GFO) [67], which attempt to identify and describe those concepts that are common across domains, such as time, space, processes, qualities, objects and so on. By committing to a common upper ontology, domain ontologies may then spe- cialize these concepts further, achieving interoperability through being related to a common core. This approach has potential, but relies on wide commitment to a common upper level ontology, which to date has not been realized. 33

In practice, as explained in Section 1.10.5, many ontologies in the same domain space co-exist without being intermapped, so widescale interoperability is not achieved. This limits the potential for web agents repurposing data, a factor that is key to the semantic web vision. The simplest way to address the interoperability problem is to encourage widespread adoption of the same ontology, or to limit the number of commonly adopted ontologies to some very low number that is manageable for mapping purposes. A major contributing factor to the plethora of ontologies describing the same domain space is the difficulty in determining whether there is a pre-existing ontology that meets the current need. When faced with the need for an ontology in a particular domain space, a designer is currently more likely to build a new ontology from scratch than to re-use an existing one, simply due to the difficulty in ensuring a pre-existing ontology is and will continue to be suited to their purposes. Therefore, giving better information about ontologies, their context and Quality of Ser- vice information may be of considerable assistance in giving designers sufficient confidence to adopt existing ontologies. Whilst the problem can be technically assisted, social components are at the heart of the trust issue and need to be tapped effectively to provide a viable solution.

1.9.6 Making Meaning Machine Processable Secondly, it is also worth considering the long-term potential for removing the human from the process of ontology mapping and mediation, so that whilst the problem is still n2, it is able to be fully automated. Informally, human mediation is required because some aspect of the ontology author’s intention is unclear or unspecified in the ontology(ies) being mapped, and it is therefore necessary to refer to annotations or some other extra-logical documentation, or consult the author directly. The clarification process takes place at the level of human comprehension, and the ontology(ies) are then mapped by humans according to this understanding. However, perhaps it is indeed possible for the ontology author to make available in the authoring process all the information that is necessary for mediation, and to do so in such a way that it becomes machine processable and there is no need for further human intervention. Such a process would require the author to specify precisely what terms mean, and correspondingly would require suitable constructs for capturing this meaning. But how should one go about defining terms in order to make every aspect of one’s intention clear? How does one establish exactly what the term is referring to in an unequivocal and machine processable manner? Such questions directly address the gap between the semantic web vision and the current reality, and open up the deeper issues inherent in it. What does it mean to make meaning machine processable? What should meaning be considered to be for the purposes of the semantic web? How do we balance the shifting nature of what words mean in a natural language with the need to make meaning precise and machine processable? How do we deal with the limitations of being confined within a symbolic system when what we want to refer to is clearly outside of that system? If meaning is grounded in embodiment and physical experience, as some argue e.g. [81], is there any hope for making it accessible to machine processes? At this level, we are not only considering the needs of the Semantic Web, but are considering questions of broad applicability to Artificial Intelligence in general. However, the Semantic Web, if it is to move closer to its stated vision, seems to demand some kind of answer to them. Thus these questions need to be carefully considered, if we are to hope to find a path forward. Chapter 1 Introduction

1.10 Included Publications

The core of the thesis is presented as a series of previously published papers that address the problems raised in the previous section, and have thereby, I hope, made some contribution to the Semantic Web’s evolution. The included publications are grouped into chapters thematically, corresponding to the problem being addressed. In summary:

• Chapter 2 Overview of Semantic Technologies is an overview of semantic technologies, which explains the ﬁeld to mainstream information technologists in order to make it more accessible and to encourage wider adoption of the technologies. In the context of this thesis, it provides background information that is relevant for setting the scene for the following chapters, which are essentially independent of each other. • Chapter 3 Integrating Topic Maps into the Semantic Web details work conducted to map the ISO’s Topic Map standards into the W3C’s Semantic Web stack of technologies, so that it might take advantage of the formal semantics and tools associated with the latter. • Chapter 4 Adding Rules to OWL Ontologies describes joint work conducted to explore the need for, and use of rules in conjunction with OWL ontologies, editors and reasoners. • Chapter 5 Controlled Natural Language Syntaxes for OWL describes joint work on designing a controlled natural language syntax for OWL 2. The initial work on Sydney OWL Syntax was presented at OWLED in 2007, where a CNL task force formed to explore various alternatives, reported in the second paper of this chapter. • Chapter 6 Encouraging Ontology Reuse details joint work on the issue of ontology reuse, its importance in semantic interoperability, and plans for a tool which would encourage reuse by leveraging social factors and networking tools. • Chapter 7 Foundational Issues in Meaning examines the issue of meaning at a fundamental level, and its importance for Semantic Web technologies.

The inclusions, context, contribution and content of each of the six chapters containing included publications is summarized below. A concluding chapter summarizing the conclusions of each publication, and ﬁnally concluding the thesis overall, follows at Chapter 8.

1.10.1 Chapter 2: Overview of Semantic Technologies Included Publications: 2.1 A. Cregan. Overview Of Semantic Technologies. In P. Rittgen (Ed), Handbook of Ontologies for Business Interactions, IGI Global, 2007.

Personal Contribution: 100 %

Context: Editor Peter Rittgen invited the contribution of this peer-reviewed book chapter, which was written over 2006 and 2007 and published in the Handbook of Ontologies for Business Inter- actions in 2007. In the context of this thesis, this chapter provides additional background 35 material describing the area of semantic technologies in general, and raises a number of important issues, some of which are addressed by later inclusions.

Contribution:

The chapter proposal was written in July 2006, when there was a strong need to write an accessible overview of semantic technologies, explaining what they were, what they could do, and describing the methods employed. Even though at this time the W3C’s OWL Recommendations [31, 102, 122, 144] had been in place for over two years, there was little awareness or use of the approach amongst the broader technical community. The Semantic Web technologies had been designed largely within research environments, had few real-life applications that a broader public were likely to encounter, and were generally perceived as somewhat unfathomable by the typical corporate technologist or IT manager. The chapter was designed to ﬁll this gap, being both an overview for technologists giving a suitable conceptual framework for reading more detailed material, and also an executive summary for technology and business managers. The emphasis is to introduce semantic technologies in an accessible way, relate them to more familiar IT technologies, explain the additional value and give tangible illustrations of it.

Content:

The chapter introduces the vision of the Semantic Web [15] as one that holds great promise for dealing with the modern plague of digital information overload [99]. Semantic technologies are introduced as a new wave of computing [110] that go beyond data exchange standards such as XML [20] to the semantic level, to enable one system to make use of the information in another system more easily and cheaply. It explains the various levels of representation (symbolic, syntactic and semantic) and the value of making the semantic level of information explicit and available for machine processing. Semantic technologies provide the means to create virtual information structures and enable data interoperability. Once interoperability is achieved, it is possible to provide more powerful and flexible information services and transactions that search, query and reason over multiple information stores linked together based on common meaning. The key strategies are explained to be, firstly the decoupling of data from applications, so that it is more easily redistributed and reused. Secondly, the use of metadata tags organized into logical structures called ontologies, which enable data stores to be described conceptually, mapped together to provide interoperability, and reasoned over via the use of the formal semantics that ontology languages provide. Ontologies [57] describe the individuals, concepts and relationships that are relevant for conceptualizing some real-world domain. Unlike traditional data modelling methods, they have more expressive language for describing concepts and relations, and a formal semantics that give inference rules for checking the consistency of knowledge bases and conducting automated reasoning. The applications and benefits that semantic technologies offer are described as being information integration and interoperability, intelligent search and services (which may be specified using OWL-S [22]), model-driven applications, and intelligent reasoning. Ulti- mately semantic technologies should enable computing to become more adaptive and au- tonomous. A case study is presented that was conducted by TopBraid consultants for the US Fed- eral Government [71], that formalized the US Federal Enterprise Architecture Reference Chapter 1 Introduction

Models as OWL ontologies so that various implementations could be checked for conformance and logical consistency, and implementations across government agencies could be inter-linking providing whole of government visibility [3]. This enabled report generation and query answering about agencies and their use of technologies, products and processes to be conducted across an overall model (RDF graph) that interlinked available information across all the agencies. The different aspects of ontologies are discussed, explaining the difference between ontology languages, the ontology structures that use them to capture knowledge about specific domains that, and the specific ontology content that may be placed into ontology structures to capture information about specific individuals and their relations. The chapter also explains the OWL ontology language constructs [144] in some detail as well as the components of the Semantic Web technology stack [17] and their status, whilst clarifying the relationship between semantic technologies generally and the Semantic Web [14,16] as their implementation with regards to the world wide web. It explains the capability for importing, merging and aligning ontologies at run-time to produce virtual information structures, the benefits of using reasoning engines, tool support for viewing, editing, querying and aligning ontologies, and the use of rules and annotation. OWL ontologies are compared to other data structures and models that various technologists might be familiar with, including traditional databases, Unified Modelling Language (UML), taxonomies and expert systems, explaining that in most cases ontologies generally leverage their value by making their semantics more executable. Lastly, the chapter considers issues and challenges. These include: 1. the need for data owners to semantically markup their data, and the lack of automated and other techniques to ease the process. 2. better tools and techniques for supporting large scale semantic applications 3. better support to enable non-logicians to participate in ontology building, particularly via the use of (controlled) natural language syntaxes [28]. 4. better standards and methods for resolving meaning without recourse to human intervention. One possible solution identified was the use of jointly developed ontologies for use as common standards within a domain, to reduce the need for pairwise ontology mapping simply by commitment to a common ontology. The second approach mentioned the need for a better underlying semantic theory that would encompass a broader view of semantics that would encompass cognitive aspects of meaning [27]. 5. better methods for dealing with incomplete, uncertain and probabilistic data, for all those commonly encountered real-world situations where statement do not neatly map to a boolean truth value. 6. appropriate methods for dealing with proof, trust and security over semantically interlinking virtual data structure. There is a need for appropriate measures to handle who can access what data and at what level, and for intelligent agents/reasoners to be able to account for and justify their actions. Some of these issues and challenges relate to later contributions within this thesis. In particular,

• With regard to Item 3, work has been progressing towards a common controlled natural language syntax for OWL 2 with the author as an active participant, and this is detailed in Section 1.10.4 and Chapter 5. 37

• With regard to Item 4, methods to encourage the adoption of common standards are addressed by Section 1.10.5 and corresponding Chapter 6, whilst Section 1.10.6 and Chapter 7 address the need to understand meaning more deeply and uncover an underlying semantic theory.

The issues and challenges raised are updated and reassessed in the concluding chapter at Section 8.1.

1.10.2 Chapter 3: Integrating Topic Maps into the Semantic Web

Included Publications: 3.1 A. Cregan. Building Topic Maps in OWL-DL. In Proceedings of Extreme Markup Languages, 2005. Available online at http://www.idealliance.org/papers/extreme/proceedings//html/2005/

Personal Contribution: 100 %

3.2 A.Cregan, An OWL-DL construction for the ISO Topic Map Data Model (presented to the ISO Topic Map committee (ISO/OEC JTC 1/SC 34, Working Group 3) and published online by XML coverpages at http://xml.coverpages.org/topicMaps.html

Personal Contribution: 100 %

Context: NICTA, Australia’s newly established national centre of Excellence for Information and Communications Technology, held an outreach workshop in late 2004 to discuss Topic Maps with Australian industry and government organizations, as part of its mission to identify and engage in use-inspired research. Some Australian companies had approached NICTA for advice and recommendations on using the ISO Topic Map standards [100], and their relationship to the W3C Semantic Web standards: OWL had become a W3C recommendation in early 2004 [31, 102, 122, 144], whilst the XML Topic Maps (XTM) 1.0 specification had been published in 2001 [126]. As a NICTA-sponsored student, my supervisors asked me to investigate this issue and determine the viability of a mapping between Topic Maps and OWL ontologies. A draft Topic Map Data Model (TMDM) [51] had been published by the ISO group responsible for the Topic Map standards (ISO/OEC JTC 1/SC 34, Working Group 3 - see notes below) in January 2005. Based on its specification, a method for building Topic Maps in OWL DL was found to be feasible, and in February 2005 a draft of the proposed OWL DL model, (an included publication at Section 3.2), was submitted to the relevant ISO committee for consideration, and was tabled for discussion at the official ISO SC34 meetings to be held in Amsterdam in May 2005. At that meeting the proposed OWL DL ontology was presented, and was judged by those present, including TMDM authors Garshol and Moore, to be a faithful representation of the Topic Map Data Model in OWL DL, implemented at an object level. As an outcome of the meeting, WG members were formally requested to provide detailed feedback in order to ensure the accuracy of the representation (Reference: http://www.jtc1sc34.org/repository/ 0625.htm). Chapter 1 Introduction

It was followed by an invited paper at the Extreme Markup Conference in August 2005 which explains the approach more fully, included at Section 3.1.

Notes on ISO SC34 WG3 This material below is sourced from the oﬃcial ISO SC34 home page at http://www.itscj. ipsj.or.jp/sc34/:

“ISO/OEC JTC 1/SC 34 is the international standardization subcommit- tee for Document Description and Processing Languages standards and technical reports related to structured markup languages (speciﬁcally the Standard Generalized Markup Language (SGML) and the Extensible Markup Language (XML)) in the areas of information description, processing and association. It has three Working Groups and Working Group 3 (WG 3) addresses Infor- mation Association. SC 34/WG 3 is responsible for producing standard architectures for information management and interchange based on ISO 8879, Standard Generalized Markup Language and related standards, including, but not limited to: • Information linking and addressing • Time-based information management • Representation of knowledge structures, indexes, etc. • Management of behavioral components in interactive documents”

Contribution: When first asked to conduct this work, there were two topic map syntaxes published by the ISO: XTM1.0 in XML [126] and another in HyTM. These syntaxes had non-trivial differences, and the original ISO Topic Map standards [100] did not explain the relationship between them, or give any indication of an underlying formal semantics. My initial attempts at mapping could only attempt to interpret the Topic Map standards and produce a suitable semantics, however, this produced multiple possible OWL ontologies. However, the publication of the Topic Map Data Model [51] in the form of an Entity-Relationship model, in January 2005, gave a means to disambiguate between the possible models. Valid Topic Maps were able to be defined as those Topic Maps that conformed to the constraints specified by the TMDM, and the TMDM Entity-Relationship model thus clearly provided a basis for building an OWL DL ontology that implemented the specified constraints. As the TMDM is a data model, whilst OWL is a language for specifying such models, this appeared to be a natural fit, although it has since been criticized for mapping at an inappropriate level, as explained at Section 8.2. The TMDM was therefore built as an OWL DL ontology, and was made publicly available for topic map authors to import and populate by adding instances to the TMDM OWL ontology, thus effectively building their own topic maps in OWL DL that were TMDM conformant. The benefit of using OWL DL as a modelling language was the ability to take advantage of OWL’s formal semantics and accompanying reasoners to ensure any topic map built was conformant to the TMDM, as well as giving access to all of OWL’s tools and querying capabilities, which were considerably more mature than those for Topic Maps. At the time, the ISO Topic Map group were also embarking on specifying a Topic Map Constraint Language [107] and Query Language [50], and the use of OWL as the native language effectively circumvented the need for these. As Topic Maps could now be authored in OWL DL, it was hoped that the existing Topic 39

Map tools would make their functionality available to Topic Maps written in OWL DL by providing users with the ability to import them directly, a relatively trivial task.

Content:

Both ISO’s Topic Map Standards and the W3C’s Semantic Web Recommendations provide the means to construct meta-level semantic maps describing relationships between information resources. Developed independently, attempts at interoperability between the original Topic Map standard and RDF [101] prior to 2005 proved challenging: the W3C set up an RDF/Topic Map task group in late 2004 to address interoperability issues between RDF and Topic Map standards, but the way forward was unclear, and the task group did not include OWL within its considerations [125]. (Note: In any case, this task group was conducted in parallel with my own initial Topic Map mapping to OWL DL, and my original proposal was submitted for consideration prior to the task force releasing its findings.) Whilst the ISO Topic Map working group continued to work on its ultimate Topic Map Reference Model [32], its drafting of an explicit Topic Map Data Model (TMDM) [51] early in 2005, combined with the advent of the W3C’s more expressive Web Ontology Language (OWL) Recommendations in 2004 [31, 102, 122, 144], provided the possibility for authoring TMDM-conforming Topic Maps directly in OWL. Recognizing that any Entity-Relationship model may be represented as an OWL DL ontology, an ontology was built to the TMDM’s specification. Written early in 2005, the two included papers present a construction of the TMDM model as an OWL-DL ontology. The presented ‘TMDM Ontology’ is a construction in OWL-DL of the Topic Map elements specified by the TMDM, making them available for use by Topic Map authors as a basis for building TMDM-compliant Topic Maps directly in OWL DL. Using OWL DL as the language for Topic Map authoring gives users access to OWL DL’s formal semantics, constraint expressivity, and suite of tools such as Protégé[41], which includes an API and capability for ontology visualisation, querying, and automated constraint checking and reasoning using Description Logic reasoners such as Racer [60]. The use of XML Schema Datatypes [18] provides datatypes where required for data values. Using the offered OWL DL “TMDM ontology” as a basis for building Topic Maps thereby ensures that they will be TMDM conformant. In contrast, whilst Topic Maps at that time had tools geared for exploration and visualization e.g. Omnigator [117], a planned constraint language (TMCL) and querying language (TMQL) specifically for Topic Maps were still at a very early stage of development, and thus this functionality was unavailable through existing Topic Map tools. By using OWL-DL as the language for Topic Map authoring, one gains access to the querying and constraint capabilities associated with OWL-DL, without requiring the use of any Topic Map language, Topic Map Constraint Language or Topic Map Query Language. By virtue of being TMDM-compliant, it is a straightforward matter to write a simple translation algorithm to convert Topic Maps written in OWL DL to other TMDM based-Topic Map authoring languages if required, thus providing users with access to Topic Map engines and tools. Illustrating by example, a Topic Map written in OWL-DL using Protégéis shown, highlighting the constraint and querying abilities provided, which clearly satisfy the key requirements set by the ISO for a Topic Map Constraint Language and Query Language. One outstanding issue regarding the interpretation of Typing is identified, and options for its resolution are discussed. Chapter 1 Introduction

1.10.3 Chapter 4: Adding Rules to OWL DL Ontologies

Inclusions: 4.1 A. Cregan, M. Mochol, D. Vrandecic, and S. Bechhofer. Pushing the limits of OWL, rules and Prot´eg´e- a simple example, in B. Cuenca Grau, I. Horrocks, B. Parsia and P. Patel-Schneider (Eds.): Proceedings of the OWLED2005 Workshop on OWL, Experiences and Directions, Galway, Ireland, November 11-12, 2005 CEUR-WS, Vol. 188, 2005.

Personal Contribution: 35 % (estimated by co-authors)

4.2 M. Mochol, A. Cregan, D. Vrandecic, S. Bechhofer. Exploring OWL and Rules - A Simple Teaching Case, International Journal of Teaching and Case Studies (IJTCS), Special Issue: Teaching Semantic Web: Integration with CS/IS curriculum and Teach- ing Case Studies, to appear.

Personal Contribution: 30 % (estimated by co-authors)

Context: This exploratory work was conducted at the European Semantic Web Summer School 2005 in the context of a mini-project undertaken by four summer school students including myself, Malgorzata Mochol, Denny Vrandecic and Antoine Zimmerman, and summer school tutor Sean Bechhofer. At that time I was interested in gaining a greater understanding of the use of rules in conjunction with ontologies. Having decided on the topic, I invited the other students to join the team, and chose Bechhofer as the tutor due to his involvement in designing the OWL language and high level of expertise. Following the success of the mini-project (it was the winner of the summer school mini- project competition) the work was written up as a paper for the inaugural OWLED workshop held in December 2005. Due to its relevance for instruction, the editor of the Interna- tional Journal of Teaching and Case Studies subsequently invited the authors to submit it as a case study for a special issue titled ‘Teaching Semantic Web’.

Contribution: The paper used a simple example OWL ontology to show the difficulties and challenges of implementing some intuitively simple classification criteria using OWL-DL and SWRL rules. The starting point for the work was to explore the boundary where it was necessary to add rules to an OWL ontology in order to accomplish the classification. Whilst the work is not novel in a theoretical sense, it provides an accessible account of the topic, including representation choices and their impact, the limits of OWL and how to supplement it with rules, the impact of the Open World assumption, and various implementation difficulties. As a teaching resource, it adds a useful contribution to the very limited amount of teaching material on Semantic Web technologies then available [73, 101, 133]. It highlights some of the difficulties commonly encountered in modelling using OWL, such as needing to backtrack in the ontology design, difficulty constructing suitable axioms in a straightforward way, and generating unanticipated results in reasoning. It also pointed 41 out where the OWL language could benefit from extensions such as property chaining and qualified cardinality restriction, which are now being included in OWL 2 [108] and the shortcomings of the tools available at the time. Its key contributions are instruction in the art of ontology building, and identification of difficulties with implementing OWL and rules on both the language level and tool level that are informative for those working on developing OWL, SWRL and associated tools and reasoners.

Content: The work is set in the context of the Semantic Web [66] vision, which is to extend the web with machine-understandable data leveraged by applications to perform tasks automatically, based on the core technologies of of RDF [34], OWL [31,144] and a rule language for which SWRL [76] is considered a prototype. The starting point for the work was to explore the boundary where it was necessary to add rules on top of an OWL ontology in order to accomplish intuitively simple classification tasks. This is pursued using Protégé[112] in combination with the Racer automated reasoner [59] and SWRL plugin. Whilst Protégé does not include a reasoner, it may be used in conjunction with any reasoner that uses the DIG (Description Logic Implementation Group) interface [6]. SWRL provides Horn-like rules for both OWL DL and OWL Lite, includes a high-level syntax for representing these rules [2] and is more powerful than either OWL DL or Horn rules alone [75]. An exercise was designed that involved classifying a number of Student groups as a GoodGroup if they satisfied all of a number of intuitively simple criteria, and as a BadGroup otherwise. Initially a simple OWL ontology was constructed as a base. The first condition: 1. Groups should have either 4 or 5 students as members was easily modelled in OWL using minimum and maximum cardinality constraints on an object property (hasMember) that connected the class Group to the class Student. The second condition: 2. Groups should have at least one member of each gender was also able to be modelled in OWL alone, but with more difficulty. Firstly we noted that we could not model it simply as a minimum cardinality since it involved values from more than one class - both Man and Woman. Our need for hasMember to have a minimum cardinality for values from each of the subclasses Man and Woman requires the use of qualified cardinality restrictions which are unsupported by OWL DL, but are now included in the OWL 2 proposals [108, 121]. We also considered using existential restrictions on GoodGroup, but had trouble formulating the condition in Protégé. We then attempted to look at the situation in reverse and specifically define groups corresponding to the two ways the criteria could be broken, namely by having all women or all men, and making these subgroups of BadGroup. However in applying classification, we noted an unforeseen consequence: due to the Open World assumption, a group with four male members was not classified as a BadGroup due to the possibility of there being a fifth female member without breaking the cardinality restraint set in Condition 1. One option was to rework our base ontology and define each group as an enumerated class composed of specific students, but this was unsupported by reasoners as it required nominals [70], and seemed unintuitive to us in any case. Instead we defined two new classes BigGroup and SmallGroup corresponding to having four or five members, and overlaid our conditions on these. This was sufficient to capture any four or five member same gender group as a BadGroup. Chapter 1 Introduction

We proceeded to consider the next three conditions: 3. Groups should have members of all different nationalities. 4. Groups should have members from all different institutions. 5. Groups should have a tutor who is the favourite tutor of all the group’s students. which required property chaining: that is, consideration of more than one property at the same time. Whilst OWL permits chaining of properties, it does not support making assertions about the equality of the objects at the end of two different properties/ property chains. Therefore it was necessary to add rules in order to implement these criteria. We were able to formulate SWRL rules as Horn rules for BadGroup. At this point we had all the rules needed to correctly classify every BadGroup.We considered this an exhaustive list, so we defined GoodGroup as any group not classified as BadGroup. However, when we applied this definition, we found the Open World Assump- tion prevented us from successfully classifying the GoodGroups, as our definition required negation as failure. As OWL’s semantics adopt the Open World Assumption, a statement cannot be assumed true simply because its negation cannot be proven [77]. We needed to start again and reformulate the rules positively for GoodGroup as a conjunction of all the criteria stated positively. In order to do this we needed to create new classes corresponding to each condition. It was possible to define a MixedGenderGroup in OWL fairly easily by using existential restrictions to define it as a conjunction of Group, existence of hasMember value Male, existence of hasMember with value Female.How- ever, the rules corresponding to Conditions 3, 4 and 5 proved to be extremely long: for instance, Conditions 3 and 4 had to include all the pairwise comparisons to ensure each member was different. We also found that there was no reasoner that could easily be plugged in to the Windows-based Protégéenvironment we were working in to operate on both OWL and SWRL, and therefore planned subsequent work using reasoners Hoolet [8] and KAON2 [42, 78]. We also noted that regrettably, the current DIG interface [6] does not specify how to exchange and reason over rules, so even though both the editor and the reasoner being used provide support for rules, they were not able to be used together seamlessly via the editors interface. This problem has been addressed in a proposal to extend the DIG interface appropriately [7], but we were forced to pursue other solutions, such as using a file based ontology exchange. We also considered the rule engine Jess, as Jess and a DIG enabled reasoner can be used together to reason over a rule-enhanced OWL DL ontology, as described in [53]. Ultimately we were able to successfully build an ontology with rules that in theory would successfully classify GoodGroup and BadGroup. Our experience highlighted some of the difficulties in modelling using OWL, illustrated several cases where it was necessary to supplement the ontology with SWRL rules in order to achieve the desired classification, and pinpointed some areas that could be addressed in language development and tool support.

1.10.4 Chapter 5: Natural Language Syntax for OWL

Inclusions:

5.1 A. Cregan, R. Schwitter and T. Meyer. Sydney OWL Syntax - towards a Controlled Natural Language Syntax for OWL 1.1, in C. Golbreich, A. Kalyanpur, B. Parsia (Eds.): Proceedings of the OWLED 2007 Workshop on OWL: Experiences and Di- rections, Innsbruck, Austria, 6-7, June 2007, CEUR-WS, Vol. 258, 2007. 43

Personal Contribution: 80 % (estimated by co-authors)

5.2 R. Schwitter, K. Kaljurand, A. Cregan, C. Dolbear, G. Hart, A Comparison of three Controlled Natural Languages for OWL 1.1, presented at OWLED, OWL Experiences and Directions, 4th International Workshop, Washington, USA, 1-2 April, 2008.

Personal Contribution: 32.5 % (estimated by co-authors)

Context: At the second OWLED, held in November 2006, it was widely recognized by attendees that there was a need to extend the approach taken by Manchester OWL syntax in substituting English words for logical operators, and to design a syntax for OWL which would enable ontologies to be constructed and read entirely as English sentences. This would be an important step in giving non-logicians the ability to understand and author OWL ontologies, and thus to make semantic technologies accessible to the greater community. On return to Sydney from attending OWLED, I invited Rolf Schwitter and Thomas Meyer to work with me to design such a syntax; Schwitter being an expert in machine- processable Controlled Natural Language, whilst Meyer was my PhD co-supervisor from 2004 through mid-2007 and an expert on Description Logics. As the group were based in Sydney, we decided on the name Sydney OWL Syntax. The first paper is a discussion paper describing the resulting Sydney OWL Syntax, covering the scope, design, and examples of Sydney OWL Syntax in use, including a call to action for others to provide feedback and collaborate on agreeing on a CNL syntax for OWL 1.1 (now known as OWL 2). Following presentation of the paper and accompanying poster at the subsequent third OWLED held in May 2007, it was discovered that there were two other parallel efforts also working on some form of controlled natural language syntax for OWL 2: Rabbit and ACE. An OWLED CNL task force including members from each effort was formed for the purpose of working jointly towards a proposal for a common CNL syntax for OWL 2. The second paper is a comparison paper comparing the three syntaxes SOS, Rabbit and ACE, presented at the fourth OWLED held in April 2008.

Contribution: The first paper in this chapter describes a proposal for a new syntax, Sydney OWL Syn- tax, that can be used to write and read OWL ontologies in Controlled Natural Language (CNL): a well-defined subset of the English language with a limited grammar and lexicon. The proposed Sydney OWL Syntax is a complete syntax enabling two-way translation and generation of grammatically correct full English sentences to and from every element of the OWL 2 functional syntax. It also offers insights into many important considerations for designing a CNL syntax for OWL. Discussion of this syntax and two comparators led to the formation of the OWLED CNL task force, tasked with agreeing on a common CNL syntax for OWL. The second paper offers a comparison of the three syntaxes in preparation for making recommendations for going forward.

Content: The first paper in this chapter describes a proposal for a new syntax that can be used to write and read OWL ontologies in Controlled Natural Language (CNL): a well-defined subset of the English language with a limited grammar and lexicon. Following OWL reaching W3C Recommendation status in February 2004, a variety of notations became available Chapter 1 Introduction for editing ontologies through tools such as Protégé[41] and SWOOP [86], including the official RDF/XML exchange syntax [34], N-triples [9] and subsequently Turtle [10] and OWL Abstract Syntax [122]. Following the lead of Manchester OWL Syntax [72] in making OWL more accessible for non-logicians, and building on the previous success of Schwitter’s PENG (Processable English) [137], the proposed Sydney OWL Syntax enables two-way translation and generation of grammatically correct full English sentences to and from OWL 1.1 functional syntax. PENG (Processable English) [137] translates specification texts into first order logic via discourse representation structures [87] and can be used for web page annotation [140]. It uses predictive interface techniques [138] to aid the writing process. However, PENG was not designed for bidirectionality or for Description Logics. Whilst informed by PENG, Sydney OWL Syntax was designed specifically for DL rather than FOL expressivity, and with bidirectionality in mind. The first paper identifies that a bidirectional mapping between a subset of OWL DL and Attempto Controlled English (ACE) has been presented in 2006 paper by Kaljurand and Fuchs [85] using a discourse representation structure as interlingua, but immediately subsequent work prepared for OWLED2007 focussed only on one direction: verbalizing OWL DL [84]. With regard to bidirectionality, it is noted that Schwitter and Tilbrook [139] had previously showed it can be achieved in a direct way using axiom schemas without need for interlingua, and this informs the approach taken by Sydney OWL Syntax. Sydney OWL Syntax, used in conjunction with OWL tools, was designed to facilitate ontology construction and editing by enabling authors to write an OWL ontology in a defined subset of English. This improves readability and understanding of OWL statements or whole ontologies, by enabling them to be read as English sentences. By providing the option of an intuitive, easy to use English syntax which requires no specialized knowledge, it was hoped that the broader community will be far more likely to develop and benefit from Semantic Web applications. Sydney OWL Syntax was scoped to be compatible with OWL1.1, now referred to as OWL 2 [108, 121] and its functional-style syntax. It covers the entire OWL language and offers bi-directional translation without loss of information. Design goals are to support non-logicians to build high quality OWL ontologies, provide English translation of OWL ontologies, take a modular approach that allows for future extensions of the OWL language, and be sufficiently detailed and precise for implementation purposes. The paper discusses a number of design choices that were faced and documents the decisions made. A key choice is that between sounding natural and staying close to the original OWL, and in general, Sydney OWL Syntax (SOS) opts towards tight binding with OWL 2 so that there is no loss of precision, whilst attempting to make expressions as natural as possible whilst honouring that choice. SOS offers one and only one translation for each OWL statement, and each statement is translated as a unit, without reference to any other statement or background linguistic knowledge. Although somewhat unnatural, it makes limited use of variables and explicit OWL constructs where necessary. In the second paper, presented at the following OWLED held in April 2008, the members of the task force compare the three controlled natural languages - Attempto Controlled English (ACE), Ordnance Survey’s Rabbit, and Sydney OWL Syntax (SOS). It identifies the need for a ‘pedantic but explicit’ paraphrase language to counter the common problems users encounter when working with OWL DL [133] and notes that paraphrasing for representing OWL in a more natural way has been suggested previously [68,80,103]. Controlled Natural Language (CNL) assists in reducing the ambiguity and complexity of full natural language [90], and existing CNLs include Attempto Controlled English (ACE) [45], PENG [137], Common Logic Controlled English [146] and Boeing’s Computer Processable Language [23]. 45

User testing has shown that CNLs can offer improvements over standard OWL syntaxes [46, 61, 88]. Apart from Lite natural language [12] and CLOnE [46], the three key CNL-based approaches to authoring OWL ontologies are ACE [44], Rabbit [63] and SOS [29]. Kaljurand’s PhD thesis [83] describes a bidirectional mapping of OWL 2 (without data properties) and a fragment of ACE, which is being used in experimental ontology editors ACE View and AceWiki [92], whilst Rabbit was developed by Ordnance Survey for authoring ontologies using a domain expert-centric approach [64], with Protégéplugin currently in development [24] and the GATE tool [30] being used to convert Rabbit into OWL. The three syntaxes are compared with reference to a common ontology, Ordnance Sur- vey’s ‘Buildings and Places’ ontology [118]. The paper discusses a number of requirements to create a workable OWL-compatible CNL. It summarises the similarities and differences of the three CNLs and makes some preliminary recommendations for an OWL-compatible CNL. Further work remains to be conducted by the Task Force to agree on a common CNL syntax for OWL 2.

1.10.5 Chapter 6: Encouraging Ontology Reuse

Inclusions: 6.1 D. Peterson, A.Cregan, R. Atkinson, J. Brisbin. n2Mate: Exploiting social capital to create a standards-rich semantic network. In Proceedings of Linked Data on the Web (LDOW2008), 2008.

Personal Contribution: 31 % (estimated by co-authors)

Context: This paper was prepared jointly with three co-authors from industry who contacted me following a talk on Semantic Technologies I gave to Australian government and industry in November 2007. They shared my insight into the need to encourage interoperability by making it easier for users to commit to common ontologies and wanted to help facilitate this, especially in the Australian government space. It was presented in April 2008 at the Linked Data on the Web workshop, co-chaired by Sir Tim Berners-Lee.

Contribution: One of the cornerstones of the semantic web is the need to interlink ontologies, vocabularies and URIs to create a dense semantic network. A sparse network lacks interoperability, and makes it diﬃcult for the semantic web to realize its full potential. There are several ways to attempt to increase interoperability. The publication included in this Chapter outlines the various approaches and focuses on one in particular: encouraging ontology re-use. The paper explains the barriers to ontology re-use and proposes a method and tool that leverages social networks to encourage ontology re-use, and thus achieve greater interoperability.

Content: Semantic Interoperability is achieved by commitment to common ontologies or by connecting ontologies via matching and alignment methods. However, intermapping n ontologies is an n2 problem, and whilst ontology mapping can be assisted by tools, it cannot currently Chapter 1 Introduction be fully automated (current precision results plateau at around 60% and whilst recall results plateau around 80% [38]) and therefore requires human mediation. There appears to be an intrinsic upper limit to automated matching, due to the current lack of techniques and constructs to get all the meaning out of people’s heads and into the ontologies we are working with. Almost invariably the matching needs to be completed and checked by the humans who built the ontologies negotiating with each other or referring to other sources of information about the ontologies. With an n2 problem where n is large, this becomes unviable. As an illustration of the sparsity of connections, the Linking Open Data (LOD) project [19, 36] referred to at Section 1.4.3, holds datasets that currently comprise over 2 billion triples, but reveal only about 3 million links [65], so overall the graph is very sparsely interconnected. One way to tackle the problem is to attempt to provide techniques and constructs that fully support the specification of meaning in a machine processable way, so that such negotiations can be fully automated and do not require human mediation or intervention. However, there is a great deal of foundational work required before such technology is likely to be possible. The publications included at Chapter 7 open up the questions that need to be addressed by such an endeavour. Even in the event of such techniques and constructs being devised, there is still the question of applying them correctly in practice. The remaining possibility is to provide interoperability simply via commitment to the same ontologies. Building consensus about which ontologies to use in a particular information space is a pragmatic approach which reduces the potential number of ontologies to be mapped to a manageable number. Thus the n2 mapping problem is addressed simply by reducing n. The key to the effort is to provide sufficient information and trust about the ontologies so that the relevant parties will commit to them and reuse them. Currently, as can be seen by searching Swoogle or examining Australian Government documents, one can find hundreds if not thousands of attempts to describe the same or very similar concept spaces with competing standards, vocabularies and ontologies. The wealth of choice makes it very difficult for those seeking to locate, apply or distinguish between existing standards, and faced with the need for an ontology describing a particular conceptual space, it can often be easier to build one’s own ontology from scratch rather than to attempt to navigate the existing plethora of standards, vocabularies and ontologies. However, this simply adds to the problem, adding yet another ‘standard’ for the next person to consider. Our theory is that it is the lack of socially-sensitized processes highlighting who is using what and why, that have led to the current unmanageable plethora of vocabularies, where it is far easier to build one’s own vocabulary than try to find a suitable, reliable existing one. Each vocabulary has its own associated attributes to do with why it was developed, what purposes it is best suited for, and how accurate and reliable it is at both a content and technical level. However, most of this information is opaque to the wider community of users. As shown in Figure 1.11, standards, vocabularies and ontologies do not exist in a vacuum, but form an intrinsic part of social and business contexts where they are related to people, projects and organizations. As Snowden points out [145], human knowledge is contextual, triggered by circumstance and need, and best drawn out through social interaction. Our argument is that there is considerable value in the development of an online facility that exposes and leverages the relationships that provide the context in which standards are developed, and provides a central space for listing vocabulary and ontology resources with their associated authority, governance and quality of service attributes. We propose to present this in a visual form and provide pivotable search facilities to enhance recognition 47

Figure 1.11 Vocabularies exist within social and business contexts

and comprehension. A sample interface screen is shown at Figure 1.12.

The facility, dubbed ‘n2mate’ provides a focal point where discourse communities can make authority claims, rate vocabularies on various parameters, register their commitment to or usage of particular vocabularies, and provide feedback on their experiences. It is designed to be a novel exploitation of social networking software to provide a lightweight, ﬂexible platform for testing the eﬃcacy of leveraging social networks to link existing registers and ‘seed’ an information space focusing on the use of standards in online information management.sing on the use of standards in online information management.

Implementation options for ‘n2mate’ include the use of existing tools and datasets, such as traditional registry technology provided by ebXML registry, automated harvesting of the relationships into a triple-store, perhaps supported by Sesame [1] visualization and facet search through tools such as Gnizr [97] and Solr [131]. Trust and Governance will be aided by the W3C’s Protocol for Web Description Resources (POWDER) [151], once available. In addition to collecting information from users, a number of existing web services and registers will be used to seed and populate the register.

The paper uses examples from the Australian context to provide clear illustration of the central arguments. Information harvested via the ‘n2mate’ facility provides the necessary context for potential users to identify, choose and commit to the most appropriate and robust vocabularies and ontologies in their sphere of activity, and it is expected that the selection en masse will coalesce around a few of the most solid and useful vocabularies. These emerging ‘de facto’ consensus standards will then form a stable semantic platform for content representation and interlinked knowledge. Chapter 1 Introduction

n2Mate Match Maker Marine

TOP 10 RESULTS POPULARITY

Marine Fishing Vocab Used by 34 Sites Marine Boating Rules AUTHORITY Reef Marine Onto Martime Warfare W3C Speciﬁcation GeoScience Ratiﬁed Boating and Fishing YOUR FRIENDS Surf Water Language Marine Mammals Billie Uses it Lake Water Onto Markus Reviewed it Finding Nemo Vocab Jo No longer uses it Marine R Us Standard USER RATINGS MORE RESULTS 5/5 from OntoGold Beware from UNFISH

58 Search Results Returned + -

Figure 1.12 Sample n2mate interface screen

1.10.6 Chapter 7: Foundational Issues in Meaning Inclusions: 7.1 A. Cregan. Towards a Science of Deﬁnition. In Proc. Australasian Ontology Workshop (AOW05), Sydney, Australia. CRPIT, 58. Meyer, T. and Orgun, M. A., Eds., ACS. 75-82, 2005.

Personal Contribution: 100 % 7.2 A. Cregan. Symbol Grounding for the Semantic Web. In Proceedings of the European Semantic Web Conference, ESWC2007: 429-442 (2007)

Personal Contribution: 100 %

Context: The papers in this chapter deal with fundamental issues in deﬁnition, meaning and symbol grounding that are raised by the Semantic Web.

Contribution: As explained at Section 1.7, the current approach to meaning taken by RDF and OWL is that of model-theoretic semantics, but these have been insufficient to make meaning fully 49 machine processable, or to determine reference unequivocally. In Section 1.9, it was stated that in order to make meaning truly machine processable, it would be necessary for the author to specify precisely what terms mean, and correspondingly would require suitable constructs for capturing this meaning. The question of how one should go about defining terms is addressed by the first publication included at this chapter. The second publication addressed the symbol grounding problem and explains its importance for the Semantic Web.

Content: The vision of the Semantic Web is to provide machine processable meaning for intelligent applications. Whilst knowledge representation structures like ontologies have well-developed formalisms such as the OWL language, the issue of determining or specifying exactly what it is that they represent is still not well-understood. However, it is crucial for validation, merging and alignment. We cannot possibly hope to judge the accuracy or applicability of a representation structure without a clear specification of what it is intended to represent. This being the case, we must either accept that our representations will have a limited applicability and lifespan, or develop methods by which we can define our terms in a robust and standardized way. The first paper of this section proposes that what is needed is a methodology for grounding the elements of ontologies: that is, a means to objectively establish what they are representing. Currently, outputs of the process of validating, aligning or merging ontologies have to be checked by domain experts in order to verify their correctness as representations. This process goes beyond checking the internal properties of the representation itself, for example its logical properties such as consistency. It also requires checking that what the ontology is stating about the world is an accurate representation of actual states of affairs in the world. (Here we permit ‘the world’ to encompass anything we want to represent, whether physical, conceptual or social.) The need for these kind of checks by domain experts indicates that the ontology(ies) alone do not contain everything that is needed for such tasks to proceed purely by machine processing and guarantee correct results. What is missing is a complete and unambiguous method for specifying what a term represents, so that it is no longer necessary to include human beings in the process. Whilst ever it is necessary to defer to humans to negotiate meaning, the process of achieving semantic interoperability is severely hamstrung. The publication of Chapter 6 explained the n2 problem and proposed a tool to assist in making the task manageable simply by reducing the number of ontologies in common use, and therefore the number of mappings humans are required to do. In contrast, the publications in Chapter 7 tackle the question from another angle and explores the potential for releasing humans from the process altogether, once the representation is in place. Such an undertaking takes us far beyond the current Semantic Web specifications and is to be considered speculative. The first paper of the section, entitled ‘Towards a Science of Definition’, investigates the nature of the representation relationship, with a view to uncovering principles for relating a term unambiguously and unequivocally to the thing that it represents. It is often claimed that definition is an art rather than a science e.g. [148], and the paper seeks to challenge this notion and investigate to what extent it can be done scientifically. The investigation draws on philosopher Richard Robinson’s 1950 analysis of definition [135] in order to shed some light on the matter and begin to draw out some useful distinctions. Firstly, Robinson notes that definition can be used to relate words with words, words with things and things with things, and concludes that only the relationship between words and things should be considered to be definition: definition is the activity of giving infor- Chapter 1 Introduction mation about what a symbol is referring to or denoting. This concurs with the question we are exploring and the notion of Symbol Grounding which is expounded in the second paper of this chapter. Secondly, Robinson distinguishes between lexical definition and stipulative definition. Lexical definition, commonly used by dictionaries, records how words relate to things in the context of language use. These social agreements belong to a particular time and community, and are pliable. In contrast, stipulative definition is the act of setting up a relationship between a symbol and a referent (the thing the sign stands for). We note that lexical definition is word-driven, and the referent attached to any particular word tends to shift over time and between usage communities. In contrast, stipulative definition seems to be driven by the referent: we attempt to circumscribe it precisely and then assign a word to it. Thirdly, Robinson offers a number of methods for relating words and things by the methods of:

1. Synonyms Uses a word-word correspondence to access an existing word-thing relationship.

2. Analysis Describes the thing being deﬁned and relies on the listener to be able to mentally construct the description.

3. Synthesis Describes where the thing being deﬁned sits in relation to other things within some known system or organization

4. Implication Uses other methods in conjunction with use of the word being deﬁned; the deﬁnition is implicit rather than explicit.

5. Denotation Mentions known examples of the thing being deﬁned, leaving the listener to abstract the connotation.

6. Ostension Unlike the previous ﬁve methods, this method does not use words alone but draws the listener’s attention to some actual thing or non-verbal representation of a thing, for instance by pointing to it.

7. Regularity Whilst the six methods above are suitable for names, where a name has a ﬁxed correspondence to a thing, other words are not names and need to be deﬁned by rule.

Whilst it has been argued that some concepts like ‘beauty’ or ‘good’ are intrinsically undefinable, this is due to difficulty in determining their structure and essential nature, due to it being unanalyzable, rather than to the inability to establish what thing the word refers to. Unanalyzable concepts may well have primary importance as motivators of human behaviour, as George Kelly’s personal construct psychology [89] would attest, but unanalyzable is not the same as indefinable. The paper argues that there are four fundamental mechanisms underlying Robinson’s methods of definition, all of which are mental in nature. Each of Robinson’s methods is aimed at a listener who interprets the speaker’s definitional activity. It is necessary to point out initially that what Robinson refers to as things, are in fact concepts of things: that is, the listener’s mental representation of things. Thus instead of using Robinson’s word-thing pairs, it is more appropriate to frame definitional activity as a Peircian semiotic triangle between (using Ogden and Richards 1923 terminology) [115, 123] (a) symbol, (b) thought or reference and (c) referent as shown in Figure 1.13. 51

Figure 1.13 Semiotic Triangle - Odgen and Richard’s 1923 version

According to the semiotic triangle, the symbol represents the referent as mediated by the mental thought or reference, and thus there is not necessarily any observable or direct relationship between the sign and the referent. The paper argues that Robinson’s methods of definition are in fact connecting the symbol with the listener’s thought or reference of the referent, rather than linking the symbol directly with the referent. Therefore it is appropriate to delineate the mental mechanisms whereby the listener is able to create connections between symbols and mental representations. We propose that the listener developmentally acquires and uses four mental mechanisms to identify and create the appropriate mental representation within what we refer to as their conceptual landscape, inspired by Gärdenfors’notion of conceptual spaces [48]. The four mental mechanisms whereby definitions are interpreted are proposed to be:

1. Example: associating words with recurring clusters of sensory stimuli in the mental landscape. This process enables language learning to commence, by associating symbols to mental constructs corresponding to elements of bodily experience.

2. Semantic Relation: abstraction is used to learn semantic relations such as ’between’ and ’above’. Learned initially in relation to the physical (sensory) environment, they can then be applied to conceptual landscapes.

3. Analysis: once the conceptual landscape is suﬃciently developed, concepts can be applied to each other to extend the conceptual landscape into purely conceptual domains such as mathematics.

4. Rules: Within the conceptual landscape, particular territories can be marked out using language and the concepts it corresponds to.

The process relies to a certain extent on the notion of commonality in the conceptual landscapes of humans, as supported by the evidence of having common perceptual and cognitive apparatus and thus similar underlying conceptual spaces [48], Rosch’s work establishing basic level categories corresponding to sensori-motor handling [136], Johnson’s experiential gestalt ‘image schema’ [81], Piaget’s common stages of cognitive development in infants and children [127], Wierzbicka’s identification of common ‘semantic primes’ across Chapter 1 Introduction languages [150], and CK Ogden’s ‘Basic English’ [116]: an exercise in successfully reducing English to a functional core of under a thousand words for instructional purposes. Lakoff and Johnson’s work ’Metaphors we live by’ [93] expounds a mechanism whereby concepts learned in relation to the physical domain are applied to conceptual domains. The second paper included in this chapter, entitled ‘Symbol Grounding for the Semantic Web’, extends the arguments advanced in the first paper. It argues that whilst the objective of the Semantic Web is to make information more easily shared and applied by making its meaning explicit [15], current Semantic Web technologies lack an adequate notion of meaning, and consequently, ontologies are logical castles in the air without firm foundations. It undertakes an analysis of meaning that draws the distinction between meaning as entailment and meaning as designation. Whilst the semantic web has well-developed notions of logical entailment, as underpinned by its model theoretic semantics, it lacks a methodology for designation: that is, for connecting the symbols it uses to their referents. The paper explores designation and draws a distinction between its two aspects: denotation and connotation, a distinction introduced by the philosopher John Stuart Mill [104]. The denotation of a term is all the individuals to which it may correctly be applied, whilst its connotation gives the attributes by which the term is defined. The semantic web, to date, has based its notion of meaning on model theory and so is essentially denotationally based at the level of its logical primitives. In contrast, it is argued that it is important to establish a correspondence between primitives and the domain that is connotationally-based and thus is an extra-logical consideration. The challenge of establishing a relationship between symbol and referent is explained to be what is known in AI circles as the ‘Symbol Grounding Problem’. Harnad’s [62] 1990 paper uses Searle’s famous Chinese Room scenario [141] to expose the inadequacy of defining symbols in terms of other symbols, asking the reader to imagine trying to learn Chinese without any pre-existing knowledge of the language, using only a Chinese-Chinese dictionary. Without some way to ‘ground’ the symbols to something outside of the dictionary, one is caught on a symbol/symbol merry-go-round and is never able to come to a halt on what anything means. The example Chinese dictionary provides a parallel for that of ontologies used on the Semantic Web: whilst they may have intricate interconnections, without some mechanism to connect the logical structure to the world they are ungrounded. Having a structure that supports logical entailment is then of dubious benefit, as there is no guarantee that logically sound results will correspond to accurate descriptions of the domain. This principle is illustrated by Narens’ work on meaningfulness in Mathematical Psychology [109]. In the context of the semantic web, the resolution of semantic conflicts generally needs to make reference to the real-world dimensions and entities being represented. The paper uses the analysis of types of semantic conflicts from Pollock and Hodgson [129] to illustrate, giving the resolution method for each and explaining why and how Symbol Grounding and the use of meaning is relevant for executing the resolution. Currently such resolutions need to be undertaken by domain experts and is unable to be automated, and it is hoped that examining these patterns in greater depth will help uncover suitable approaches that would allow such resolutions to be automated. The paper also considers the mapping constructs currently provided by OWL, explaining that they are limited to specifying that two or more classes, properties or individuals are the same or different. Such identity or difference needs to be determined outside of OWL, and although tools and heuristics can support the process, there is no formal basis for a machine to determine this. The paper argues that the use of class extensions and URIs is inadequate for the following reasons. Noting the distinction between intensive and extensive class definitions, it is clear 53 that simply having the same extension is inadequate to show that two classes (or properties, when considering pairs) are the same under a connotative(intensive) notion of meaning. Furthermore, class membership can be inherently fuzzy [48], for example, in determining the class of all red objects, and a complete specification of meaning clearly needs to be able to support graded class membership, as well as be able to determine whether a previously unseen instance qualifies for class membership or not. It is commonly argued that reference may be adequately established for the Semantic Web by the use of URIs. However, with regard to establishing reference, two classes, individuals or properties with the same URI cannot necessarily be concluded to be representing the same referent, if what is being represented is outside the information system. There would need to be some one to one correspondence in place between URIs and referents to make this workable. However, the case where what is being represented is some information resource that can be dereferenced (and compared directly with another dereferenced URI if necessary), is clearly a case where one may clearly establish the sameness or difference of the referent by use of the URI.

2 Overview of Semantic Technologies

2.1 Overview of Semantic Technologies

Title of Publication: Overview of Semantic Technologies Type of Publication: Book Chapter Appears In Book: Handbook of Ontologies for Business Interaction Editor: Peter Rittgen Publisher: IGI Global Publication Date: 2007 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Personal Contribution: 100%

Permission to reproduce this publication has been obtained from the copyright holder, IGI Global.

55 Chapter 2 Overview of Semantic Technologies 57

Chapter I Overview of Semantic Technologies

Anne M. Cregan National ICT, Australia University of New South Wales, Australia

ABSTRACT

Semantic technologies are a new wave of computing, using explicit representation of meaning to enable data interoperability and more powerful and ﬂexible information services and transactions. At the core of semantic technologies are ontologies, which capture meaning explicitly and may be used to manipulate and reason over information via its semantics. Unlike traditional data schemas or models, ontologies are capable of representing far more complex relations, may be linked directly to the data they describe, and have a formal logical semantics, facilitating automated deductive reasoning. This chapter introduces the vision of semantic technologies, and provides an overview of the approach and the techniques developed to date. It provides both an executive summary and an orienting framework for reading more technical material.

INTRODUCING THE VISION touted for ages will ﬁnally materialize. (Berners-Lee & Fischetti, 1999, p. 169) I have a dream for the Web [in which computers] become capable of analysing all the data on the Web—the Technology visionaries like Sir Tim-Berners Lee, content, links, and transactions between people and the inventor of the World Wide Web, have long dreamed computers. A “Semantic Web,” which should make of such a seamless information technology platform this possible, has yet to emerge, but when it does, the (Berners-Lee & Fischetti, 1999) to support distributed day-to-day mechanisms of trade, bureaucracy and business and government and personal interactions, our daily lives will be handled by machines talking as well as other information-based activities like to machines. The “intelligent agents” people have research, learning, and entertainment. The beneﬁts

Overview of Semantic Technologies

of sharing and using knowledge seamlessly, globally, and on demand hold great promise for the future of HARNESSING SEMANTICS economics, government, health, the environment, and all areas of human life. Semantic technologies, which Typically, each IT system reflects the unique missions, are designed to process information at the level of its work flows, and vocabularies of its own organization. meaning, hold the key for delivering this vision. Differences in syntax, structure, and the concepts The amount of worldwide digital data generated used for representation prevent the interoperability of annually is now measured in exabytes (1018 bytes), information across systems and organizations. Whilst (Lyman, & Varian, 2003) providing access to un- middleware and data exchange standards like XML precedented amounts of information. While methods (Bray, Paoli, Sperberg-McQueen, Maler, & Yergeau, and technologies to store data and retrieve it reliably 2006) address some of the problems, they provide only and securely over distributed environments are well- a partial solution. The main obstacle in achieving ef- developed and generally highly effective, the ready ficient and seamless system integration is the lack of availability of vast amounts of data is, in itself, not effective methods for capturing, resolving, and using enough. Each data store is designed within its own meaning, a field referred to as “semantics.” organization or business unit for a specific purpose, To date, information processing has been primar- and the resulting vocabularies, data formats, data ily at the syntactic or symbol-processing level, whilst structures, data value relationships, and application the semantic level—the level of the meaning of the processing vary considerably from one system to an- information—has been relatively inaccessible to other. Faced with information overload and a spectrum machine processes. The knowledge of exactly what of incompatibility, most organizations are experiencing the data means resides in the mind of the database a constant struggle to find, assemble, and reconcile architect, system designer, or business analyst, or, if even a portion of the potentially relevant and useful made explicit, in a document or diagram produced data, even within the enterprise itself, and the potential by these people. Such documentation is not in an benefits of leveraging the knowledge implicit in this executable form and without a direct function in the data are largely untapped. live system it quickly becomes out of date. On the Semantic Technologies are a new wave of comput- other side of the coin, the understanding of the needs ing (Niemann, Morris, Riofrio, & Carnes, 2005) that and wants of the information consumer resides in enable one system to make use of the information their mind, and traditionally there has been no way resident in another system, without making funda- for them to represent this directly or to match their mental changes to the systems themselves or to the needs with the system. way the organization operates. In the same way that Semantic technologies provide the capability to a universal power adaptor enables an Australian ap- handle information on the basis of its meaning, or pliance to be plugged into a PowerPoint in Europe, semantics. The core idea of semantic technologies the U.S., or Asia without the need to change the local is to use logical languages to make the structure and power grid, semantic technologies enable semantic meaning of data explicit, and to attach this informa- interoperability for IT systems with different data tion directly to the data, so that at run-time, automated structures, formats, and vocabularies, without chang- procedures can determine whether and how to align ing the core systems themselves. By providing more information across systems. By enabling this “semantic effective ways to connect systems, applications, and interoperability” across systems, a linked virtual data data, greater capabilities like intelligent search, au- structure is created, where the relevant data can be tomated reasoning, intelligent agents, and adaptive searched, queried, and reasoned over across multiple computing become possible, and the potential to native data stores based on its common meaning. leverage existing information for far greater benefits becomes realizable.

2 59

Overview of Semantic Technologies

KEY STRATEGIES OF SEMANTIC for searching and intelligent processing over the virtual TECHNOLOGIES data store created. The key strategies used by semantic technologies The goals of semantic technologies are twofold: firstly, are: to make distributed, disparate data sources semantically interoperable so that data can be retrieved and • Tagging physical data with metadata describing aligned automatically and dynamically on demand, the data. Metadata is unlimited, in the sense and secondly, to provide techniques and tools to en- that it can describe anything about the data. able machines to intelligently search, query, reason, Additionally, because it links directly to the and act over that data. data it is about, the tag provides a handle for Semantic technologies capitalize on the availability data identification and retrieval. of data in sharable, processable electronic form. Some • Metadata tags are organized into logical struc- of these forms (e.g., databases and XML documents) tures called ontologies, which capture the logical contain structured data, and some contain less struc- and conceptual relationships between the tags, tured or unstructured data (e.g., text documents and and provide a semantic map overarching the Web pages). Semantic technologies can work with data data. in any form, providing it can be directly electronically • Aligning and mapping ontologies produces a linked into an ontology through some form of unique semantic map over all the data sources, creating identifier. Ontologies are explicit, machine-readable semantic interoperability, and providing the pos- specifications of the structure and meaning of data sibility for coordinated and seamless searching, concepts, enabling automated processes to map and querying, and processing over the virtual data reconcile the data into a conceptually cohesive whole structure.

Figure 1. Semantic technologies overview

3 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

• As ontologies are underpinned by formal log- of meta-data over the same data, there is no limit to ics, they support automated reasoning over the kind or amount of metadata that may be used to the amassed data. Semantic interoperability describe and organize the same information. Complex thus provides a basis for semantic brokers and relationships within and between the various metadata semantic services. Intelligent agents then may can be harnessed and used for knowledge processing. compose these services to perform more complex tasks on behalf of the user. Ontologies

As shown in Figure 1, each level builds succes- Ontologies are the key component of semantic tech- sively upon the previous one, while emphasizing the nologies, whether for the Semantic Web or other ap- decoupling of data from applications for greater reuse plications. The word was borrowed from philosophy, and modularity. but as applied to Semantic Technologies, it is commonly defined as “the specification of a conceptual- Data ization” (Gruber, 1993), and may be thought of as an explicit conceptual model representing some domain Rather than replacing existing database technol- of interest. ogy, semantic technologies allow data to continue Ontologies organize metadata tags, capturing the to physically reside in its native environment, while logical and conceptual relationships between them, and providing improved access to the data via a con- electronically linking each tag directly to the data or ceptual virtual layer. By making the meaning of resource it represents. Typically ontologies describe the data explicit, it may be harvested more easily the individuals, concepts, and relationships that are for new uses. The data simply needs to be linked relevant for conceptualizing some real-world domain. to a metadata tag via some form of unique identi- The kind of knowledge they capture include: fier. For World Wide Web resources, unique resource identifiers (URIs) perform this function. • The concepts of the domain, and relations between the concepts such as broader, narrower, Metadata and disjointed. These set up the basic terminology of the domain. Metadata is data about data. XML, for instance, is • Properties that relate concepts to each other a common standard used to attach metadata tags to and to data fields, specifying the nature of the raw data. Metadata can be used to capture anything relationship, constraints on the relationship, and at all about data: its format, syntax, structure, se- ranges for data values. mantics, pragmatics, or any other relevant aspect. • Assertions or facts about individuals in the do- Metadata about format and syntax can be used to main; for example, that a particular individual guide processes which physically link and retrieve the is an instance of a particular concept. data, while metadata about the data’s structure and meaning can guide semantic alignment of the data. Ontologies are closely related to existing data Pragmatic metadata can be used to capture information modeling methodologies, but enable more explicit, about how the data can be used in action. Semantic richer descriptions, with more emphasis on the mul- Technologies represent metadata in a form suited tiplicity of relationships and on precise formulation to logical manipulation, and are thus a tool which of logical constraints. One of the key principles of may be applied in any or all of these scenarios. As semantic technologies is to decouple information ontologies support the co-existence of multiple kinds from applications, so that it can be redistributed and

4 61

Overview of Semantic Technologies

re-used by other applications, both inside and outside Intelligent Agents the enterprise. Whilst current methodologies implicitly reference logical relationships, ontologies capture these Finally, intelligent agents can use semantic brokers to explicitly, decouple them from the application, and find and compose services to undertake complex tasks make them available for machine processing. on behalf of users. The modularity of data, logic, and For instance, a coded application procedure may application supports the composition and redeployment make use of the programmer’s knowledge about the of each element for new and innovative uses. way years, months, weeks, days, and hours are related in order to process temporal data, without actually making this knowledge explicit in a way that can be APPLICATIONS AND BENEFITS reused by other applications, or redeployed for un- forseen purposes. In contrast, an ontology captures The innovations that semantic technologies offer such knowledge explicitly, removing the need for it simplify the process of achieving interoperability to be coded in at the application level, making the between data sources, paving the way for vastly im- knowledge available for automated reasoning, and proved searching, querying, and reasoning over the supporting reuse by other applications. As ontologies amassed data. are the key enabler for semantic technologies, they The semantic interoperability community of prac- are examined in depth in the section titled “Explor- tice (SiCoP) forecasts that in the near term, semantic ing Ontologies.” technologies will deliver the capabilities of information integration and interoperability, intelligent search, and Semantic Interoperability semantic Web services, and in the longer-term, will deliver model-driven applications, adaptive autonomic Mapping and aligning ontologies provides a cohesive computing, and intelligent reasoning (Niemann et semantic view of multiple data sources, enabling al., 2005). Each of these applications brings its own searching, querying, and reasoning across them as specific benefits. though they were a single data store. Mechanisms for one ontology to import and use another at run time are Information Integration and provided, as well as tools for the semantic alignment Interoperability of ontologies. Aligned ontologies are connected via explicit mapping of the entities in one ontology via Typically, an organization needs to work with and semantic relationships to entities in the other ontology. reconcile multiple data sources, including disparate Such alignment can be human-mediated or semi-auto- systems within the enterprise or between different mated, using heuristics and matching algorithms. organizational systems in the supply chain, across an industry, between government organizations, or Semantic Brokers and Services on the Web. The ability to seamlessly integrate these into a cohesive whole for search, querying, retrieval, Semantic brokers and services take advantage of and reasoning is clearly of great benefit. When busi- semantic data interoperability to provide intelligent ness units or parts of the supply chain are not cur- search and other reasoning-based services over the rently connected, or when a corporate merger takes interlinked data. The use of ontologies supports model- place, the ability to connect data at a virtual semantic driven applications to access and process executable level, rather than having to physically merge it, is a models of the domain. powerful means to expedite operational efficiency and effectiveness.

5 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

Semantic technologies reduce the cost and effort with a Web ontology language service specification involved in integrating or aligning heterogeneous data (OWL-S), according to the OWL Services Coalition sources by obviating the need for system architects to (2004). OWL-S provides a core set of constructs for perform pairwise mappings between each and every describing the properties and capabilities of a Web system in order to achieve interoperability. Using service. Itself an ontology, it enables better discovery semantic technologies to appropriately represent the of relevant Web services and a more flexible framework data at a meta-level, each system may only need to for composing and using them. be mapped once to achieve interoperability with other systems. Having data mediation take place at the Model-Driven Applications meta-level, based on explicit logical representation and automated inferencing, thus reduces the time and cost Model-driven applications enable software applica- involved in achieving system interoperability. tions to directly access and process actionable models of a domain. These models explicitly capture the Intelligent Search entities, logical relationships, and business rules of the domain. By separating logic and business rules Most search capabilities are currently based on string from applications, they can be maintained centrally matching, and make limited use of conceptual rela- and explicitly. When domain logic such as business tions like synonyms and contextual information, let rules change, the domain model can be updated with alone more complex semantic relations. As a result, the new logic and the change automatically flows most searches deliver only a subset of the available through to the relevant applications. This enables relevant data and a large amount of irrelevant data. software developers to produce software applications Sifting through search results to find the accurate to support and execute business processes more quickly “hits” is time-consuming and tedious. Additionally, and easily. Less code maintenance is required and the the relevant data sources may not present the data business is able to be more responsive to changes in in a way that is easily processable for the intended the business environment. usage, so significant effort is needed to adapt the data into the right format, structure, terminology, Adaptive and Autonomic Computing measurement units, currency, and so on. In contrast, Intelligent search, combined with interoperability Adaptive and autonomic computing capabilities en- capabilities, enable searches to be conceptual-driven. able applications to diagnose and forecast system Context-sensitive, preference-driven searching can problems and system administration needs. The use be performed across combined data sources to find of self-diagnostics and support for complex systems all the relevant information, and, additionally, the planning are helpful for system administrators, and approach supports flexibility in specifying the form allow them to maintain reliable systems with less cost results should take, with retrieved data automatically and effort involved. transforming accordingly. Intelligent Reasoning Semantic Web Services The ability to reason effectively over a virtual data store The World Wide Web has an existing suite of Web ser- using meaning directly is clearly very powerful, and vice standards based on universal description discov- creates all kinds of new possibilities in terms of what ery and integration (UDDI), Web services description machines can do for us and for business. Ontologies language (WSDL), and simple object access protocol currently provide a formal logical semantics to enable (SOAP), and Semantic Web services combine these automated reasoning over the amassed data. Whilst

6 63

Overview of Semantic Technologies

traditional databases already have an underlying with a view to each agency aligning its architecture logical base supporting their existing functionality, to a common reference model, enabling architects which ensures, for instance, that querying is sound in other agencies to more readily understand ar- and complete, ontologies enable more sophisticated chitectural components and to identify possibilities reasoning techniques to be used while still ensuring the for collaboration and re-use. The specific goals fidelity of the results. Some of these techniques require of the FEA architecture framework included: greater processing power than databases typically use, and work is proceeding on optimization algorithms to • Elimination of investments in redundant IT ensure response time is not compromised. In the future, capabilities, business processes, and capital as reasoning tasks become more intelligent, techniques assets for probabilistic and other kinds of reasoning are • Identification of common business functions likely to appear. Safeguards are required to ensure across agencies and reuse of business processes, reasoning is conducted correctly and transparently data, and IT-components for time and cost sav- with the ability to provide accountability, including ings human-understandable explanations and justifications • A simpler way for agencies to determine whether for any conclusions reached. potential new IT investments were duplicating efforts of other agencies, eliminating unneces- sary expenditure CASE STUDY: U.S. FEDERAL • A means for agencies to evolve the FEA business GOVERNMENT reference model as their needs and situations changed Semantic Technologies have extensive applications for interoperability, integration, capability reuse, These goals had the potential to save the U.S. accountability, and policy governance within and Federal Government many millions of dollars annu- across government agencies (Hodgson & Allemang, ally, while significantly improving the quality and 2006). The U.S. Federal Government has made use of effectiveness of government services. semantic technologies to improve cost-effectiveness and service quality across the U.S. Federal Govern- Formalizing the FEA Reference ment agencies, as described in detail in Allemang, Models Using Ontologies Hodgson, and Polikoff (2005) and outlined below. The reference models of the FEA were written in Federal Enterprise Architecture natural language and presented as PDF files. Whilst Reference Models they could be read by anyone, the alignment process could not be implemented or verified without sig- In the U.S. in 2004, a federal enterprise architecture nificant subjective interpretation, which is prone to (FEA) designed by the U.S. Office of Management and ambiguity and errors. Creating formal representations Budget (OMB) to facilitate cross-agency analysis and of the reference models provides objective criteria for identify duplicative investments, gaps, and opportuni- conformance. However, the reference models were ties across U.S. Federal Agencies was released. The exceedingly complex and could not be represented by framework for providing these benefits comprised simple lists and hierarchies. Luckily, by early 2004, five reference models relating to performance, busi- the OWL (Web ontology language) standard (Smith, ness, services, technology, and data. These models Welty, & McGuinness, 2004) had been formally recom- were conceived by researching and assembling the mended by the World Wide Web Consortium (W3C), current practices of the various government agencies the Web’s governing body. By using OWL ontologies,

7 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

it was possible to fully capture the models formally, presidents’ e-government initiatives, was able to ensure the conformance and logical consistency of provide project-specific guidance for preparing implementations, and provide a basis for combining business cases, ensuring project compliance the implementations of different agencies into a uni- with the FEA, knowledge of related initiatives fied whole. and possible duplication, and candidate federal, The work of creating the ontologies was performed state, and local partners for the project. by TopQuadrant Consultants over a 3 month period. • Ability to dynamically generate cross-refer- By creating a set of OWL ontologies to cover the ence tables showing multidimensional agency five reference models, plus bridging and reference relationships and capabilities, through use of a ontologies, they were able to create an ontology-based “model-browser” directly linked to the relevant system to support an automated advisor to answer data, ensu r ing an up-to-date view over all infor- questions such as: mation gathered directly from the information source. • Who is using which business systems to do what? • Who is using what technologies and products EXPLORING ONTOLOGIES to do what? • What systems and business processes will be Broadly speaking, an ontology is any specification affected if we upgrade a software package? of a conceptualization, and, in this broad sense, can • What technologies are supporting a given busi- include virtually any kind of model or representation, ness process? including taxonomies, entity-relationship diagrams, • Where are components being re-used or where flowcharts, and so on. In recent years, ontologies have could they be re-used? drawn from the disciplines of artificial intelligence, • What are the technology choices for a needed particularly knowledge representation & reasoning, component? and formal logics, evolving the ability to represent • How is our agency architecture aligned with the more complex relationships supported by an under- FEA? lying formal semantics. This section explores those capabilities. The Semantic Web ontology language An ontology graph was produced, which captured (OWL) is currently the most well-developed language the rich relationships connecting the concepts stated for building ontologies, and the examples and descrip- across the five FEA reference models. These relation- tions used may be taken to reflect OWL unless stated ships provided a basis for understanding and reasoning otherwise. Please note, however, that OWL is not over the overall model. Some of the resulting benefits confined to use on the Web: being XML-based it may included: be implemented as widely as XML itself.

• Answering the listed questions through use of Expressing Knowledge model querying and automated reasoning. For instance, automated graph traversal reasoning The typical constructs used by ontologies include was used to infer “line-of-sight” between dif- classes (also known as concepts), instances (or indi- ferent enterprise entities. viduals), and properties (or relations), which have a • Context-speciﬁc information: a “capabilities complex set of possible roles, interrelationships, and advisor”, using a semantic engine to advise dif- constraints. ferent stakeholders on the capabilities available Instances correspond to individual things that or in development to support the FEA and the U.S. have associated properties, whilst classes are various

8 65

Overview of Semantic Technologies

Figure 2. Typical ontology constructs

Class Class

groupings over those things and properties are the class Person may have subclasses Male and connections between them. Figure 2 shows an example Female defined to be disjoint from each other, of an ontology illustrating these notions. so that no Person may be an instance of both Male and Female. A set of subclasses may • Classes contain instances, for example, the class also give complete coverage of the class they Female contains specific individual Mary belong to—for instance, it can be specified that Smith. the two classes Male and Female completely • Classes are typically related to each other by cover the class Person, so that every Person subclass relations, meaning that the instances must be an instance of either Male or Female; in one class are a subset of another; for example, there can be no Person who is neither Male Male is a subclass of Human. Subclasses nor Female. inherit the properties of all their superclasses; • Classes may have properties which connect for example, if Engineer is a subclass of them to specific literal values or individuals; Technical Profession, and Technical for example, a Person may have a specific Profession is a subclass of Occupation, age which is a non-negative integer and have then Engineer inherits all the properties of a specific relationship to other individuals. For both Technical Profession and Occu- example, a Person can be a familial relative of pation. This entails that instances of subclasses another Person. While the property is defined are automatically classified as instances of the on the class, note that it applies to the individuals classes above, for example, if John Smith is in the class, rather than to the class itself—that a Male, he is automatically an instance of the is, it is each individual Person who has an age class Human also, inheriting any properties of value, not the class Person itself. Human. • Properties may have specific domains and • There can be distinct sets of class-subclass hi- ranges. For example, “husband of” is a property erarchies that overlap; that is, ontologies are not with domain Male and range Female. This just a tree (hierarchy) but a graph. For instance, means that the Husband Of property may Engineer can be a subclass of both Techni- only apply to an individual who is an instance cal Profession and of Person. of the class Male and may only connect that • Classes may be disjoint from each other, that is, individual to an individual who is an instance have no instances in common. For example, the of the class Female.

9 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

Figure 3. Example ontology

Integer

• Cardinalities and other property characteristics In building an ontology, there are potentially many like subproperty of, transitivity, being the inverse design decisions in choosing how to represent the of another property, and so on may be speciﬁed. domain to be made to ensure the ontology will best For instance, Has Age is given a cardinality suit the stated purpose. More than one model may be of exactly one: a Person has exactly one age. considered to be “correct”, but usually some designs Has Husband would have a cardinality of will provide the desired functionality more readily than maximum one: a Female may have no more others. As experience and understanding develops, than one husband, but may have no husband. ontology engineering is emerging as a research area Has Wife is the inverse property of Hus- and profession in its own right. band Of: if a certain Male is the husband of a Female, then that Female is the wife of the Components Male. Has Husband is also a subproperty of Familial Relative Of, and thus inherits Ontologies may be understood in terms of language, from and specializes this property. structure and content components. While closely in- tertwined, each component performs a separate and distinct function.

10 67

Overview of Semantic Technologies

Figure 4. Ontology components

Ontology Language Ontology Model Structure

The ontology language provides the fundamental Using the ontology language, a model is built to rep- modeling constructs for building specific ontologies. resent information about a domain of interest. This It includes language constructs relating to classes, structure is like a template or stencil, specifying the properties, instances, and other formal constructs concepts and the logical relationships and constraints reflecting the various interrelationships and constraints they must satisfy in every specific case. For instance, these may have. Grammatical rules specify how these the example in Figure 3 defines classes Person, Male, may be combined. Every ontology written in the ontol- and Female and their relationships. In description ogy language uses these constructs, independently of logics, this part of the ontology is referred to as the the domain being modelled. T-box, as it is where the terminology is defined. The particular language used is chosen by the on- Ontology model structure is usually static in real- tology builder for its ease of use, expressivity, logical time processing (excepting the provisions for auto- properties, and tool support. OWL is a very expressive mated merging and importation between ontologies), ontology language, based on a kind of formal logic but ontology authors or owners may choose to adapt known as description logics. Several subspecies or and extend it as often as they wish, usually in a way “flavors” of OWL with different expressive and logical that is backwards compatible with previous versions properties are available (OWL-Full, OWL-DL, and of the ontology, unless the entire conceptualization OWL-Lite). OWL essentially extends the constructs is radically changed. Ontologies for business are of the resource description framework (RDF) and likely to be reasonably shallow and relatively static, RDF-schema. whereas ones describing intricate research domains

11 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

like medical science may need progressive clarifica- multiple data sources. Elements of the ontology lan- tions, extensions, and revisions as the underlying guage, model structure, and content can physically understanding of the research area evolves and the reside anywhere: the ontology language may reside in conceptual model changes. a W3C namespace, linked in by its URI, the ontology structure can live on an analyst’s desktop in another Ontology Content namespace, and the data can live in a corporate data base. For instance, John Smith’s age may reside in a Ontology Content pertains to the object level of the human resources database, while the ontology contains ontology, corresponding to specific facts, individuals, a unique identifier providing a direct link to this data. and data values populating the ontology model struc- All that is needed is a way of uniquely specifying the ture. In the example in Figure 3, this would include address/location of the data or resource, via a URI or specific individuals like John Smith, his gender, some other mechanism. This approach ensures that age, and relationships. In description logics this is data can be maintained centrally and applications referred to as the A-box, as it is where assertions are always access the current information. made. Ontology content reflects and conforms to the ontology model structure, for example, when John Ability to Import, Merge, and Align at Smith is asserted to be Male, he is automatically a Run-time Person, because Male has been defined as a subclass of Person at the structural (terminological) level. Additionally, ontologies can import and build on other Depending on the tools that are used to construct ontologies, providing the ability to reuse and extend and edit the ontology, in some cases it will not be pos- ontologies. This can occur at the design phase, but sible to insert inconsistent or nonconformant data, and can also occur at run-time, merged based on match- in others errors will automatically be identified. For ing URIs or identifiers: if two data resources have the instance, an attempt to assert that John Smith is same URI, they are assumed to be the same, and a both Male and Female will either not be permitted or combined ontology structure is generated on this basis. will be flagged as an error, if the ontological structure There are also language constructs within ontology has specified that these two classes must be disjoint, languages to explicitly specify that one information and, therefore, can have no instances in common. resource is the same as another, even though they may In some ontology languages, the structure and have different identifiers, for example, synonymous content level are not kept separate: for instance, in concepts in different ontologies. OWL-Full, a class may also be an instance, while OWL-DL and OWL-Lite do not allow this. While it Connection to Formal Logics gives more freedom of expression, this has ramifications for its inferencing capabilities, as the underlying Typically, ontology languages are designed to have logic is no longer tractable. In contrast, the OWL-DL what is called a “formal semantics,” which give infer- and OWL-Lite flavors of OWL are well-behaved in ence rules for drawing valid conclusions from an exist- every respect. ing knowledge base. This ensures that starting from a knowledge base that includes only propositions that are Features of Ontologies true, and following only the specified rules of inference to deduce more propositions, will be guaranteed to Virtual Structures generate only statements which are also deductively true. Under certain circumstances, this process can An ontology is a virtual conceptual structure over also be guaranteed to produce every possible logical distributed physical resources, dynamically linking

12 69

Overview of Semantic Technologies

conclusion (RDF and the OWL-Lite and OWL-DL • Combining multiple ontologies into a consistent subspecies of OWL have this property). logical framework gives the potential to generate OWL has a description logic formal semantics, logical conclusions that could not be made from where description logics are a well-behaved fragment any component ontology alone. of first order logic. The ontology engineer has some • The data described by ontologies may be queried choice in selecting a suitable underlying description in a large variety of different and novel ways. logic via the language used to build the ontology. This Automated inference procedures can deduce the choice should optimize the relative expressiveness and answers, even though they are not specifically inferencing capabilities, as in any formal logic there asserted in the knowledge base, simply because is a tradeoff between logical expressivity (freedom they can be logically derived from the ontology’s in what can be stated) and decidability (the ability to specification of logical relationships and con- determine, using a terminating procedure, whether a straints. This optimizes re-use, and enables data proposition is or is not a valid inference). to be used in novel ways unanticipated by the data owners or even by the ontology designers. Inferencing with Reasoning Engines Rules Inferencing over ontologies based on formal logics can be automated using reasoning engines, the im- It is a matter for debate whether rules are part of or plications being: separate from ontologies, and depends on the specific language used and the logical constructs it supports. • Two or more ontologies can be joined automati- Generally speaking, rules can be used alongside ontolo- cally to create a logically consistent, combined, gies as a cohesive whole, providing additional logical interconnected framework. Given the nodes to be constraints above and beyond what is represented in matched, the ontologies organize themselves to the ontology as logical axioms. This can be useful for form a cohesive whole that is logically consistent making business rules explicit, providing the ability across all the component ontologies. to apply different sets of business rules to the same • Implicit knowledge can be generated automati- ontology and its data. cally from the knowledge representation; that is, it is possible to generate statements which Tools must be true, given what has been asserted, even though it has not itself been explicitly asserted. Ontologies need to be supported with tools for Because ontologies have an underlying logical viewing, editing, querying, and alignment. Because basis, they can be used with automated reasoners ontologies are often inherently complex, ontology to derive new information which is implicit in editors often provide multiple views over the ontol- the ontology but not explicitly represented. From ogy. Display may be either text-based, graphical, or the example in Figure 3, it can be determined a combination of both, and editors will usually offer that if John Smith is the husband of Mary a variety of views from different perspectives, for Smith, then Mary Smith is the wife of John example, by properties, classes, individuals, and so on. Smith, because husband of and wife of are inverse properties, thus it is a logical necessity. Annotation Similarly, if Mary Smith is the wife of John Smith, she is not anyone else’s wife because of Ontology languages provide the means to annotate the cardinality constraint on the wife of property, elements of an ontology with descriptive text, directly and furthermore, John Smith must be Male linked to those elements. due to the range constraint.

13 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

Figure 5. Screenshots from the Protégé ontology editing tool developed by Stanford Medical Informatics (Free, open source software available at http://protege.stanford.edu/)

14 71

Overview of Semantic Technologies

Ontologies vs. Other Data Structures for performance reasons it is not be advisable unless and Models working with a very small data set. Some data base vendors such as Oracle now support some aspects Ontologies are generally compatible with other of semantic technologies in conjunction with more methods, and rather than replacing them, leverage traditional data base technology. Rather than replac- their value. A brief comparison of the key differences ing databases, ontologies are best used in conjunction between ontologies and other technologies follows. with then to providing a conceptual virtual view over the data, enabling interoperability and leveraging its Ontologies vs. Data Models value.

While a data model may be the outcome of a conceptual Ontologies vs. Taxonomies analysis and provides the design for a database, the data model itself is not directly linked to the data, whereas Taxonomies and ontologies are related, but whereas an ontology is an explicit map of data, directly linked a taxonomy has a tree structure, ontologies have a to the data. Updating a data model, such as an entity- graph structure due to their ability to support multiple relationship diagram, does not automatically generate inheritance and to link via properties. Also, ontologies new knowledge about instance data or adapt the way have a much richer ability to capture relationships than the data links to other data sources but updating an the basic taxonomical “is-a” relation. Taxonomies may ontology can potentially do so. Data models are not be viewed as very simple ontologies. as expressive as ontologies: ontology languages are richer and treat relationships as “first-class” constructs. Ontologies vs. Expert Systems Data models can usually be constructed as simple ontologies in a straightforward manner. Ontologies may be considered as weak expert systems, in the sense that they make knowledge explicit, can Ontologies vs. Unified Modeling Lan- use rules, and support deductive capabilities. However, guage (UML) ontologies are currently more geared to capturing knowledge than to decision making per se. Intelligent UML is a specific modeling language, and compared agents using ontologies may, in the future, incorporate to current ontology languages, it provides more con- some of the capabilities envisioned for expert xystems. structs because it is intended not only for data modeling, but for modeling processes, use cases, and so on. However, it is not linked directly or dynamically to THE SEMANTIC WEB the data and does not support automated reasoning. Work is proceeding on a formal semantics for UML, a Semantic Technologies and the UML standard notation for ontologies, and executable Semantic Web capabilities, and it is likely that a convergence between ontologies and UML will be reached at some point. In While “semantic technologies” is an umbrella term en- the short-term, a tool for importing UML data models compassing all those technologies that seek to explicitly directly to an ontology editor is certainly feasible. specify, harness, and exploit meaning for automated processing, the Semantic Web is a specific application Ontologies vs. Databases of this idea to the World Wide Web. The World Wide Web Consortium (W3C), in collaborative association While data can be stored within an ontology, ontol- with many researchers and other organizations, has ogy tools are generally not optimized to do this, and developed a suite of complementary technologies

15 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

which together comprise the “Semantic Web.” In one better enabling computers and people to work in coop- sense, the Semantic Web is narrower than semantic eration” (Berners-Lee, Hendler, & Lassila, 2001). technologies generally, as not all semantic technologies necessarily make use of the W3C-endorsed recom- The Semantic Web coalesced as a specific vision mendations. However, in scope, the Semantic Web is for the World Wide Web in 1998, initiated by Bern- broader and probably more challenging than any other ers-Lee himself (Berners-Lee, 1998). However, many application, as it potentially interlinks data across of the principles on which semantic technologies are the breadth and depth of the entire Web, and needs based pre-date the Web itself, coming from diverse to handle the constant ebb and flow of available data areas such as artificial intelligence, formal logics, sources across an unlimited and constantly evolving database theory, information modeling, and library subject domain. The Semantic Web thus has to be science. The Semantic Web has been an effective more robust, or less brittle, than any other application catalyst to crystallize the efforts of many research of semantic technologies, as any brittleness is likely and industry groups into a cohesive and coordinated to produce cracks very quickly in such a demanding effort, and currently represents the most highly devel- environment. oped and complete approach for delivering Semantic Technology. The Semantic Web suite of standards is Idea not confined for use only on the Web: it is equally applicable to enterprise systems for organizing internal “The Semantic Web is an extension of the existing Web data or across private data networks coordinating in which information is given well-defined meaning, information between multiple participants.

Figure 6. Semantic Web “layer cake” (Berners-Lee & Swick, 2006)

16 73

Overview of Semantic Technologies

Table 1. Description and status of Semantic Web ‘layer cake’ elements, as of August 2007

Element Description

Unicode The basic character set encoding (pre-existing). Status: Operational

URI Universal Provides a mechanism for uniquely identifying and locating current and future resources on the Web. Resource Status: Operational Locator XML Extensible XML provides a syntax for structuring data and tagging it, without specifying or constraining the structure or MarkUp Language tags.

XML Schema is a language for restricting the structure of XML documents. Status: Operational

RDF RDF is a simple data model for referring to objects (known in RDF as resources) and specifying how they are Resource related. An RDF-based model can be represented in XML syntax. Description RDF Schema is a vocabulary for describing properties and classes of RDF resources with a semantics for Framework generalization-hierarchies of such properties and classes. and Status: Operational; signiﬁcant Database vendor support implemented RDF-S RDF Schema OWL OWL adds more vocabulary for describing properties and classes, such as relations between classes (e.g., Web Ontology disjointness), cardinality (e.g., “exactly one”), equality, richer typing of properties, characteristics of properties Language (e.g., symmetry), and enumerated classes. Status: Operational; Further extensions in development

RIF Certain kinds of logical constraints cannot be implemented by OWL alone. Rule languages provide a means Rule Interchange to implement these and are potentially very useful for encoding business rules. RIF working group currently Format active at W3C, developing a framework for rule interchange. Status: Candidates are under consideration, including SWRL and Rule-ML

SPARQL Provides the ability to query RDF. Similar in nature to SQL, but SPARQL allows for a query to consist of triple RDF Query patterns (for RDF triples), conjunctions, disjunctions, and optional patterns. Language Status: Candidate Recommendation

Unifying Logic A logical framework providing Formal Semantics for inferencing. OWL currently has a Description Logic basis. Status: DL Formal Semantics for OWL (2004); Horn Logics proposed for RIF; continuing evolution.

Proof Logical conclusions by themselves are not convincing. This layer provides justiﬁcation of inferences made, giving logical grounds for inferences. Status: In development

Trust Once a basis of logic and proof is set up, it leads to an environment of trust for conducting transactions. Status: A social variable, to be engendered by the technologies in development, especially Proof and Cryto

Crypto Support privacy and security. Status: In development

User Interface and Provide the semantic technology to the user through appropriate user interfaces and applications. The W3C has Applications emphasised the need for more well-designed UIs to encourage the spread of Semantic Technologies. Status: Mechanisms to embed RDF in existning Web are in development.

17 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

Relationship to the World Wide Web Large-Scale Semantic Markup of Existing Data The Web’s governing body, the World Wide Web Consortium (W3C), believes the Web can only reach Se m a nt ic Te ch nolog ie s r ely on d at a ow ne r s t o se m a nt i- its full potential when data can be shared, processed, cally markup their data and, to date, there is no way to and used by automated tools as well as people, and automate the process. There is an element of critical furthermore can be used by programs that have been mass here: if only a few sources are marked up, not designed independently of each other and the original as much value is delivered. However, the benefits of data sources. The Semantic Web does not replace the sharing data effectively among even a few sources can existing Web, but builds on it, enabling better interop- be quite considerable. Whether it becomes universal erability and further capabilities. While the existing remains to be seen, but key players in the IT industry Web focuses on uniquely identifying resources (URIs), are starting to embracing Semantic Technologies. The displaying information using HTML, and publishing successful and widespread adoption of bottom-up documents online, the Semantic Web focuses on data tagging in Web 2.0 applications such as FlickR and and interlinking it, ultimately supporting intelligent del.icio.us has shown the potential of the approach Web services/agents. and the willingness of participants to do tagging to support virtual communities. While existing tagging Semantic Web Components is essentially unorganized, Semantic Technologies can provide structure, logic, and reasoning to make The Semantic Web is comprised of several layers, sum- tagging far more powerful. marized by Tim Berners-Lee’s now famous “Semantic Web Layer Cake” shown in Figure 6, which has been Large Scale Data Manipulation and revised several times since it first appeared. Querying The lower layers are already well-established Web standards used in the existing Web (URI, Unicode, Tools and techniques for supporting large scale ap- XML), while the higher layers are specific to the plications need further development. While ORACLE Semantic Web, and build on the platform provided and other vendors currently support RDF triple stores, by the existing technologies. Table 1 contains a brief further integration with existing database technologies description of each element and its current status. is needed. Some semantic techniques and algorithms Full details may be found by following the links at need further optimization to ensure computing re- the W3C’s Web site (www.w3.org). sources can adequately support them.

Ontology Building by NonExperts ISSUES AND CHALLENGES A new initiative known as Sydney OWL Syntax helps Semantic technologies range from being emergent to nonlogicians to build ontologies by offering the op- being quite well-developed and mature. Many of the tion of using a simple English syntax for building and key issues are social, rather than technological. The reading ontologies, instead of having to use formal exposition in this chapter has concentrated on the logical or XML-based notations (Cregan, Schwitter, vision of semantic technologies. While key technical & Meyer, 2007). This is to be supported by a guided components have been delivered, widespread adoption interface. requires addressing several issues.

18 75

Overview of Semantic Technologies

Standards and Methods for Proof, Trust, and Security Resolving Meaning The capability of accessing and dynamically lining Interoperability relies on representing data explicitly data must be balanced with appropriate measures to and mapping it with other representations. However, h a nd le who c a n a c c e s s wh at d at a at wh at level , e n s u r i ng it is not always obvious whether it is appropriate for privacy, security, and the protection of digital rights. the elements of different ontologies to be mapped to Challenges also include ensuring that intelligent agents each other and the constructs currently available in will be accountable for their actions and able to provide ontology languages for mapping them are somewhat useful and understandable explanations of the chain limited. The ideal scenario would include a more of reasoning underlying their decisions. descriptive mapping for data transformations and mediation without human intervention. One approach is to gather stakeholders to jointly develop ontologies CONCLUSION for a domain, for use as a common standard. It is not imperative that everyone use the standard, only that While there are issues and challenges to be addressed they map their own ontology to it as a kind of “lingua and further developments in the pipeline, semantic franca”. This avoids the need for pairwise mappings of technologies are already sufficiently developed to be every ontology needed for interoperability, as each can applied in real-world scenarios, as shown by the case simply be mapped once to the common standard. study, to achieve interoperability and other benefits. On the other hand, in order to truly automate the Looking to the future, Semantic Technologies hold resolution of meaning in a way that is dynamic and great promise for delivering more intelligent informa- adaptable, it is necessary to have a solid understand- t ion se r v ic e s e n a bl i ng much mor e ef fe c t ive s u p p or t for ing of the underlying theory of semantics—not just finding, analyzing, and using knowledge, hopefully formal semantics, but cognitive semantics, situated leading to the emergence of what might be called meaning, the identification of semantic primitives, “pragmatic technologies” which use this knowledge and symbol grounding strategies. Work on upper level as a basis for effective, informed, automated action. ontologies fits in this space, as well as the author’s own work on symbol grounding for the Semantic Web (Cregan, 2007). ACKNOWLEDGMENT

Dealing with Incomplete, Uncertain, NICTA is funded by the Australia Government’s and Probabilistic Data Department of Communications, Information, and Technology and the Arts and the Australian Research Real world data tends to be imperfect and is not always Council through Backing Australia’s Ability and the suited to deductive reasoning. A W3C incubator group ICT Center of Excellence program. It is supported by [URW3] has formed to investigate bridging the gap its members the Australian National University, Uni- between the reasoning capabilities currently provided versity of NSW, ACT Government, NSW Government, by Semantic Web technologies and what is needed and afﬁliate partner University of Sydney. to deal effectively with incomplete, uncertain, and probabilistic data.

19 Chapter 2 Overview of Semantic Technologies

Overview of Semantic Technologies

REFERENCES Cregan, A.M., Schwitter, S., & Meyer, T. (2007). Sydney OWL syntax - Towards a controlled natural Allemang, D., Hodgson, R. & Polikoff, I. (2005). language syntax for OWL 1.1. Retrieved June 6, 2007, Federal reference model ontologies (FEA-RMO), from http://www.ics.mq.edu.au/~rolfs/sos version 1.1. TopQuadrant. Retrieved June 6, 2007, Gruber, T.R. (1993). A translation approach to portable from http://www.topquadrant.com/documents/ ontologies. Knowledge Acquisition, 5(2), 199-220. TQFEARMO.pdf Hodgson, R., & Allemang, D. (2006). Semantic Berners-Lee, T. (1998, October). Semantic Web technologies for e-government (SiCoP Working roadmap (W3C Draft 14). Retrieved June 6, 2007, from document). Retrieved June 6, 2007, from colab. http://www.w3.org/ DesignIssues/Semantic.html cim3.net/file/work/SICoP/2006-02-09/Proposals/ Berners-Lee, T., & Fischetti, M. (1999). Weaving the RHodgson.pdf Web. San Francisco: Harper. Lyman, P., & Varian, H.R. (2003). How much Berners-Lee, T., Hendler, J., & Lassila, O. (2001, May). information 2003? Retrieved June 6, 2007, from http:// The Semantic Web. Scientiﬁc American. www.sims.berkeley.edu/how-much-info-2003

Berners-Lee, T., & Swick, R. (2006). Semantic Niemann, B., Morris, R.F., Riofrio, H.J., & Carnes, E. Web development ﬁnal technical report (Tech. Rep. (Eds.). (2005). Introducing semantic technologies & No. AFRL-IF-RS-TR-2006-294). Retrieved from the semantic Web (SICoP White Paper Series Module http://www.scribd.com/doc/122258/semantic-web- 1). Retrieved June 6, 2007, from http://colab.cim3. development-2006 net/ﬁle/work/SICoP/WhitePaper/

Bray, T., Paoli, J., Sperberg-McQueen, C.M., OWL Services Coalition (2004). OWL:S semantic Maler, E., & Yergeau, F. (Eds.). (2006, August 16). markup for Web services (W3C Draft). Retrieved June Extensible markup language (XML) 1.0 (4th ed.) (W3C 6, 2007, from at http://www.daml.org/services/owl- Recommendation). Retrieved June 6, 2007, from s/1.0/owl-s.pdf http://www.w3.org/XML/Core/#Publications Smith, M.K., Welty, C., & McGuinness,D.L (Eds.). Cregan, A.M. (2007). Symbol grounding for the (2004, February 10). OWL Web ontology language Semantic Web (2007). In Proceedings of ESWC07 (pp. guide (W3C Recommendation). Retrieved June 6, 429-442). Retrieved June 6, 2007, from http://www. 2007, from http://www.w3.org/TR/2004/REC-owl- eswc2007.org/ guide-20040210/. Latest version available at http:// www.w3.org/TR/owl-guide/

20 3 Integrating Topic Maps into the Semantic Web

3.1 Building Topic Maps in OWL-DL

Title of Publication: Building Topic Maps in OWL-DL Type of Publication: Conference Paper Appears In: Proceedings of Extreme Markup Languages 2005 Publication Details: Published Online www.idealliance.org/papers/extreme/proceedings/ Publication Date: 2005 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Personal Contribution: 100%

77 Chapter 3 Integrating Topic Maps into the Semantic Web 79

Extreme Markup Languages 2005® Montréal, Québec August 1-5, 2005

Building Topic Maps in OWL-DL

Anne Cregan University of New South Wales

Abstract Both ISO's Topic Map Standards and the W3C's Semantic Web Recommendations provide the means to construct meta-level semantic maps describing relationships between information resources. Developed independently, attempts at interoperability between the original Topic Map standard and RDF have proved challenging. However, ISO 13250's drafting of an explicit Topic Map Data Model (TMDM) early in 2005, combined with the advent of the W3C's more expressive Web Ontology Language (OWL) Recommendations in 2004, together now provide the possibility for authoring TMDM- conforming Topic Maps directly in OWL. OWL provides the ability to express the TMDM constraints explicitly and to ensure that Topic Maps authored in OWL conform to the TMDM. This paper presents a construction of the TMDM model as an OWL-DL ontology. This “TMDM Ontology” is a construction in OWL of the Topic Map concepts modelled by the TMDM, making them available for use by Topic Map authors as a basis for building TMDM-compliant Topic Maps directly in OWL. Using OWL-DL as the language for Topic Map authoring gives users access to OWL's formal semantics, constraint expressivity, and suite of tools such as Protégé, which includes an API and capability for ontology visualisation, querying, and automated constraint checking and reasoning using Description Logic reasoners. The approach described does not require the use of any Topic Map language, or Topic Map Constraint or Query language, although a simple algorithm could translate the OWL-authored Topic Maps directly into other TMDM-based Topic Map authoring languages if required, providing access to Topic Map engines and tools. Illustrating by example, a Topic Map written in OWL-DL using Protégé is shown, highlighting the constraint and querying abilities provided, which overlap many of the requirements set by the ISO for a Topic Map Constraint Language and Query Language. One outstanding issue regarding the interpretation of Typing is identified, and options for its resolution are discussed. Chapter 3 Integrating Topic Maps into the Semantic Web

Building Topic Maps in OWL-DL Table of Contents 1 Motivation...... 1 2 Scope of the Approach...... 1 3 Background...... 1 3.1 The ISO's Topic Map Tradition...... 1 3.1.1 Topic Map Origins...... 1 3.1.2 Incomplete Specification of Topic Maps...... 2 3.1.3 A Formal Data Model for Topic Maps...... 2 3.1.4 The TMDM is not the last word on Topic Maps ...... 2 3.2 The Semantic Web Initiative...... 2 3.2.1 Berners-Lee’s Vision...... 2 3.2.2 Formal Semantics...... 2 4 Interoperability Issues...... 2 4.1 Topic Maps and RDF...... 2 4.1.1 Interoperability Challenges...... 2 4.2 TMDM and OWL...... 3 4.2.1 New Interoperability Possibilities...... 3 4.2.2 TMDM to OWL? ...... 3 4.2.3 OWL to TMDM?...... 3 5 TMDM to OWL Mapping Investigation ...... 3 5.1 The TMDM side...... 3 5.1.1 What does the TMDM give us? ...... 3 5.1.2 But wait, lurking in the back of the TMDM closet...... 4 5.2 The OWL side ...... 5 5.2.1 Flavours of OWL...... 5 5.2.2 Which flavour of OWL tastes best to make Topic Maps?...... 5 5.2.3 What can be asserted in OWL-DL?...... 6 5.2.4 OWL-DL can represent E-R models!...... 6 5.3 Using OWL-DL to build the TMDM...... 6 5.3.1 The Approach In a Nutshell...... 6 5.3.2 Why would Topic Map authors want to use OWL-DL to build TMs?...... 6 6 Current Status...... 7 6.1 Work conducted to date...... 7 6.2 Functionality Implemented and Outstanding...... 7 6.3 Intention...... 7 7 Designing and Building the TMDM Ontology...... 7 7.1 The Design Phase...... 7 7.1.1 Design Choices ...... 8 7.1.2 Identifying Individuals...... 8 7.1.3 Filling in the rest of the Model...... 8 7.1.4 A note about the use of Topics for Typing other Topic Map Objects...... 8 7.2 The Ontology Model...... 8 7.3 Additional Constraints Implemented...... 9 7.3.1 Internal and External Occurrences...... 10 81

7.3.2 At least one Locator for each Topic...... 10 7.3.3 Variant Name Scope must be a true Superset of Topic Name Scope ...... 10 7.4 Implementation...... 11 8 Examples...... 11 8.1 Topics and Subjects...... 11 8.2 Occurrences...... 13 8.3 Topic Names and Variant Names...... 14 8.4 Associations...... 15 9 OWL capabilities for TM authoring...... 17 9.1 Consistency Checks and Automatic Inferencing...... 17 9.2 Additional Constraints...... 17 9.3 Querying...... 17 9.4 Visualisation...... 20 10 Outstanding Issues...... 22 10.1 The OWL-Full approach...... 22 10.2 The "Quick Fix" Workaround...... 22 10.3 The Extra Java code approach...... 22 10.3.1 Using Object Properties...... 22 10.3.2 Generating Instances of Association and Association Roles...... 23 11 Conclusions...... 23 12 Appendix A...... 23 Bibliography...... 28 The Author...... 29 Chapter 3 Integrating Topic Maps into the Semantic Web 83 Building Topic Maps in OWL-DL

Building Topic Maps in OWL-DL Anne Cregan

§ 1 Motivation Anne Cregan, the author of this paper, is a PhD student enrolled at the University of New South Wales, and endorsed by the National Information and Communications Technology Australia (NICTA) Centre of Excellence. Part of NICTA’s mandate is to assist Australian industry to achieve best practice in the use of IT. In mid-2004, Australian industry representatives from the software and analytics industries requested NICTA’s assistance in understanding the relationship between ISO Topic Maps and the W3C’s Semantic Web Recommendations, and exploring possibilities for interoperability. As part of this effort, Anne was asked to investigate the possibility of defining a mapping between the two standards. Being within NICTA’s Knowledge Representation and Reasoning Program, which has a core capability in the use of formal logics for reasoning, she was encouraged to consider formal semantics in assessing possible solutions. This effort fits within the broader NICTA-wide priority challenge "From Data to Knowledge". § 2 Scope of the Approach This paper presents a mapping and construction of the TMDM model as an OWL-DL ontology, as a proof- of-concept of the viability of building Topic Maps based on the TMDM directly in OWL-DL. This approach gives TM authors access to OWL-DL's capabilities and tools for expressing constraints, doing constraint checking and automated reasoning, and for the querying and visualisation of ontologies. There is significant overlap with the capabilities currently provided by Topic Map engines and envisaged to be provided by the Topic Map Constraint and Query Languages in development. The approach presented herein should not be considered to be a final or definitive solution for TMDM/ OWL interoperability. The TMDM itself is still officially at Draft status and in the process of being finalised. One aspect of the specified required functionality of the TMDM is not fully implemented in this construction (see Section 10 Outstanding Issues), and the remainder has been subjected only to rudimentary testing. Furthermore, this approach has no status as a Standard or Recommendation with either the ISO or the W3C, although both bodies have indicated receptiveness to the concept and a willingness to work on interoperability, and the ISO Topic Map Working Group has kindly reviewed an earlier proposal outlining the approach, and indicated its willingness to provide continued assistance in providing feedback on OWL representations for the TMDM. By virtue of being implementation as an OWL-DL ontology, a Description Logic Formal Semantics is provided but the specifics of the Semantics are not detailed herein, and interested parties would need to consult the relevant OWL documentation. Whilst OWL and its tools provide Querying and Constraint capabilities which significantly overlap with the requirements specified for a Topic Map Query Language and Constraint Language, it is not claimed that they necessarily meet all such requirements. The TMDM ontology and examples as built using Protege are freely available at http:// www.cse.unsw.edu.au/~annec/ to anyone who may wish to download them for closer inspection and evaluation, and feedback via email to [email protected] is welcomed. § 3 Background

3.1 The ISO's Topic Map Tradition 3.1.1 Topic Map Origins Topic Maps are an ISO/IEC standard [TM] (ISO 13250-1) for mapping both web and real-world information resources, by reifying real-world resources as "subjects", and creating "topic" constructs to capture their characteristics and relationships with other topics and subjects. Dubbed the "GPS of the information universe", they are akin to an electronic back-of-book index, facilitating information navigation and retrieval in on-line environments. In 2001, the first two standard syntaxes for Topic Maps were published by the ISO: XTM1.0 in XML [XTM1.0] and another in HyTM.

Extreme Markup Languages 2005® page 1 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

3.1.2 Incomplete Specification of Topic Maps Although there were non-trivial differences between the two Topic Map syntaxes, the original standard did not explain the relationship between them, nor did it make the underlying data model explicit. Additionally, both syntaxes failed to specify what Topic Map implementations were required to do in certain situations, and neither supported constraints such as: "A person must be born in a place" or "A person must have at least one name". Requirements subsequently drafted by the ISO Topic Map Working Group (WG3) for a Topic Map Constraint Language [TMCL] and a Topic Map Query Language [TMQL] both needed a clear description as to how Topic Map constructs are to be evaluated.

3.1.3 A Formal Data Model for Topic Maps To address these issues, the WG3 commenced work on a Topic Map Data Model in May 2001, to date culminating in the official Draft ISO 13250-2 TMDM [TMDM] published in January 2005, which clarifies the representation and gives Topic Maps a formal data model for the first time, in Entity-Relationship model form. The TMDM document is currently (as at June 2005) at draft status, but close to its full completed form, whilst work on the specification of TMCL and TMQL standards, based on the Draft TMDM document, is progressing rapidly.

3.1.4 The TMDM is not the last word on Topic Maps The reader is asked to note that within the ISO Topic Map Working Group, the TMDM is not considered to be definitive of all forms of Topic Maps as described by the ISO 13250-1 Standard, and the question of distilling the essence of the Topic Map paradigm and giving its formal representation is being worked on at a more abstract level, with the Topic Map Reference Model [TMRM] effort leading the charge. However, as this work is still maturing, this paper limits its scope to the consideration of Topic Maps as they are described by the TMDM. 3.2 The Semantic Web Initiative 3.2.1 Berners-Lee’s Vision The Semantic Web is the vision of Tim Berners-Lee, and is a World-Wide Web Consortium (W3C)-led initiative with the goal of providing technologies and standards for semantic markup of online information resources, enabling improved web navigation and supporting intelligent web services. It is an XML-based technology, using Resource Description Framework [RDF]] and Web Ontology Language [OWL] layers superimposed on XML to provide more expressive representations of the characteristics of, and relationships between, logical entities. Further layers and functionalities are planned as part of the overall vision.

3.2.2 Formal Semantics The original RDF Recommendations did not specify a formal semantics, but later work has provided this, and OWL has built on to this further. OWL-DL, a formal W3C recommendation finalised in 2004, is a subset of the full OWL language which has a Description Logic (DL) equivalent semantics, providing decidable reasoning. Tools to support the generation and use of OWL-DL ontologies are well-developed, notably including Stanford University’s Protégé [Protege] package, which provides a GUI that generates OWL code and gives access to modules for ontology visualisation, querying, consistency checking and automated reasoning. § 4 Interoperability Issues

4.1 Topic Maps and RDF 4.1.1 Interoperability Challenges The ISO and W3C initiatives were originally developed independently, sheerly through lack of awareness of each other's work. On discovery of this around 2001/2002, many felt that as both the Topic Map and Semantic Web standards provide the means to construct meta-level maps of information entities, there must be synergies to be exploited. However, previous attempts at achieving interoperability between Topic Maps and RDF have been problematic for the following reasons, amongst others:

• Neither Topic Maps nor RDF had stable formalized data models at the time • Topic Maps use n-ary relations to express associations between topics, whilst RDF is composed of triples which express binary relations.

page 2 Extreme Markup Languages 2005® 85 Building Topic Maps in OWL-DL

• Topic Maps use two kinds of locators: both direct and indirect, whilst RDF has only one kind of locator mechanism. • The concept of scoping, which is fundamental to the Topic Map paradigm, has no obvious equivalent in RDF. • The use of Topics as Types within Topic Maps has been problematic — should Topics be interpreted as classes, instances, or both? How does one retain the flexibility that the Topic Map Paradigm offers in this respect, where any Topic can take on a multiplicity of uses, whilst still providing an unambiguous formal model for representation? It has proved to be a complex and non-trivial issue which does not easily lend itself to a general solution appropriate for every case. A full analysis is available in a survey of RDF/Topic Map Interoperability Proposals [RDFTM], recently published by the W3C's RDF/Topic Maps Interoperability Task Force. This task force, which is part of the W3C's Best Practice initiative, includes representatives from both the ISO Topic Map Working and the W3C, and is currently working to develop a standard for interoperability between RDF and Topic Maps. 4.2 TMDM and OWL 4.2.1 New Interoperability Possibilities Many of the problems in achieving interoperability between RDF and Topic Maps appear to stem from RDF being insufficiently expressive to capture the full semantics of TM constructs in a straightforward way. The lack of a formal TM specification prior to the TMDM coming into being has also been a considerable hindrance. So the release of an official ISO Draft of the TMDM, combined with the advent of the more expressive OWL language as an official W3C Recommendation, is potentially of huge benefit for interoperability efforts, and begs an investigation of the fit between the two.

4.2.2 TMDM to OWL? The TMDM is fundamentally a data model, whilst OWL is a language. It therefore seems natural to investigate whether the model defined by the TMDM can be represented and implemented using the semantics provided by the OWL language. That investigation is the primary focus of this work. The mapping produced is an Object Mapping as opposed to a Semantic Mapping, in that it takes the letter of the TMDM specification as bible and represents it fully in OWL, and produces a result where core TMDM constructs are mapped not to core OWL constructs, but to an OWL ontology built using them. Topic Maps are then authored in OWL simply by importing the TMDM ontology and populating the ontology with instances. If constructed in this way, such Topic Maps are quite intuitive to build, and are directly translatable to other Topic Map languages based on the TMDM. Although not covered here, there is also the possibility for another OWL representation for Topic Maps, also being considered by the author: one which does not adhere to the letter of the TMDM but produces the desired functionality using a simpler, more natural OWL ontology.

4.2.3 OWL to TMDM? The question of whether ontologies written in OWL in a more general sense can be mapped to Topic Map languages or models based on the TMDM is also very worthy of consideration, but unfortunately cannot yet be fully investigated. It would require a complete specification of the TMCL, as OWL clearly provides the means to express constraints that are beyond the scope of the TMDM’s expressivity, and fall squarely within the ken of the associated TMCL standard. The TMCL specification is taking shape rapidly, whilst the TMRM in development may also inform the issue, but such an investigation is still premature at the time of writing. § 5 TMDM to OWL Mapping Investigation

5.1 The TMDM side 5.1.1 What does the TMDM give us? The drafted TMDM offers some significant benefits over previous Topic Map specifications. The majority of it is represented as Entity-Relationship diagrams with explanatory text, and the clear delineation of Topic Map constructs into Entities, Relationships and Attributes is a significant and most welcome contribution. The directions and cardinalities of relationships between Topic Map entities are clearly

Extreme Markup Languages 2005® page 3 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

represented. Type and parent relationships are stated separately and explicitly. The different kinds of relationships Topics may have with Subjects are clearly represented. Data attributes are given for the logical entities, which include values, data types and locators such as URI references. The draft also introduces two additional constraints to ensure the integrity of Topic Maps based on the model. For ease of reference, all the E-R diagrams provided by the TMDM specification plus the two additional constraints have been collated into a single diagram shown at Figure 1. Note that "Topic Map Object" can be any of the other Topic Map Objects shown (Association, Occurrence, Topic, etc).

Figure 1

The TMDM Model, showing collated E-R diagram and two additional constraints

5.1.2 But wait, lurking in the back of the TMDM closet However, lurking in the last section of the TMDM specification (at least in the current draft) under the rather innocuous title of “Published Subjects”, are hidden two of the most fundamental Topic Map constructs : "Type-Instance" and "SuperType-SubType". In this section they are specified to be special instances of Associations with specific types between role-playing Topics which play special instances of Association Roles with specific types. The types are grounded by being set to specified Published Subject Indicators for "Type-Instance" and "SuperType-SubType". This is obviously much heavier and

page 4 Extreme Markup Languages 2005® 87 Building Topic Maps in OWL-DL

more complex than standard ontological Class-Instance and SuperClass-SubClass relations. In terms of the Data Model given up to this point, they are able to be interpreted simply as special instances of its constructs; rather unwieldy and convoluted to be sure, but not particularly problematic. However, unlike any other instances of Associations in the TMDM, instances of these special types of Associations are required to have the functionality of ontological Class-Instance and SuperClass-SubClass relations, in that instances of "SuperType-SubType" Associations must be transitive, and instances of "Type-Instance" Associations, although not transitive in themselves, must ensure that any Instances of a Type are also Instances of any SuperTypes of that Type. That is, if A is an Instance of Type B, and B is a SubType of SuperType C, then A must be an Instance of C also. On consideration, this seems to entail that Topic Map processing is required to create new instances of Associations and Association Roles on the fly in order to capture the flow-on effects implied here. Whilst this is rather inconvenient, it is not necessarily insurmountable for an OWL construction if a little extra Java code is attached. This issue is put aside for the moment and revisited in detail in Section 10 Outstanding Issues. 5.2 The OWL side OWL is the most recently developed, and most expressive layer of the W3C’s Semantic Web Recommendations to date. It is a Web Ontology Language, designed for the representation of classes containing individuals which have relations with each other and data attributes. It also supports the expression of constraints on these entities.

5.2.1 Flavours of OWL There are currently three “flavours” of OWL which vary in their level of expressivity. OWL-Lite has the least expressivity and is a subset of OWL-DL. Both OWL-Lite and OWL-DL have a Description Logic equivalent semantics. Description Logics are a decidable fragment of First Order Logic. Being decidable means that it is possible to ask whether certain things are the case and be guaranteed of getting an answer in a finite number of steps. For Description Logics, the kinds of things that can be requested are:

• computing the inferred superclasses of a class • determining whether or not a class is consistent (a class is inconsistent if it cannot possible have any instances) • deciding whether or not one class is a subclass of another class Automated Description Logic reasoners are available to perform these inferencing services and others for OWL-DL and OWL-Lite. The full OWL language, OWL-Full, supports additional expressivity beyond OWL-DL and OWL-Lite, such as allowing a thing to be both an individual and a class at the same time, but sacrifices the decidability of reasoning. This is largely due to the ability OWL-Full has to create self-referential loops: a class is allowed to contain itself as an instance. OWL-Full does not have a Description Logic semantics, and while it is certainly still possible theoretically to perform many inferencing tasks effectively on OWL-Full, currently the majority of OWL tools and reasoners do not support OWL-Full, whilst they provide good support for the OWL-DL and OWL-Lite sublanguages.

5.2.2 Which flavour of OWL tastes best to make Topic Maps? Whilst OWL-Full is a more natural medium for expressing the notion that Topics may be both Classes and Instances, if one's aim is to build a practical, functioning ontology with full support for querying and constraint checking, at this time OWL-DL is the preferred OWL sublanguage to use, as it gives the highest level of expressivity whilst still being decidable, having a well-defined semantics and being well-supported by OWL tools. However, OWL-DL may not always be the flavour of choice in future. Since the main purpose of Topic Maps is to enable humans to find things, rather than machines to reason about them, and OWL-Full's looser expressivity is a better fit for this, it may not ultimately be an issue that OWL-Full is undecidable, as long as there are functioning tools to provide the kind of consistency checks and querying needed by Topic Maps. For effective navigation, most of this needs only to be done at a local scale - that is, looking only at constructs within a local area of the TM, and how they are related to their neighbouring constructs. It should be possible to do this in most cases without coming up against the issue of undecidability. As an

Extreme Markup Languages 2005® page 5 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

analogy, consider one of M.C.Escher's "strange-loop" drawings like "Ascending and Descending" which may be found at http://www.mcescher.com/indexuk.html : If you look at only a localised part of the picture it looks perfectly sensible, even though when you look at the whole picture, you realise there's an impossible loop. Because of the way the human mind works, it may be argued that it is perfectly reasonable to implement such loops in Topic Maps, because they reflect the way humans think about information, and having loops in Topic Maps doesn't stop them from being useful maps showing where things are and how they relate to each other. In fact, the loop can be considered to be a valid part of the map!

5.2.3 What can be asserted in OWL-DL? In OWL-DL, Classes and Individuals need to be kept separate: no thing can be both a Class and an Instance of a Class. Classes may be related via a SuperClass-SubClass hierarchy, and may also be declared to be disjoint from each other, that is, to have no overlapping Instances. Individuals may be asserted to be Instances of Classes, and may be related to each other via OWL's Object Properties. They may also be given data attributes, using OWL's Datatype Properties. Both Object and Datatype Properties may be constrained on domain, range, and cardinality, and given various properties such as transitivity and symmetry. Properties may be constrained not only globally, but with respect to specific Classes. OWL Object Properties are one-directional, but may be linked to a corresponding property in the opposite direction using OWL's built-in "inverseOf" relation construct.

5.2.4 OWL-DL can represent E-R models! The commonality between OWL-DL's expressiveness and standard Entity-Relationship models renders OWL capable of representing a simple Entity-Relationship model by the following method:

• Entities are mapped to OWL Classes • Instances of Entities are mapped to OWL Individuals belonging to those OWL Classes • Relationships between Entities are mapped to OWL Object Properties linking OWL Classes The relationship’s cardinality may be captured within the Object Property, as may its directionality, through domain and range specification. To represent bi-directional relations, an additional Object Property in the opposite direction is created and declared to be an inverse property of the original. • Attributes of Entities are mapped to OWL Datatype Properties linking OWL Classes to Data Value Ranges. Attribute Values of the Instances of Entities correspond to specific values from the Data Value Ranges for individual within the OWL class that corresponds to the Entity. These Properties may also have explicit cardinality. It is not claimed that such a mapping is the only possible way one might map an E-R diagram to OWL- DL; merely that it is a reasonable, natural and straightforward approach. The question of mapping UML (the means to represent E-R diagrams) to OWL is a research issue in its own right. 5.3 Using OWL-DL to build the TMDM 5.3.1 The Approach In a Nutshell The method above gives a straightforward algorithm for mappingsimple E-R diagrams to OWL-DL. As the TMDM is largely specified as such a simple E-R model (refer to Figure 1), those parts of it can be directly deposited into OWL-DL using the method given above. The TMDM also has a couple of extra constraints to do with cardinality of relations, and whilst these take a little more thought, they may also be implemented in OWL-DL. With the TMDM constructs now constructed in OWL-DL, Topic Map authors may then build their own Topic Maps in OWL-DL simply by importing the OWL ontology that defines these constructs, and populating it with instances.

5.3.2 Why would Topic Map authors want to use OWL-DL to build TMs? The advantage of using OWL-DL for Topic Map authoring based on the TMDM ontology is that the resulting Topic Maps are guaranteed to conform with the TMDM, as all the conditions for comformance are implemented within the TMDM ontology and thus passed on to the Topic Map ontologies based on it. The author can also use OWL’s expressivity to express additional constraints on their Topic Maps in a way that closely corresponds with the desired functionality of a TMCL, and can use the existing, freely available OWL tools to query their Topic Maps in a way that closely corresponds with TMQL functionality. By virtue of using OWL-DL, the Topic Map author gains the benefits of:

page 6 Extreme Markup Languages 2005® 89 Building Topic Maps in OWL-DL

• Access to a well-developed suite of free, open-source APIs, visualisation, querying and reasoning tools : For instance, Protege combined with the RACER [RACER] reasoner provide the means for both querying Topic Maps and enforcing constraints, meeting many of the ISO's requirements for a Topic Map Query Language and a Topic Map Constraint Language • A clearly defined Description Logic formal semantics for the resulting Topic Maps. Semantic Web reasoning services and Intelligent Agent services are likely to become increasingly sophisticated and capable, and having a formal semantics helps open the door to these. Whilst OWL tools do not currently include an engine specifically for the navigation of Topic Maps, the resulting Topic Maps would be directly translatable to any other TMDM-based Topic Map languages which become available, as the representation is equivalent. If the method is widely accepted, it is likely that a means to import Topic Maps written in OWL to commercially available Topic Map engines would be made available, providing TM authors with all the advantages of both worlds. By the same token, it would also be possible to port Topic Maps written in other TMDM-based TM languages into OWL. § 6 Current Status

6.1 Work conducted to date Using the method described above, the TMDM E-R model shown at Figure 1 was translated into an OWL- DL ontology summarised at Figure 2, and implemented using Protege. Some sample Topic Maps were built as OWL-DL ontologies in Protege by importing the TMDM ontology to define the Topic Map constructs and populating it with instances. The construction in Protege is able to successfully utilise Protege's querying plugin and RACER reasoning plugin for constraint checking, as well as the other plugins that Protege provides, including those for Ontology visualisation. The OWL-DL construction for the TMDM was submitted to the ISO TMDM authors for review and comment, and discussed with the WG3 at meetings held in May 2005. They agreed it was indeed a faithful representation of the TMDM, implemented at an object level. 6.2 Functionality Implemented and Outstanding Note that the whilst the TMDM’s "Type-Instance" and "Supertype-Subtype" relationships as described in Section 7 of the TMDM Draft document may be fully asserted as instances of the ontology built here, the flow-on functionality is not currently implemented: that is, the transitivity of instances of Supertype- Subtype Associations, and the upward inheritance of Instances within Type-Instance associations to the SuperTypes of a Type expressed in Supertype-Subtype relations. Some approaches for implementing this functionality are discussed in Section 10 Outstanding Issues. With this exception, all other desired functionality required by the TMDM is believed to be captured in the ontology presented in the following section. 6.3 Intention If the outstanding issue can be resolved satisfactorily, and the resulting OWL syntax accepted by the relevant Standards Bodies as a valid syntax for the TMDM, it would then be made available online at an appropriate location, for Topic Map authors to import to OWL. As described, by providing the TMDM constructs for authoring Topic Maps in OWL-DL, users are enabled to author Topic Maps in OWL-DL which are TMDM compliant. This method for Topic Map construction has no reliance on XTM or any other TM languages: no XTM references are necessary, either in the TMDM ontology or in the user- defined Topic Map ontologies. It may be reasonably expected though, that if an OWL syntax were accepted as a standard for authoring Topic Maps, then Topic Map engines and tools would develop the ability to import Topic Maps written in this syntax, as they are doing with RDF. § 7 Designing and Building the TMDM Ontology The following sections describe the design and construction of the TMDM as an OWL-DL ontology in detail. 7.1 The Design Phase The method used corresponds to that described above for mapping Entity-Relationship Models into OWL- DL, but is here explained at a micro level with specific reference to the TMDM model shown at Figure 1. Visually comparing Figure 1 and Figure 2 should also aid the reader's understanding.

Extreme Markup Languages 2005® page 7 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

7.1.1 Design Choices As with the construction of any OWL ontology, some design choices must be made. For reasons of implementability discussed above, it is desirable, at least currently, to stay within the expressivity of OWL- DL rather than to use OWL-Full, and to enable this, the TMDM ontology needs to maintain a clear distinction between Classes and Individuals.

7.1.2 Identifying Individuals The first task is to identify the nature of the individuals in user Topic Maps, keeping in mind that whilst OWL's Object and Data Properties are defined using Classes as their domains and ranges, the properties ultimately link the Individuals within those Classes. That is, if we want to be able to assert that something has Object and Datatype Properties with particular values, then it needs to be an Individual, not a Class. Accordingly, the candidates for Individuals are taken to be the individual Topics, Occurrences, Associations and so on that Topic Map authors would ultimately want to define relationships between, and characteristics for.

7.1.3 Filling in the rest of the Model Having decided that these are to be the Individuals, the classes "Topic", "Occurrence" and "Association" and so on are created to contain these individuals. These are subsequently used as domains and ranges to define OWL Object Properties between individuals, providing the general framework defining which groups of Individuals may be related to which other groups of Individuals, and in what way (cardinality of relations, etc). The Topic Map author will ultimately import these these imported Classes and Properties and instantiate them with her own Topics, Associations, linkages between them and so on to build her own Topic Maps. The attributes of each entity in Figure 1 (source locator, etc) are translated to OWL Datatype Properties belonging to Individuals within the OWL Classes. Constraints may be implemented as restrictions on the OWL classes.

7.1.4 A note about the use of Topics for Typing other Topic Map Objects A distinguishing feature of Topic Maps is that Topics are used as Types for other Topic Map Objects, whilst also being used in all the ways that regular Topics are. Normally in an ontology, a type would be represented as an OWL Class, and Individuals would be typed by being asserted to belong to the Class. However, if this were to be done here, Topics would need to be represented as both OWL Classes and OWL Individuals, as they need to be able to be Types whilst still having Objects and Datatype Property values in the way individuals do. This would require the use of OWL-Full. However, as OWL-DL provides the expressivity to define complex relations between individuals using Object Properties, it is preferable to represent Types as individual Topics, and to represent the "type" of Topic Map Objects, not through class membership, but as an Object Property between the Topic Map Object and the Topic which types it. The issue of Typing for Topics themselves is a more complex issue addressed at Section 10 Outstanding Issues. 7.2 The Ontology Model The OWL-DL model for the TMDM used for construction is presented at Figure 2. As there is not yet a standard graphical notation for representing OWL ontologies, RDF graph notation has been extended: Classes are represented as ovals; the nesting of an oval within a larger oval indicates a subclass-superclass relation; arrows represent object properties, labelled with the Object Property's name and cardinality, and indicating direction by pointing from Domain to Range. To avoid cluttering the diagram, Datatype Properties and complex constraints are omitted, SuperProperty constructions are not shown and not all Inverse Properties are shown. Note that there are no Individuals in the TMDM ontology: these are defined in the process of Topic Map authoring. The TMDM ontology can be considered to be a template or mould of exactly the right dimensions for TM authors to fill up with instances, turning out perfectly formed Topic Maps.

page 8 Extreme Markup Languages 2005® 91 Building Topic Maps in OWL-DL

Figure 2

The TMDM Ontology Model, showing Classes and Object Properties

Note the correspondences between Figures 1 and 2: Each entity (box) in Figure 1 has been translated to a class (oval) within Figure 2. Relationships between the entities in Figure 1 are captured as Object Properties in Figure 2. Where there is a bi-directional relation in Figure 1, two relations have been given in Figure 2, related as inverses. For additional ease of use, many additional inverse relations have also been defined. Logically, these are already implied by the TMDM, and by making the inverse forms available, users are given more flexibility to write their own Topic Map ontologies in OWL in the most convenient way, as they can start from either end. 7.3 Additional Constraints Implemented The Ontology described above and shown in Figure 2 was built in OWL-DL using Stanford University's Protégé package for Ontology Engineering. In order to achieve all the required functionality specified by

Extreme Markup Languages 2005® page 9 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

the TMDM, some additional constraints were needed that are worthy of additional commentary. These are described below.

7.3.1 Internal and External Occurrences Although Topics may have any number of Occurrences, each Occurrence may have only one value, but it may be either a string (indicating an internal Occurrence) or a URI (indicating an external Occurrence). Correspondingly, two different Datatype Properties were attached to Occurrences:

• “occurrenceTextString” which maps to a string value, used for asserting internal occurrences and • “occurrenceTextLocatedAt” which maps to a URI value, used for asserting external occurrences. By using the appropriate Datatype Property, one may distinguish between internal and external Occurrences and map each to the appropriate Data Type. However, both are defined as subproperties of an “OccurrenceText” Datatype Property, which is defined as a functional property (that is, it may only map to one Data Value), ensuring that there is only ever one Data Value (which may be either a string or a URI) for each instance of the class Occurrence. A similar situation for Variants applies, dealt with using a "variantName" superproperty for "variantNameString" "VariantNameLocatedAt" Datatype Properties.

7.3.2 At least one Locator for each Topic According to the mapping described previously, the three locators relevant to Topics were mapped as Datatype Properties with domains as follows:

• SourceLocator: Topic Map Object • SubjectIdentifier: Topic • SubjectIndicator: Topic According to the constraint specified by the TMDM, it is necessary to ensure that for all Topics, at least one of these locators had a value. However, the non-empty locator can be of any of the three kinds available for Topics. To achieve this functionality, a superproperty Datatype Property called “Locator” containing these three Datatype Properties was created. Then a restriction was placed on the class of Topics with respect to the "Locator" Datatype Property, requiring it to have a minimum cardinality of 1. This ensures that for all Topics, at least one of the three locator sub-Properties MUST be non-empty. As the restriction only applies to the class "Topic", not to all classes in the domains of the Properties, it does not affect other TM objects using the "Source Locator" Datatype Property.

7.3.3 Variant Name Scope must be a true Superset of Topic Name Scope The TMDM places the following constraint on Variant Scopes: the set of scopes of each Variant of a Topic Name must be a true superset of the set of scopes of its parent Topic Name. This can be interpreted to mean that a Variant name must inherit all the scopes of its parent Topic Name plus have at least one of its own in addition to this. According to the methods previously described, an Object Property “isVariantOf” was defined with domain "Variant" and range "TopicName", and with a cardinality of 1, as a variant must belong to exactly one parent Topic Name. Additionally, the "hasScope" Object Property, of unlimited cardinality, relates both Variant and Topic Name to the class "Scope" : a subclass of "Topic" containing those Topics being used as Scopes. By making "isVariantOf" a subproperty of "hasScope", and including "TopicName" in the range of "hasScope", a variant may be considered to be scoped by its Parent Topic Name. By then making the "hasScope" property transitive, the Variant inherits the scopes of its parent Topic Name. To ensure that Variant will have a least one scope relating it directly to the class "Scope", another subproperty of "hasScope" is defined with has domain "Variant" and range "Scope", and a minimum cardinality of 1 is placed on this property. Voila! To tidy up, as only variants are allowed to be scoped by Topic Names, and other Topic Map Objects are not, we limit the hasScope Property on the other classes which use it to only allow values from the class "Scope", and not from "TopicName".

page 10 Extreme Markup Languages 2005® 93 Building Topic Maps in OWL-DL

7.4 Implementation The OWL-DL code generated by Protege for the TMDM ontology described above is included at the end of this paper and is also available at the author's website, along with the Protege ontology itself and examples. § 8 Examples In this section, the OWL TMDM constructs defined above are used to author an example Topic Map in OWL-DL which describes an Italian Opera domain. The example ontology's header (not shown) sets up namespaces, makes the standard XML, RDF and OWL inclusions, and imports the TMDM Ontology described above. Use of the prefix "tm" in the OWL code shown below indicates a construct is sourced from the imported TMDM ontology. The names used as identifiers for instances of Topic Map elements are completely arbitrary and up to the author to choose, the only requirement being that they must be unique within the authored Topic Map ontology. 8.1 Topics and Subjects Say that within the Italian Opera Topic Map we would like to create a topic called “Puccini” to represent the composer Giacomo Puccini. This is done simply by:

• creating an instance of the class “Topic” from the TMDM ontology, • calling it “Puccini”, • and giving its Subject Identifier Property an appropriate value to identify the subject that this Topic identifies: in this case the URL of an online wiki about the composer Giacomo Puccini. The corresponding OWL code is shown below:

http://en.wikipedia.org/wiki/Puccini If using a GUI OWL tool such as Protégé, the OWL code is generated automatically by undertaking the actions described above. Protege generates an input form based on the Properties defined for the Individual, and the user can simply assign values to the Properties by select from previously defined entities, or enter new data using this form. A Protege screenshot used to assign values to the "Puccini" Topic is shown below.

Extreme Markup Languages 2005® page 11 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

Figure 3

page 12 Extreme Markup Languages 2005® 95 Building Topic Maps in OWL-DL

Protege Screenshot : Entering Values for the Topic "Puccini" in the Italian Opera Topic Map

Processing Points to Note:

• The value given to the Topic's "Subject Identifier" property gives the Topic's "Locator" SuperProperty a value, thus satisfying the requirement that "Locator" has a minimum cardinality of 1. When running consistency checks, an error would be issued for every Topic does not have at least one value for "Locator". • As the Topic is now defined, it can be used later in the OWL ontology by using the ID "Puccini" to refer to it ( rdf:about = "Puccini" ) 8.2 Occurrences We would now like to assert two occurrences for the topic “Puccini” in our Italian Opera Topic Map:

• Puccini’s date of birth, which is a string, and • a website where Puccini’s operas can be found, which is a URL This is done by:

1. Firstly, asserting that the Topic “Puccini” has an Occurrence called “Puccini_Date_Of_Birth”, where:

• The "type" property of the Occurrence is an instance of the class “Occurrence Type”,called “Date_of_Birth” • The "occurrenceTextString" property of the Occurrence is the string “1858-12-22” (Puccini’s birthdate) 2. A second assertion that the Topic “Puccini” has another Occurrence called “Puccini_Web_Site”, where:

• The "type" property of the Occurrence is an instance of the class “Occurrence Type”,called “Web_Site” • The "occurrenceLocatedAt" property of the Occurrence is the URL “http://www.puccini.it/” The corresponding OWL code is:

1858-12-22 http://www.puccini.it/

Processing Points to Note:

• Consistency checks ensure that each occurrence has only one data value for "occurrenceText", although it can be either a string (when using "occurrenceTextString" subproperty) or a URI (when using "occurrenceTextLocatedAt" subproperty).

Extreme Markup Languages 2005® page 13 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

• Conversely, we may assert as many Occurrences as we like, or none at all for the Topic, since it has no cardinality restriction. Each such assertion links the instance of Topic to a separate, newly created instance of Occurrence. • Instances of "OccurrenceType" are being generated in the code above within the "type" declaration for each Occurrence. Since Occurrence is a subclass of Topic, these are automatically also Topics, and once created, may be referred to and used in all the ways that Topics are, and inherit all the constraints and properties that Topics have. • An Occurrence is only allowed to have at most one "type" property and it must be from the "OccurrenceType" subclass of Topic. If we try to give the Occurrence more than one value for its type or to link it to any other class, the consistency checks will fail. • Because of the way the inverse properties are set up, every “hasOccurrence” asserted between a Topic and an Occurrence automatically generates an “isOccurrenceOf” property between the Occurrence and Topic in the opposite direction. 8.3 Topic Names and Variant Names Say we would now like to give the “Puccini” Topic a Topic Name (also "Puccini") with scope of Normal_Name, and an associated Variant Name which has the full name “Giacomo Puccini” with scope of Long_Name. To do this, we would simply:

• create an instance of TopicName called “Puccini_Topic_Name” with topicNameString set to “Puccini” and Scope of “Normal_name”. • assert that this Topic Name has a Variant using the “hasVariant” Object Property, and create an instance of Variant which has "variantNameString" of “Giacomo Puccini”, and Scope of “long_name”. The corresponding OWL code is:

Puccini Giacomo Puccini

Processing Points to Note:

• When “belongsToTopic” is declared between an instance of Topic Name and an instance of Topic, the inverse property “hasTopicName” automatically links the Topic and the Topic Name in the opposite direction. • A Topic Name is only allowed to belong to exactly one parent Topic, so if we tried to assert any more “belongsToTopic” properties linking it to other Topics, or if we did not assert any links to a parent Topic, the consistency checks would produce an error. • Conversely, as there is no limit to the number of Topic Names a Topic may have, we are free to assert as many values of “hasTopicName” for each Topic as we like, or none at all. • When “hasVariant” is declared between an instance of Topic Name and an instance of Variant, the inverse property “isVariantOf” automatically links the Topic and the Topic Name in the opposite direction

page 14 Extreme Markup Languages 2005® 97 Building Topic Maps in OWL-DL

• A Variant is only allowed to belong to exactly one parent TopicName, so if we tried to assert any more “isVariantOf” properties linking it to other TopicNames, or if we did not assert any links to a parent TopicName, the consistency checks would produce an error. • Conversely, as there is no limit to the number of Variants a Topic Name may have, we are free to assert as many values of “hasVariant” for each Topic Name as we like, or none at all. • Instances of the class “Scope” are being generated within the code above within the "hasScope" declarations. Since Scope is a subclass of Topic, each instance of Scope is automatically a Topic, and once created, may be referred to and used in all the ways that Topics are and inherit all the constraints and properties that Topics. • Through the transitivity of the "hasScope" property, which is a superproperty of “isVariantOf”, the Variant inherits all the scope elements of its parent Topic Name. After inferencing processes are completed, the Variant will be linked to Scopes which include both “Long_name” and any scopes belonging to “Puccini_Topic_Name”; in this case “Normal_name”. Thus the Variant will fulfil the constraint of having a true superset of the scope of its parent Topic Name. If we did not use the VariantScope property to link the Variant to least one instance of Scope, this consistency check would fail. • Although not shown in this example, TopicNames may linked via the "type" property to at most one instance of TopicNameType. If we attempt to link via the type property more than once or to anything other than a TopicNameType, the consistency checks will fail. Variants are not allowed to have any type at all, and thus the "type" property does not include the class "Variant" as part of its domain. If we were to try to declare a "type" for an instance of "Variant", consistency checks would indicate an error. When using a GUI such as Protégé, the input forms do not provide an option to select a "type" for an instance of Variant, as this class is not within the domain of the property. 8.4 Associations Say we would now like to assert an association between topics Puccini, Giacosa and the opera Tosca, to indicate that Puccini and Giacosa wrote Tosca, with Puccini being the composer, Giacosa the librettist and Tosca the Opera. This is done in the following way:

• Create an instance of the class “Association” and call it “Puccini_and_Giacosa_Wrote_Tosca" • Using the “type” Object Property, link this instance of "Association" to an instance of “AssociationType” called “Written_By” • Using the “hasRole” Object Property”, link the instance of "Association" to three newly created instances of "AssociationRole":

1. Firstly to an AssociationRole called “Librettist_is_Giacosa” which indicates that the Topic “Giacosa” plays the Role of “Librettist” in this association. This is achieved by linking the "type" property of the AssociationRole instance to an instance of AssociationRoleType called “Librettist”, and the "playedby" Property to a newly created instance of Topic called “Giacosa” 2. Secondly to an AssociationRole called “Composer_is_Puccini” which indicates that the Topic “Puccini” plays the Role of “Composer” in this association. This is achieved by linking the "type" property of the AssociationRole instance to an instance of AssociationRoleType called “Composer” and the "playedby" Property to the previously created Topic instance called “Puccini” 3. Thirdly to an AssociationRole called “Opera_is_Tosca” which indicates that the Topic “Tosca” plays the Role of “Opera” in this association. This is achieved by linking the "type" property of the AssociationRole instance to an instance of AssociationRoleType called “Opera”, and the "playedby" Property to a newly created instance of Topic called “Tosca”. The corresponding OWL code is shown below:

Extreme Markup Languages 2005® page 15 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

Processing Points to Note:

• When “hasRole” is declared between an instance of Association and an instance of AssociationRole, the inverse property “roleInAssociation” automatically links the instances of AssociationRole and Association in the opposite direction. • An AssociationRole is only allowed to belong to exactly one parent Topic, so if we tried to assert any more “roleInAssociation” properties linking it to other Associations, or if we did not assert any links to a parent Association, the consistency checks would produce an error. • Conversely, Associations must have at least one AssociationRole, so if we did not declare at least one link between this instance of Association and an instance of Association Role using either “hasRole” or its inverse somewhere within the ontology, consistency checks would fail. • When “playedBy” is declared between an instance of AssociationRole and an instance of Topic, the inverse property “playsRole” automatically links the Topic and the AssociationRole in the opposite direction. • “playedBy” must be linked to exactly one Topic and consistency checks will fail if this is not the case. • Conversely, Topics may be linked to any number of AssociationRoles via the "playsRole" property. • Each Association and AssociationRole may be linked by the "type" property to at most one instance of AssociationType and AssociationRoleType respectively. If we attempt to link via the "type" property more than once or to anything other than an instance of the appropriate class, the consistency checks will fail. When using a GUI such as Protégé, the input forms do not provide an option to select a "type" value for an instance of Variant other than from the appropriate class. • Guess what? We forgot to give at least one value of Locator for our new Topics "Giacosa" and "Tosca". We will generate an error indicating this when we run the consistency checking.

page 16 Extreme Markup Languages 2005® 99 Building Topic Maps in OWL-DL

§ 9 OWL capabilities for TM authoring

9.1 Consistency Checks and Automatic Inferencing As described throughout the examples above, by using a reasoner such as RACER with the Protege application, consistency checks and automatic inferencing to ensure TMDM compliance are available. These are activated simply by selecting the appropriate services from the Protege menu. 9.2 Additional Constraints The TMDM ontology is set up in such a way to enable the user to easily add the kind of constraints described in the use cases of the TMCL requirements. For instance: "Topic T can be used for Typing Occurrences, not for typing anything else". Topic T should be declared as an instance of the "OccurrenceType" subclass. Occurrences are only allowed to get values for the "type" property from this Subclass, and no other Topic Map construct is allowed to get its "type" from this Subclass. To rule out the possibility that any Topic might be an instance of both SubClasses at the same time, the user can add an axiom to force the SubClasses within the "Type" subclass of "Topic" to be disjoint from each other. "Occurrence of type O can only be used within scope S": The user should create a subclass of Occurrence, which has "type" property set to an instance of OccurrenceType representing "O", and "hasScope" property mapped to an instance of Scope representing "S". Instances added to this subclass of Occurrence will automatically be of type "O" and have Scope "S". If the intention is that this is the only value of Scope these Occurrences may have, then the "hasScope" property for the created subclass should have its cardinality set to 1, to prevent it being linked to other Scopes via the "hasScope" property. It should be mentioned that many of the TMCL Use Cases are related to "Topics of Type T...... ": these are not discussed here due to the issue discussed in the Section 10 Outstanding Issues. It is worth noting though, that if the TMDM model were to be adapted to allow Topics to have types in the same way that Occurrences, Associations and so on have types, these cases would be able to be treated in a similar manner to the above examples. Putting aside the above issue, the kinds of constraints required by the TMCL would not cause the authored Topic Map to go from OWL-DL to OWL-Full, although the user is not prevented from using the full available OWL buffet of constraint expressivity to add their own more creative constraints, which might cause their ontology to move into OWL-Full territory. To assist, Protege has a function which will tell the user which OWL-sublanguage the ontology is currently using, so that if one desires to stay within OWL- DL, one may check that this is the case when adding constraints. It is also likely that creative constraints would be at least partially lost when porting the Topic Map ontology to a Topic Map application based on a TMDM/TMCL platform, as there would be no way to translate these into TMCL. 9.3 Querying Topic Maps authored using the TMDM ontology can be queried in Protege using a special Querying Plugin Tool which has a GUI frontend. Complex queries can be constructed by nesting simpler ones, and may make use of OR and AND logic through "Match Any" and "Match All" buttons provided. The format simple queries in Protege is: [Class] [Slot] contains [Individual/Data value] The choices are presented from those that are validly available - that is if one chooses the [Class] to be "Topic", then the choices for [Slot] are only those properties where "Topic" is within the domain of the Property. For instance for the Italian Opera Topic Map above, one can construct the following three simple queries:

1. Retrieve all instances of class [Topic] where slot [playRole] contains [Association Role X] 2. Retrieve all instances of class [AssociationRole] where slot [roleInAssociation] contains [Association Y] 3. Retrieve all instances of class [Association] where slot [type] contains [Association Type Y] and then nest them, the third one being innermost. By setting the value of AssociationType to "written_by" in the innermost query, the query finds all Topics which play Roles in Associations of Type "written_by". Based on the Italian Opera Topic Map ontology

Extreme Markup Languages 2005® page 17 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

described in the previous section, the result would be the Topics "Puccini, "Giacosa" and "Tosca" because they all play roles in the only Association of type "written by" (namely the Association "Puccini_and_Giacosa_Wrote_Tosca"). A query screenshot from this query being run in Protege is shown below:

page 18 Extreme Markup Languages 2005® 101 Building Topic Maps in OWL-DL

Figure 4

Extreme Markup Languages 2005® page 19 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

Querying in Protege Using Querying Plugin

9.4 Visualisation Protege provides several plugins for visualisation of ontologies, for instance Jambalaya. It provides various views of the ontology, and one may navigate through the Ontology by clicking on the Classes and Instances one wishes to view. The user may choose for views to be filtered to only contain certain constructs, and to be presented in different ways eg radially, as a tree structure, etc. A sample screenshot produced by Protege's Jambalaya plugin is shown below, showing only one of the smorgasbord of possible views:

page 20 Extreme Markup Languages 2005® 103 Building Topic Maps in OWL-DL

Figure 5

Extreme Markup Languages 2005® page 21 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

Visualising the TMDM Ontology Using Protege's Jambalaya Plugin

§ 10 Outstanding Issues Up to this point we have implemented a solution which appears to successfully support Topic Map authors to build TMDM compliant Topic Maps as OWL ontologies. However, having put aside the issue of the functionality of "Type-Instance" and "SuperType-SubType" Associations as specified in the TMDM earlier, we must now return to see what our options are to get over the line in implementing ALL the required TMDM functionality. Within the context of the approach presented in this paper, the choices to complete the functionality seem to break into three camps, which are described below. 10.1 The OWL-Full approach In this approach, we do not implement Type-Instance and SuperType-SubType as instances of Associations, but instead use OWL’s builtin instanceOf and subclassOf relationships to express these relations in a normal ontological manner. This requires the use of OWL-Full, since any topic may be treated as either a class or an instance in these relationships. (Even if we could separate into classes and instances with no overlap, there is another reason for going beyond OWL-DL: we need to be able to assert characteristics of topics which are acting as classes without those characteristics being inherited by the topics which are their instances — for example, we want a topic’s subject identifier to belong to it only and NOT to all the instances it is being used to type. However, there is no means to assert Data or Object Properties on classes, but only on the individuals contained within them. Thus any property asserted to on a class must be inherited by every individual in the class). We dispense with the TMDM notion of building associations and roles with specific types to represent these relationships, and consider the “Type-Instance” and “SuperType-SubType” sufficiently grounded through being builtin OWL constructs. In theory this produces all the required functionality, however, in practice we now lack tools which support our ontology well enough to do all the consistency checking and querying we would like. However, better support for OWL-Full is on the W3C’s agenda and if we revisit the issue in future we may find that the tools we need have become available. We live with the fact that our Topic maps may be undecidable ie that sometimes our queries don’t terminate or that not all consistency checking can give conclusive results. Since we are creating self-referential loops in this solution, this seems only fair. 10.2 The "Quick Fix" Workaround As with the previous approach, we drop the notion of these relations being special Associations, but here Topics remain individuals at all times so that they can have all the Object and Data Properties they need. To support the Type behaviour, the user may create Class hierarchies within the Class "Topic", in the normal OWL manner, using OWL's InstanceOf and SubClassOf builtin properties. These classes may then simply be connected to the corresponding Topics which we want to have this Typing behaviour via a specially created ObjectProperty for this purpose. It is a little ugly and creates an extra step for the user in querying and constraint processing, but should mimic the desired functionality, and is within OWL- DL. 10.3 The Extra Java code approach In this approach, we use a Protege plugin such as JESS to write extra Java code to produce the required functionality. For authoring this would then have to be downloaded by the TM author along with the TMDM ontology, and run in a similar way to the running of consistency checking and automatic inferencing (effectively it defines additional automatic inferencing). The functionality is produced by creating the flow-on relationships required, which are then available for subseqent functions such as querying and visualisation. Within this approach there are two paths, depending on how close we want to stay to the TMDM.

10.3.1 Using Object Properties For this path, as with the approaches above, we dispense with the TMDM notion of building associations and roles with specific types to represent these relationships. We define the “Type-Instance” and “SuperType-SubType” relations not by using OWL built-in constructs, but as special Object Properties

page 22 Extreme Markup Languages 2005® 105 Building Topic Maps in OWL-DL

between Topics, in a similar way to the way Types are defined for other Topic Map objects in the previous sections. We remain in OWL-DL, as all Topics are always Individuals. Taking this approach, we are able to use the built-in transitivity property within OWL to get the required behaviour for the "SuperType-SubType" Object Property, and just need a little extra code to ensure that Instances of Types get passed upwards to their SuperTypes.

10.3.2 Generating Instances of Association and Association Roles On this path, we stay with the letter of the TMDM specification and express “Type-Instance” and “SuperType-SubType” relations as special instances of associations as the TMDM states, then use extra Java code as above to exhibit the required behaviour. This means creating extra instances of Associations and Association Roles on the fly to accommodate the flow-on relations. The TMDM ontology as it has been described in previous sections can easily accommodate the representation through the use of subClasses in combination with constraints, and the definition of special instances of Topics for the types. Constraints are used to ensure there are exactly two Roles of two particular Types in each instance of an Association which has its type set to one of these special Association Types. Special subclasses are created to ensure the Subject Identifiers for the Types of the Roles and Associations are set appropriately. The full construction in Protege is available at the authors website. The additional Java Code is then required to produce the extra instances of Associations and Association Roles which make the SuperClass-Subclass Association transitive, and to ensure that Instances of Types in "Type-Instance" Associations get passed upwards to any Topics playing the roles of "SuperType" in instances of "SuperType-Subtype" Associations where where the original Topic is playing the role of "SubType". § 11 Conclusions This paper presents a proposal for an OWL-DL formalisation of the ISO drafted standard Topic Map Data Model. It illustrates the fit between OWL and the TMDM, and the potential for the use of OWL-DL as a medium for representing the TMDM constructs, providing the basis for end-users to author Topic Maps in OWL-Dl. The use of OWL-DL and its associated tools for constructing Topic Maps provides significant advantages over previous Topic Map representations in terms of explicit specification, formal semantics, constraint checking and querying capabilities. An outstanding issue relating to the implementation of Type-Instance and SuperType-SubType relations as specified in the TMDM was identified, and some approaches for implementing the required functionality were outlined. § 12 Appendix A TMDM Ontology: OWL-DL code generated by Protege:

Extreme Markup Languages 2005® page 23 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

1 1 1 1

page 24 Extreme Markup Languages 2005® 107 Building Topic Maps in OWL-DL

1 1 1 1 1

Extreme Markup Languages 2005® page 25 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

page 26 Extreme Markup Languages 2005® 109 Building Topic Maps in OWL-DL

Extreme Markup Languages 2005® page 27 Chapter 3 Integrating Topic Maps into the Semantic Web Cregan

Bibliography [OWL] Mike Dean, Guus Schreiber, editors , OWL Web Ontology Language Reference, W3C Recommendation, 10 February 2004 http://www.w3.org/TR/2004/REC-owl-ref-20040210/ [Protege] Stanford University The Protege Ontology Editor and Knowledge Acquisition System, Open Source Ontology Editor protege.stanford.edu/ [RACER] Volker Haarslev, Ralf Moller, Michael Wessel, RACER, Semantic Middleware for Industrial Projects based on RDF/OWL http://www.sts.tu-harburg.de/~r.f.moeller/racer/ [RDF] Dave Beckett, editor, RDF/XML Syntax Specification (Revised), W3C Recommendation, 10 February 2004 http://www.w3.org/TR/rdf-syntax-grammar/ [RDFTM] Steve Pepper, Fabio Vitali, editors, RDFTM: Survey of Interoperability Proposals, W3C RDF/ Topic Maps Inter-operability Task Force publication, Editors Draft, 24 February, 2005 http:// www.w3.org/2001/sw/BestPractices/RDFTM/survey-2005-02-24 [TM] M. Biezunski, M. Bryan, S. Newcomb, editors, ISO/IEC 13250, Topic Maps (Second Edition)ISO 13250 International Standard, 22 May 2002http://www.y12.doe.gov/sgml/sc34/document/0322.htm/ [TMCL] Graham Moore, Mary Nishikawa, Dmitry Bogachev, Topic Map Constraint Language (TMCL) Requirements and Use Cases ISO 13250 Editors Draft, 16 October, 2004 http://www.jtc1sc34.org/ repository/0548.htm [TMDM] Lars Marius Garshol, Graham Moore, Topic Maps - Data Model ISO 13250 Final Committee Draft, 10 January, 2005 http://www.isotopicmaps.org/sam/sam-model/ [TMQL] Lars Marius Garshol, Robert Barta, TMQL RequirementsISO 13250 Draft, 7 November, 2003 http://www.y12.doe.gov/sgml/sc34/document/0448.htm

page 28 Extreme Markup Languages 2005® 111 Building Topic Maps in OWL-DL

[TMRM] Patrick Durusau, Steve Newcomb, editors Topic Maps - Reference ModelISO 13250 Committee Draft, 12 February, 2005 http://www.isotopicmaps.org/TMRM/ [XTM1.0] Steve Pepper, Graham Moore, editorsXML Topic Maps (XTM) 1.0 TopicMaps.Org Specification, 6 August, 2001http://www.topicmaps.org/xtm/1.0/

The Author Anne Cregan University of New South Wales, Knowledge Representation and Reasoning Program, National Information and Communications Technology Australia (NICTA) Centre of Excellence and Artificial Intelligence Laboratory, School of Computer Science and Engineering Sydney NSW Australia 2052 [email protected] http://www.cse.unsw.edu.au/~annec/ Anne Cregan is a PhD student in NICTA's Knowledge Representation and Reasoning Program. Prior to PhD enrolment she worked as a Technology and Business Consultant, Data Miner and Marketing Modeller. She was General Manager of two Internet-related companies in 1998-2000. She has a first class honours degree from the University of Sydney in Cognitive Psychology and Human Intelligence.

Extreme Markup Languages 2005® Montréal, Québec, August 1-5, 2005 This paper was formatted from XML source via XSL by Mulberry Technologies, Inc.

Extreme Markup Languages 2005® page 29 Chapter 3 Integrating Topic Maps into the Semantic Web 113

3.2 An OWL DL Construction for the ISO Topic Map Data Model

Title of Publication: An OWL DL Construction for the ISO Topic Map Data Model Presented to: ISO/OEC JTC 1/SC 34, Working Group 3 Amsterdam, May 2005 Published: Online at http://xml.coverpages.org/CreganTMs-OWL200505.pdf Peer Reviewed: By the ISO working group - see Section 1.10.2 Contributing Author(s): Anne Cregan Personal Contribution: 100% Chapter 3 Integrating Topic Maps into the Semantic Web 115

DRAFT ONLY 16/05/05 An OWL DL construction for the ISO Topic Map Data Model

Anne Cregan [email protected]

Knowledge Representation and Reasoning Program, Artificial Intelligence Group National Information and Communication Technology School of Computer Science and Engineering, Australia (NICTA) Centre of Excellence University of New South Wales AUSTRALIA

Abstract tionships with other topics and subjects. Dubbed the “GPS of the information universe”, they are akin to an Both Topic Maps and the W3C Semantic Web electronic “back-of-book” index, supporting information technologies are meta-level semantic maps de- navigation and retrieval in on-line environments. XML scribing relationships between information re- Topic Maps, or XTM [XTM 01] is the foremost accepted sources. Previous attempts at interoperability be- representation of the ISO standard, developed as XML tween XTM Topic Maps and RDF have proved DTDs by the TopicMaps.Org consortium. Recent problematic. The ISO’s drafting of an explicit ISO/IEC work has concentrated on defining Topic Maps' [TMDM 05] Topic Map Data Model combined intended semantics, resulting in the drafting of a Topic with the advent of the W3C’s XML and RDF- Map Data Model has been drafted. [TMDM 05]. based Description Logic-equivalent Web Ontol- ogy Language [OWLDL 04] now provides the 1.2. The Semantic Web means for the construction of an unambiguous semantic model to represent Topic Maps, in a The Semantic Web is a W3C-led initiative with the goal form that is equivalent to a Description Logic of providing technologies and standards for the semantic representation. markup of information resources, thus enabling improved web navigation and supporting intelligent web services. This paper describes the construction of the pro- Like XTM, it also is an XML-based technology, with posed TMDM ISO Topic Map Standard in OWL Resource Description Framework (RDF) and Web Ontol- DL (Description Logic equivalent) form. The ogy Language (OWL) layers superimposed on XML to construction is claimed to exactly match the fea- provide more expressive representations of the character- tures of the proposed TMDM. The intention is istics of, and relationships between, logical entities. OWL that the topic map constructs described herein, DL, a formal W3C recommendation finalised in 2004, is a once officially published on the world-wide web, subset of OWL Full which provides a description logic may be used by Topic Map authors to construct equivalent semantics for OWL ontologies. their Topic Maps in OWL DL. 1.3. The Interoperability Goal

The advantage of OWL DL Topic Map construc- Although the Topic Map standards have been developed tion over XTM, the existing XML-based DTD independently of the W3C's Semantic Web initiatives, it standard, is that OWL DL allows many con- has long been felt that as both Topic Maps and the Se- straints to be explicitly stated. OWL DL’s suite mantic Web have the same goal to be meta-level maps of of tools, although currently still somewhat imma- information entities, there must be synergies to be ex- ture, will provide the means for both querying ploited. This paper describes how OWL DL may be used and enforcing constraints. This goes a long to create a Topic Map ontology whose constructs match way towards fulfilling the requirements for a those in the draft Topic Map Data Model. The OWL DL Topic Map Query Language (TMQL) and Con- Topic Map ontology described herein: straint Language (TMCL), which the Topic Map • Provides the Topic Map Constructs to allow the user Community may choose to expend effort on ex- to create their own OWL DL Topic Maps, using the tending. Additionally, OWL DL has a clearly constructs in the proposed TMDM. defined formal semantics (Description Logic ref) • Enables the user to draw on OWL’s querying and constraint tools & capability to support validity and con- 1. Introduction sistency checking which both enforces the intended semantics of the TMDM, and allows the user to implement additional constraints for their own user- 1.1. The Topic Map Tradition defined ontologies, by using the constructs provided in Topic Maps are an ISO/IEC standard [TM 02] for map- OWL DL. ping web and “real-world” information resources, by rei- • Is Description Logic equivalent, and enables the con- fying real-world resources as “subjects”, and creating struction of user-defined Topic Maps which have a “topic” constructs to capture their characteristics and rela- formal Description Logic semantics.

Page 1 of 11 Chapter 3 Integrating Topic Maps into the Semantic Web

DRAFT ONLY 16/05/05 The conversion of existing XTM topic maps to OWL DL one direction only, they may be linked to the correspond- is a separate issue, addressed in the author’s previous ding property in the opposite direction by the use of the work but not covered herein. “inverse” relation construct.

2. Topic Map Data Model 3. Construction of OWL DL model 2.1. Motivation In 2001, two ISO 13250 standard syntaxes for Topic 3.1. Design Choices Maps were established, one in HyTM and another in In order to construct the TMDM in OWL DL, some de- XML [TM 02]. However, the standard did not explain sign choices need to be made. how the two syntaxes related to each other, and did not make the common underlying data model explicit. There Firstly, we need to decide what the final individuals in are non-trivial differences between the syntaxes, and both Topic maps will be, recalling that whilst Object and Data fail to specify what implementations are to do in a number Properties are defined with classes as their domains and of situations. Additionally, neither standard supports con- ranges, they actually exist between the individuals within straints of the form “A person must be born in a place”, those classes. Accordingly, the candidates for individuals “A person must have a least one name” etc. are taken to be those individual Topics, Occurrences, As- sociations and so forth that users would ultimately want to A draft of requirements for both a Topic Map Constraint define within their Topic Maps. Language and Query Language have since been created, and both require a clear description of how Topic Map The constructs Topic, Occurrence and Association and so constructs are to be evaluated. Accordingly, the ISO forth should therefore be defined as classes which will Topic map team have addressed this with the Topic Map contain those individuals. We may then define the rela- Data Model work that commenced in May 2001, and so tions between these and other Topic map constructs as far has culminated in the current official draft ISO 13250- Object Relations, with domains and ranges being appro- 2 Topic Maps Data Model [TMDM 05] published in priately defined classes. The user will ultimately instanti- January 2005. ate these relations in their own ontologies, but our OWL DL ontology will provide the general framework which 2.2. The Topic Map Data Model (TMDM) defines which entities may be related to which other enti- For ease of reference the TMDM specification has been ties, in what ways they relate (cardinality etc), and the compiled by the author and all aspects of the model are attributes they may have. shown graphically at Figure 1. (Note that “Topic Map Object” can be any of the other Topic Map Objects shown We also explain here why we have opted to define Topics (Association, Occurrence, TopicName, etc)1). as individuals, even though they are used as types. Nor- mally in an OWL ontology, a type would be represented In the author’s opinion, the proposed TMDM makes sig- as a class of individuals. However, as the TMDM re- nificant improvements to the previous Topic Map Data quires that types are also themselves Topics which may be Model implied, but never explicitly stated, by the XTM used in all the ways that regular Topics are, we must opt DTD. It clearly states the directions and cardinalities of to reflect type relations as Object Relations between indi- relationships, and introduces two constraints. It gives vidual Topic Map Objects and the individual Topics data attributes for its logical entities, which include loca- which are their types. tors such as URI references. It also states type and parent relationships separately and explicitly. As the “Type” and “Scope” of a Topic Map Object are filled by Topics, we have defined these as specific sub- 2.3. Fit with OWL DL classes of “Topic”. We have preserved the ability to sepa- The fit between the TMDM and OWL DL is a good one. rate Association Types, Occurrence Types, Topic Name OWL DL has been designed for the representation of Types and Association Role Types, as these groups are classes containing individuals, which may be related via likely to have large non-overlapping components. The Object Properties, and may also have attributes repre- ability to separate these constructs will prove useful for sented by Data Properties. OWL enables the building on implementing some of the desired TMCL requirements, ontologies with class-subclass relations, additionally al- as well as for supporting the user in managing user- lowing both Object and Data Properties to be constrained defined types within their Topic Maps. We have not, on Domain, Range, Cardinality and other properties such however, opted to preserve any separation of Scope sub- as transitivity. Although OWL properties are defined in classes according to the kind of TM Object relating to it, as these as expected to differ little between TM Objects, 1 Although technically, if a Topic reifies another Topic, they as typically Scope relates to the division of subject do- will be forced to merge into one Topic in the merge processing mains more generally.

Page 2 of 11 117

DRAFT ONLY 16/05/05 3.2. The OWL DL Model references are necessary, either in the TMDM ontology or The proposed OWL DL model is presented pictorially at in user-defined Topic Map ontologies.

Figure 2. As the author is not currently aware of any Section 5 works through some examples showing how a standard graphical notation for representing OWL or other user-defined Topic Map would reference and use the ontologies beyond RDF graphs, she has devised her own OWL DL constructs defined in Section 4. notation as a natural extension of RDF notation. 4.1. OWL document Header Classes are represented as ovals, and the nesting of an oval within a larger oval indicates a subclass-superclass After the standard XML, RDF and OWL inclusions are relation. Arrows represent object properties, labelled with made, this ontology is named “Topic Map Data Model the Object Property’s name, and cardinalities, and indicat- Ontology”: ing direction by pointing from the Domain to the Range. To avoid cluttering the diagram, data properties are omitted in Figure 2, as are constraints. Topic Map Data Model Ontology Each entity (box) in Figure 1, has been translated to an (Note: indicative name only, subject to ISO approval). oval (class) within Figure 2. For instance, the oval “Topic” denotes an OWL class which will contain indi- 4.2. Classes viduals Topics as instances. As explained previously, Firstly, OWL classes are created for each TMDM Entity: subclasses are created for the various Types and for Scope. Although shown not overlapping, we have opted not to define them as disjoint, so the individuals therein may in fact be common to multiple subclasses if the user wishes ie the same Topic could be used for both an Asso- ciation Type and an Occurrence Type, as well as a Scope. This could also be prevented by additional OWL code to define these subclasses as disjoint. Scope and types are defined as subclasses of Topic. Type Relationships between the entities in Figure 1 are cap- also has further subclasses within it: tured as Object Properties in Figure 2. Where there is a bi-directional relation in Figure 1, two relations have been given in Figure 2, and related as inverses. For additional ease of use, many additional inverse relations have also been defined. Logically, these are already implied by the TMDM, and will ultimately allow the user more convenience in writing code for their own ontologies.

The attributes of each entity in Figure 2 (source locator, etc) will translate to OWL Data Properties of the individuals within the OWL classes (not shown at Figure 2). Constraints will be implemented as restrictions on OWL classes. Full details are given in the following section.

4. Construction of OWL DL ontology

This section gives the full OWL construction corresponding to Figure 2, plus Data Properties and Constraints, thus fully capturing every aspect of the proposed TMDM as detailed in Figure 1. Individuals to be declared in user-defined OWL TM The intention is that the code here defined, once finalized documents will belong to at least one of the classes de- and approved by the relevant Standards Bodies, will be fined. For instance a user-defined topic called “mytopic” the definitive representation of the Topic Map Data would be declared thus: Model in OWL DL syntax. It will be published as an online document, and included by import in all User- Within this section, further restrictions are added onto Defined Topic Maps written in OWL which wish to use these classes via constraints on the OWL properties re- the Topic Map Data Model constructs. Note that no XTM lated to them.

Page 3 of 11 Chapter 3 Integrating Topic Maps into the Semantic Web

DRAFT ONLY 16/05/05

M !R##$ $

4.3. Object Properties 4.3.1 scope This section defines the relationships between individuals Occurrence, Topic Name, Variant and Association may within the Topic Map classes. As OWL properties are all be related to any number of Scopes.2 Thus we define a directional, inverses are also defined to give the complete Property “scope” which has all these classes as its domain construction. By making these logically equivalent forms and Scope as its range. There is no need to add any car- available, users are given more flexibility to write their dinality restriction. own Topic Map ontologies in the most convenient way. 2 Note the OWL convention that Classes commence with a capital letter, whereas properties do not, so “Scope” is a class, whereas “scope” is a property.

Page 4 of 11 119

DRAFT ONLY 16/05/05

%& & ' ! (#$ $ 4.3.2 type Individuals within the classes TopicName, Occurrence, Association and AssociationRole may each have 0 or 1 types. Accordingly, “type” is defined as a functional property from individuals within the union of these classes to the “Type” class: Variant Scope Constraint: NOTE: There is a constraint on Variant Scope as follows: The value of the scope of each individual variant item must be a true superset of the value of the scope of its parent topic name. //** TO BE ADVISED: the best way to implement this is still under consideration by the author **//

Page 5 of 11 Chapter 3 Integrating Topic Maps into the Semantic Web

DRAFT ONLY 16/05/05

A functional property relates each item in its domain to Conversely, an individual Topic may have any number of one and only one item in its range. Thus if a Topic Map TopicNames. This is defined as an inverse property of individual has a user-defined type property, it is able to the “belongsToTopic” property defined above, giving be related to only one individual from the Type class. It users the option to define Topic/TopicName relations may also have zero types in a user-defined Topic Map if using either Topic or TopicName as the starting point, as no type property is declared for it. is most convenient.

Additional restrictions are set to force each relevant Topic appropriate kind within the “Type” class:

Each Occurrence has exactly one parent Topic, enforced via a Cardinality constraint on the “occurrenceOfTopic” Object Property:

”&xmls;nonNegativeInteger”>1 Conversely, a Topic may have any number of Occur- Each Variant has exactly one parent TopicName, en- 4.3.3 parent

Each TopicName must have exactly one parent Topic. This is enforced via a cardinality restriction on the property relating TopicName to Topic: ”&xmls;nonNegativeInteger”>1

Conversely, a TopicName may have any number of Vari- ants, defined as an inverse property of hasVariant: 1

Page 6 of 11 121

DRAFT ONLY 16/05/05

tion, enforced via a Cardinality constraint on the “roleInAssociation” Object Property:

4.3.5 Reification and any Topic may be used to reify one only Topic Map Object:

”&xmls;nonNegativeInteger”>1

Conversely, each Association must have at least one As- 4.3.6 Parent Topic Map sociationRole, set up as a minimum cardinality restriction on the inverse property “hasRole”: Lastly, object properties are created to capture the notion that each Topic and Association belongs to exactly one parent TopicMap : ”&xmls;nonNegativeInteger”>1

4.3.4 Topics Playing Association Roles Every AssociationRole must be played by exactly one 1 ”&xmls;nonNegativeInteger”>1 ”&xmls;nonNegativeInteger”>1 This concludes the definition of OWL object properties. 4.4. Data Properties Conversely, each Topic may play any number of Asso- ciationRoles, captured by playsRole, defined as an in- This section defines the relationships between individuals verse property of playedBy: within the Topic Map classes, and data values, of specified datatypes. The standard XML Schema Part 2: Datatypes [XMLS2 04] are used (denoted “&xmls” here). As per the TMDM specification, the locators are given as

Page 7 of 11 Chapter 3 Integrating Topic Maps into the Semantic Web

DRAFT ONLY 16/05/05 being of type “string”, but the option of forcing them to 4.4.6 Topic Locator Constraint datatype = ”&xmls; any- be URI references by setting The TMDM specifies the following constraint on Topic: URI” should probably be considered, as this is the stan- at least one of the three attributes subjectLocator, sub- dard for locating resources on the Semantic Web. jectIdentifier and sourceLocators must be present. This is enforced in OWL by creating a minimum cardinality 4.4.1 baseLocator restriction on class Topic with regard to the Locator A TopicMap may have a base locator that is a string: property defined earlier. As “Locator” encompasses the other three properties bound by the constraint, the restriction ensures at least one of them must be present.

1 As there is a constraint over the three Topic attributes (see also 4.4.6), we create a data property “Locator” that will contain the three data properties sourceLocator, sub- jectLocator, and subjectIdentifier as subproperties: 4.4.7 topicNameString

(The datatype is given by the Property’s range): Every topic map object may have any number of source locators which are URIs:

Variant and Occurrence may have either: • Name or text (respectively) as a string value OR • a locator where the relevant name or text resides. The TMDM distinguishes between the two cases based on 4.4.4 subjectLocator datatype; however, as this is not implementable in OWL A Topic may have any number of subject Locators of without extra processing, we have opted to separate the type anyURI (all required to point to the same “thing” two cases into different Data Properties: but enforcing this is outside the scope of the TMDM):

4.4.5 subjectIdentifier A Topic may have any number of subject Identifiers:

Page 8 of 11 123

DRAFT ONLY 16/05/05

5. Examples In this section, we use the OWL TMDM ontology as de- The “hamlet” example from the XTM1.0 specification was used as the inspiration and the information presented herein corresponds very closely with it. Firstly, we create a Topic “hamlet”, and state that it be- Assume that the Topic Map Data Model Ontology laid out longs to the “Shakespeare_Topic_Map” TopicMap. in Section 4 resides at “http://somelocation/tmdm” Then we create an individual TopicName, which has topicNameString of “Hamlet, Prince of Denmark”, and 5.1. User TM document Header state that this TopicName belongs to the Topic “hamlet”. Note that we could have opted to define the TopicName Set up “tm” namespace for tmdm, make standard XML, first, and then define the Topic, connecting it to the Topic RDF and OWL inclusions are made, import the “Topic Name using the “hasTopicName” Object Property. Map Data Model Ontology” and name the user ontology: This would give an exactly equivalent construction.

We then create the OccurrenceType An example user-defined Topic Map for “plain_text_format”. Since OccurrenceType has been Shakespeare, constructed in OWL DL using the TMDM OWL DL defined in the TMDM ontology as a subclass of topic, all constructs OccurrenceTypes are automatically Topics. We state Shakespeare Topic Map Ontology that it belongs to our Shakespeare TopicMap. Then we create an individual Occurrence, which has a The ontology can contain more than one Topic Map, so topicNameString of “Hamlet, Prince of Denmark”, and we name the individual Topic Map. give it a type of “plain_text_format”. Since in the TMDM ontology we stated that when the domain of the “type” Object Property was “Occurrence” the range must be “OccurrenceType”, OWL consistency checks will produce an error if this is not the case.

Our user-defined Topics and Associations will belong to Using the occurrenceOfTopic ObjectProperty, we state this Topic Map. that this Occurrence is of the Topic “hamlet”. Lastly we use occurrenceTextIsLocated to give the URI location Note the use of the prefix “tm”, which indicates the con- where the full text of hamlet is located. Note that this struct is from the imported TMDM ontology. must be a valid URI or an error will be produced.

5.2. Topics, Topic Names, Types, Occurrences 5.3. Locators Say we would like to create a topic “hamlet” which has: • A topic name String: “Hamlet, Prince of Denmark” OWL consistency checks will still however produce an • An occurrence of type “plain-text-format” which error on the above code, because we have neglected to resides at the Gutenberg.org ftp site specify at least one locator for our Topics.

locator SubjectIdentifier The OWL code which achieves this is given below: A such as can be set as follows (similarly for baseLocator, sourceLocator and sub- jectLocator):

The Tragedy of Hamlet, http://www.topicmaps.org/xtm/1.0/ Prince of Denmark country. xtm#DK/

Here we have set the subjectIdentifier for a Topic “dk”

Page 9 of 11 Chapter 3 Integrating Topic Maps into the Semantic Web

DRAFT ONLY 16/05/05 5.4. Associations and Association Roles easily do so in the same manner as it is done for each As- sociationRole type Suppose we would now like to capture the authorship , and OWL would check that the given to the Association belonged to the Association- relationship between the work “hamlet”, the author “Wil- Type class (subclass of Topic). liam Shakespeare” and the type of work “play”.

5.5. Scope To do this we define an Association which has three association roles for the work, author and type of work. Say we would like to give the Topic “hamlet” a different Firstly we set these up as Association Role Types, then topicNameString depending on whether we are referring create the Association which has roles of these types. As- to “Hamlet” the play, or “Hamlet” the character. We do suming that we have already set up the topics “hamlet”, this by creating Scope individuals for “play” and for “William Shakespeare” and “play”, we simply slot “character” and assigning the respective scopes to these in to the relevant AssociationRole using the Object TopicNames which have different Topic name Strings. Property playedBy: Scopes are automatically Topics as Scope is a subclass of Topic in the TMDM ontology.

The Tragedy of Hamlet, Prince of Denmark Hamlet

Note that both TopicNames belong to the same Topic. If we wanted to refer to “hamlet” the geographical township, this would refer to a different subject, and should therefore be created as a different Topic. (We probably would put it in a different Topic Map too.) Note that scope can be used with Associations, AssociationRoles and Oc- currences also.

5.6. Variants Consistency checks enforce the following OWL restric- Say we would like to give our “hamletTopicName” Topic tions: • Name from the previous “scope” example at 5.5 some Association Roles can only be played by topics Variant names for display under different circumstances. • An Association must have at least one Role • A Role must belong to exactly one Association • Note the constraint on Variant Scope as follows: Types of Association Roles must come from the “AssociationRoleType” class (subclass of Topic) The value of the scope of each individual variant item must be a true superset of the value of the scope of its Note that in this example, we have opted not to give our parent topic name. Association an associationType property, but we could

Page 10 of 11 125

DRAFT ONLY 16/05/05

So any variants of “hamletTopicName” must have at least all the scopes that apply to “hamletTopicName”, References plus at least one more. Note in the example the scope

“play” is added to each Variant, as this is the scope of “hamletTopicName”. Additionally we specify the [OWLDL 04] OWL Web Ontology Language Reference, World- scopes which define the parameters within which the Wide Web Consortium (W3C) Recommendation, 10 February Variant is to be used. 2004, available at http://www.w3.org/TR/2004/REC-owl-ref- 20040210/

[TMDM 05] Topic Maps – Part 2: DataModel Draft ISO/IEC International Standard (ISO/IEC JTC 1/SC34 ISO/IEC FCD 13250-2), 10 January 2005, available at http://www.isotopicmaps.org /sam/sam-model/ [TM 02] Topic Maps Second Edition, ISO/IEC International Standard (ISO/IEC 13250), 19 May 2002, available at 3250-2nd-ed-v2.pdf

[XTM 01] XML Topic Maps (XTM) 1.0, TopicMaps.Org Speci- fication, 06 August 2001, available at http://www.topicmaps.org/xtm/1.0/xtm1-20010806.html [XMLS2 04] XML Schema Part 2: Datatypes Second Edition, World-Wide Web Consortium (W3C) Recommendation, 28 October 2004, available at http://www.w3.org/TR/2004/REC- xmlschema-2-20041028/

6. Conclusions and Future Work

This paper has presented a proposal for an OWL DL formalisation of the ISO drafted standard Topic Map Data Model. The constructs defined herein provide the basis for building user- defined Topic Maps in OWL DL or OWL Full.

The author believes that the use of this formalism and associated tools will meet the majority of the requirements for a Topic Map Query Language (TMQL) and a Topic Map Constraint Lan- guage (TMCL). Some additional extensions to the OWL tools may be needed to meet all the requirements specified, but the use of the OWL platform and existing tools as a base should significantly reduce the workload required.

The author intends in immediate future work to analyse the extent to which OWL DL and associated tools may satisfy the TMQL and TMCL use cases specified.

DISCLAIMER: Please note that this document is currently at DRAFT stage only, and is NOT an official proposal. The purpose of this document is to illustrate the fit between OWL and the TMDM, and the potential for the use of OWL DL as a standard for Topic Map implementation. After discussion with the ISO/IEC, it may or may not be revised and put forward as an official proposal for an OWL standard for Topic Map implementation. At the very least, revisions are likely to be needed before such an event might take place.

Page 11 of 11

4 Adding Rules to OWL DL Ontologies

4.1 Pushing the limits of OWL, Rules & Prot´eg´e

Title of Publication: Pushing the limits of OWL, rules and Prot´eg´e - a simple example Type of Publication: Workshop Paper Appears In: Proceedings of the OWLED2005 Workshop on OWL, Experiences and Directions, Galway, Ireland, November 11-12, 2005 CEUR-WS, Vol. 188, 2005. Publication Date: 2005 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Malgorzata Mochol Denny Vrandecic Sean Bechhofer Personal Contribution: 35% (estimated by co-authors)

127 Chapter 4 Adding Rules to OWL DL Ontologies 129 Pushing the limits of OWL, Rules and Prot´eg´e A simple example

Anne Cregan1,2, Malgorzata Mochol3, Denny Vrandeˇci´c4, and Sean Bechhofer5

1 University of New South Wales, Australia, [email protected] 2 National ICT Australia (NICTA) 3 Freie Universit¨at Berlin, Germany, [email protected] 4 AIFB, Universit¨at Karlsruhe (TH), Germany, [email protected] 5 University of Manchester, UK, [email protected]

Abstract. The Semantic Web brings powerful languages for creating models, based on Description Logics and Rules. These languages are used within ontology engineering tools and applied in strong and efficient reasoners. Working within the context of a Summer School, we explored these technologies by creating a small and easy to understand example, working firstly with OWL alone, and introducing SWRL rules as needed, to automatically classify a number of instances according to some intuitively simple criteria. We present the example OWL ontology, SWRL rules, and the issues and problems we encountered. Our experience highlights the capabilities and limitations of OWL and of SWRL, not in theoretical but in practical terms, and points to the need for better tool support, but is primarily a lesson in ”traps for young players” in terms of formulating classifications, that we hope will be instructive for Semantic Web students following in our footsteps, and their tutors.

1 Introduction

The Semantic Web[4] vision is to extend the Web with machine-understandable data, forming a global distributed knowledge store which application may leverage to perform tasks automatically. The base technologies for realizing this vision are: Uniform Resource Identifiers(URIs) as a global identification mechanism for resources; the Resource Description Framework(RDF) as a basic data model, together with its XML-based serialisation syntax for publishing data on the Web[16]; the Web Ontology Language(OWL) which extends RDF with more expressive knowledge representation[15, 14]; and a yet to be defined rule language, for which the Semantic Web Rule Language(SWRL) may be considered a prototype[7]. The development of the Semantic Web is a joint effort of scientific and business institutions around the globe, led by the World Wide Web Consortium(W3C). This paper describes the experiences of a group of students(Cregan, Mo- chol, Vrandeˇcić) and their tutor(Bechhofer) in a mini-project conducted at the ThirdEuropeanSummerSchoolonOntologicalEngineeringandtheSemantic Web(SSSW05) held in Spain in July 20051. The students, all PhD students at 1 http://babage.dia.fi.upm.es/sssw05/ Chapter 4 Adding Rules to OWL DL Ontologies

their respective institutions, formed a project group with the aim of discover- ing the limits of OWL in relation to the use of rules. Specifically, we wanted to discover what tasks are beyond the expressivity of OWL and require the use of rules, and what additional capabilities do rules provide? We used Protégé ontology software[12], the Racer automated reasoner[3] and SWRL in our investigation: SWRL is based on a combination of the OWL DL and OWL Lite sublanguages of OWL together with the Unary/Binary Datalog RuleML sublanguages of the Rule Markup Language. To facilitate the investigation, we devised a simple example problem which we regard as a realistic, though small, use case for Semantic Web technologies. The paper describes the challenges that arose and our learning process in applying various conditions to our example domain. We hope that our experience will be helpful both as a teaching resource, adding to the (currently) limited amount of tutorial material available for teaching Semantic Web technologies to students (see [6, 11, 13]), and also, by highlighting the problems we experienced, will help tool implementers and standards bodies to identify issues inherent in their approaches, and initiate discussion of possible solutions. The material discussed is available on http://www.aifb.uni-karlsruhe.de/WBS/dvr/rove.Itmayalso be useful as an early test case for reasoners or editors implementing SWRL. The paper is structured as follows: Firstly, we specify the general scenario of our use case in section 2. Then we describe the ontology we constructed, and present some limitations of the expressivity of OWL in section 3, and of SWRL in section 4. Section 5 deals with problems we experienced using the tools, and outlines some directions for future work. Lastly in section 6, we summarise our learning experience, the issues we encountered and their significance.

2 Scenario

All participating students at SSSW05 were asked to form a group of four or five students to conduct a mini-project. At the outset, the organizers asked the students to make the groups as mixed as possible, to ensure everyone had a diverse collaborative experience contrasting with their usual work/study activities. The organizers also impressed on the students that having fun was a very important part of the activity (no doubt as it enhances learning!). Every student belonged to exactly one group. Each group had a unique name, and was led by exactly one summer school tutor. Every tutor led at least one group, and some led more than one group. All participants in the mini-project were either tutors or students, and no-one was both a student and a tutor. Due our chosen area of investigation we named our mini-project group“ROVE ” (Rules for Ontology Validation Effort). For the ROVE project, we chose a simple and readily accessible domain for applying ontology representation and rules: the student groups taking part in the Summer School mini-projects. In order to discover the limits of OWL’s abilities, and the capabilities provided by adding rules, we attempted to formally define and implement the informally stated conditions the organizers had placed on group formation. 131

2.1 Conditions We decided that the following conditions reflected desirable group formations: Condition #1 : Groups should have either 4 or 5 members Condition #2 : Groups should have at least one member of each gender Condition #3 : Group members are of all different nationalities Condition #4 : Group members are from all different institutions Condition #5 : Groups should have fun. We decided it would be fun if the tutor of the group were the favourite of all the students in the group, so we asked all students to nominate which tutor was most attractive (see sec 4). Our goal was to formalize these conditions and use a classifier to automatically categorize all groups as either a GoodGroup - one that fulfills all conditions, or a BadGroup - one which does not satisfy one or more of the stated conditions: Group is a GoodGroup iff Cond1 ∧ Cond2 ∧ Cond3 ∧ Cond4 ∧ Cond5 Group is a BadGroup iff¬Cond1 ∨¬Cond2 ∨¬Cond3 ∨¬Cond4 ∨¬Cond5

3 ROVE Ontology 3.1 Classes and Instances Initially we built a simple ontology (see Fig. 1) containing disjoint classes Person, Country, Institution,andGroup. Person was divided into disjoint subclasses Tutor and Student exactly partitioning the class. Person was also divided by gender into disjoint subclasses Man and Woman, also as an exact partition.

Fig. 1. ROVE-ontology Chapter 4 Adding Rules to OWL DL Ontologies

3.2 Properties

Object Properties were then set up as follows: each person has a Nationality (hasNationality), and works at an Institution(worksAt). Each Group is led by a Tutor (ledBy) and has members (hasMember), only from the class Student. Each Student is a memberOf exactly one Group, and has a favorite Tutor (at- tractedTo). Data Properties related Persons and Groups to name strings.

3.3 Asserting Conditions using OWL

Using this ontology, we then tried to formalise and implement as many of our stated conditions as possible using only OWL expressivity.

Condition #1 : Groups should have either 4 or 5 members

This condition was easily implemented by setting minimum and maximum cardinalities on the property hasMember which related Groups to Students:

Group ≥4 hasMember ≤5 hasMember

As all the groups satisﬁed this condition and cardinality conditions are easily formulated within OWL, we simply asserted this condition on Group.

Condition #2 :Groupsshouldhaveatleastonememberofeachgender

This condition could not be modelled in OWL in the same way as Condition #1, as it requires hasMember to have a minimum cardinality for values from each of the subclasses Man and Woman. We used existential restrictions to accomplish the task, by defining GoodGroup as a subclass of Group where hasMember has (owl:)someValuesFrom Man and (owl:)someValuesFrom Woman.However,we wanted to capture that this condition was only one of those to be satisfied by a good group, so we did not want to make it a sufficient condition. Yet we also wanted to ensure that all members of the group must be either male or female, and there was no way to do this using a necessary condition. We then approached the problem in reverse by specifying conditions for being a bad group rather than a good group. Note that not satisfying any one of our five conditions is sufficient for classification as a bad group. Protégésupports explicit specification of either ‘Necessary’, or ‘Necessary and Sufficient’ asserted conditions (see Figure) through an interface box, but if one has conditions which are Sufficient without being Necessary, as in this case, this can only be implemented by adding a subClass. We felt it would be more intuitive for Protégéto support the assertion of subsumption axioms through the interface box also. By introducing two new concepts to represent groups with all male members or all female members: ManGroup and WomanGroup,where:

WomanGroup ≡ Group ∀hasMember.W oman ManGroup ≡ Group ∀hasMember.Man 133 and creating BadGroup as a subclass of Group, it was then simple to implement these two sufficient conditions for being a “bad group” by making WomanGroup and ManGroup subclasses of BadGroup: WomanGroup BadGroup ManGroup BadGroup However, having set this up, we noticed in automatic classification that one of the groups was not classified as a BadGroup, although it consisted of four male students. This was due to OWL’s Open World Assumption: the group, having four male members, could still potentially have a fifth member who may be female, without breaking the cardinality restriction, thus the reasoner could not classify the group as a BadGroup. To solve this problem we considered several alternatives. We had to dismiss the possibility of defining groups by enumeration, as this included nominals and was not supported by the available reasoners [5]. Had we pursued this path, we would have needed to rework our ontology to model group membership as a class instead of a relation, and this did not seem intuitive to us in any case. Instead, we decided to state the size of the groups explicitly by creating two new concepts: BigGroup (group with exactly 5 members) and a SmallGroup (group with exactly 4 members) as disjoint subclasses of Group, and then assert our other conditions onto these concepts:

BigGroup ≡ Group ≥5 hasMember ≤5 hasMember SmallGroup ≡ Group ≥4 hasMember ≤4 hasMember Group ≡ SmallGroup BigGroup

Conditions #3-#5 These conditions required consideration of more than one property at a time: for example, both a student’s nationality and group membership. Whilst OWL has a relatively rich set of class constructors, expressivity for properties is much weaker. Whilst OWL permits chaining of properties, it does not support making assertions about the equality of the objects at the end of two diﬀerent properties/property chains. Since conditions #3-#5 require precisely this kind of assertion, it was not possible to formulate them using only OWL.

4ROVERules

In the next step, we explored the use of rules to implement our conditions, using the Semantic Web Rule Language(SWRL)[7] plugin to ProtégéwiththeROVE ontology described above. Rules are constructed in the form of an implication between an antecedent (body) and a consequent (head): whenever the conditions specified in the antecedent hold, then the conditions specified in the consequent must also hold. In SWRL rules, one may assert equivalences as well as implications. SWRL provides Horn-like rules for both OWL DL and OWL Lite, includes a high-level syntax for representing these rules[1] and is more powerful than either OWL DL or Horn rules alone[8]. Chapter 4 Adding Rules to OWL DL Ontologies

Condition #3 : Groups should have members of all diﬀerent nationalities

To classify bad groups in regard to this condition we constructed a rule “Same- NationalitiesRule” which states that if a group has any two members with the same nationality it is a BadGroup.Speciﬁcally,ifgroupg has member s and member s has nationality n and group g has another member x who is diﬀerent from member s and member x also has nationality n then g is a BadGroup. SWRL notation:

hasMember(?g,?s) ∧ hasNationality(?s, ?n) ∧ hasMember(?g,?x) ∧ hasNationality(?x, ?n)∧ differentFrom(?s, ?x) → BadGroup(?g)

Condition #4 : Groups should have members from all diﬀerent institutions

The same approach applies as was used for diﬀerent nationalities:

hasMember(?g,?s) ∧ worksAt(?s, ?i)∧ hasMember(?g,?x) ∧ worksAt(?x, ?i)∧ differentFrom(?s, ?x) → BadGroup(?g)

Condition #5 : Groups should have fun

To stress-test OWL’s abilities to represent and reason with compositions of properties and situations where there were multiple properties that connected the classes, we devised a complex criteria for“fun group”: a fun group is one where all the students in the group are attracted to the tutor leading their group. We also had the ulterior motive of providing amusement for both tutors and students when the project work was presented on the last day of the summer school. Each student was asked which tutor they were most attracted to: if the student asked for clariﬁcation as to what “attracted to” meant, they were told that they may interpret it as they wished. As the school had more male tutors than female (5 as opposed to 2) and more male students than female (35 to 20), then assuming a traditional interpretation of “attraction” there is likely to be a bias in the data, favoring the student body’s attraction to the female tutors, at least in terms of raw numbers. However, while the results were a source of much amusement for the participants (the ROVE presentation included poll results for “Most attractive Summer School tutor”), they were inconsequential for the real purpose of the exercise (although perhaps not to the tutors!). We formulated the rule to capture that any group that has a member who is attracted to someone other than the group’s tutor is a bad group:

hasMember(?g,?s) ∧ attractedT o(?s, ?t) ∧ hasT utor(?g,?x)∧ differentFrom(?t, ?x) → BadGroup(?g)

Incidentally, our other ulterior motive in formulating all the conditions was to surreptitiously ensure that the ROVE group was the only group that satisﬁed all conditions, and we are pleased to say that we successfully achieved this. 135

4.1 OWL: Not Bad does not equal Good At this point we thought we had now captured all our conditions adequately using OWL and SWRL, and would be able to automatically classify the groups as good or bad. Whilst we were able to classify BadGroups, we had no way do the converse: that is, to automatically classify GoodGroups. Stating that GoodGroups are all groups that are not bad groups will not work: not being classified as a bad group only indicates that the group’s status as a good or bad group is unknown. There is no way to specify that the list of criteria is exhaustive and use negation as failure, as OWL’s semantics adopt the open world assumption: a statement cannot be assumed true just because its negation cannot be proven[9]. Therefore, in order to classify GoodGroups as such, we needed to reformulate all the prior conditions in a positive form, thereby setting up necessary and sufficient conditions for being a GoodGroup. By positive form, we mean a formulation designed to satisfy the condition, rather than to violate it. For instance, with the nationality condition (Condition #3), the positive form ensures all members of the Group have different nationalities, whereas the negative form simply tests whether any two members have the same nationality. We found it particularly challenging to grasp the asymmetric implications of using the positive and negative forms of rules, as intuitively we were only attempting to view the situation from the opposite side. We felt that a symmetric approach would be easier to use, as could be provided by tools, for example a wizard creating closures over rules automatically. Maybe such a pattern should be expressible in SWRL itself, in order to increase interoperability. By formulating positive rules for our conditions, we were able to define Good- Groups as exactly those Groups which satisfied all the rules: GoodGroup ≡ MixedGenderGroup InternationalGroup InterInstitutionalGroup FunGroup This is done by creating an OWL axiom using class descriptors defined in SWRL. However, formulating the rules in positive form in SWRL was extremely onerous for most of our conditions.

Condition #2 :Groupsshouldhaveatleastonememberofeachgender

Creating a MixedGenderGroup was easy: we were able to take the pure-OWL approach previously described in section 3.3: MixedGroup ≡ Group ∃hasMember.Male ∃hasMember.F emale

Condition #3 : Groups should have members of all diﬀerent nationalities

To define this condition positively, we constructed an InternationalGroup rule containing as antecedent a pairwise comparison of nationalities of every member to test whether they were all different. There need to be two forms of the rule for BigGroup and SmallGroup, as 5-member groups require 5 different nationalities, not 4 : since every binary combination has to be considered, 4-member Chapter 4 Adding Rules to OWL DL Ontologies

groups require only 6 pairwise comparisons, whilst 5-member groups require 10 n−1 comparisons ( 1 n comparisons for n members). To show the complexity of such a rule we have shown it below for a 3-member group only:

Part of the rule Explanation Meaning SmallGroup(?x) if x is a small group ∧ and hasMember(?x, ?a)∧ hasMember(?x, ?b)∧ group x has members: a, b and c Group x has 3 members. hasMember(?x, ?c) ∧ and differentFrom(?a, ?b)∧ member a is different differentFrom(?a, ?c) from members b and c Group x has 3 ∧ and different members differentFrom(?b, ?c) member b is different from member c (equality to a was checked above) ∧ and hasNationality(?a, ?e)∧ member a has nationality e and Each group member has hasNationality(?b, ?f)∧ member b has nationality f and nationality (no information hasNationality(?c, ?g) member c has nationality g regarding different or same nationalities). ∧ and differentFrom(?e, ?f)∧ nationality e is different from differentFrom(?e, ?g) nationalities f and g Each of the 3 group members ∧ and has different nationalities. differentFrom(?f, ?g) nationality f is different from nationality g (equality to e was checked above) → then GoodGroup(?x) the group x is a good group Such a rule is long and unwieldy to write with the SWRL plugin: we lamented the lack of pre-defined predicates in SWRL, as a simple “all-different” predicate to enter the arguments and have SWRL take care of the pairwise comparisons wouldhavebeenagreataid. Condition #4 : Groups should have members from all different institutions Analogously to the above, InterInstitutionalGroup is another huge rule, requiring separate forms for 4 and 5 member groups.

Condition #5 : Groups should have fun

Although the intuition behind the deﬁnition of a FunGroup was easy - a group where all its members were attracted to the every tutor leading the group - formalizing the rule in positive form again proved to be an extremely tedious task, resulting in a huge and hard to maintain rule.

5 Tool support

The first and most important problem was the lack of an appropriate reasoner that could be plugged into a rule-enriched ontology created with Protégé. Racer supported reasoning within OWL, but to use rules in automatic classification obviously requires a reasoning engine which is able to operate on both OWL ontologies and SWRL rules, and none were available for the Windows-based 137

Summer School environment. However, we plan to use such reasoners, like Hoo- let[2] or KAON2 [10] and describe the results on the ROVE website. Further problems related to the ontology engineering environment where we identified two issues with ProtégéandonewithSWRL: Issue #1: entering data proved very tedious when making instances of a superclass belong to an orthogonal subclass partition: in our example, Person was partitioned into Student/Tutor subclasses as well as Man/Woman, and when entering the second SuperClass in Protégé we could not simply select a number of instances and assign them to a class, but had to edit each instance individually. Issue #2: the tabs containing “necessary” and “necessary & sufficient” conditions left us wondering why there is no simple “sufficient” tab in Protégéaswellto create subsumption axioms, as described in section 3. Issue #3: as described in section 4.1, entering SWRL rules to create a closure for a class description with regard to a superclass was very tedious and error-prone. Even in our simple use case, the rules quickly become very large and therefore difficult to edit, let alone maintain and extend. A real-life example is likely to produce rules that are not manageable manually, so creating (and especially maintaining) such closures automatically needs to be made possible.

6 Conclusions

In this simple example of our Summer School mini-project, we showed the difficulties and challenges of implementing some intuitively very simple class constraints. The key learning experience were the limitations of the expressivity of OWL. By going beyond OWL and using SWRL, we were able to formulate several further conditions using rules, but found it challenging to fully understand theimplicationsoftheOpenWorldAssumptionandthelackofNegationas Failure. Without the input of our tutor, we might easily have fallen into the trap of believing we had captured all our conditions, when in actual fact we had not. Constructing and editing long rules in SWRL to address this was particularly difficult, and we strongly encourage developers to introduce better support for rule construction using templates and built-in predicates. We hope our experience will provide a lesson for Semantic Web students following in our footsteps, and will provide insights for Semantic Web tool developers. The ROVE material is available in different stages at http://www.aifb.uni-karlsruhe.de/WBS/dvr/rove for use in training and tool testing.

Acknowledgements We thank Antoine Zimmermann, who also was a member of the ROVE team, Enrico Motta and Asuncion Gomez-Perez, the organizers of the SSSW2005, and all the students, tutors and speakers. Congratulations to Natasha Fridman-Noy who was the Summer School’s ‘most attractive tutor’. Yes Asun, now it’s time to go to the bus. Research reported in this paper has been partially ﬁnanced by NICTA (www.nicta. com.au), the EU project SEKT (www.sekt-project.com), the NoE KnowledgeWeb (knowledgeweb.semanticweb.org), and the German Ministry of Research (BMBF) Chapter 4 Adding Rules to OWL DL Ontologies

project Knowledge Nets (nbi.inf.fu-berlin.de/research/wissensnetze) of the In- terVal (interval.hu-berlin.de) Berlin Research Centre for the Internet Economy (www.internetoekonomie.net).

References

1. R. Aggarwal. Semantic web services languages and technologies: Com- parison and discussion. LSDIS Lab, University of Georgia, 2004. citeseer.ist.psu.edu/700682.html. 2. S. Bechhofer. Hoolet. Department of Computer Science, University of Manchester, 2004. http://owl.man.ac.uk/hoolet/. 3. V. Haarslev and R. Möller.Racer: An owl reasoning agent for the semantic web. In Proc. of the International Workshop on Applications, Products and Services of Web-based Support Systems, in conjunction with 2003 IEEE/WIC International Conference on Web Intelligence, Halifax Canada, Oct 13, pages 91–95, 2003. 4. J. Hendler, T. Berners-Lee, and E. Miller. Integrating applications on the semantic web. Journal of the Institute of Electronic Engineers of Japan, pages 676–680, 2002. 5. J. Hladik. Reasoning about nominals with FaCT and RACER. In Proc.ofthe 2003 International Workshop on Description Logics (DL2003), CEUR-WS, 2003. 6. M. Horridge, H. Knublach, A. Rector, R. Stevens, and C. Wroe. A practical guide to building OWL ontologies using the Protégé-OWL plugin and CO-ODE tools, 2004. University of Manchester. 7. I. Horrocks, P. Patel-Schneider, H. Boley, S. Tabet, B. Grosof, and M.Dean. Swrl: A semantic web rule language combining owl and ruleml. DARPA DAML Program, 2003. http://www.w3.org/Submission/2004/SUBM-SWRL-20040521/. 8. I. Horrocks, P. F. Patel-Schneider, S. Bechhofer, and D. Tsarkov. Owl rules: A proposal and prototype implementation. Conditionally accepted for publication. Journal of Web Semantics, 2005. 9. I. Horrocks, P. F. Patel-Schneider, and F. van Harmelen. From SHIQ and RDF to OWL: The making of a web ontology language. J. of Web Semantics, 1(1):7–26, 2003. 10. U. Hustadt, B. Motik, and U. Sattler. Reducing SHIQ− description logic to disjunctive datalog programs. In D. Dubois, C. Welty, and M.-A. Williams, editors, Proceedings of the KR2004, pages 152–162. AAAI Press, 2004. 11. F. Manola and E. Miller. Resource Description Framework (RDF) Primer. W3C Recommendation 10 February, 2004. At http://www.w3.org/TR/rdf-primer/. 12. N. Noy, R. Fergerson, and M. Musen. The knowledge model of Protégé-2000: Combining interoperability and flexibility. In R. Dieng and O. Corby, editors, Proc.ofthe12thEKAW, volume 1937 of LNAI, pages 17–32. Springer, 2000. 13. A. Rector, N. Drummond, M. Horridge, J. Rogers, H. Knublauch, R. Stevens, H. Wang, and C. Wroe. OWL pizzas: Practical experience of teaching OWL-DL: Common errors & common patterns. In E. Motta, N. R. Shadbolt, and A. Stutt, editors, Proc.ofthe14thEKAW, volume 3257 of LNCS, pages 63–81. Springer, 2004. 14. M. K. Smith, C. Welty, and D. McGuinness. OWL Web Ontology Language Guide, 2004. W3C Recommendation, http://www.w3.org/TR/owl-guide/. 15. W3C. Owl web ontology language-reference. LSDIS Lab, University of Georgia, 2004. http://www.w3.org/TR/owl-ref/. 16. W3C. W3c: Rdf/xml syntax specification (revised). LSDIS Lab, University of Georgia, 2004. http://www.w3.org/TR/rdf-syntax-grammar/. 139

4.2 Exploring OWL & Rules

Title of Publication: Exploring OWL and Rules - A Simple Teaching Case Type of Publication: Journal Article Accepted for International Journal of Teaching and Case Studies (IJTCS), Publication In: Special Issue: Teaching Semantic Web: Integration with CS/IS curriculum and Teaching Case Studies. Editor: Miltiadis Lytras Publisher: Inderscience Peer Reviewed: Yes Contributing Author(s): Malgorzata Mochol Anne Cregan Denny Vrandecic Sean Bechhofer Personal Contribution: 30% (estimated by co-authors) Chapter 4 Adding Rules to OWL DL Ontologies 141

International Journal on Teaching and Case studies, Vol. x, No. x, xxxx 1

Exploring OWL and Rules - A Simple Teaching Case

Malgorzata Mochol1, Anne Cregan2,3, Denny Vrandeciˇ c´4, and Sean Bechhofer5

1 Freie Universitat¨ Berlin, AG Netzbasierte Informationssysteme Takustr. 9, 14195 Berlin, Germany, [email protected] 2 University of New South Wales, Sydney, NSW 2052, Australia 3 National ICT Australia (NICTA), 223 Anzac Parade Kensington NSW 2052 Australia, [email protected] 4 Universitat¨ Karlsruhe (TH), Institut AIFB Englerstrasse 11, 76131 Karlsruhe, Germany, [email protected] 5 University of Manchester, Oxford Road, Manchester M13 9PL, UK [email protected]

Abstract: This case study explores current Semantic Web technologies through creating a small ontology using an Ontology Editor, and applying Axioms, Rea- soning and Rules to automatically classify data in the ontology, according to some intuitively simple pre-set criteria. The study includes designing and building a suitable ontology to represent a simple knowledge domain using an ontology editor, representing Logical Constraints using constructs available within the OWL-DL language, and then illustrating that some of the classification criteria set for the task cannot be represented within OWL and require logical rules to be added using a rule language to achieve the desired functionality. An automated reasoner is used for classification, and the impact of the Open World Assumption on classification results is carefully examined. The case supports hands-on exercises using current Semantic Web Tools and Languages to represent a simple knowledge domain, giving exposition of the underlying logical formalisms.

Keywords: Semantic Web, ontology, OWL, rules, reasoner, SWRL

Reference to this paper should be made as follows: Mochol, M., Cregan, A., Vrandeciˇ c,´ D., and Bechhofer, S. ‘Exploring OWL and Rules - A Simple Teach- ing Case’, International Journal on Teaching and Case studies, Vol. x, No. x, pp.xxx–xxx.

Biographical Notes: Malgorzata Mochol is a Graduate Research Assistant at the Institute for Com- puter Science of the Free University Berlin, Group of Networked Information Systems where she works primarily on ontology matching and application of Semantic Web technologies in e-business. She earned her degree in Computer Science at the Technical University Berlin in 2003. Anne Cregan is a PhD Student with the Knowledge Representation and Rea- soning Group of the Artiﬁcial Intelligence Lab in the School of Computer Sci- ence and Engineering at the University of New South Wales, Australia. She is also sponsored by the National Centre of Excellence for Information and Com- munication Technologies, Australia (NICTA). Denny Vrandeciˇ c´ is a Researcher at the AIFB, working primarily on collaborative ontology engineering and evaluation. He received his Master in Computer Chapter 4 Adding Rules to OWL DL Ontologies

2 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

Science and Philosophy at the University of Stuttgart. Sean Bechhofer is a Lecturer in the School of Computer Science of the Uni- versity of Manchester. He graduated in Mathematics from the University of Bris- tol in 1988 and has worked in Manchester since 1992, primarily in the area of tools and infrastructure to support the Semantic Web.

1 SUMMARY OF EDUCATIONAL ASPECTS OF THIS PAPER

This case study provides practical teaching material for introducing undergraduate students to the essential aspects of Semantic Web languages, tools and reasoning. It provides subject matter and resources to provide student experience in designing, building and reasoning with ontologies, with emphasis on giving a concrete context for students to explore the interaction between ontologies as defined by the W3C’s Web Ontology Language (OWL)a, rules as defined by rule languages such as the Semantic Web Rule Language (SWRL)b, and the impact of their underlying logical formalisms (e.g. OWL’s Open World Assumption), in relation to achieving desired outcomes in automated classification using Description Logic Reasoning. Currently available Semantic Web editors, tools and reasoners which may be applied to the task, and the authors’ experiences with them, are described and compared. The study is intended to be used in conjunction with the materials provided online at http://km. aifb.uni-karlsruhe.de/projects/rove.

2 INTRODUCTION

This case study provides practical teaching materials for introducing undergraduate students to Semantic Web Languages, Tools and Reasoning. It provides:

• a description of a scenario and task which requires the use of Semantic Web technologies to achieve an automated classiﬁcation according to set criteria

• a summary of currently available Semantic Web tools which may be used

• a possible solution to the task

• problems likely to be encountered

• links to other resources

• teaching guidelines

All tools, materials and resources referred to are available online through the ROVE website at http://km.aifb.uni-karlsruhe.de/projects/rove.

ahttp://www.w3.org/TR/owl-features bhttp://www.w3.org/Submission/SWRL 143

Exploring OWL and Rules - A Simple Teaching Case 3

2.1 Motivation

The key questions motivating this case study are:

• What are the limits of OWL’s expressivity?

• When does OWL need to be supplemented with rules to provide additional functionality?

• What additional capabilities do rules provide?

It is not intended to answer these questions in a formal way – such answers are readily available – but rather to allow the students to test what can and what cannot be expressed with a certain language with regards to a speciﬁc use case, and thus increase their understanding about the limitations of the languages.

2.2 Learning Outcomes

The material provided by this study gives practical hands-on experience in:

• Designing an ontology to represent a simple domain

• Using ontology editors (e.g Protege, SWOOP) to build a simple ontology and populate it

• Adding axioms to the ontology to capture logical properties of the domain

• Using OWL-DL and understanding how its constructs represent ontologies and their logical constraints

• Visualising and querying ontologies using tools

• Using automated reasoners (e.g. Racer, Fact, Pellet) to perform automated classiﬁ- cation tasks

• Investigating the limits of the OWL ontology language: ﬁnding out what it can and cannot express

• Supplementing ontologies with rules in a rule language such as SWRL to provide additional expressivity

• Investigating the behaviour of Description Logics and automated reasoners

• Understanding the logical implications of OWL’s ”Open World Assumption”

The case study and questions posed provide an expose of current Semantic Web technologies and tools, and executing the task will provide students with a tangible illustration of some of the more elusive aspects of Description Logics. Chapter 4 Adding Rules to OWL DL Ontologies

4 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

2.3 Using this Teaching Case

Practical exercises based on this case study would work well as a precursor to, or in parallel with a more technical formal treatment of the Semantic Web field. The case study itself is self-contained and does not require previous knowledge of Seman- tic Web technologies, although it is advanced enough to provide interest and challenges for those students who may have some previous experience in building and using ontologies. Some previous experience in formulating Necessary & Sufficient conditions would be advisable, although practical exercises based on this case study may be adapted to suit the level of the audience, from a step-by-step tutorial with tools provided, to a research- oriented task which simply sets out the required classification criteria and encourages students to find and explore various tools which are available to achieve the desired results.

2.4 Background

The material contained in this case study originated as a mini-project within the 3rd Euro- pean Summer School on Ontological Engineering and the Semantic Web (SSSW05), held in Spain in July 2005. The mini-project, named ”ROVE” (Rules for Ontology Validation Effort) was initiated by a group of post-graduate students to explore the limits of functionality of the Web Ontology Language(OWL), and the additional functionality which could be achieved by adding rules to OWL ontologies. The authors of this case study are three of the four post-graduate students of the ”ROVE” group (Cregan, Mochol and Vrandeciˇ c)´ and the group’s tutor (Bechhofer), all PhD research students at their respective institutions, whilst Bechhofer, the group’s tutor, has played a key role in authoring Semantic Web languages and tools for some time. The task was originally conducted using tools readily available at the time: Proteg´ e´ with OWL-DL and SWRL Plug-in, used in conjunction with RACER for automated reasoning. A thorough description of the insights gained is given in [4].

2.5 Potential Variations and Local Adaptations

Instance Population and Conditions: The ontology presented captures information about people and groupings of people, and the task set is to automatically classify each group of people according to whether it meets certain intuitively simple criteria imposed on group membership and structure. The knowledge domain is therefore readily understood, and the task easily grasped, so that the students can fully concentrate on exploring the representation itself. The case was originally constructed with instance data provided by the students and tutors at the summer school undertaken by the authors, and pre-existing populated ontologies (partially anonymized for public availability) in various stages corresponding to amendments described in the paper, are available at the ROVE website http://km.aifb.uni-karlsruhe.de/projects/rove. However, for best results the the students should be allowed to build their own ontologies from scratch, to give an appreciation of the design decisions involved. The case study can easily be modiﬁed for local use by populating with details of the student group being trained, which is probably more interesting for the students and connects them better to the given tasks. In this case, the criteria set on group 145

Exploring OWL and Rules - A Simple Teaching Case 5

membership will usually need to be adapted to suit the students (eg using course enrolment instead of nationality, etc). Creativity is of course encouraged: there are any number of possible conditions for testing which would be both instructive and potentially entertaining.

Tools: The case study refers to currently available tools and languages (OWL, SWRL, Protege, etc), but as this is a developing ﬁeld with tools and languages rapidly evolving, the exploration of new tools and standards becoming available is encouraged. The described task and variants on it could readily be conducted with any available ontology editor using some ontology language, rules, and an automated reasoner. In the original task, the authors experienced numerous technical problems using the available tools (further described in Section 4) and we believe this case study provides a useful base case for tool testingc.

2.6 Organization of the Case Study

The case study is organized as follows: Section 3 describes the task set for the exercise including background of the scenario, the conditions for testing, and some key considerations in approaching the task. Section 4 contains a brief description of some available tools (editors and reasoners) which may be used to conduct the task. This is followed by the speciﬁcation of a possible solution (Section 5) which shows “step by step” how to satisfy the requirements of the task. The steps include building an ontology to cover as many of the stated conditions as possible within OWL, showing that it is not possible to capture all the requirements within the ontology itself, and then adding rules to capture the remaining conditions. The classiﬁcation results highlight the impact of OWL’s Open World Assumption, and a clear exposition of all the rules needed to complete the task gives a good insight into reasoning with Description Logics. Section 6 describes some practical problems likely to be encountered in using the tools and completing the task, and provides some suggestions for dealing with these. Throughout the case study, reference is made to various resources including reasoners, editors and languages, and Section 7 provides a list sort by category, including websites for each resource. The teaching guidelines are presented in Section 8 which is followed by the brief conclusion closing the case study in Section 9.

3 DESCRIPTION OF THE TASK

3.1 Scenario

The 3rd European Summer School on Ontological Engineering and the Semantic Web (SSSW05) was held in Spain for a week in July 2005. The student group at SSSW05 was made up of some 60 post-graduate students, both male and female, having many different nationalities and originating from many different educational institutions. At the outset, all participating students at SSSW05 were asked to form themselves into a group of four or ﬁve students to conduct a mini-project of their own choice. The summer school organizers asked the students to form groups as diverse as possible, in terms of having mixtures of

cCheck the ROVE website http://km.aifb.uni-karlsruhe.de/projects/rove for links to tools and other materials. Chapter 4 Adding Rules to OWL DL Ontologies

6 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

nationalities, institutions and genders in each group. This was to provide a collaborative experience contrasting with usual work and study activities. The organizers also impressed on the students that having fun was a very important part of the activity (no doubt as it enhances learning!).

3.2 Goals & Requirements

The authors, joined by Antoine Zimmerman, formed a project group with the intention of investigating the use of rules in conjunction with OWL. Sean Bechhofer, due to his expertise in the area, was invited to be the group’s tutor. Concerning the chosen area of research, the mini-project group was named “ROVE” (Rules for Ontology Validation Effort). For the ROVE project, we chose a simple and readily accessible domain for applying ontology representation and rules: the student groups taking part in the summer school mini-projects. In order to discover the limits of OWL’s abilities, and the capabilities provided by adding rules, we attempted to formally deﬁne and implement the informally stated conditions the organizers had placed on group formation.

We decided that the following conditions reﬂected desirable group formations:

Condition #1 Every group should have either 4 or 5 members Condition #2 Every group should have at least one member of each gender Condition #3 Members of a group should all be different nationalities Condition #4 Members of a group should all be from different institutions Condition #5 Groups should have fun. We decided it would be fun if the tutor of the group were the favourite of all the students in the group, so we asked all students to nominate which tutor was most attractive to them.

Our goal was to formalize these conditions and use a classiﬁer to automatically categorize all groups as either a GoodGroup - one that fulﬁlls all the stated conditions, or a BadGroup - one which does not satisfy one or more of the stated conditions. Stated logically:

Group is a GoodGroup iff Cond1 ∧ Cond2 ∧ Cond3 ∧ Cond4 ∧ Cond5

Group is a BadGroup iff ¬Cond1 ∨¬Cond2 ∨¬Cond3 ∨¬Cond4 ∨¬Cond5

3.3 Ontology Design and Construction

In building an ontology there are potentially many design decisions to be made to ensure the ontology will best suit the stated purpose. There is not necessarily a “right” and “wrong” way to do it, but usually some designs will provide the desired functionality more readily than others. The fundamental ontology constructs provided by OWL are classes, instances and properties, where: • classes contain instances, e.g. the class Person contains speciﬁc individual Mary. 147

Exploring OWL and Rules - A Simple Teaching Case 7

• classes may have subclasses which inherit their characteristics; e.g. Person can be subclassed as Male or Female, where any individual male or female belongs to the class of Person as well as to the appropriate gender subclass; • classes may themselves be subclasses of superclasses, e.g. the class Person could belong to a superclass of Mammal; • instances may have properties which connect them to specific values (DataProper- ties) or individuals (ObjectProperties), e.g. a person may have a specific age (data) and have a specific relationship to other individuals e.g. a person is the child of another person; • property relationships may have logical conditions imposed on them: e.g. the property “child of” from the previous example has an assigned domain (the class Person, and range (also the class Person), restricting the classes of individuals related by the property. Other logical axioms such as cardinality and transitivity may be imposed, and there may also be subproperties defined. We began by constructing a simple ontology to contain data about the students, tutors and project groups. Designing an ontology commences with identifying the key domain entities to be modelled, and selecting the most appropriate representations for these within the modelling language (classes, properties, instances, etc). For instance, in the domain being modelled, a person’s gender could potentially be represented either as subclasses over the class Person, or a data property mapping each instance of Person to a value of Male or Female, or an object property, should the designer choose to model gender itself as a class having instances Male and Female. Such design choices set the logical foundations for ontology construction, so it is important to choose wisely at the outset. In designing our ontology, we determined firstly that we should build a class “Person” whose instances would correspond to individual people. As our desired classifications depended on characteristics of individual people within groups, we needed to capture each person’s: gender, nationality, educational institution, membership of a mini-project group, status as a student or tutor, and if a student, which tutor was their favourite. In determining the best representation, we were guided by both the classification goals and the logical properties of the domain to be captured: • All participants in the mini-project were individual persons • All participants were either tutors or students • No-one was both a student and a tutor • All participants were either male or female • No-one was both male and female • Every student belonged to exactly one group. • Each group had a unique name • Each group was led by exactly one summer school tutor • Every tutor led at least one group, and some led more than one group. Some possible representations which adequately capture these conditions are suggested in Section 5, following a consideration of some of the Semantic Web tools available for approaching the task. Chapter 4 Adding Rules to OWL DL Ontologies

8 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

4 AVAILABLE TOOLS

One major advantage of Semantic Web technologies are their standardization on interchange formats and thus their interoperability. In principle it should not matter which tools are used for editing the ontologies and later for reasoning with them, as long as they are using an OWL representation. This advantage (and its practical limitations) can be demon- strated in class by allowing the students to take tools of their own choice. In this section we will discuss some available and popular tools that can be used. We are aware that several further tools exist but cannot provide an extensive list here: furthermore as this is a rapidly evolving ﬁeld, additional tools are regularly becoming available. Download links and references to additional resources regarding these tools are given in Section 7. All the tools brieﬂy described here are in active development at the time of writing, and it is recommended to use a recent version of each.

4.1 Ontology Editors

Proteg´ e´ is currently the best known ontology development environment and was used in the original development of the ROVE ontology. It has been in active development by the SMI (Stanford Medical Informatics) group of the University of Stanford since the early 1990s. A number of Semantic Web plug-ins have been developed specifically for use with Proteg´ e.´ In particular, an OWL plug-in available for use with Proteg´ e´ produces ontologies in the OWL language. This plug-in was developed largely by the Co-ode project in Manchester http://www.co-ode.org. The SWRL plug-in for Proteg´ e´ supports the use of rules. The OWL and SWRL plug-ins provide enough expressivity to state all required axioms. The user interface of Proteg´ e,´ although intuitive, is not based directly on OWL, so teaching OWL with Proteg´ e´ requires consideration of how Proteg´ e´ translates into the OWL language and conversely, how constructs within OWL are represented in Proteg´ e.´ For example, the notion of “Necessary & Sufficient conditions” in the Proteg´ e´ interface is translated to semantically equivalent OWL subsumption axioms, which may not appear obvious on the first glance. Proteg´ e´ itself does not include a reasoner but may be used with any external reasoner which uses the DIG [2] (Description Logic Implementation Group) interface. Proteg´ e´ has an extensive range of plug-ins for visualization and querying of ontologies, and exploration of these is encouraged.

SWOOP is an ontology editor based on a browser-inspired interface, developed by the MINDSWAP group of the University of Maryland. By default, SWOOP ships with the Pellet reasoner, and is integrated more tightly with it than is possible via the DIG interface. It provides useful features like explanations (cf. Fig. 1), which are especially helpful for learners. At the time of writing, the last stable release of SWOOP (Version 2.2) is now more than a year old, and does not offer support for rules. However, SWOOP is in active development and far more advanced releases are available, but users are warned that they may be unstable.

Both editors are freely available for download. 149

Exploring OWL and Rules - A Simple Teaching Case 9

Figure 1 Swoop editor

4.2 Reasoners

It is currently standard practice for ontology editors to use the DIG interface to enable connection to a reasoner of the user’s choice. The following reasoners all offer a DIG interface, and can thus be combined with the editors as wished. However, let us ﬁrst note a regrettable limitation of the current DIG interface [2], in that it does not specify how to exchange and reason over rules. Thus even though both the editor and the reasoner being used provide support for rules, they are not able to be used together seamlessly via the editor’s interface. This problem is addressed in a proposal to extend the DIG interface appropriately [3], but for now other solutions must be applied, like using a ﬁle based ontology exchange.

RACER also known as RacerPro, is a commercial reasoner developed by RACER Sys- tems. It is currently available for free for educational and scientific purposes, and requires registration. Registration may take a while, so make sure to have the li- censes available before class. Although RacerPro provides almost complete support for OWL-DL it still has some limitations: (i) individuals in class expressions (so-called nominals) are only ap- proximated (although at this time RacerPro is the only optimized OWL system that supports reasoning about individuals); (ii) it cannot currently process user-defined datatype types given as external XML Schema specifications (although all required datatypes of OWL-DL are properly supported); and (iii) RacerPro 1.9 does not employ the Unique Name Assumption, required by OWL-DL, i.e. it is not possible to state that two different URIs denote the same individual. However, UNA can be enabled globally to maximise efficiency. None of these limitations are issues for the exercises of this case study. Version 1.9 of RacerPro offers support and an integrated development environment Chapter 4 Adding Rules to OWL DL Ontologies

10 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

for reasoning over ontologies enhanced with SWRL rules. KAON2 is a reasoner developed jointly by FZI, the University of Karlsruhe, and the Uni- versity of Manchester. It is available freely for scientific and academic purposes. It implements the OWL-DL standard without nominals (which is not needed within the tasks described here). It features SWRL reasoning with so called DL-safe rules, which is sufficient for the given tasks. Pellet is developed by the MINDSWAP group of the University of Maryland. It accompa- nies the SWOOP editor, and, although tightly integrated into it, can also be used as a standalone reasoner. It implements the full OWL-DL standard and also the DL-safe set of SWRL rules. Although the Pellet reasoner comes with a warning that it is not optimized for speed, the ROVE ontology does seem to be small enough to support interactive use with reasonable response times. Hoolet is an implementation of an OWL-DL reasoner developed by the University of Manchester that uses a first order prover. It consists of a graphical front end that allows loading of ontologies and rule sets, along with a reasoner. The prototype provides a useful tool but only for small examples and is for Linux only. Jess is a rule engine. [5] describes how JESS and a DIG enabled reasoner can be used together in order to reason over rule enhanced OWL-DL. Thus the shortcoming of the current DIG interface can be avoided.

5 POSSIBLE SOLUTION

After the short introduction, brief deception of some editors and reasoners, and description of the requirements and outcomes of the exercise it is time for the presentation of a possible answer to the given problem. Nevertheless before searching for the solution to the task the problem should be split into small subproblems which allow the students to better understand the approach, its goals and the final solution. In particular it means that the students should recognize the limitation of OWLs abilities and the capabilities provided by adding rules. Due to this goal, in the first step the students will build an ontology which describes the group issue as far and detailed as possible. In the second step, the open questions which could not be covered by the ontology are to be defined using rules.

5.1 Building an Ontology

The ontology requires the representation of each student and tutor taking part in the mini- project activity, the groups themselves, and attributes including: (i) for persons: nationality, associated institution, and gender,(ii) for groups: group membership (of students) and group leadership (of tutors). Some of these terms can be deﬁned as disjoint concepts within an ontology which should describe the domain of summer school taking into account the conditions described in the Section 3: Person, Country, Institution, and Group (see Fig. 2). Furthermore, Person is divided, on the one hand, into disjoint subclasses Tutor and Student partitioning the class completely and on the other hand by gender, into disjoint subclasses Man and Woman, also as a complete partition. To describe a person in the context of the summer school the person’s name, nationality, and the corresponding institution, at which the person works, needs to be deﬁned by 151

Exploring OWL and Rules-ASimple Teaching Case 11

Figure 2 ROVE-ontology building (object and data) properties hasName, hasNationality and worksAt, respectively. Furthermore, each student is a member of exactly one group (memberOf) and has exactly one favorite tutor (attractedTo). Additionally, the summer school project groups have names (hasLongName, hasShortName), are led by a tutor (ledBy) and have members (hasMember) only from the class Student. To ﬁnd the limitations of OWL, the students should ﬁrst try to formalise and implement as many requirements (cf. Section 3) as possible using the developed ontology and staying in the framework of the OWL-DL.

Condition #1 : Groups should have either 4 or 5 members

This condition can be easily implemented by setting minimum and maximum cardinalities on the property hasMember which related Groups to Students:

Group ≥ 4hasMember≤5hasMember

With the ROVE data, all the groups satisfy this condition and cardinality conditions are easily formulated within OWL.

Condition #2 : Groups should have at least one member of each gender

As seen above, to satisfy Condition #1 the relationship between a Group and each Student may be represented with the property hasMember (inverse to memberOf) and inserted cardinality restrictions. This method cannot be applied to the Condition #2 since it requires hasMember to have a minimum cardinality for values from each of the subclasses Man and Woman (so called qualiﬁed cardinality restrictions, which are not available in OWLDL, but are included in the proposal for OWL 1.1 [6]). To cope with this problem one can use existential restrictions to accomplish the task by deﬁning GoodGroup as a subclass of Group where: hasMember has (owl:)someValuesFrom Man and Chapter 4 Adding Rules to OWL DL Ontologies

12 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

(owl:)someValuesFrom Woman. However, it needs to be captured that this condition was only one of those to be satisfied by a good group (cf. Sec.3.2), so it must not be a sufficient condition for a good group. To face this issue, one could approach the problem in reverse way by specifying conditions for being a bad group (BadGroup) rather than a good groupd. In order to state a sufficient conditione for a class BadGroup a class ManGroup, which is a group with male members only, must be introduced: ManGroup ≡ Group ∀hasMember.Man The same applies for WomanGroup:

WomanGroup ≡ Group ∀hasMember.W oman

The class BadGroup is defined as a subclass of Group. Now it is simple to implement these two sufficient conditions for being a “bad group” by making WomanGroup and ManGroup subclasses of BadGroup. This means that every ManGroup and every WomanGroup is also necessarily a BadGroup, thus rendering the necessary & sufficient conditions of the subclass sufficient conditions of the superclass.

WomanGroup BadGroup ManGroup BadGroup

However, having set this up, the automatic classification will not classify one of the groups as a BadGroup with the ROVE data, although it consists of four male students. This is due to OWL’s Open World Assumption: the group, having four male members, could still potentially have a fifth member who may be female, without breaking the cardinality restriction, thus the reasoner could not classify the group as a BadGroup. There are several ways to handle this problem: one could define groups by enumeration (but not the restrictions many reasoners have regarding nominals [7]) or the size of the groups could be stated explicitly by creating two new concepts: BigGroup (group with exactly 5 members) and a SmallGroup (group with exactly 4 members) as disjoint subclasses of Group(other conditions will be asserted onto these concepts):

BigGroup ≡ Group≥5hasMember≤5hasMember SmallGroup ≡ Group≥4hasMember≤4hasMember Group ≡ SmallGroup BigGroup

Conditions #3-#5

These conditions require consideration of more than one property at a time, for example both a student’s nationality and group membership. Whilst OWL has a relatively rich set of class constructors, expressivity for properties is much weaker. Whilst OWL permits chaining of properties, it does not support making assertions about the equality of the objects at the end of two different properties/property chains. Since conditions #3-#5 require precisely this kind of assertion, it is not possible to formulate them using only OWL. From the ﬁve requirements deﬁned in Section 3 only two of them can be implemented using OWL. Here we summarize the experiences the students should have by now:

dNote that not satisfying any one of the five conditions is sufficient for classification as a bad group. eNote that Proteg´ e´ supports explicit specification of either “Necessary” or “Necessary & Sufficient” asserted conditions through the interface but no conditions which are “Sufficient” without being “Necessary”. 153

Exploring OWL and Rules-ASimple Teaching Case 13

• the requirement on each group to have 4 or 5 members can be expressed using cardinality on the hasMember property for groups

• sufﬁcient conditions for a BadGroup can be stated by deﬁning the additional concepts WomanGroup and ManGroup

• the requirements #3-#5 are not expressible in OWL, as they need property chaining.

5.2 Deﬁning Rules

As not all requirements can be implemented using OWL alone, in the next step the students investigate some solutions using rules. For this purpose the usage of the Semantic Web Rule Language (SWRL) [9] (available with a plug-in to Proteg´ e,´ or natively in SWOOP) within the ROVE ontology described above will be explored. Rules are constructed in the form of an implication between an antecedent (body) and a consequent (head): whenever the conditions speciﬁed in the antecedent hold, then the conditions speciﬁed in the consequent must also hold. SWRL provides Horn-like rules for both OWL-DL and OWL Lite, includes a high-level syntax for representing these rules[1] and is more powerful than either OWL-DL or Horn rules alone[8]. In SWRL one may assert equivalences as well as implications.

The first step towards the application of the SWRL to the group scenario is the definition of the conditions 3 and 4 which reflect that mini-project groups should have members of all different nationalities, and be from different institutions, i.e. no two members of the group should have the same nationality, or be from the same institution.

Condition #3 - groups should have members of all different nationalities

At this point it would be intuitive to come up with a definition for a group in which all members are from different nationalities - InternationalGroup (which, in turn, would be another necessary condition for a GoodGroup). Nevertheless, a correct definition of such a group will be a very challenging task since one has to define different rules for different types of groups: one rule for groups with four members, and another for those groups which have five members. This results in a high number of necessary comparisons between the nationalities of the group members. Since every pairwise combination has to be considered, the rule for 4 member groups takes into account 6 different comparisons, n−1 whilst the 5 member groups needs 10 (it is 1 n comparisons for n members, thus growing quadratically). To overcome this problem, instead of detection good groups, the students should construct a rule which determines bad groups. This means, a rule that states if a group has any two members with the same nationality it is a BadGroup,is needed (“SameNationalitiesRule”):

• Natural language: if the group has two members with the same nationality it is a BadGroup.

• Explanation of SWRL notation with natural language: If group g has member m1 and member m1 has nationality n and group g has another member m2 who is different from member m1 and member m2 also has nationality n (member m1 and member m2 have the same nationality n ) then the group g is a BadGroup. Chapter 4 Adding Rules to OWL DL Ontologies

14 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

• SWRL notation:

hasMember(?g,?m1) ∧ hasNationality(?m1, ?n)∧ hasMember(?g,?m2) ∧ hasNationality(?m2, ?n)∧ differentFrom(?m1, ?m2) → BadGroup(?g)

Condition #4 - groups should have members from all different institutions, that is, no two members of the group should be from the same institution

The same approach applies to different institution as was used for different nationalities:

hasMember(?g,?m1) ∧ worksAt(?m1, ?n)∧ hasMember(?g,?m2) ∧ worksAt(?m2, ?n)∧ differentFrom(?m1, ?m2) → BadGroup(?g)

Condition #5 : Groups should have fun

The criteria for a “fun group” was chosen in order to learn more about OWL’s and SWRL’s abilities to represent and reason with compositions of properties and situations where there were multiple properties that connected the classes. The deﬁnition is: a fun group is one where all the students in the group are attracted to the tutor leading their group. One further motive here is to provide some amusement for the students. The following rule provides a formalization of the criteria:

hasMember(?g,?m) ∧ attractedT o(?m, ?t1) ∧ hasT utor(?g,?t2)∧ differentFrom(?t1, ?t2) → BadGroup(?x) which captures that any group that has a member who is attracted to someone other than the group’s tutor is a bad group.

5.3 OWL: Not Bad does not equal Good

At this point the students may think they have captured all requirements adequately using OWL and rules, as they are now able to automatically classify all bad groups. However, it is not possible to classify good groups yet! Stating that good groups are all groups that are not bad groups relies on “negation as failure” which is not supported by OWL. Thus whilst automatic classification may be used to identify bad groups, it is not able to identify good groups simply on the grounds of not being bad. Not being classified as a bad group only indicates that the “group’s status” as a good or bad group is unknown: there is no way to specify that the list of criteria for bad groups is exhaustive. In order to classify GoodGroups as such, one needs to reformulate all the prior requirements as “positive rules”, defining sufficient conditions for classification as a GoodGroup. The positive form of the rule often becomes long and unwieldy, as discussed with the InternationalGroup. The same applies for InterInstitutionalGroupf,having another huge rule with more than 20 conjunctions in its body. Also a rule to define a FunGroup is needed. Although the intuition behind the definition of a FunGroup is easy - a group where its members all were attracted to the very tutor leading the group -

fInterInstitutionalGroup - a group where all members are from different institutions 155

Exploring OWL and Rules-ASimple Teaching Case 15 formalizing the rule again is an extremely tedious task, again leading to a huge and hard to maintain rule. Many students will ﬁnd it particularly challenging to fully understand the implications of the asymmetric structure of the positive and negative forms of rules, as intuitively they are only attempting to view the situation from the opposite side. On the other hand, a MixedGenderGroup is much more easily described (actually, this is possible in pure OWL again):

MixedGroup ≡ Group ∃hasMember.Male ∃hasMember.F emale

Finally, having formalized the requirements as rules and being able to classify whether a given group passed each rule, a GoodGroup can be deﬁned as the intersection of all groups meeting each single requirement. Here an OWL axiom can be created, using class descriptors actually deﬁned in the SWRL part of the ontology.

GoodGroup ≡ MixedGenderGroup InternationalGroup InterInstitutionalGroup FunGroup

6 PROBLEMS

In Section 4 some technical problems that may be encountered with the usage of the current tools have been already mentioned. Besides such problems – that will make the students aware of the current state of the art in Semantic Web technology – they may also run into issues like performance problems with the reasoners. This will allow the teacher to explain why some constructs or combination of constructs show undesirable properties for reasoning, and how they can be avoided. Besides these practical problems, this chapter points out an issue students have faced when understanding the conceptualization and semantics of the ROVE ontology, and trying to achieve the goals outlined in Section 3. The main problem for most students seems to be to understand the Open World Assumption and its consequences. This is easily demon- strated while solving Condition #2 above: although the student has already defined that a ManGroup is a group where all members are male, there are still male groups (with only male members) which were not classified as a ManGroup. This happened due to the fact that if we consider a group with 4 members without explicit specification that this group do not have the fifth member the reasoner assumes that the hypothetical fifth member could be female. This result is usually a surprise to the students, who by this point are likely to believe they have adequately covered all the stated conditions. In our experience, the struggling to understand the needed formalization and capture the sufficient conditions for GoodGroup turns out to be a great chance for focussed discussions and real insights into how Description Logics work. As we experienced ourselves, the tutor can easily help with a few well placed hints, and still allow the students to come to most of the solution themselves, thus deepening the learning experience.

7 RESOURCES FOR TEACHERS AND STUDENTS

This section offers a list of links with further resources regarding the topics covered in this paper. Instead of typing these links by hand, you can also go to the ROVE homepage on Chapter 4 Adding Rules to OWL DL Ontologies

16 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

http://km.aifb.uni-karlsruhe.de/projects/rove and ﬁnd an up to date electronic list of the given resources.

7.1 Reasoners

Hoolet - implementation of an OWL-DL reasoner that uses a ﬁrst order prover

• download: http://owl.man.ac.uk/hoolet/

RACER/RacerPro - the ﬁrst OWL Reasoner

• documentation: http://www.sts.tu-harburg.de/˜r.f.moeller/ racer/ • download: http://www.racer-systems.com/de/index.phtml? lang

KAON 2 - an OWL reasoner

• download: http://kaon2.semanticweb.org/ • publications: http://www.aifb.uni-karlsruhe.de/Publikationen/ showPublikationenProjekt?id_db=62 • command line tools to work with owl ontologies: http://owltools.ontoware. org

Pellet - an open-source Java based OWL-DL reasoner

• documentation: http://www.mindswap.org/2003/pellet/ • download: http://www.mindswap.org/2003/pellet/download. shtml

FaCT++ - OWL-DL reasoner

• download: http://owl.man.ac.uk/factplusplus/

7.2 Ontology Editors

Proteg´ e´ - an ontology editor and a knowledge-base editors:

• documentation: http://protege.stanford.edu/ • download: http://protege.stanford.edu/plugins/owl/index. html • tutorial: http://www.co-ode.org/resources/tutorials/ProtegeOWLT pdf

Swoop - a Hypermedia-based Featherweight OWL Ontology Editor

• download: http://www.mindswap.org/2004/SWOOP/ 157

Exploring OWL and Rules-ASimple Teaching Case 17

7.3 Language speciﬁcations

OWL - Web Ontology Language

• Documentation: http://www.w3.org/TR/owl-features/ • http://www.w3schools.com/rdf/rdf_owl.asp • http://www.cs.vu.nl/˜frankh/postscript/OntoHandbook03OWL. pdf • experiences with teaching OWL [10]

SWRL - Semantic Web Rules Language

• current proposal: http://www.w3.org/Submission/SWRL/ • XML cover pages on SWRL: http://xml.coverpages.org/ni2004-05-21- html

8 TEACHING GUIDELINES

8.1 Teaching Items

The case study can be used as supporting material in the undergraduate teaching of the basics within the Semantic Web field. The study highlights a practical situation considering the interaction between an ontology defined in the Web Ontology Language (OWL) and rules as well as the issue of the Open World Assumption. Since it is an easy to understand scenario with clear defined requirements the teaching case can be used for teaching both groups of students: students that are not familiar with Semantic Web technologies (whereas some basic knowledge and experience in building ontologies would be very helpful) as well as during the lessons with students with some previous knowledge in this field. Depending of the level of knowledge we would recommend different duration/amount of the lesson items: students with no experiences - As the students have no previous knowledge about building ontologies a breakdown of the case into two subproblems is recommended:

1. In the first teaching item the students should concentrate on building an ontology which satisfies the first two requirements (Condition #1 & #2). They should learn about the concepts, data and object properties, individuals, etc. and work with ontology editors like Proteg´ e.´ 2. In the second item, since the students have already built/worked with an ontology, the notion of rules and their development should be introduced. At this point the students can start to analyze the remaining conditions (Conditions #3- #5) and trying to formalize them using at first only OWL-DL and then the rules (in particular SRWL). The main issue within this item, apart from the rules, is the Open World Assumption. students with some experiences - If the case study is used to work with students that have already some basic (general) knowledge about ontologies or even some experiences in developing own ontologies they can, from the beginning on, analyze all given Chapter 4 Adding Rules to OWL DL Ontologies

18 M. Mochol, A. Cregan, D. Vrandeciˇ c,´ S. Bechhofer

requirements. This means the two abovementioned teaching items can be integrated into one.

After two or one teaching item respectively, the students should be able to work with OWL ontologies, deﬁne rules (in SWRL) and be able to explain the problem of the Open World Assumption. If the time allows, it is recommended to let the students formalize their own class, instead of taking the summer school ontology for the ground data. It is usually more fun to work on data the students can relate to. The fastest way to approach this, is to allow each student, or group of student, to formalize their own data, and then merge the data.

8.2 System Requirements

Most of the tools are written in Java and can thus run on several different operating systems. It is suggested to test the chosen applications with the local installation of Java, as some of the suggested tools require Java 5. Currently available Semantic Web tools have usually been developed within research institutions, and have not been not optimized with regards to performance. Reasoners especially may make heavy claims on both memory and processor time. Also some advanced features of the ontology development environments, like the explanation module shown in Figure 1 may require considerable resources. For these reasons, the system requirements and response times will depend heavily on the chosen tools, and the size of the ﬁnal ontology. Our experience with these exercises has shown that most commonly available hardware can deal adequately with ontologies of the size used in the examples presented here (i.e. less than 100 instances and only a small number of classes and properties). However, it is recommended that the teacher run some preliminary tests using the local hardware and chosen tools prior to running tutorials, to ensure that the tools will work in the local environment with reasonable response times.

9 CONCLUSION

This paper gave a brief overview of a Semantic Web teaching case presenting, by means of an easy-to-understand example, some limitations of OWL-DL and ﬁrst steps in using SWRL rules. To facilitate practical exercises some currently available ontology editors and additional Semantic Web tools, which may be used to construct and work with ontologies and rules to complete the described task, were described. The experiences with ontology development and the usage of different tools enabled the authors to highlight some commonly-experienced problems and limitations of the available technologies. In the context of the easy summer school scenario the difﬁcult issue of the “Open World As- sumption” was pointed in a way which should be readily accessible to students and may encourage their interest in the underlying formalism. In conclusion, the authors hopes that this teaching case will serve to assist “young players” and their guides in the practical navigation of Semantic Web Technologies whilst avoiding some common traps and pitfalls. 159

Exploring OWL and Rules-ASimple Teaching Case 19

Acknowledgements

We would like to thank Antoine Zimmermann, who was a member of the original ROVE mini-project team at SSSW2005; the organizers of the SSSW2005: Enrico Motta and Asuncion Gomez-Perez; and all the students, tutors and speakers who took part. Research reported in this paper has been partially ﬁnanced by NICTA (http://www. nicta.com.au), the EU project SEKT (http://www.sekt-project.com), the NoE KnowledgeWeb (http://knowledgeweb.semanticweb.org), and the Ger- man Ministry of Research (BMBF) project Knowledge Nets (http://wissensnetze. ag-nbi.de) which is a part of the InterVal (http://interval.hu-berlin.de) Berlin Research Centre for the Internet Economy (http://www.internetoekonomie. net).

References and Notes

1 R. Aggarwal. Semantic Web Services Languages and Technologies: Comparison and Discussion. LSDIS Lab, University of Georgia, 2004. http://citeseer.ist.psu.edu/700682. html. 2 S. Bechhofer, I. Horrocks, P. F. Patel-Schneider, and S. Tessaris. A proposal for a Description Logic interface. In Proc. of the Description Logic Workshop (DL’99), pages 33–36, 1999. CEUR Workshop Proceedings http://ceur-ws.org/Vol-22/. 3 S. Bechhofer, T. Liebig, M. Luther, O. Noppens, P.F. Patel-Schneider, B. Suntisrivaraporn, A.- Y. Turhan, and T. Weithoner.¨ DIG 2.0 – Towards a Flexible Interface for Description Logic Reasoners. In B. Cuenca Grau, P. Hitzler, C. Shankey, and E. Wallace, editors, Proc. of the 2nd Workshop OWL Experiences and Directions 2006, 2006. 4 A. Cregan, M. Mochol, D. Vrandecic, and S. Bechhofer. Pushing the limits of OWL, rules and Protege - a simple example. In B. Cuenca Grau, I. Horrocks, B. Parsia, and P. Patel-Schneider, editors, OWL: Experiences and Directions, Galway, Ireland, 2005. 5 C. Golbreich and I. Atsutoshi. Combining SWRL rules and OWL ontologies with Protege OWL plugin, Jess, and Racer. In R. Fergerson and N. Noy, editors, Proceedings of the 7th International Protege Conference, 2004. 6 B. Cuenca Grau et al. The OWL 1.1 extension to the W3C OWL Web Ontology Language. University of Manchester, 2006. 7 J. Hladik. Reasoning about nominals with FaCT and RACER. In Proc. of the 2003 International Workshop on Description Logics (DL2003), CEUR-WS, 2003. 8 I. Horrocks, P. F. Patel-Schneider, S. Bechhofer, and D. Tsarkov. OWL Rules: A Proposal and Prototype Implementation. Conditionally accepted for publication. Journal of Web Semantics, 2005. 9 I. Horrocks, P.F. Patel-Schneider, H. Boley, S. Tabet, B. Grosof, and M.Dean. SWRL: A Semantic Web Rule Language Combining OWL and RuleML. DARPA DAML Program, 2003. http: //www.w3.org/Submission/2004/SUBM-SWRL-20040521/. 10 A. Rector, N. Drummond, M. Horridge, J. Rogers, H. Knublauch, R. Stevens, H. Wang, and C. Wroe. OWL Pizzas: Practical Experience of Teaching OWL-DL: Common Errors & Common Patterns. In E. Motta, N. R. Shadbolt, and A. Stutt, editors, Proc. of the 14th EKAW, volume 3257 of LNCS, pages 63–81. Springer, 2004.

5 Controlled Natural Language Syntaxes for OWL

5.1 Sydney OWL Syntax

Title of Publication: Sydney OWL Syntax - towards a Controlled Natural Language Syntax for OWL 1.1 Type of Publication: Workshop Paper Appears In: C. Golbreich, A. Kalyanpur, B. Parsia, editors, Proceedings of the OWLED 2007 Workshop on OWL: Experiences and Directions, Innsbruck, Austria, 6-7, June 2007, CEUR-WS, Vol. 258, 2007. Publication Date: 2007 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Rolf Schwitter Thomas Meyer Personal Contribution: 80% (estimated by co-authors)

161 Chapter 5 Controlled Natural Language Syntaxes for OWL 163 Sydney OWL Syntax - towards a Controlled Natural Language Syntax for OWL 1.1

Anne Cregan1,2, Rolf Schwitter3, and Thomas Meyer1,2

1 NICTA, [Anne.Cregan,Thomas.Meyer]@nicta.com.au 2 University of New South Wales, Australia 3 Macquarie University, [email protected]

Abstract. This paper describes a proposed new syntax that can be used to write and read OWL ontologies in Controlled Natural Lan- guage (CNL): a well-defined subset of the English language. Following the lead of Manchester OWL Syntax in making OWL more accessible for non-logicians, and building on the previous success of Schwitter’s PENG (Processable English), the proposed Sydney OWL Syntax enables two-way translation and generation of grammatically correct full English sentences to and from OWL 1.1 functional syntax. Used in conjunction with OWL tools, it is designed to facilitate ontology construction and editing by enabling authors to write an OWL ontology in a defined subset of English. It also improves readability and understanding of OWL statements or whole ontologies, by enabling them to be read as En- glish sentences. It is hoped that by providing the option of an intuitive, easy to use English syntax which requires no specialized knowledge, the broader community will be far more likely to develop and benefit from Semantic Web applications. This paper is a discussion paper covering the scope, design, and examples of Sydney OWL Syntax in use, and the authors invite feedback on all aspects of the proposal via email to [email protected]. Working drafts of the full specification are available at http://www.ics.mq.edu.au/~rolfs/sos.

1 Introduction Following OWL reaching offical W3C recommendation status, a variety of notations for OWL class, property and individual descriptions and axioms became available through various tools, most notably Protégé[8] and SWOOP [6]. As noted in [3], these ranged from the officially recommended RDF/XML exchange syntax [2], through to a Description Logic style syntax, with Turtle/N-Triples [1] and the OWL Abstract Syntax [9] somewhere between the two extremes of ver- bosity and the specialized logical notation known as “squiggles” to non-logicians. The experience of experts such as the Manchester Group in delivering OWL tutorials and workshops for domain experts identified that for the vast majority of non-logicians, none of the existing OWL syntaxes were suitable for writing class expressions and other types of axioms: they were either too verbose, or else the logical notation was intimidating and inconvenient to use. Manchester OWL Syntax [3] addressed these problems by providing an alternative syntax designed to be concise, without DL symbols, and quick and easy to read and write. Chapter 5 Controlled Natural Language Syntaxes for OWL

Manchester OWL Syntax has had substantial success and is reported to be the preferred syntax for non-logicians [3]. Discussion following its presentation at OWLED 2006 identified the potential, as a future goal for OWL, to extend the approach even further, to provide a syntax representing OWL in full English sentences. With a view to building on the previous success of Schwitter’s Pro- cessable English (PENG) [11], which translates Controlled English to first-order logic, Cregan formed a small working group comprising the three Sydney-based authors (Cregan, Schwitter and Meyer) to design such a syntax. The resulting proposed Sydney OWL Syntax is presented herein, and feedback is invited from all interested parties. As there are many design decisions to be made in such an undertaking, a large part of the paper is devoted to covering the design choices identified, and giving the rationale for the choices made. The syntax itself is presented via examples throughout the paper, but as space does not permit the inclusion of the emerging specification, readers should also consult the documentation available at http://www.ics.mq.edu.au/~rolfs/sos. 2 Background 2.1 Manchester OWL Syntax Manchester OWL Syntax [3] is largely based on the German DL syntax and shares its compactness. Its key differentiating features are the replacement of special logical symbols such as ∃, ∀ and ¬ with the more intuitive keywords some, only, and not; the use of infix rather than prefix notation for keywords used in restrictions, preventing a misreading of class expressions found to be common amongst non-logicians; and the introduction of keywords such as ValuePar- tition facilitating common ontology design patterns. Manchester OWL Syntax has been reported to be well-received by non-logicians [3] and is the default syntax for Protégé-OWL and the commercially released OWL ontology editor TopBraid Composer 1. In general, non-logicians have found it easier to grasp, remember and use than DL syntax. Although needing some training to re-align their natural interpretation of keywords to the correct OWL/DL interpretation, it successfully lowered the barrier for reading and interpreting ontologies. Limitations: Although able to represent complete ontologies, Manchester OWL Syntax has been primarily designed for presenting and editing class expressions via tools, and representation / tool support for property and individual expressions seems to have had less focus. In addition, whilst certainly lowering the barrier, a syntax closer to English, with semantics matching a natural English interpretation could potentially remove it altogether.

2.2 PENG (Processable ENGlish) PENG (Processable ENGlish) [11] is a machine-oriented controlled natural language (CNL) designed for writing unambiguous and precise speciﬁcation texts for knowledge representation. Whilst easily understood by speakers of the base language, it has the same formal properties as an underlying formal logic language

1 http://www.topbraidcomposer.com/ 165 and thus is machine-processable. It can be used, for example, for annotating web pages with machine-processable information [12]. PENG covers a strict subset of standard English, and is precisely defined by a controlled grammar and lexicon. Specification texts written in PENG are incrementally parsed using a unification-based phrase structure grammar, and translated into first-order logic via discourse representation structures [7]. Standard first-order logic reasoning services are applied for reasoning tasks including consistency and informativity checking, and question answering. As a brief example, the following sentences are written in PENG: 1. If X is a research programmer then X is a programmer. 2. Bill Smith is a research programmer who works at the CLT. 3. Who is a programmer and works at the CLT? Sentence (1) describes a subclass relationship, sentence (2) asserts factual knowledge about a domain, and sentence (3) is used to query the terminological and factual knowledge expressed in (1) and (2). Standard first-order logic (FOL) query processing returns the answer Bill Smith. The writing process of PENG is facilitated by predictive interface techniques: after the author enters a word form, the authoring tool displays look-ahead information indicating the available choices for the next word form, ensuring adherence to the lexicon and grammar. The author does not need to learn or remember the rules of the controlled natural language as these are taken care of by the authoring tool. Limitations: The grammar of PENG is first-order equivalent and therefore more expressive than OWL 1.1. It is informed by FOL rather than DL considerations. In addition, the grammar has not been designed with bidirectionality in mind: PENG sentences are translated into FOL but not from FOL backwards into PENG. For these reasons, Sydney OWL Syntax, whilst informed by the learnings and experience of PENG, has essentially been designed from scratch. With regard to bidirectionality, Kaljurand and Fuchs [4] have presented a bidirectional mapping between a subset of OWL DL and Attempto Controlled English using a discourse representation structure as interlingua, but in recent work [5] they focus on one direction only: the verbalisation of OWL DL. Schwit- ter and Tilbrook [13] previously showed that there is no need for an interlingua and that bidirectionality can be achieved in a direct way using axiom schemas.

3 Scope

Sydney OWL Syntax has been scoped as follows:

1. OWL 1.1 compatible Unlike the 2004 OWL recommendation which uses a frame-like syntax convenient for manipulating ontologies by hand: ObjectProperty(hasAncestor domain(person) range(person)) the emerging OWL 1.1 [10] has a functional-style syntax which breaks such axioms apart and makes them easier to manipulate programmatically: Chapter 5 Controlled Natural Language Syntaxes for OWL

ObjectPropertyDomain(hasAncestor person) ObjectPropertyRange(hasAncestor person) Sydney OWL Syntax takes OWL 1.1 functional syntax as the normative form for expressing OWL ontologies and the base form for translations. Combining readability and processability, it expresses the same information as: If X has Y as an ancestor then X is a person. If X has Y as an ancestor then Y is a person. See Section 6 for considerations of conciseness in the design. 2. Coverage of the entire OWL language Anything that can be expressed in OWL 1.1 may be expressed in Sydney OWL Syntax. It provides complete coverage of all axioms and assertions that may be made in OWL 1.1, for example subproperty relations: If X has Y as a parent then X has Y as an ancestor. and property chains (= role composition): If X owns Y and Y has Z as a part then X owns Z. 3. Two-way translation Any OWL 1.1 ontology may be represented in Sydney OWL Syntax and conversely, ontologies constructed in Sydney OWL Syntax can be fully represented in any other OWL 1.1 syntax, without loss of information. The writing of ontologies in Sydney OWL Syntax is to be supported with interactive functionality such as look-ahead information, to assist the user and enforce syntactic validity. 4 Design goals The key design goals of Sydney OWL Syntax are: 1. Support non-logicians to build quality OWL ontologies Support domain experts and analysts, particularly those without a logical background, to write good quality OWL ontologies. We assume that users are literate in English, and have at least an average ability to use a computer and think and express themselves logically in the normal sense of the word, but no specific knowledge of any formal notation is assumed. 2. Provide English translations of OWL ontologies Provide English translations of OWL ontologies which can be read and understood by English-speaking persons, without the need to refer to any other ontology syntax or representation. As with any ontology syntax, it is the re- sponsibility of the author(s) to choose sensible and appropriate names for user-defined classes and properties. 3. Modularity for future flavours of OWL As OWL is an evolving language, and it is likely that new flavours of OWL corresponding to various formal logics will emerge, one of the design goals is a modular approach which facilitates contracting and expanding the syntax in correspondence with the logical operators to be included. For instance, words such as must, may and cannot are not currently used, as they correspond to 167

notions of permissibility and obligatoriness used by deontic logics. At some stage OWL may have a flavour based on a deontic logic, so these words are kept in reserve for that scenario. 4. Implementable by OWL tools Provide a specification which is sufficiently detailed and precise for implementation in ontology tools, as an alternative syntax to Manchester Syntax, OWL Abstract Syntax, and/or the other existing syntaxes. We note however that as OWL 1.1 functional syntax is not fully backwards-compatible with previous OWL syntaxes, the same applies for Sydney OWL Syntax. 5 Design choices Whilst developing the syntax, several design decisions were encountered and choices made. We believe these decisions are ones which would be encountered by any effort to translate between a formal and a controlled natural language, and give the rationale for the choices made for Sydney OWL Syntax.

5.1 Naturalness versus closeness to OWL A key decision was how natural we wanted the language to be. We observed a fundamental tradeoﬀ between naturalness and closeness to OWL: on the one hand, the language could be more natural, but would lose its binding to OWL and thus become ambiguous and open to interpretation. This would seem to defeat the purpose of building an ontology as it is expressly for explicit logical representation of a domain. On the other hand, one can bind very tightly to OWL but this can result in some unnatural sounding English expressions, as there is often no exact or at least succinct equivalent in English for an OWL construction. For example, hasFather is a FunctionalObjectProperty. does not sound like a natural English expression, as ﬁrstly, it is an artefact of the ontology itself, and secondly, it uses abstract terms that are unknown to non- specialists. In contrast, Sydney OWL Syntax uses the terms of the application domain to convey the meaning without the need for any opaque encoding: If X has Y as a father then Y is the only father of X. In general we have opted towards tight binding with OWL 1.1 functional syntax whilst endeavouring to make the expressions as natural as possible.

1. One or many CNL translations? In natural language there are many ways to say the same thing - did we want to try to support all or a collection of them in translating a given OWL statement? For example, for the expression SubClassOf(male, person) should we support both If X is a male then X is a person and Every male is a person or allow one and only one CNL representation? We decided that for the ﬁrst cut of Sydney OWL Syntax, there would be only one. Design choice: OWL syntax corresponds uniquely to Sydney OWL Syntax. That is, there is only one Sydney OWL Syntax form for each OWL form, Chapter 5 Controlled Natural Language Syntaxes for OWL

chosen to maximise succinctness and precision.

However, we appreciate the potential usefulness of supporting different modes of natural language expression for ontology construction purposes. For example, we have chosen to represent disjointness with the succinct mutually exclusive, but for clarification purposes it may be helpful to offer an expanded CNL translation such as female or male is the case but not female and male. One option is that such modes could be handled through the interface without formally being part of the syntax. We also note that uniqueness of form refers to the syntactical form of the OWL statement, not its logical status - in some cases two OWL expressions may be logically equivalent but use different syntax, for example: DisjointUnion(person male female) and DisjointClasses(male female) EquivalentClasses(person ObjectUnion(male female))

In this case, each syntactically distinct OWL expression corresponds to its own Sydney OWL Syntax equivalent: The class person is equivalent to male or female, and male and female are mutually exclusive. and The classes male and female are mutually exclusive. The class person is fully defined as anything that is a male or a female.

2. How explicit should the OWL constructs be? One of the fundamental design decisions we faced was whether or not to talk about the ontology constructs themselves in the syntax. For instance, would a class axiom like subClassOf(male, person) translate to something like There is a class called male which is a subset of a class called person or to a statement about the domain itself, like Every male is a person? In the latter case, some OWL statements, such as class male, would have no translation at all in CNL as they as artefacts of the modelling process and don’t assert anything about the domain itself. Design choice: We opted to have limited explicit references to OWL constructs like classes and properties. As a consequence, some OWL axioms are not translated at all, but the knowledge is captured in Sydney OWL Syntax implicitly. Using a parsing process, any implicit concept in Sydney OWL Syntax can be unpacked into a corresponding OWL axiom. E.g., the translation of Every male is a person back to OWL produces a class male declaration and a subset axiom. Overall, all information from OWL is captured in Sydney OWL Syntax and given a Sydney OWL Syntax translation, the original OWL statements can be regenerated.

3. Correspondence between OWL constructs and CNL constructs To facilitate modularity in respect of the addition or removal of logical op- 169

erators and constructions, we have carefully chosen grammar and lexicon to correspond tightly with the underlying logic, the aim being to implement as much modularity as possible within the boundaries of using natural grammar. For instance, the word only in If X has Y as a son then Y is the son of only X. is reserved for use in expressing functional or inverse functional properties, and not in any other context. By virtue of this tight binding, a person with familiarity with both OWL and Sydney OWL Syntax can read an ontology represented in Sydney OWL Syntax and recognise the OWL constructs via the words and phrases used. Design choice: Where possible, each OWL construct has its own distinct natural language keyword or phrase.

4. Use of linguistic and other background knowledge Anaphoric reference: In natural language, it is common to refer to concepts introduced in previous statements via pronouns and deﬁnite noun phrases to refer to previously introduced entities. Note that to use the pronoun “he” requires previous knowledge of the referent being male. In an OWL context, such references require logical processing of other statements. In OWL ontologies statements are not necessarily in any order, so the entire ontology would need to be parsed. Number agreement: Linguistic background knowledge is also commonly used in natural language. For instance, the knowledge that the correct plural of mouse is mice is necessary to refer to Three blind mice instead of Three blind mouses. Adding an “s” to create plurals is a useful default rule but not always correct. However, we can use morphological rules and a list of exceptions as best approximation. Design choice: Each OWL statement is translated as a unit, without reference to any other statement in the ontology, or any other background or linguistic knowledge. Processing and using knowledge from outside the OWL statement vastly compounds the complexity of processing, thus has been avoided at the expense of providing anaphoric reference and safe number agreement. 5. Use of variables Whilst not a design preference, we found that some OWL statements could not be expressed clearly in CNL without using variables. If you need convincing, try as a test to express succinctly and unambiguously in English without using variables, the example involving role composition in section 3: If X owns Y and Y has Z as a part then X owns Z. One option we considered is to use phrases such as “something” and “something else” as pseudo-variables. But then one ends up with howlers such as the following: If something owns something else and that something has another thing as a part then the original something owns that something. Design choice: We decided to minimise the use of variables but found it impossible to do without them completely. Chapter 5 Controlled Natural Language Syntaxes for OWL

5.2 Complex constructs Design choice: Complex class deﬁnitions are supported through an approach which supports nesting of expressions to any level. We plan to support expressions which use nesting up to three levels, for example: The class old lady is partly defined as anything that has only cats as a pet and has some animal as a pet or has only gardeners as a lover.

5.3 Extra language support for user-deﬁned terms In building ontologies it is very common to use has and is combined with some other word or phrase when naming properties, e.g. hasAge; isMotherOf etc. Design choice: Sydney OWL Syntax supports special processing of property names for has and is and their grammatical variants, providing camel case is used e.g. isMotherOf not ismotherof. This provides a more natural translation.

5.4 Definitions Constructing correct definitions is challenging in OWL, since authors often fail to make a definition complete rather than partial. To address this problem we use the two markers fully defined as and partly defined as to indicate the logical status. For example, the following statement:

The class adult is fully defined as any person that has at least 20 as an age.

claims that the concept adult is fully deﬁned by a set of necessary and suﬃcient conditions. The translation of this statement results in the subsequent functional- style syntax representation:

EquivalentClasses(adult ObjectIntersectionOf(Person DataAllValuesFrom(hasAge DatatypeRestriction(Datatype(xsd:nonNegativeInteger) owl:minInclusive "20"^^xsd:int))))

6 Design consequences 6.1 Tight binding to functional-style syntax

In general, the OWL 1.1 functional syntax requires more statements to express the same thing than the previous frame-like notation. The consequence of tight binding to the former is that Sydney OWL Syntax also has more statements. For instance, using the original OWL frame-like notation as a starting point, it would have been easier to translate the example given in Section 3 into one sentence rather than two: If X has Y as an ancestor then X is a person and Y is a person. 171

6.2 Bidirectionality and context sensitive grammar Sydney OWL Syntax is bidirectional, thus each statement translates into OWL functional-style syntax and vice versa, with the exception of statements of explicit OWL constructs, which have no Sydney OWL Syntax translation. An elegant way to achieve bidirectionality is to use a definite clause grammar and generate the output format during the parsing process [13]. In general, bidirectional translation requires a context-sensitive grammar. This may be illustrated as follows: If X has Y as a parent then Y has X as a child. expresses an inverse relationship between two properties. Note that in the antecedent, the grammar needs to store the variable X in the subject position and the variable Y in the object position, whereas in the consequent their positions must be switched, otherwise we have a subproperty relationship. Additionally, the grammar has to provide a mechanism to absorb the auxiliary verb has and the prepositional objects parent and child into an OWL property name. To achieve bidirectionality, Sydney OWL Syntax will use a context-sensitive grammar which can store the required elements and employ an axiom schema which is instantiated during parsing: InverseObjectProperties(Prefix1:Property1 Prefix2:Property2) For this example the schema looks as follows after parsing: InverseObjectProperties(a:[has,parent],a:[has,child]) and this output can easily be transformed into the final format: InverseObjectProperties(a:hasParent a:hasChild). In the ideal case the same grammar should accept this output and generate the original input sentence. 7 Conclusion and future work Above we have set out the scope, design goals, decisions and choices informing the emerging Sydney OWL Syntax specification. The authors invite feedback on every aspect of the proposal, via email to [email protected]. In parallel with collecting feedback from interested parties and moving towards a stable specification, we plan to start work on a demonstrator. In conclusion we note some features we envisage for tool interfaces. The writing of an ontology in Sydney Syntax is to be supported by a predictive text editor which generates look-ahead information while a specification text is written [12, 14]. Thus the user does not need to learn the rules of the Sydney Syntax explicitly, since the writing process is guided by the text editor. Such a text editor will be able to be used either in TBox mode, to express terminological axioms, or in ABox mode, to assert factual information about a specific domain. Once a set of terminological axioms has been specified, the resulting user-defined terminology can be used in ABox mode to specify instance data. From the terminological information available in the ontology, the text editor becomes “ontology-aware”, harvesting TBox input to generate new lookahead information guiding the writing process in ABox mode. Acknowledgements Research reported in this paper has been partially financed by the Macquarie Univer- sity Centre for Language Technology (http://www.clt.mq.edu.au). We thank Phillip Chapter 5 Controlled Natural Language Syntaxes for OWL

Quinn and Matthew Horridge for their assistance in producing examples of OWL 1.1 functional syntax, and the members of the [email protected] mailing list for their contributions. NICTA is funded by the Australia Government’s Department of Com- munications, Information and Technology and the Arts and the Australian Research Council through Backing Australia’s Ability and the ICT Centre of Excellence program. It is supported by its members the Australian National University, University of NSW, ACT Government, NSW Government and aﬃliate partner University of Sydney. References

1. D. Beckett. New syntaxes for rdf. Technical Report, 2004. Institute for Learning and Research Technology, Bristol. 2. D. Beckett. Rdf/xml syntax specification (revised). W3C Recommendation 10 February, 2004. At http://www.w3.org/TR/2004/REC-rdf-syntax-grammar- 20040210/. 3. M. Horridge, N. Drummond, J. Goodwin, A. Rector, R. Stevens, and H. H. Wang. The manchester owl syntax. In Proc. of the 2006 OWL Experi- ences and Directions Workshop (OWL-ED2006), 2006. available at http://owl- workshop.man.ac.uk/acceptedLong/. 4. K. Kaljurand and N. E. Fuchs. Bidirectional mapping between owl dl and attempto controlled english. In LNCS 4187, pages 179–189, 2006. 5. K. Kaljurand and N. E. Fuchs. Verbalizing OWL in Attempto Controlled English. In Proceedings of OWLED07, 2007. 6. A. Kalyanpur, B. Parsia, B. Cuenca-Grau, and J. Hendler. Swoop: A ’web’ ontology editing browser. Journal of Web Semantics, (4(2)), 2005. 7. H. Kamp and U. Reyle. From Discourse to Logic. Dordrecht: Kluwer, 1993. 8. N. Noy, R. Fergerson, and M. Musen. The knowledge model of Protégé-2000: Combining interoperability and flexibility. In R. Dieng and O. Corby, editors, Proc. of the 12th EKAW, volume 1937 of LNAI, pages 17–32, Juan-les-Pins, France, 2000. Springer. 9. P. F. Patel-Schneider, P. Hayes, and I. Horrocks. Owl web ontology language, semantics and abstract syntax, 2004. W3C Recommendation 10 February 2004, available at http://www.w3.org/TR/owl-semantics/. 10. P. F. Patel-Schneider and I. Horrocks. Owl 1.1 web ontology language overview, 2006. 11. R. Schwitter. English as a formal specification language. In Proceedings of the Thirteenth International Workshop on Database and Expert Systems Applications (DEXA 2002), pages 228–232, 2002. 12. R. Schwitter and M. Tilbrook. Annotating websites with machine-processable information in controlled natural language. In Advances in Ontologies 2006, Pro- ceedings of the Second Australasian Ontology Workshop (AOW 2006), pages 75–84, 2006. 13. R. Schwitter and M. Tilbrook. Let’s talk in description logic via controlled natural language. In Proceedings of the Third International Workshop on Logic and Engineering of Natural Language Semantics (LENLS2006), pages 193–207, 2006. 14. C. W. Thompson, P. Pazandak, and H. R. Tennant. Talk to your semantic web. volume 9, pages 75–79, 2005. 173

5.2 A Comparison of three CNLs for OWL 1.1

Title of Publication: A Comparison of three Controlled Natural Languages for OWL 1.1 Type of Publication: Workshop Paper Appears In: Kendall Clark and Peter F. Patel-Schneider, editors, Proceedings of OWL: Experiences and Directions (OWLED 2008 DC), 4th International Workshop, Washington, USA, 1-2 April, 2008 Publication Date: 2008 Peer Reviewed: Yes Contributing Author(s): Rolf Schwitter Kaarel Kaljurand Anne Cregan Catherine Dolbear Glen Hart Personal Contribution: 32.5% (average of estimates by co-authors) Chapter 5 Controlled Natural Language Syntaxes for OWL 175

A Comparison of three Controlled Natural Languages for OWL 1.1

Rolf Schwitter1, Kaarel Kaljurand2, Anne Cregan3, Catherine Dolbear4,and Glen Hart4

1 Macquarie University & NICTA, Australia – [email protected] 2 University of Zurich, Switzerland – [email protected] 3 NICTA, Australia – [email protected] 4 Ordnance Survey, Southampton, UK – {Catherine.Dolbear|Glen.Hart}@ordnancesurvey.co.uk

Abstract. At OWLED2007 a task force was formed to work towards a common Controlled Natural Language Syntax for OWL 1.1. In this paper members of the task force compare three controlled natural languages (CNLs) — Attempto Controlled English (ACE), Ordnance Survey Rab- bit (Rabbit), and Sydney OWL Syntax (SOS) — that have been designed to express the logical content of OWL 1.1 ontologies. The common goal of these three languages is to make OWL ontologies accessible to people with no training in formal logics. We brieﬂy introduce these three CNLs and discuss a number of requirements to an OWL-compatible CNL that have emerged from the present work. We then summarise the similarities and diﬀerences of the three CNLs and make some preliminary recommendations to an OWL-compatible CNL.

1 Introduction

The mathematical nature of description logics makes it difficult for non-logicians such as domain experts to understand and author OWL-based ontologies. This forms a significant impediment to ontology creation and reuse. If domain experts’ knowledge is to be represented and verified, an easily understandable syntax for writing ontologies is needed. [22] list the problems that users encounter when working with OWL DL and identifies the need for a ‘pedantic but explicit’ paraphrase language. This need was partially met by Manchester syntax [14], which paraphrased the logical symbols with English glosses and improved domain experts’ understanding and ability to author ontologies. In 2007 three new offerings appeared that enabled OWL ontologies to be rendered in English paraphrases: Attempto Controlled English (ACE), Ordnance Survey’s Rabbit (Rabbit), and Sydney OWL Syntax (SOS). The purpose of such OWL syntaxes is not to replace the graphical user interface generally used for ontology building, although these syntaxes can be used in this way if a text-based approach is desired. Instead, they complement the GUI by enabling the author (= domain specialist or knowledge engineer) to Chapter 5 Controlled Natural Language Syntaxes for OWL

understand and write the most appropriate axioms, as well as providing a means to output the built ontology as a readable piece of text for sharing with others interested in the domain knowledge that the ontology captures. With three new Controlled Natural Language (CNL) syntax alternatives represented at OWLED2007, it was decided to create a task force including members from each eﬀort for the purpose of comparing these approaches and working towards a common Controlled Natural Language. This paper is written by key members of the task force. It compares the ACE, Rabbit and SOS controlled English syntaxes for OWL 1.1 using concrete examples, discusses similarities and diﬀerences between the renderings, and makes some initial recommendations.

2 Controlled Natural Languages for the Semantic Web

A controlled natural language is an engineered subset of a natural language with explicit constraints on grammar, lexicon, and style. These constraints usually have the form of writing rules and help to reduce both ambiguity and complexity of full natural language [18]. Over the last decade, a number of controlled natural languages have been designed and used for writing software speciﬁcations, for supporting the knowledge acquisition process, and for knowledge representation — among them Attempto Controlled English [8], PENG Processable English [23], Common Logic Con- trolled English [25], and Boeing’s Computer-Processable Language [3]. Since the early days of the Semantic Web, simple teaching languages (for example Notation 3) have been used that are equivalent to RDF in its XML syntax, but easier to ‘scribble’ when getting started [2]. There are other languages [13,20,15] that have been suggested in order to represent OWL in a more natural way. However, the major shortcoming of these approaches is that they lack any formal check that the resulting expressions are unambiguous. In this sense, a better approach is based on controlled natural languages that typically have a formal language semantics and come with a parser that can convert the statements into the OWL representation so that the natural language version becomes the primary human interpretable representation. ACE, Rabbit, and SOS are three controlled natural languages that have been designed to be used as interface languages to OWL ontologies. Apart from these languages there exist other CNL-based approaches to authoring OWL ontologies [1,9] but we will not further discuss these languages.

2.1 ACE, Rabbit, and SOS

ACE is a subset of English designed to provide domain specialists with an expressive knowledge representation language that is easy to learn, read and write [7]. ACE is deﬁned by a small number of construction rules that deﬁne its syntax and a small number of interpretation rules that disambiguate constructs that in full English might be ambiguous. 177

In [16], a bidirectional mapping of a fragment of ACE to OWL 1.1 (without data properties) is described. This mapping captures all semantically diﬀerent OWL constructs as diﬀerent ACE sentences, but often there are many possibilities for expressing the same OWL axiom. For example, all sentences in

John likes no man that owns a car. No man that owns a car is liked by John. Every man that owns a car is not liked by John. If a man owns a car then it is false that John likes the man.

map to the same OWL SubClassOf -axiom. On the other hand, the mapping does not differentiate between all syntactic forms that OWL offers, i.e. syntactically different OWL constructs can end up the same in ACE (given that they are semantically equivalent). This mapping has been fully implemented and is being used in experimental ontology editors ACE View [16] and AceWiki [19].

Rabbit is a controlled natural language developed by Ordnance Survey with the help of domain experts for the purpose of authoring ontologies [11]. It has so far been used by domain experts to develop two medium-scale ontologies containing about 600 concepts for ‘Buildings and Places’ and Hydrology, using most of the expressivity of OWL 1.1 (namely ALCOQ and SHOIQ respectively for the ontologies). This practical implementation experience has enabled Ordnance Survey to tailor the design of the CNL, concentrating on those constructs and models of knowledge that are frequently required by ontology authors, or where the authors most commonly make errors. Rabbit was developed as part of a wider methodology for authoring ontologies using a domain expert-centric approach [12]. A Protégé 4 plugin is currently being developed in cooperation with the University of Leeds [4] to implement the Ordnance Survey methodology. This allows domain experts to author ontologies in Rabbit. The GATE natural language processing tool [6] is being used to implement a backend to the tool to convert Rabbit into OWL. The fundamental principles underlying the design of Rabbit are: (a) to allow the domain expert, with the aid of a knowledge engineer, to express their knowledge as easily and simply as possible and in as much detail as necessary; (b) to have a well defined grammar and be sufficiently formal to enable those aspects that can be expressed as OWL to be systematically translatable and to enable other non-DL based applications to access this knowledge.

SOS is a controlled natural language that has been designed from scratch to fulﬁll the requirements of a modern high-level interface language to OWL 1.1 [5]. The key design goals are: (a) supporting non-logicians to write OWL ontologies in a well-deﬁned subset of English, and (b) expressing existing ontologies in the same subset of English. SOS uses the terms of the application domain plus some other terms to convey the meaning of the information. SOS enforces a one-to- one mapping between controlled natural language and OWL Functional-Style Chapter 5 Controlled Natural Language Syntaxes for OWL

Syntax (FSS). That means SOS does not allow to say the same thing in different ways. Furthermore, the language uses only limited references to OWL constructs like classes and properties. SOS uses only very little linguistic knowledge in order to deal with plural forms (e.g. ‘confluences’) and compound constructions (e.g. ‘has ... as a part’). A particularly interesting feature of SOS is the use of variables – as known from high school math textbooks – which enables the expression of certain axioms in a very compact and natural way. To support the writing of definitions, the language provides specific constructs (‘fully defined as’ and ‘partly defined as’) that indicate the logical status of a definition. In principle, SOS supports nesting of expressions to any level but deep nesting results in structures which are difficult to understand by people. Therefore, it is recommended that authors limit the depth of nesting up to three levels using an authoring tool (similiar to [24]). In order to achieve bidirectional translations between SOS and OWL FSS, experiments were conducted with logic programming techniques which allow us to generate formulas in OWL FSS notation during the parsing process.

2.2 Requirements, design choices, scope

Several requirements and design choices for an OWL-compatible controlled natural language (CNL OWL) have emerged from the work done on ACE, Rabbit, and SOS. A controlled natural language for OWL should offer authors who write, modify or view OWL ontologies an improved usability over the existing OWL syntaxes. This improved usability is gained by defining a fragment of English and its precise mapping into OWL in such a way that the mapping preserves the intended meaning of the English constructs. There are two main requirements that are in slight conflict with each other — the need to see OWL as a fragment of English (semantically), and the need to cope with OWL’s design in order to make a straightforward mapping to and from OWL possible. Firstly, we try to define a language which is a subset of English and does not use any formal notations. In places where this requirement conflicts with the requirement to provide a straightforward translation between CNL and OWL, we may tolerate minor formal-looking additions like variables for anaphoric reference, brackets and indentation for grouping, etc. The results of user evaluation need to decide on the exact balance. The syntax of CNL OWL should be defined by a closed class of function words, an open class of content words, and a small set of grammar rules presented using linguistic notions like ‘phrase’, ‘subject’, and ‘negation’. A limited amount of morphological variation is supported, e.g. ‘mouse’ and ‘mice’ have the same lemma, etc. The description of CNL OWL should not be significantly longer than the descriptions of other OWL syntaxes. Secondly, the designed language and its associated translation programs should support a two-way mapping to a standard OWL syntax for which we 179 have chosen the OWL 1.1 Functional-Style Syntax5 (FSS). Related to this point are the two questions whether the CNL should allow for expressing OWL axioms in alternative ways in order to offer more flexibility for the author and whether the language should allow for representing several OWL axioms as one CNL sentence to increase compactness. While the focus of our work is on writing OWL ontologies in CNL, providing access to existing OWL ontologies and viewing entailed axioms in CNL is also important. We have decided to cover all of OWL 1.1 without extra-logical features like annotations. At the first step we ignore data properties and namespaces, as those are hard to express in natural language alone and would require including more formal-looking notations.

3 Comparison

This section examines a set of OWL 1.1 axioms and their renderings in ACE, Rabbit, and SOS, discussing the similarities and differences between the respective approaches. The axioms originate from a domain ontology for ‘Buildings and Places’ authored by domain experts at Ordnance Survey [21]. The full ontology contained over 600 concepts; we have used a subset6 that covers all different axioms types of OWL 1.1 except one, where we have constructed an artificial case.

OWL AsymmetricObjectProperty(ObjectProperty(is-larger-than)) ACE If something X is larger than something Y then Y is not larger than X. RAB The relationship "is larger than" is asymmetric. SOS If X is larger than Y then Y is not larger than X.

There are two key differences between these renderings: firstly, SOS and ACE use variables, whilst Rabbit does not. Secondly, Rabbit speaks on a meta-level whereas SOS and ACE speak on the object level: that is Rabbit speaks about the ontology and the nature of its properties, whilst SOS and ACE attempt to frame the phrasing as a statement about things in the domain. The meta-level versus object-level difference is a recurring one throughout the examples and a key design choice to be addressed. While in ACE each variable is introduced as an apposition to the indefinite pronoun ‘something’, SOS does not do this and is thus less verbose.

OWL SubClassOf(OWLClass(river-stretch), ObjectMaxCardinality(2, ObjectProperty(has-part), OWLClass(confluence))) ACE Every river-stretch has-part at most 2 confluences. RAB Every River Stretch has part at most two confluences. SOS Every river stretch has at most 2 confluences as a part.

5 http://www.w3.org/2007/OWL/wiki/Syntax 6 http://code.google.com/p/owl1-1/downloads/list Chapter 5 Controlled Natural Language Syntaxes for OWL

All three syntaxes present ‘confluence’ in its plural form ‘confluences’, this requires linguistic knowledge. Differences between syntaxes reflect different choices in presenting the ‘has-part’ predicate. Rabbit has opted to use upper case to indicate class names, whilst SOS and ACE do not. This makes it easier for the author to recognise which part is the class name, but looks unnatural when read as an English sentence. Unlike ACE and Rabbit, SOS breaks the ‘has part’ predicate apart and nests the cardinality (‘at most 2’) within it. ACE and Rabbit keep this predicate in one piece but ACE adds a hyphen.

OWL SubClassOf(OWLClass(factory), ObjectSomeValuesFrom(ObjectProperty( has-part), ObjectIntersectionOf([ObjectSomeValuesFrom( ObjectProperty(has-purpose), OWLClass(manufacturing)), OWLClass(building)]))) ACE For every factory its part is a building whose purpose is a manufacturing. RAB Every Factory has a part Building that has Purpose Manufacturing. SOS Every factory has a building as a part that has a manufacturing as a purpose. The use of ‘a manufacturing’ in SOS and ACE is unnatural. This is due to the initial authoring choice by the domain experts at Ordnance Survey to nominalise all processes and use only a small set of properties (e.g. ‘has-purpose’, ‘applies- to’) in the ontology. An interesting alternative is to use transitive verbs (e.g. ‘manufactures something’) instead of nominalisations (e.g. ‘has-purpose manufacturing’) in order to describe processes. Note that the use of simple transitve verbs can also avoid other unnatural renderings (e.g. ‘comprise’ instead of ‘has- part’).

OWL EquivalentClasses([OWLClass(petrol-station), OWLClass(gas-station)]) ACE Every petrol-station is a gas-station. Every gas-station is a petrol-station. RAB Petrol Station and Gas Station are equivalent. SOS The classes petrol station and gas station are equivalent. In this example, SOS uses the meta-level by referring explicitly to classes, whilst ACE and Rabbit use the object level. ACE’s approach produces a sentence for each pair of equivalent classes, which will be unwieldy to process when going from text to OWL. Rabbit’s statement is ambiguous as it is not entirely clear what the nature of the meta-level predicate ‘equivalent’ is (although the presence of capitalization may help the reader conclude it is the classes themselves). OWL SubClassOf(OWLClass(bourne), OWLClass(stream))) ACE Every bourne is a stream. RAB Every Bourne is a kind of Stream. SOS Every bourne is a stream. SOS and ACE produce exactly the same minimal ‘is a’ rendering, whilst Rab- bit uses the construct ‘is a kind of’. All three syntaxes use an explicit universal quantifier ‘every’ rather than the indefinite article ‘a’ or the definite article ‘the’. 181

OWL SubClassOf(ObjectSomeValuesFrom(ObjectProperty(has-part), OWLClass(water)), ObjectSomeValuesFrom(ObjectProperty(contain), OWLClass(water))) ACE Everything whose part is a water contains a water. RAB Everything that has a Part that contains some Water will also contain some Water. SOS Everything that has some water as a part contains some water.

These examples illustrate that mass nouns are diﬃcult to handle without additional linguistic knowledge. Note also that Rabbit uses the construction ‘will also’ which may be interpreted as having a temporal reading, whilst ACE and SOS have been careful to avoid temporal constructions, as they are not intended in the underlying OWL constructs.

OWL DifferentIndividuals([Individual(Scotland), Individual(England)]) ACE Scotland is not England. RAB England and Scotland are different things. SOS Scotland and England are different individuals.

Here, ACE uses negation more explicitly (‘is not’) compared to Rabbit and SOS that both use the expression ‘different individuals’. Rabbit makes the choice of referring to England and Scotland as different ‘things’ whereas SOS refers to different ‘individulas’.

OWL SubObjectPropertyOf(SubObjectPropertyChain([ObjectProperty(has-part), ObjectProperty(contain)]), ObjectProperty(contain)) ACE If something X has-part something that contains something Y then X contains Y. RAB Everything that has a Part that contains something will also contain that thing. SOS If X contains Y and Y has Z as a part then X contains Z.

Both SOS and ACE are based on an ‘If...then’ construction whereas Rabbit’s rendering uses a more complex construction and avoids using variables.

OWL EquivalentClasses([OWLClass(source), ObjectIntersectionOf([ObjectUnionOf([OWLClass(spring), OWLClass(wetland)]), ObjectSomeValuesFrom(ObjectProperty(feed), ObjectUnionOf([OWLClass(river), OWLClass(stream)]))])]) ACE Every source is a spring or is a wetland, and feeds something that is a river or that is a stream. Everything that is a spring or that is a wetland, and that feeds something that is a river or that is a stream is a source. RAB Every Source is defined as: Every Source is a kind of Spring or Wetland; Every Source feeds a River or a Stream. SOS The classes source and spring or wetland that feed some river or some stream are equivalent.

SOS refers to classes explicitly whereas ACE does not. ACE uses multiple clauses and stays completely on the object level. Rabbit uses the ‘is deﬁned as’ Chapter 5 Controlled Natural Language Syntaxes for OWL

construction and a series of clauses separated by semi-colons in order to structure the complex statement, but this works only in the case of intersection but not with union.

4UserTesting

Different forms of user testing [10,9] present evidence supporting our argument that controlled natural languages can offer improvements over standard OWL syntax. This was found to compare favourably with OWL as represented by the Protégé ontology editor, although no distinction was made between evaluation of the software tool which encapsulates the language and testing of users’ comprehension of the language itself. [17]’s user testing also confirms that natural language interfaces are useful, in this case, for querying the semantic web. Ordnance Survey has initiated a programme of user testing of Rabbit to evaluate how easy Rabbit is to understand. In the first phase of user testing, 31 sentences were shown to 223 participants (geography undergraduates), asking them to chose one of a selection of answers explaining what each Rabbit sentence meant. The answer choices were created to indicate why participants were getting the answer wrong. The order was randomised to ensure there was no bias. Similarly the subject of the ontology was an imaginary insect chosen to ensure the participants would have minimal background knowledge. Thirteen of the sentences were answered correctly by 75% or more of participants, with a large group near to the 75% acceptance mark. These sentences were deemed sufficiently understandable by most participants. They include the structures using ‘exactly’, ‘at least’, ‘at most’, ‘1 or more of A or B or C’ (to indicate non-exclusive or), that, ‘eats is a relationship’, and ‘only A or B or nothing’ (to indicate the universal quantifier). ‘is an instance of’ wasn’t well understood, nor was the structure ‘is a kind of’, although it was unclear whether this was due to Rabbit’s original use of the indefinite article to start the sentence. Comprehension of reflexivity, irreflexivity, asymmetry, transitivity and inverses was tested, using the same ‘if...then’ structure employed by SOS and ACE, with mixed results. Asymmetry, reflexivity and irreflexivity were understood, while transitivity and inverses were not. This might be because it was not always clear whether users really understood that these characteristics applied to the relationships on a global scale, or if they assumed that they were only valid at a local level when dealing with the connection between the two concepts in the supplied example. This kind of issue needs further testing (with a control group), along with validation of the CNL against the Manchester Syntax, which is being addressed in our second phase of testing, currently underway.

5 Discussion and Conclusions

Although there are clearly diﬀerences between the three CNLs, there is considerable overlap between them and therfore much common ground to build on. 183

There are four principle areas of difference. The first, least important and most easily resolvable concerns style. For example, ACE chooses to hyphenate noun phrases: river-stretch, whereas Rabbit and SOS allow River Stretch and river stretch (the capitalisation Rabbit being another minor difference). Secondly there are differences in approach in how to express certain constructs. This is most apparent with examples such as where the natural English form assumes the reader will understand the meaning of a phrase due to the context. So where in English a speaker might say ‘a river has a bank’ all three CNLs have found the need to be explicit about the interpretation of ‘has’. ACE and Rabbit both opt for ‘has-part/has part’ whereas SOS chooses to place the phrase ‘as a part’ at the end of the clause. Probably the biggest area of difference is where the CNLs represent mathematical constraints such as transitivity. Here there is really no good solution and here the approaches are most different. Rabbit’s approach has been to assume that no solution will really work and so requires the reader to be educated in the meaning of such constructs or be aided by a tool. SOS and ACE both try variations on the theme of explain-through-example and tool support. Lastly, while Rabbit explicitly endorses the cooperation between domain experts and knowledge engineers, ACE does not and tries to eliminate knowledge engineers altogether, whereas SOS is neutral in this question. We conclude that there is sufficient commonality between the three CNLs described here to provide a good base from which to proceed. Looking to the future, it is our intention to systematically resolve the differences that exist, guided, where possible, by user testing.

Acknowledgment

The authors of Rabbit would like to thank Martina Johnson for her assistance in preparing and analysing the human subject tests. This research on ACE has been funded by the EC and SER within the 6th Framework Program project REWERSE number 506779 (cf. http://rewerse.net). All authors would like to thank Nobert E. Fuchs for useful comments on a previous version of this paper and Rolf would like to thank Norbert for hosting him while being on sabbatical. Special thanks go to three anonymous reviewers of OWLED2008 DC for their useful comments.

References

1. Raﬀaella Bernardi, Diego Calvanese, and Camilo Thorne. Lite Natural Language. In IWCS-7, 2007. 2. Tim Berners-Lee. Notation 3 - A readable language for data on the Web. 1998. http://www.w3.org/DesignIssues/Notation3.html. 3. Peter Clark, Philip Harrison, Thomas Jenkins, John Thompson, and Richard H. Wojcik. Acquiring and Using World Knowledge Using a Restricted Subset of English. In FLAIRS 2005, pages 506–511, 2005. Chapter 5 Controlled Natural Language Syntaxes for OWL

4. Confluence project, 2007. http://www.comp.leeds.ac.uk/confluence/. 5. Anne Cregan, Rolf Schwitter, and Thomas Meyer. Sydney OWL Syntax — towards a Controlled Natural Language Syntax for OWL 1.1. In OWLED 2007, 2007. 6. Hamish Cunningham, Diana Maynard, Kalina Bontcheva, and Valentin Tablan. GATE: A framework and graphical development environment for robust NLP tools and applications. In Proceedings of the 40th Anniversary Meeting of the ACL, 2002. 7. Norbert E. Fuchs, Kaarel Kaljurand, and Gerold Schneider. Attempto Controlled English Meets the Challenges of Knowledge Representation, Reasoning, Interoper- ability and User Interfaces. In FLAIRS 2006, 2006. 8. Norbert E. Fuchs, Uta Schwertel, and Rolf Schwitter. Attempto Controlled English — Not Just Another Logic Specification Language. In LOPSTR’98, 1999. 9. Adam Funk, Valentin Tablan, Kalina Bontcheva, Hamish Cunningham, Brian Davis, and Siegfried Handschuh. CLOnE: Controlled Language for Ontology Edit- ing. In ISWC 2007, 2007. 10. Christian Halaschek-Wiener, Jennifer Golbeck, Bijan Parsia, Vladimir Kolovski, and Jim Hendler. Image browsing and natural language paraphrases of semantic web annotations. In SWAMM Workshop, Edinburgh, Scotland, 2006. 11. Glen Hart, Catherine Dolbear, and John Goodwin. Lege Feliciter: Using Structured English to represent a Topographic Hydrology Ontology. In OWLED 2007, 2007. 12. Glen Hart, Catherine Dolbear, John Goodwin, and Katalin Kovacs. Domain On- tology Development. Technical report, Ordnance Survey, 2007. 13. Daniel Hewlett, Aditya Kalyanpur, Vladimir Kolovski, and Chris Halaschek- Wiener. Effective Natural Language Paraphrasing of Ontologies on the Semantic Web. In End User Semantic Web Interaction Workshop (ISWC 2005), 2005. 14. Matthew Horridge, Nick Drummond, John Goodwin, Alan Rector, Robert Stevens, and Hai H. Wang. The Manchester OWL Syntax. In OWLED 2006, 2006. 15. Mustafa Jarrar, Maria Keet, and Paolo Dongilli. Multilingual verbalization of ORM conceptual models and axiomatized ontologies. Technical report, Vrije Universiteit Brussel, February 2006. 16. Kaarel Kaljurand. Attempto Controlled English as a Semantic Web Language.PhD thesis, Faculty of Mathematics and Computer Science, University of Tartu, 2007. 17. Esther Kaufmann, Abraham Bernstein, and Lorenz Fischer. NLP-Reduce: A “na¨ıve” but Domain-independent Natural Language Interface for Querying On- tologies. In ESWC 2007, 2007. 18. R. I. Kittredge. Sublanguages and controlled languages. Oxford University Press, 2003. 19. Tobias Kuhn. AceWiki: A Natural and Expressive Semantic Wiki. In Semantic Web User Interaction at CHI 2008: Exploring HCI Challenges, 2008. 20. Chris Mellish and Xiantang Sun. Natural Language Directed Inference in the Presentation of Ontologies. In ENLG, Aberdeen, Scotland, August 8–10th 2005. 21. Ordnance Survey. Buildings and Places, 2008. http://www.ordnancesurvey.co. uk/ontology/v1/BuildingsAndPlaces.owl. 22. Alan L. Rector, Nick Drummond, Matthew Horridge, Jeremy Rogers, Holger Knublauch, Robert Stevens, Hai Wang, and Chris Wroe. OWL Pizzas: Practi- cal Experience of Teaching OWL-DL: Common Errors & Common Patterns. In EKAW 2004, 2004. 23. Rolf Schwitter. English as a Formal Specification Language. In DEXA 2002, 2002. 24. Rolf Schwitter, Anna Ljungberg, and David Hood. ECOLE — A Look-ahead Editor for a Controlled Language. In EAMT-CLAW03, pages 141–150, 2003. 25. John F. Sowa. Common Logic Controlled English. Technical report, 2004. Draft, 24 February 2004, http://www.jfsowa.com/clce/specs.htm. 6 Encouraging Ontology Reuse

6.1 n2mate

Title of Publication: n2Mate: Exploiting social capital to create a standards-rich semantic network Type of Publication: Workshop Paper Appears In: Proceedings of the Linked Data on the Web workshop (LDOW2008) at WWW2008, Beijing, China, April 22, 2008. Publication Date: 2008 Peer Reviewed: Yes Contributing Author(s): David Peterson Anne Cregan Robert Atkinson John Brisbin Personal Contribution: 31% (average of estimates by co-authors)

185 Chapter 6 Encouraging Ontology Reuse 187

n2Mate: Exploiting social capital to create a standards-rich semantic network

David Peterson Anne Cregan Rob Atkinson John Brisbin BoaB interactive National ICT Australia CSIRO Land & Water BoaB interactive 2/84 Denham St. 223 Anzac Parade Lucas Heights Research 2/84 Denham St. Townsville, QLD Australia 4810 Kensington NSW Australia 2052 Laboratories Townsville, QLD Australia 4810 +61 7 4724 2933 +61 2 8306 0458 Private Mail Bag 7, Bangor +61 7 4724 2933 [email protected] [email protected] NSW 2234, Australia [email protected] [email protected] ABSTRACT 1. Social and technical context A significant boost on the path towards a web of linked, open data is the establishment and promotion of common semantic resources The current emergence of a data web has re-focussed our attention including ontologies and other operationalised vocabularies, and on standards. To be truly effective, the semantic web needs to their instance data. Without consensus on these, we are evolve towards a minimum number of ontologies, highly re-used, hamstrung by the famous “n-squared” mapping problem. In and densely interlinked, rather than a sparse network with addition, each vocabulary has its own associated attributes to do minimal interoperability. with why it was developed, what purposes it is best suited for, and 1.1 The standard problem with standards how accurate and reliable it is at both a content and technical level, but most of this information is opaque to the general The project to link open data can be realised through explicit community. declarations by one data source in relation to another. These “hard” linkages provide a high degree of certainty, but make data Our theory is that it is the lack of socially-sensitised processes maintenance exponentially difficult as the number of hard highlighting who is using what and why, that have led to the linkages grows. current unmanageable plethora of vocabularies, where it is far easier to build your own vocabulary than try to find a suitable, Standards, understood as nodes of agreed meaning, provide a reliable existing one. more scalable approach to data linking. By agreeing to use the same term to describe similar ideas in our different data, we We therefore suggest that there is considerable value in the establish an implicit (semantic) linkage between our data. The development of an online facility that performs the function of project to conceive, negotiate, and promote standards, however, providing a space listing vocabulary and ontology resources with has proven to be even more difficult than the maintenance of hard their associated authority, governance and quality of service linkages. attributes. Presenting this in a visual form and providing pivotable search facilities enhances recognition and comprehension. It is often noted, with some irony, that the great thing about standards is that there are so many to choose from...and if you Additionally, and critically, the facility provides a focal point can’t find one you like, you can always create your own. . where discourse communities can make authority claims, rate vocabularies on various parameters, register their commitment to While these sentiments provide excellent platforms for pub-based or usage of particular vocabularies, and provide feedback on their oratory, the realities are not so easily dismissed. Application experiences. Through social interaction, we expect the most solid designers, knowledge seekers, and agencies with a mandate to and useful vocabularies to emerge and form a stable semantic interoperate are all too familiar with the significant resource platform for content representation and interlinked knowledge. drains that occur when standards are hard to locate, difficult to Our strategy is to become sufficiently enmeshed in the native apply, or confusing to distinguish between. information habits of people and their derivative institutions to Standard vocabularies and data definitions have been quietly reveal and collect their standards-seeking needs and activities with multiplying in traditional media since ancient Sumer (ca. a minimum of effort on their part. Wikipedia, Cuneiform) but in more recent times the Semantic This paper describes a pilot facility testing the theory above. Web has inspired a hyperbolic growth in contributions to the Dubbed “n2Mate”, it is a novel exploitation of social networking standards project. For instance, a search in Swoogle on the word software to provide a lightweight and flexible platform for testing “address” returns 12,834 semantic web documents; on “book” it the efficacy of leveraging social networks to link existing registers returns 19,601 (at 2008-01-24). For someone seeking to exercise and ‘seed’ an information space focussing on the use of standards the efficiencies of knowledge reuse, this wealth of choice is in online information management. simply overwhelming and self-defeating. The current state of affairs reveals semantic fragmentation, not semantic integration The paper uses examples from the Australian context to provide and knowledge creation. clear illustration of the central arguments. Even within a narrow domain like the Australian government, Keywords there are a wealth of terminologies and metadata “standards” Registers, vocabularies, standards, linking density, rdf graph, available for government agencies to consider. It is not clear if a social networking, knowledge re-use, n2Mate, n-squared whole of government survey of standards has ever been undertaken, but informal observation suggests that there are hundreds of attempts to describe very similar concept spaces. Chapter 6 Encouraging Ontology Reuse

1.2 Does anyone have a wheel like mine? AGIMO and others have a role in promoting the use of common approaches, but it is increasingly difficult to track which standards People have been trying to standardise themselves in one way or apply to which set of problems. another for quite some time. The most obvious benefit of this In general, there is an issue about the scalability of any approach instinct toward standardisation is communication efficiency, a for improving interconnectedness. We believe that the most direct input to the rate of knowledge creation. By speaking the promising strategy is to utilise registers to hold metadata about same language, we can communicate and collaborate far more standards and their implementation, including records of effectively. Yet the barriers to standardisation appear to take on organisations, projects, standards, controlled vocabularies (and new forms as fast as we evolve knowledge. associated people and roles). A network of such registers, coupled In our present age the benefits of information interoperability are through normal web services mechanisms, has the potential to now well understood, if only through their absence. Most people form a semantic fabric that addresses the business-level needs of and institutions involved in project scoping, information product people and institutions. Whilst this is potentially a vast development, and online service provision clearly grasp the power undertaking, the bulk of target information already exists, and of knowledge re-use and the cost efficiencies of standards-based there are already a great many people actively tasked with interoperation. This assertion is supported by the existence of an identifying, using and promoting standards. These people are entire government department whose mandate is to promote likely to be receptive to an effort such as n2Mate. effective and efficient information sharing, governance structures, A network of registers, supported by a “register of registers” tools, methods and re-usable technical components across the addresses the most important questions: who is doing what, which Australian Government. standards are relevant, who can I talk to, what is the governance The Australian Government Information Management Office model for these artefacts, and how trustworthy is the source. (AGIMO) published a Government Architecture Reference Model Through a richly populated network of registers, these become that discusses “...a repository of architectural artefacts (including questions any organisation can rapidly address, and in doing so standards, guidelines, designs and solutions) that may be utilised can promote commonality of approach within and amongst by agencies to deliver an increasing range of Whole of various discourse communities. Government services.” 1.4 Socially-sensitive metadata In practice, however, we find that the task of identifying and verifying the suitability of existing artefacts is simply too time- One of the dark secrets of the machine-based knowledge project is consuming. As a consequence, there are a great many ontologies the enormous loss of content as we move from people’s minds to and informal vocabularies used by a very limited number of their documents and datasets. David Snowden, amongst many organisations or agencies, with a great sparsity of intermappings others, has pointed to the impossibility of “collecting” knowledge between them, even though there is a very large amount of from people without providing a meaningful context: crossover in terms of content. “Human knowledge is deeply contextual, it is triggered by More globally, the Linking Open Data (LOD) project [1] holds circumstance and need, and is revealed in action. .... to ask datasets that currently comprise over 2 billion triples but reveal someone what he or she knows is to ask a meaningless question in only about 3 million links (SWEO, 2007), so overall the graph is a meaningless context. Tacit knowledge ... comes about when our very sparsely interconnected [2]. skilled performance is punctuated in new ways through social interaction” [3]. In many ways the current situation is akin to a train network that has millions of stations (nodes) covering the same area A socially-sensitised strategy provides the meaningful context and (knowledge domains) but with a great sparsity of tracks familiar atmosphere that people require before they can (or will) (mappings) between stations, and hardly any trains and passengers reveal their knowledge in a useful way. (services, publishers, agents, users) running on the vast majority We suggest there is a cluster of persistent problems in complex of them. information spaces that can be socially characterised as follows: Our experience with efficient rail networks shows that we want to Who and what: reach a necessary minimum of stations interconnected with an optimised number of tracks, and attract a maximum number of Owner: Who owns it? trains to utilise the infrastructure. This obviously gives us a far Creation: Who created it? more robust and useful semantic network to traverse. Maintenance: Who is responsible for maintaining it? In related research, it should be possible to show how the density of interconnectedness in the RDF graph improves the efficiency Domain: which domains is it relevant to? This will of machine process operation without producing a debilitating include a number of different ways of considering level of ambiguity. We would argue that the degree of domains. interconnectedness implemented between ontologies can be taken Usage: Who uses it? as a proxy indicator of interoperability across the knowledge domain. Endorsement: Who endorses it? This will include various parameters and a rating system. 1.3 Scalable register networks Processes: What Business, Government or other As we have argued, there are many technical standards and processes is it used in? What role does it play? common policies in use across a wide range of government Governance: Who is in charge of it? Who has formally activities, but the very number of such activities and standards is agreed to support, maintain, and implement it? in itself posing a significant challenge. 189

Quality of Service Parameters: 2.2 Instance Data Provenance: What guarantees are there that the The facility needs to be designed around a sufficient minimum of information is accurate and verified? predicates that embody the “business logic” of the facility and Currency: How often is it updated? What guarantees are establish the semantic armature we require for inferencing. there that it is up to date? We propose the following [shows predicate] as a starting point: Availability: What guarantees are there regarding the Organisations are [responsible for] people, projects, availability of the vocab, dereferencing considerations standards, and vocabularies Other Considerations: People are [associated with] Projects How does it relate to other standards in the space? Projects are [implemented by] Standards User experiences Standards are [expressed with] vocabularies 2. Social architectures and semantic networks Trust or utility of Standards are [ranked by] People The principle social platform techniques we seek to exploit Using these indicative predicates as a starting point, we can include: answer a matrix of discovery questions through faceted visualisation. In each search operation, the user can rotate to a Popularity Rankings: number of times a standards artefact facet of interest to continue the discovery process. is referenced (implemented). Authority Badges: mechanism to advertise an authority I know someone like me [PersonName] > What projects claim over a standards artefact. are they associated with? Related to (“Friends of a Standard (FOAS)” ): linkages Those projects are like mine [ProjectName] > What from standards artefacts to their cohort of implementers. standards are used in them? Trust ratings: showing satisfaction with the custodian of a Those standards are of interest [StandardName] > How standards artefact. can I decide which one is most appropriate for me? Hero worship: most interlinked, most trusted, most useful The logic described here is possible because we have imposed a Each of these techniques have corresponding interface strategies limited set of predicate types. These types are native to the that provide a powerful social platform in which people (and n2Mate facility. To take advantage of existing social networks institutional roles) can operate quite naturally. that utilise other predicate types, Semantic Web vocabularies such as SIOC and FOAF will be used. Each of these techniques also forms a search facet that can be traversed with high efficiency faceted search and browsing tools. The facility will also consider structured lists of resources, like a list of country names available from the same address, to itself be 2.1 Use Case a kind of register. For instance, many applications need a list of A simple use case will help us set the stage for describing the every valid country name for users to input their address technical architecture proposed. information. The ability to reference an external source that is authoritative, accurate, up-to-date and reliably available and A researcher is preparing her research plan on a section of the derefenceable reduces the need for application maintenance. Great Barrier Reef. Although she is an experienced marine scientist, she is new to the GBR and to her host research facility. The metadata held in these registers can be typed according to She suspects she should be using: existing conceptualisations. For example, the National Data Network draws on ideas from the Metadata Open Forum to standard naming conventions for the GBR regions; classify their metadata as: Discovery metadata; Quality metadata; standard identifications for the particular reefs; and Definitional metadata. standard data sampling techniques appropriate to the Australian tropics; We note that the semantic register network can also list web standard data formats, enumerators, and vocabularies in services in addition to typical standards artefacts such as her datasets; ontologies and vocabularies. standard citations of agencies, programmes, and people We intend to specifically tune this facility to the needs of referenced in her work; government and community agencies that have a mandate to standard metadata fields and vocabularies to describe participate in the creation and maintenance of highly effective her research output; approaches to service improvement. standard project management practice in reporting on her project’s progress. 3. Implementation options In the absence of a useful standards locator, it’s not likely that she A demonstrator version of n2Mate can be established using will achieve a high standard of conformance to the norms of her readily available tools and datasets so that a more detailed critique discourse community. can be pursued with a minimum of upfront overhead. In this In the absence of a socially-sensitised register space, it is not section we discuss some of the more promising approaches. likely her discourse community is actively sharing their experience and wisdom with standards. Chapter 6 Encouraging Ontology Reuse

3.1 Key components what ontologies Sun Microsystems is using, they select Sun from the ‘Who is Using’ facet. The other facets instantly re-order and The registration process, and maintaining a network of linked re-number themselves and the user is free to further refine the objects, is the function of traditional registry technologies, such as results by selecting additional facets. ebXML Registry. Navigating and efficiently querying the contents Faceted search visualisation can be negotiated through cluster and relationships is not well supported by this environment. maps (eg, Aduna) with a high degree of efficiency. It is proposed to automate the harvesting of object relationships from the “Register of Registers” into a triple-store. This is the Semantic interpretation: MOAT same pattern found in data-mining, where transactional database MOAT (Meaning of a Tag) could serve as the basis for giving content is restructured into generalised query-oriented structures. extended quality of information to free form folksonomy tagging. For our purposes, automated discovery of patterns is not the This will allow users of the bookmarking system to have the focus: fast, efficient visual presentation is essential. flexibility of folksonomy and the interlinked structure of the Users will be parsing through extensive data structures, and may Semantic Web. The added benefit is that MOAT is a distributed need to propose and refine their discovery logic in quick, system and can tap into other servers to give extended meaning to exploratory sorties. free-form tags. Triple-store: Sesame Sesame could provide backend triple store, graph manipulation, RDF inferencing, and remote SPARQL endpoint access. Policy layer: PLING The development of robust approaches to policy negotiation is being driven by a W3C discussion group at the moment. The n2Mate project could field test various strategies for handling issues of personal privacy, information reuse, and access control. Trust and Governance: POWDER POWDER is the W3C’s Protocol for Web Description Resources, currently in development. Governance: is related to the idea of trust. In the context of this project, we want to explore two aspects of governance: 1. How to make it easy for agencies who have a mandate to be an authority for some asset to discharge their duty in an efficient and useful way. 2. How to provide users with a suite of trust measures that will allow them to evaluate the qualities of a particular asset in relation to their needs. POWDER seeks to develop a mechanism through which structured metadata can be authenticated and applied to groups of web resources. POWDER provides us with a means to both Figure 1: n2Mate Conceptual architecture retrieve information about a block of Web Resources and authenticate that this information may be attributed to the owners of the information. Visualisation and facet search: Gnizr + Solr We want a tool that thinks natively in URIs and triples. Gnizr is 3.2 Testing the system with existing resources an open source front end that handles user account management, There are already many semantically rich registers implicit in the bookmarking, tagging, and semantic search operations of government, including the identifier of government Every object stored by gnizr is a bookmark (URI), and the agencies, registers of company names, standards recognised by folksonomy tag interface is SKOS enabled. Standards Australia, legislation and regulations, management areas for land, water, soils, health etc. This represents a wealth of Solr is an open source enterprise search server based on the entities about which assertions can be made, to create a Lucene Java search library, with XML/HTTP and JSON APIs, hit semantically rich environment. highlighting, faceted search, caching, replication, and a web administration interface. Semantic Web data can be roughly broken down into 3 levels:[2] Solr could be used to facet the data into searchable and 1. Vocabulary / Ontology browseable components. For example, if users are interested in 191

2. Individual occurrence of those terms and actual A bridging space between government, business, community, instances of non-information resources academia and science knowledge assets to enhance 3. The links that tie the vocabularies to their occurrences broadscale interoperability. All three of these need to be captured with adequate provenance A genetic algorithm to breed, select, and hybridise various data to bootstrap n2Mate. standards artefacts such as ontologies, services, and trust The following web services can be utilised to populate/update authorities. information as well as add important metadata to the Register of In conclusion, we suggest that there is currently a significant level Registers component of n2Mate. of inefficiency in the applied domain of project scoping, Watson: A gateway to the Semantic Web, focusing on: information product development, and online service provision semantic data quality; relations between ontologies; due to the inadequacy and irrelevance of existing knowledge access to semantic data registers. Talis Schema Cache: Cross-linked and navigable index We further suggest that a promising solution strategy involves of ontologies and vocabularies. using the power of social networks, coupled with semantic discovery and visualisation tools, to create a socially-sensitised Swoogle: Search engine for Semantic Web artefacts semantic network of standards registers. Sindice: Indexes the RDF web and pulls out the triples. From there it essentially creates a reverse lookup. 5. REFERENCES Falcons: Currently indexing 34,566,728 objects (2008- 5.1 Citations 02-01), Provides bi-directional resource linking. Ping the Semantic Web: archives the location of [1] C. Bizer, T. Heath, D. Ayers, and Y. Raimond. Interlinking recently created/updated, web-accessible RDF Open Data on the Web (Poster). In 4th European Semantic Web Conference (ESWC2007), pages 802–815, 2007. 3.3 Data harvesting and processing [2] M. Hausenblas, W. Halb, Y. Raimond, and T. Heath. What is n2Mate can leverage existing search engine services, such as the Size of the Semantic Web? - Metrics for Measuring the those listed above, to collect data instances from target registers Giant Global Graph, 2007. and sources. Many of these have or are developing APIs that [3] Snowden, Dave. Information vs Knowledge. facilitate direct access to their collections and service points. http://www.rkrk.net.au/index.php/Information_Vs_Knowledg e Where well-formed registers and artefact collections exist already, n2Mate could establish harvesting relationships (presumably 5.2 Resources through appropriate API arrangements). OWL files, RDF data dumps, and SPARQL endpoints could be pointed to the n2Mate Aduna: http://www.aduna-software.org system for automated data fetching and processing. AGIMO, GovDex http://www.agimo.gov.au/services/GovDex Falcons: http://iws.seu.edu.cn/services/falcons/ Additionally, trust algorithms would be created from graph Gnizr: http://www.gnizr.com inferencing, metadata and social data to further guide the Metadata Open Forum: http://metadataopenforum.org/ prospective n2Mate user, allowing them to more quickly MOAT: http://moat-project.org/ determine what is the best artefact to use in their situation. This National Data Network: http://www.nationaldatanetwork.org/ will be an evolving process that will occur over time as the quality Ping the Semantic Web: http://pingthesemanticweb.com/ of data and user interactions flows back and forth. Policy Language Interest Group: http://www.w3.org/Policy/pling/ 4. CONCLUSION POWDER: http://www.w3.org/2007/powder/ Sesame: http://sourceforge.net/projects/sesame/ The unique aspect of this proposal is that it leverages the hidden Sindice: http://sindice.com formal and informal knowledge networks created by existing SIOC: http://sioc-project.org/ business processes, and marries this information with social solr: http://lucene.apache.org/solr/ networking models to provide a useful way of organising and SWEO IG: navigating the wealth of available information. It uses the http://esw.w3.org/topic/SweoIG/TaskForces/CommunityProjects/ community of people using vocabularies to empower others, LinkingOpenData starting with the places where agreements already exist. Swoogle: http://swoogle.umbc.edu TALIS Cache: http://schemacache.test.talis.com/ The n2Mate provides a tool that encourages use of standardised Watson: http://watson.kmi.open.ac.uk/Overview.html artefacts by exposing existing registers, leveraging social networks and building a central reference point for users that will 5.3 Special thanks…. assist them to identify relevant semantic assets for their needs, choose amongst them, and feel confident about their utilisation. Renato Iannella provided background thinking on the Policy Aware Web. Alan Ruttenberg of the Science Commons and Tom Further, research into the strategy proposed should provide Heath of the Linking Open Data project provided encouragement contributions to related projects, such as the development of: and wisdom from their broad experience. Steve Matheson from A lightweight mechanism revealing the state of the Australian Bureau of Statistics corroborated our intuition that interconnectedness in and between discourse communities. social platforms could play an important role in standards adoption.

7 Foundational Issues in Meaning

7.1 Towards a Science of Deﬁnition

Title of Publication: Towards a Science of Deﬁnition Type of Publication: Workshop Paper Appears In: Proceedings of the Australasian Ontology Workshop (AOW05), Sydney, Australia. CRPIT, 58. Meyer, T. and Orgun, M. A., Eds., ACS. 75-82, 2005 Publication Date: 2005 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Personal Contribution: 100%

193 Chapter 7 Foundational Issues in Meaning 195

Towards a Science of Definition

Anne M. Cregan1,2 1Artificial Intelligence Group 2 Knowledge Representation and Reasoning Program School of Computer Science and Engineering Kensington Laboratory, Sydney University of New South Wales National ICT Australia

223 Anzac Parade Kensington, NSW 2052, Australia [email protected]

Abstract “grounding” the elements themselves: that is, of objectively establishing exactly what the elements of the The vision of the Semantic Web is to provide machine- ontology are representing. However, when attempting to processable meaning for intelligent applications. Whilst validate, align or merge formalisms, this information is knowledge representation structures like ontologies now crucial. We cannot hope to determine whether any one have well-developed formalisms, the issue of determining ontology, let alone an alignment or merging of more than or specifying exactly what it is that they represent is still one ontology, is correct, without judging whether the not well-understood. However, it is crucial for referring relations are correct, in the sense of matching validation, merging and alignment, as we cannot possibly the intentions of the authors of the ontology(ies). hope to judge the accuracy or applicability of a representation structure without a clear specification of But how are such intentions to be clearly stated? As there what it is intended to represent. This being the case, we is currently no clear process for this kind of specification, must either accept that our representations will have a such a judgement usually requires human intervention, limited applicability and lifespan, or develop methods by often in the form of a dialogue or negotiation between the which we can define our terms in a robust and various parties responsible for each ontology. As is to be standardized way. Building on philosopher Richard expected, the exercise is often labour-intensive, and it Robinson’s analysis, it is argued that ‘definition’ is in fact may be quite difficult to establish exactly what is meant the isolation of territories within conceptual landscapes, by a term if the underlying mind-sets are divergent. using the four mechanisms: example, semantic relation, Yet the vision of the semantic web is to make meaning analysis, and rule. These mechanisms are related to machine-accessible and processable. Therefore, we need cognitive processes like abstraction and categorization. to establish a method by which we can formally state the We speculate that there is a common semantic ground meanings of terms, so that no further human intervention which forms the initial basis for symbol grounding, and is is required. Such a methodology would constitute a then extended through cognitive mechanisms. Some “science of definition” starting points for identifying common semantic ground and points of divergence from it are suggested.. However, the principles by which one can relate a term, whether a symbol, or a word from natural language, Keywords: Semantic Web, Ontology, Definition. unambiguously and unequivocally to a thing, have not yet been established. Or stated another way, there is not yet a 1 Why do we need a Science of Definition? sufficient understanding of the representation relationship. In fact, many are unsure whether it is even A key consideration in constructing ontologies, or indeed possible to establish such principles. It is often claimed any form of knowledge representation, is to be clear (eg Thomas 2005) that definition is an art, not a science, about what it is, exactly, that the elements of the and surveying the field up to this point, one has to agree. representation are to represent, and to be able to But if this is always to be the case, we are looking at a communicate this precisely. Without this, any claim that very challenging road ahead for the semantic web, if the structure is a representation at all is tenuous at best. multiple parties must engage in personal negotiations Yet, whilst formal knowledge structures and ontology every time interoperability of metadata needs to be languages are now well-developed in terms of their achieved. ability to express the relationships between and attributes of their elements, and to apply reasoning algorithms on The purpose of this paper, then, is to investigate whether this basis, there is currently no accepted methodology for there can in fact be a science of definition, and if so, on what principles would it be based? The paper is structured as follows: Section 2 explores what it means to define something, and the different methods by which Copyright © 2005, Australian Computer Society, Inc. This this may be done, drawing heavily on philosopher paper appeared at the Australasian Ontology Workshop Richard Robinson’s analysis (1950). Section 3 argues (AOW2005), Sydney, Australia. Conferences in Research and that Robinson’s methods isolate territories within Practice in Information Technology (CRPIT), Vol. 58. T. conceptual landscapes, using the mechanisms of example, Meyerand M. Orgun, Eds. Reproduction for academic, not-for semantic relation, analysis, and rule, and relates these to profit purposes permitted provided this text is included. underlying cognitive processes like abstraction and Chapter 7 Foundational Issues in Meaning categorization. It claims that common semantic ground or logical structure, merely rendering it more verbose and forms an initial basis for symbol grounding, which is later therefore more difficult to understand. extended through cognitive mechanisms. Starting points On the other hand, a relation between things and things for investigating this common semantic ground and the (A3) is the main means by which scientific understanding points of divergence from it are suggested. Conclusions is gained: its very purpose is to make logical as to the plausibility of a science of definition are drawn contributions. Robinson refers to A3 as ‘real definition’, in Section 4. but his ultimate conclusion is that it should not be referred to as “definition” at all, because it is in fact the 2 Dimensions of Definition analysis, synthesis and improvement of concepts. That is, It is commonly accepted that in setting up any formal it is the activities of finding and describing structure system, such as an ontology, one must carefully define within concepts, finding and describing concepts as part one’s terms. Therefore, let us firstly examine the concept of larger structures, and improving the fidelity of these of definition itself, to see what it has to offer us, and finding and describing activities. It encompasses the alternately, how it might mislead us in our endeavour. purpose of scientific endeavour very well, but does not tell us what we mean by something. For example (mine The concept of definition has had a long and tortured not Robinson’s), whilst scientific theory may constantly history dating back to the ancient Greeks, and in its long revise and improve our understanding of the nature of history, many strands have become tangled together, so light, it does not tell us what is meant by ‘light’: this is that ‘definition’ has come to mean many things to many something that is established before the scientific people. In untangling the various strands, the work of investigations begin. If we establish that light Oxford philosopher Richard Robinson (1950) is corresponds to a certain band of wavelengths in the particularly enlightening, and this section draws heavily electromagnetic spectrum, this is certainly useful to on Robinson’s analysis to isolate the dimensions of know, but does not define what light is: further down the definition and draw attention to some salient points. track we might hypothesise that light is energetic Robinson recognized that the word ‘definition’ had been subatomic particles called ‘photons’ and not just a wave, used by many people for many different things, often in and yet we are still talking about exactly the same entity ways which were inconsistent, even by the same person. ‘light’. From an analysis of usages of the term, he distilled Thus in Robinson’s ‘real definition’, we are not defining several key distinctions, which we will refer to here as a thing, but attempting to describe its relationships with dimensions. other things. Whilst we may create a formal system with the aim of capturing logical or empirical relations 2.1 Dimension A:What is Being Related? between things, we must necessarily in such a system use The first distinction drawn by Robinson is that symbols to represent these things. Indicating what the ‘definition’ has been used in attempts to relate completely symbols refer to extends outside the system itself to the different kinds of entities. He found that ‘definition’ has domain of things through Word-Thing relations (A2). been used variously to describe relations between: Robinson’s conclusion is that only relations between

words and things (A2) give information about what a A1. Words and other Words symbol is referring to, or denoting. He recommends that Eg “ Let i be the square root of minus 1.” only this activity should be considered to be “Definition”.

He then goes on to discuss further ways in which Word- A2. Words and Things Thing relations may be distinguished. Eg “ ‘Red’ is the colour of my shirt.”

A3. Things and Things 2.2 Dimension B: (Word:Thing) Are we Eg “ E = mc2 ” Reporting an Existing Use of a Word, or Establishing a New Usage? Robinson refers to A1 and A2 as ‘Nominal Definition’ and to A3 as ‘Real Definition’. Although he refers to Robinson’s analysis of “Word-Thing” (A2) definition ‘words’ in the categories above, he intends this to include revealed two variants: all written and spoken symbols. By analysing the three kinds of relation, Robinson concludes that what is B1. Lexical : the reporting of an existing or past actual attempted under the guise of ‘definition’ is a usage of a term eg “ ‘compound fracture’ is a medical fundamentally different activity in each of the three cases. term meaning a broken bone which punctures the skin” Whilst all three activities are valuable and useful, Robinson ultimately concludes that only A2 should lay B2. Stipulative : the specification of how a term is to be claim to being called “definition”. used in future in a particular context eg “ where for the purpose of our analysis ‘grue’ means ‘green before the Robinson’s reasoning may be summarised as follows: A year 2020 and blue after the year2020’ relation between words and other words (A1) is a substitution relation which may give brevity and Lexical definitions seek to pinpoint what was meant by convenience, but makes no logical contribution : it could some word to someone at some time. These definitions be omitted and make no difference to the representation 197 reflect the way language is used: the definitions may be however, that there are non-verbal methods of definition, intuitive or systematic, but often are not, as they have an such as the ostensive method, and that by including such organic, evolutionary aspect to them as the word methods all words are definable. Using these methods it meanings shift and adapt over time. Dictionaries may be is possible to go beyond the lexicon, and we will argue considered to be compilations of lexical definitions, as that these non-verbal methods are essential for the they are intended to assist listeners/readers and grounding of language and critical for the initial stages of speakers/writers in using the language being defined. (first) language learning. Obviously when contemporary dictionaries are published, Secondly, it has been argued that some particular there is an assumption that the entries will continue to be concepts of a philosophical nature, like “beauty” or accurate at least for the foreseeable future, but lexical “good” are intrinsically indefinable. Robinson’s response definition should always be recognised as being in the is that this is not claimed because we cannot establish context of who used the term and when: it is an historical what they refer to, but because we are encountering report. Meanings shift over time, words come in and out difficulties when attempting to understand their structure of usage, whole languages come and go. Dictionary and essential nature, which may (perhaps) be definitions of words are not god-given or absolute: they unanalysable. Thus, such a claim is confusing ‘real are a social agreement at a particular time and place definition’ with ‘nominal definition’, but unanalysable amongst particular people, and thus belong to history does not equal indefinable. Although one may be unable rather than science. to establish valid Thing:Thing (A3) relations for such concepts, we may certainly establish what the word refers However, in particular fields such as mathematics, or to by using a Word:Thing (A2) method of definition. (It where a high level of precision is needed, stipulative is my personal belief that such unanalysable concepts are definition may be used to specify that a certain term will of fundamental importance in understanding the ultimate be used only to mean a certain precise thing within a motivators of human behaviour, in a manner akin to given context. This may be a broadening or narrowing of George Kelly’s personal construct psychology (1955). an existing sense of a term; a completely new usage of an existing term; or may involve the creation of a totally new In all, Robinson identified seven methods by which term. It is a specification of how the author intends to use Word:Thing “definition” could be established. He did the term in future within a particular work or context, and not claim the division was “exhaustive, exclusive, or the binds the author to making good on that intention. only useful one”: each method is simply a means of communicating to a learner what the word refers to, ie Whilst Robinson makes it clear that he considers indicating what the word means. Not all methods are ‘definition’ to be a relation between Words and Things, applicable in all cases, and in some cases a combination there is an interesting point to be made that Robinson of methods might be needed. The seven methods of does not identify explicitly: do we define words, or Word:Thing definition Robinson described follow. things? It seems that lexical definition takes an existing word and attempts to circumscribe the things to which it applies, whilst stipulative definition circumscribes some C1. The Method of Synonyms area within the domain of things and assigns it to a This method gives the meaning of a word by giving the particular word. Robinson gives his exposition of learner a synonym for the word, with which she is already methods of definition (see Section 2.3) independently of familiar eg “ ‘Chien’ means ‘dog’ ”. this distinction: he does not differentiate between methods for defining words and for defining things. Each of Robinson’s methods has been shown diagrammatically in an effort to illustrate this point, and we will later argue for a theory of definition in which mapping in both directions is essential.

2.3 Dimension C: (Word:Thing) Methods of Definition Figure 1: The Method of Synonyms Robinson argues that all words are definable, in the sense of it being possible, at least theoretically, to establish Whilst the method requires a Word-Word correspondence what they indicate(d) to some one at some time. In fact, (A1), Robinson includes it as a Word-Thing method (A2), for something to qualify as a word at all, it must have had as the synonym is being used to identify a thing. some meaning to someone at some time, otherwise it would be only a noise or a meaningless sequence of C2. The Method of Analysis letters. Thus, in Robinson’s view, there are no This method consists of referring the learner to a thing by indefinable words. giving an analysis of it: eg “An ‘octagon’ is a polygon Let us briefly cover Robinson’s arguments against with eight sides”. If learner knows the meanings of the indefinability. Firstly, some have argued that since other words, and is able to construct the concept in the dictionaries merely relate words to other words, certain way intended by the speaker, the learner will then know words must be indefinable: there must be some both what constitutes an octagon, and what the word indefinable core vocabulary. Robinson points out ‘octagon’ refers to. Chapter 7 Foundational Issues in Meaning

If one knows the meanings of all the other words, the meaning of ‘diagonal’ is implied. In this method the word being defined is used rather than mentioned. Consequently, unlike the other methods, this method is not based on an equivalence relation. Usually however, the statement could be reworded into another method of

definition which is based on an equivalence relation. Figure 2: The Method of Analysis C5. The Denotative Method This method thus has the advantage of imparting knowledge of the thing via the analysis, at the same time This method defines a word by mentioning known as it giving the referring relation for the word. However, examples of things the word applies to. eg “ A ‘bird’ is it is useless in cases where analysis is not applicable, for things like swans, robins, geese and hens.” instance for words referring to particulars. Particulars, such as a person, cannot be defined by a list of specific characteristics, no matter how long, as it could always logically belong to some other particular.

C3. The Method of Synthesis This method identifies what a word refers to by indicating the relation of the thing indicated to other things eg “ ‘orange’ is the colour in between yellow and red”. Figure 5: The Denotative Method

Sometimes a thing is exactly a finite collection of examples, so a denotative definition is complete: for instance, this is the case connectives in symbolic logic, which are wholly defined by their truth tables. Usually though the list is incomplete and the learner is left to deduce an underlying connotation: that is, to take the necessary steps to abstract the concept, so that a Figure 3: The Method of Synthesis previously unseen example could be classified. It would The thing is indicated by saying where it can be found, or seem that such definitions would often require an iterative what causes it: it is assigned its place in a system of process of trial and error involving revision, broadening, relations and synthesised as a part of a whole with other narrowing etc, in order to ascertain the exact denotation things. Robinson notes that this method is indicative of the word. In this example, the learner might initially only; it is like saying “John is the tallest man in the room” hypothesise, for instance, that a bird is anything that has to identify which is John. He therefore seems to imply wings, and subsequently revise this notion when informed that in order to be applicable, this method requires a that a bat is not a bird. background context, which seems to involve some intended frame of reference with a particular structure. C6. The Ostensive Method Robinson also notes that the method enables great Unlike the five methods above, this method does not rely precision in indicating things where the two previous on words alone but indicates a thing by drawing attention methods would fail, and is particularly useful for primary to it in some way, for instance by pointing to an actual modalities of sense like ‘green’, ‘soft’, ‘sweet’ and thing, drawing the thing, or using demonstrative words ‘middle C’. either in context or in the absence of the physical object C4. The Implicative Method eg“ ‘Geese’ are the kind of bird we saw yesterday at the lake”. This method defines a word by putting it in a context that defines its sense: one may determine its meaning via the meanings of the other words eg “ A square has two diagonals, each of which divides the square into two right-angled isosceles triangles”.

Figure 6: The Ostensive Method

In other respects, the ostensive method resembles the denotative method as it works through giving examples. Figure 4: The Implicative Method 199

C7. The Regular Method 3.2 Definition as the isolation of areas in the Robinson points out that the six methods above are conceptual landscape suitable for names, where a name is “a word appointed to Our argument is as follows: Firstly, let us recognize that mean always some one and the same thing, whether a there is no relationship between symbols and things, other particular or a general thing”. However, some words are than that mediated by us humans. A tree is only a ‘tree’ if not names, for example, eg demonstrative words like humans make it so. The symbol-making process involves ‘this’, ‘I’, ‘yesterday’; conjunctions and logical operators having some mental representation of a thing, and giving like ‘and’, ‘not’, ‘but’ ‘who’, and words that do not have it a name of our choosing. The name then has ‘meaning’ any common quality, but all have a particular relationship by virtue of its connection to the conceptual landscape. in common eg “Swiss citizenship”. In this cases the By setting up the same mappings between names and regular method is used: the meaning of these words is mental representations in others, we set up a basis for given by rule eg “ ‘I’ is to be used by each utterer to communicating the contents of individual mental indicate himself”. In practice, the rules may be either landscapes: once the mapping between word and given explicitly, or deduced by the learner through conceptual landscape is set up in the mind of the listener, exposure to examples. the word may be used by a speaker to evoke the corresponding mental territory in the mind of the listener. 3 Charting the Conceptual Landscape Once evoked by a word, the mental territory provides a frame of reference for processing other information We will argue that underlying Robinson’s methods of provided by the speaker. In this way, the process can be definitions are four fundamental mechanisms which are bootstrapped: once some initial mappings are achieved, mental in nature. These are: example, analysis, semantic they may be used in constructing others. relation to some already grounded thing, and rule. Firstly however, let us clarify whether Robinson’s methods of The method gets off the ground initially because most definition actually relate words with things, or with our human mental landscapes are similar enough prior to the concepts of things. commencement of language learning that we can identify common landmarks, and use them as reference points. 3.1 Things or Concepts? Consequently, language learning is the process of associating symbolic markers (words) with areas in the Robinson’s methods of definition were stated to be Word- conceptual landscape. The process commences in early Thing (A2) relations. Although Robinson did not note it, childhood, where language learning draws heavily on the by working through some examples as presented in the ostensive method: conceptual representations of things previous section, it seems clear that the mapping can go directly experienced are mapped to words, under the either way: some methods map from Words to Things; guidance of a language speaker, usually a parent. By others from Things to Words, and some seem to do both tracing the stages of language learning, we are able to at the same time. identify four mechanisms and explain why it is that However, are the Word-Thing relations really mapping Robinson’s methods of definition are effective. words to and from the domain of things, or are they actually mapping words to and from the domain of 3.2.1 Examples concepts of things? • Acquiring rudimentary abstraction If we can establish our claim that Robinson’s methods are • Acquiring symbolisation underpinned by the four mechanisms of example, analysis, semantic relation to some already grounded The process of language learning commences by relating thing, and rule, then clearly the latter three are mental example stimuli to symbols. Initially, the conceptual operations and must therefore be mapping words and landscape arises naturally through interactions with the concepts. In the case of example, if the thing is not world, and is directly related to physical sensations. The physically present, then the learner must be relating a infant perceives a collection of colours, lines, sounds and word to some mental placeholder that is standing for the smells which arise in clusters, and learns through sensori- thing. On the other hand, if the thing is physically motor experience that these clusters are associated with present, there must still be some mental placeholder if the particular physical objects. mapping with the word is to last when the thing is no The infant then learns to recognise components of the longer physically present. clusters which provide a similar sensory experience eg Therefore, we claim that Robinson’s methods relate redness. By identifying and isolating common elements words not to things, but to concepts of things. We will of the naturally arising conceptual landscape, the infant now proceed to illustrate that his methods of definition begins to develop a rudimentary ability to abstract from are in fact instruments for isolating or marking out a piece experience. Through the association of example stimuli of a conceptual landscape using the four mechanisms. with spoken words, via a process of trial, error and Some of the mechanisms also involve mental operations correction, the infant can learn that ‘swan’ is a particular to construct and transform conceptual landscapes prior to cluster of sensory stimuli, and that ‘red’ is a specific such an isolation. component present in several sensory stimuli.

Chapter 7 Foundational Issues in Meaning

In this way, words become associated with a mental ‘square’. In this way, we are able to move step by step constructs, and the infant learns that symbols may be used away from the naturally arising domain of sensory stimuli to denote an element of their conceptual landscape: the into purely conceptual domains like mathematics. ability to symbolize is acquired. Having experienced Robinson’s synthesis method works because the mental success in identifying and using particular components of landscape for some underlying concept has already been sensory stimuli, like colours, these dimensions become marked out, and we can use our knowledge of its features reinforced within the conceptual landscape, and the infant to isolate or build a new construct within it or relative to gives more attention and weight in future. it, using semantic relations. This may be either pre- or post-analysis. In the ‘orange’ example, the word 3.2.2 Semantic Relations ‘between’ indicates not just the presence of a semantic • Extending rudimentary abstraction to acquire the relation, but the existence of an underlying idea with ability to recognise and use semantic relations structure: in this case, a continuum or scale by which colours are organised: we must draw on our conceptual • Acquiring Categorisation representation of colour and how it works to isolate a The basic mechanism of abstraction learnt earlier is then mental construct within it. used to extend the conceptual landscape, and language Thus analysis requires the use of semantic relations itself is a major enabler in this process. The naturally within conceptual domains that have already been arising conceptual features now marked out and mapped out. Note, conversely, that Robinson’s synonym associated with words are used to build higher order method uses an established syntactic relation to a mental constructs, further extending the conceptual conceptual territory to evoke the mental territory and map landscape. a new symbol to it. For instance, the ability to abstract common features from various sensory situations facilitates the learning of basic 3.2.4 Rules semantic relations like ‘between’, ‘above’ etc. The child • Acquiring the ability to circumscribe conceptual develops the ability to recognise semantic relations and territories, and use them to classify new examples use them as links between different mental territories. Through experience, trial and error, more sophisticated The rule method is a method for marking out a piece of mental territories are able to be marked out based on such an existing conceptual territory, which may be at any semantic relations, and along the way, words become the level of abstraction: the rules give borders that define the markers and signposts of these territories. territory, which may then be applied explicitly to determine whether a mental construct falls inside the The ability to understand and use semantic relations leads conceptual territory or outside it. to the ability to deduce rudimentary rules which support the categorisation of previously unseen examples. For instance, the denotative bird example requires the 3.3 Common Semantic Ground marking out of a new mental territory - ‘bird’ - to contain We have shown that words are intermediaries for existing territories already associated with words which communicating mental landscapes, and are effective were mapped to sensory clusters : ‘swan’, ‘hen’ etc. because the underlying landscapes can be marked out by Children learn by experience that in such cases there is reference to common features, or what we might call usually some reason to group the things together, perhaps ‘common semantic ground’. The argument assumes that the possession of some common feature(s), like having initially there are common reference points which are wings. Note that at this stage, categorization is based not naturally arising, enabling the process to commence so on just one dimension of sensory experience like that we can all acquire basic concepts like ‘redness’ and ‘redness’, but on an emergent quality based on several ‘swan’. It seems reasonable to postulate this, as without dimensions eg ‘has wings’, so semantic relations are some commonality of experience, there would be no basis required. on which to develop communication at all, and perhaps, one might argue, no reason to communicate. In the 3.2.3 Analysis example discussed, we were able to pair a set of stimuli with a word (‘swan’), and to isolate a particular element • Acquiring completely abstract concepts of several sets of stimuli with a word (‘red’). This does • Using semantic relations within abstract domains not necessarily mean that everyone has exactly the same subjective experience of these stimuli; however, As the conceptual landscape is built up, it is possible to everyone has to have the same kinds of mappings in order move away from physical stimuli altogether and use the for this to work. previously abstracted concepts as raw material to build new totally abstract concepts, with language assisting the For instance, if I experience that ‘orange’ is, on some process. This is the essence of Robinson’s analysis dimension of my sensory experience, between the method, which gives a method for building a new concepts I have learnt as ‘red’ and ‘yellow’, but you construct based on existing ones. experience it as being between the concepts ‘hot’ and ‘cold’, then we will be unable to understand each other. In the examples for ‘octagon’ and ‘diagonal’, we are But if you experience it as between ‘red’ and ‘yellow’, it given instructions for building these mental constructs by would not matter whether your subjective experience of using the existing mental constructs ‘polygon’ and 201 these concepts was like my subjective experience of ‘hot’ a. Human perceptual and cognitive apparatus : and ‘cold’, as long as there was some homomorphic There is commonality in the way we perceive sensory mapping between our experiences, such that the internal stimuli, as evidenced by the ability to identify specific semantic relations hold in both cases. Thus the theory deficits such as colour blindness or hearing loss in requires some predictable, common kind of relation individuals, and the ability to identify a common between the stimuli and the nature of the conceptual underlying topography with which we can explain landscape it evokes across individuals, but does not different mappings of perceptual words: for instance require that each individual has exactly the same colour terms in different languages. Gardenfors’ theory subjective experience. of conceptual spaces (2000) identified the structure of The issue of homomorphic semantic relations becomes shared spaces underlying perception of colour and taste, clearer when we consider more complex phenomena, like for instance. events and higher order constructs. With events, for b. “Basic level” categories: instance, if you perceive time as a linear sequence moving in a particular direction and I perceive it as a Rosch’s work (1976) showed that there is a level of basic jumbled sequence with no particular direction, then our categories of objects determined by overall gestalt conceptual landscapes cannot be homomorphic, and we perception without distinct feature analysis. These basic will not be able to understand each other: words like level categories are predicated on common methods of ‘before’ and ‘after’ will make sense to you but not to me, sensori-motor handling. or at least not in the same way, as I will not be able to abstract the common feature which marks these out c. Challenges and Experiences of surviving in the within my conceptual landscape. Consequently, I will physical environment: also be unable to use past, present or future tense in Johnson (1987) showed that there are common patterns of language in any consistent way. organisation arising from the physical environment which he called ‘image schema’. These schema, which include In extremely abstract scenarios, like law or politics, it is frame-like templates for experiential gestalts such as clear that without agreed common ground, accurate ‘container’ and ‘path’, arise directly from physical communication is impossible. In this case, the common ground is not naturally arising but is almost entirely experience and the need to operate effectively within it, and might be postulated to be common to all peoples. constructed: clearly a legal or political system does not arise as a natural part of the conceptual landscape, but is d. Early stages of Cognitive Development : based on many abstractions and analytic constructs. Even to define ‘guilt’ and ‘innocence’ requires a great Piaget (1972), for instance, was able to identify several deal of pre-existing mental architecture, and given exactly common stages of normal cognitive development. Eac the same information, does not lead to predictable stage is characterised by the distinctions and realisations categorisation of the guilty and the innocent by different the child is able to make: eg differentiating self from individuals, (eg different jurors) because things like value objects; realizing things continue to exist when no longer judgements and the relative importance of various factors, present (‘object permanence’), etc. The cognitive stages which may be weighted differently by different pass from sensori-motor through to formal operational. individuals, come into play. The earlier stages, through to concrete operational, which is usually achieved by age 12, may be postulated to be So conceptual landscapes clearly do have significant common, whilst the highest stage (formal operational) is individual and cultural differences, according to the characterized by purely abstract thinking and is likely to particular ways the culture or the individual has built up be divergent. higher order constructs. But at the same time, there is clearly a basic level where there is common ground, e. Words common to all languages: otherwise we would be unable to communicate, and If the same words arise ubiquitously across languages, totally unable to create highly abstract shared domains they would seem to mark out some common conceptual like political and legal systems. territory. Anna Wierzbicka’s work on Semantic Primes Therefore, in order to establish a science of definition, we (1996) identifies 61 such words including ‘I’, ‘you’, ‘this’ must firstly establish what and where the common ground and basic notions of time and space, which she has found is, and alternately, where and how the points of departure to be present in all the languages she has investigated. from it arise. In this way, we can establish a basis for f. The “semantic core” of English: marking out conceptual landscapes in a systematic way, which will support our endeavour of establishing robust In the 1930’s, CK Ogden was able to identify a core methods for definition. vocabulary of 800 words of English by which, he claimed anything of a practical everyday nature could be commun As a first pass to identifying a shared conceptual icated. Ogden made the reduction by eliminating verbs landscape at the most primitive level, let us consider what to basic operations like ‘have’ ‘do’ and ‘be’, and reducing could be common to all peoples of all cultures? In the nouns by means of a dimensional analysis he called the endeavour of uncovering the nature of the underlying ‘panoptic method’: a word like ‘puppy’ could be removed common conceptual landscape, we suggest the following as it could be expressed with ‘young’ on the age avenues may be fruitful for investigation: dimension in combination with ‘dog’. In all, Ogden identified twenty semantic dimensions. Chapter 7 Foundational Issues in Meaning

3.4 Higher Order Constructs and Points of 5 Acknowledgements Departure This work is part of the NICTA-wide priority challenge In ‘Metaphors We Live By’, Lakoff & Johnson provide a “From Data to Knowledge”. I would like to thank my convincing argument that higher level understandings and supervisors for allowing me the freedom to pursue the abstractions develop as metaphors from lower ones. For ideas presented in this paper. I would also like to thank instance, an image schema like a path, learnt in relation to James Nicholson, without whom none of this would be the physical domain, may then applied to an abstract, possible, and my brother John Cregan and soon-to-be non-physical domain such as planning to achieve a goal. sister-in-law Nadja for their love and support. Investigation of the abstraction process, and the foundations on which higher order mental constructs are 6 References built, may assist in identifying the points of divergence in conceptual landscapes. Gärdenfors, P. (2000): Conceptual Spaces: The Geometry of Thought. Cambridge Massachussetts, MIT Press. Some divergences may be due to lack of differentiation, Johnson, M. (1987): The Body in the Mind: The Bodily or to underlying value judgements. On the first count, for Basis of Meaning, Imagination and Reason instance, there is evidence that autistics do not . Chicago, differentiate between animate and inanimate objects, and University of Chicago Press. thus do not develop the folk concepts of belief and desire Kelly, G.A. (1955). The Psychology of Personal which are used to understand behaviour, so they never Constructs. New York: Norton. develop this part of the conceptual landscape. With more abstraction there are more possible differentiations, Lakoff, G., & Johnson, M. (1980): Metaphors We Live and if people lack the ability or interest to make these, By. Chicago, University of Chicago Press. particular abstract domains, like higher mathematics, will Ogden, C.K. (1930): Basic English: A General be inaccessible. On the second count, some ideologies, Introduction with Rules and Grammar. London: Paul value judgements or social beliefs like ‘all men are Treber & Co., Ltd. created equal’ may be pivotal in constructing the mental landscape, and conceptual landscapes may diverge from Piaget, J (1972): The Principles of Genetic Epistemology these kinds of ontological commitments. New York: Basic Books. It is worth noting that each individual’s understanding of Robinson, R. (1950): Definition. London: Oxford the world is necessarily built on certain assumptions: University Press. without them we cannot possibly hope to build many of Rosch, E., Mervis, C.B., Gray, W.D., Johnson, D.M., & our most common concepts. We should not apologize for Boyes-Braem, P. (1976): Basic objects in natural or hide this, but openly acknowledge and identify them. categories. Cognitive Psychology, 8: 382–439. However, we should also attempt to use them sparingly. Particularly when attempting stipulative definition, which Thomas, S. (2005) : The Importance of Being Earnest involves building and marking out new conceptual about Definitions. BTW 2005, 560-577. Available at territories, we should attempt to use as few ontological http://www.informatik.uni-trier.de/~ley/db/conf/btw/ commitments as possible: we certainly want to be able to btw2005.html#Thomas05 represent mental activities such as beliefs, hypotheses and Wierzbicka, A. (1996): Semantics: Primes and value judgements at the appropriate level, but should be Universals. Oxford: Oxford University Press. careful not to build these in at the foundations.

4 Conclusion In conclusion, let us review whether a science of definition is achievable in light of our discussion. Development of a ‘science of definition’ would require an understanding of common semantic ground and the ways in which divergences from it are made. We have made some progress in delineating the ways in which the conceptual landscape is naturally structured, and the means by which abstract constructs may overlaid or built within it. Taking a cognitive viewpoint, we have suggested some paths for investigation to determine where common semantic ground may be located and how and where divergences from it arise. We have mentioned some avenues, including empirical studies, which may be of use in inferring the elements and structure of the conceptual landscape. By putting together the various strands, it may well be possible to describe an underlying shared conceptual space which would provide a viable basis for robust methods of definition. 203

7.2 Symbol Grounding for the Semantic Web

Title of Publication: Symbol Grounding for the Semantic Web Type of Publication: Conference Paper Appears In: Proceedings of the European Semantic Web Conference, ESWC2007: 429-442 (2007) Publication Date: 2007 Peer Reviewed: Yes Contributing Author(s): Anne Cregan Personal Contribution: 100% Chapter 7 Foundational Issues in Meaning 205

Symbol Grounding for the Semantic Web

Anne M. Cregan1,2

1 National ICT Australia (NICTA) 2 CSE, University of New South Wales, Australia [email protected]

Abstract. A true semantic web of data requires dynamic, real-time interoperability between disparate data sources, developed by different organizations in different ways, each for their own specific purposes. Ontology languages provide a means to relate data items to each other in logically well- defined ways, producing complex logical structures with an underlying formal semantics. Whilst these structures have a logical formal semantics, they lack a pragmatic semantics linking them in a systematic and unambiguous way to the real world entities they represent. Thus they are intricate "castles in the air", which may certainly have pathways built to link them together, but lack the solid foundations required for robust real-time dynamic interoperability between structures not mapped to each other in the design stage. Current ontology interoperability strategies lack such a meaning-based arbitrator, and depend instead on human mediation or heuristic approaches. This paper introduces the symbol grounding problem, explains its relevance for the Semantic Web, illustrates how inappropriate correspondence between symbol and referent can result in logically valid but meaningless inferences, examines some of the shortcomings of the current approach in dealing effectively at the level of meaning, and concludes with some ideas for identifying effective grounding strategies.

Keywords: Ontology Alignment, Semantic Interoperability, Semantic Web, Symbol Grounding

1. Introduction

The purpose of the World Wide Web is to share and leverage information. But information is only ultimately useful if it produces some result in the real world, either in the physical environment, or in someone’s state of understanding. Raw unprocessed data is not very helpful in this regard, as it requires significant human effort, and the application of implicit human knowledge to understand it and process it appropriately to produce tangible benefits. It is generally agreed that machines should be doing more of the work of turning data into knowledge in a way that supports the production of results for human benefit. The purpose of the Semantic Web is to address this; its stated objective being to make information more easily shared and applied, by making its meaning explicit [1]. The implicit assumption is that once meaning is represented explicitly, machines will be able to align and process Chapter 7 Foundational Issues in Meaning

data according to its meaning, thus turning it into knowledge, and supporting web services and intelligent agents to produce real-world results on our behalf.

However, this implicit assumption has not yet been thoroughly investigated. To date, information processing has been based on a symbolic processing paradigm, and to process information at a semantic level requires a fundamental paradigm shift. New methodologies, processes, and criteria for judging success are needed. Many of the techniques for aligning or reconciling meaning are already known from programming, but not at a mature level where meaning is made explicit and machine processing does the rest: it requires a human being to analyze the meaning and devise and implement appropriate code to do the necessary transformations.

How are we to start making inroads into this new semantic territory? As an initial step, taking a good look at the really hard questions should help focus the effort, and provide foundations for this new information processing paradigm.

These hard questions include but are not limited to the following:

1. What is meaning? 2. What do we need to do to make meaning explicit? 3. What is the appropriate way to process meaning? 4. How can we judge whether we have been successful in representing and processing meaning at a semantic level? 5. Will the current Semantic Web approach, based on the Web Ontology Language OWL [10], produce the right kind of representations, and support the right kinds of processes, to achieve the results being sought, or is a key component of the solution missing?

Spanning from the very philosophical to the very practical is necessary because the issue of meaning is a fundamental philosophical issue, whilst the goals of the Semantic Web are very practical. By their very nature as a “specification of a conceptualization” [3], creating ontologies involves bridging between the realm of IT/ Engineering, and the realm of Cognitive Science/Philosophy. It is hoped that such as investigation can uncover the foundations for such a bridge, providing a basis not only for the Semantic Web but for the Pragmatic Web it will ultimately support.

Organization

The paper is organized as follows: − Section 1 introduces the challenge being undertaken − Section 2 relates meaning to both entailment and designation, looks at symbolization and introduces the symbol grounding problem − Section 3 explains why symbol grounding is relevant for the Semantic Web in its aim to achieve dynamic real time interoperability, and extensional approaches and URIs are not sufficient in themselves to provide adequate symbol grounding. − Section 4 considers next steps in identifying suitable symbol grounding strategies for the Semantic Web and concludes. 207

2. Meaning and Symbol Grounding

What is meaning? The greatest philosophers and thinkers have considered this question for the last several thousand years, but as yet there seems to be no definitive answer. What are the implications for the Semantic Web, which is being built around the keystone of making meaning explicit and machine-processable? Is it ever really going to get off the ground, or perhaps do so initially but quickly collapse under its own weight for lack of good foundations? It seems somewhat foolhardy to attempt to devise explicit well-defined procedures for operating at the level of meaning, without attempting to lay good foundations by stating what meaning is taken to be.

Whilst a conclusive answer to the question is unlikely (isn’t that what makes a good philosophical question after all?) and Semantic Web researchers are, generally speaking, practical people who want results in reasonable timeframes and certainly don’t want to get bogged down in the vagaries of philosophy, I believe that as part of the construction of Semantic Web technologies, for purely practical reasons, there should be some attempt to state what we take meaning to be for the purposes of the Semantic Web.

A clear conception of meaning for the purposes of the Semantic Web should, at the very least, assist researchers in devising appropriate and precise procedures and methods for making meaning explicit, which then has the flow-on effect of supporting practitioners to build appropriate semantic models representing their respective domains, and will make such models better suited for interoperability. It also provides a theoretical basis for semantically processing the information captured by such models.

Whilst many modeling errors have been identified and are well understood e.g. [8], there is still quite a spectrum of “correct” models available for modeling any given domain. The ontology builder has considerable discretion in make design choices. Are some of the resulting models better than others? Intuitively the answer is yes, and depends on the intended function of the ontology. However, we are still seeking a more precise understanding of the nature of this dependence, and at the moment there is no one clear guiding methodology for building domain models. Whilst there are obviously several factors at play, the model’s effectiveness in making meaning explicit should certainly be considered a key criterion.

Beyond this, an analysis of meaning also offers insights into the overall Semantic Web approach and whether it will ultimately be able to deliver on its promises. Capturing meaning is clearly a fundamental component, but are the current suite of Semantic Web standards and technologies adequate to the task of capturing machine- processable meaning to produce the outcomes being sought, or will they ultimately fall short? If we want to ultimately build a “Pragmatic Web”, that delivers tangible real-world benefits, we need to make sure the foundations are firm enough to support this. Chapter 7 Foundational Issues in Meaning

2.1. What is Meaning?

Without getting too bogged down in philosophy, let’s take a practical approach to home in on what meaning is, by identifying what it is that we really want when we ask the meaning of something. In everyday life, we generally don’t ask the meaning of concrete things like a chair, or a train, or a person, or a pet. Such things have no meaning: they just are. We ask the meaning of actions and events, policies and such like, in which case we are generally trying to identify the relevant entailments, or we ask the meaning of symbols, in which case we want to know what they designate, or stand for. When we ask about meaning, we are usually asking for one of two things: either for entailment, or for designation.

2.2. Entailment and Designation

Entailment: What are the logical consequences of some action, event or state? Examples: − If I take this promotion does it mean I will be able to afford the house? − If Serena wins this point, does that mean she wins the match? − If my business is registered as a public company, does that mean we are required to have annual audits conducted?

Designation: What is being referred to? What does the symbol symbolize? Examples: − What’s the meaning of “verisimility”? − What does that sign mean? − What do you mean by giving me that wink? − What does the green line on the graph mean?

Designation gives the referent being represented by some kind of symbol: a word, a street sign, a gesture, a line on a graph. It uses symbols to point to something; a convenience originating from the need to identify and communicate something that does not have a local physical existence, is abstract, or is an internal state and not directly accessible. Designation is the back end of symbolization: it establishes the referent, or what the symbols symbolize.

Symbolization

Note that there are (at least) two related senses of symbolization. In the first sense, a recognizable concrete thing is used to stand for a more abstract intangible thing e.g. a dove is used to symbolize peace. It is usually chosen as a symbol because it has some kind of real-world historical or mythological relationship with the abstract thing, or evokes it through some other kind mental or perceptual association, or it can simply be a matter of convention. In this sense, a non-verbal meaning relation is pre- established and the symbolization makes use of it to evoke the intended referent. This should not be confused with symbolization as used in this setting, which is being referred to as “designation” for the purposes of clarity within this paper. 209

In our setting, symbolization refers to the scenario where a mark, character, sound, avatar or some such arbitrary thing is used to designate some physical or conceptual thing. In this case, the symbol is an arbitrary physical token, designed by humans specifically for the purpose of representation, and does not usually have a meaning in and of itself (Although in the case of avatars, some recognizable topographical resemblance may exist, and thus their form may be argued not to be completely arbitrary). In this kind of symbolization (designation), the essential question is how the relationship between an arbitrary symbol and its intended referent is to be established. This question has been identified in Artificial Intelligence Research as the “Symbol Grounding Problem”.

2.3. Denotation and Connotation

Designation itself has two aspects: denotation and connotation, a distinction introduced by J.S. Mill [5]. To illustrate by example, the denotation of a term such as ‘woman’ refers to all the individuals to which may correctly be applied, whilst the connotation consists of the attributes by which the term is defined e.g. being human, adult and female. Connotation determines denotation, and in J.S. Mill, is taken to be meaning, whereas terms like proper names e.g. ‘Mary’, which have denotation if there is someone so called, are taken to lack meaning as they have no connotation, as no attributes define ‘Mary’.

2.4. Relevance to the Semantic Web

Both entailment and designation have relevance for the Semantic Web: entailment relating to what can be concluded from what is already known, and designation relates to establishing the connection between symbols in a formal system and what they represent. There is already a very significant body of work around entailment for the Semantic Web [10], based on description logics providing an underlying formal semantics for the various flavours of OWL.

However, designation has had less attention to date. OWL’s formal semantics have a set-theoretic basis, where a set (‘concept’ or ‘class’ in DLs) is essentially defined by its extension - clearly a denotational approach. However, meaning based on denotation is less than adequate for the needs of the Semantic Web, as will be explained below. The consequence is that the entailment parts of the Semantic Web have no theoretical basis for anchoring to anything in the real world, and are thus floating castles in the air.

To explain: an OWL ontology is made up of a set of logical axioms, themselves composed of primitive objects, predicates and operators, combined via formation rules into well-formed formulae. Unless some kind of faithful and appropriate correspondence is established between the primitives and whatever they are intended to represent outside the formal logical system, any entailment produced by the system will not result in reliable conclusions that correspond to the actual state of affairs in Chapter 7 Foundational Issues in Meaning

the real-world domain of interest. Establishing a correspondence between the primitives (which are effectively just symbols or symbol strings once they are inside the logical system), and the domain is an extra-logical consideration. The question of how the relationship between the symbol and the referent is to be established has been identified in Artificial Intelligence Research as the “Symbol Grounding Problem”.

2.5. The Symbol Grounding problem

The Symbol Grounding Problem, as described, for instance, by Harnad [4] relates to the inadequacy of defining symbols using only other symbols, as is commonly done in a dictionary or a formal logical system. In his exposition, Harnad takes Searle’s [9] famous Chinese Room scenario, originally used by Searle to illustrate the difference between mechanical symbol manipulation, which merely simulates mind, and a true understanding of intrinsic meaning, which necessarily involves processing at the semantic level.

The scenario involves a machine hidden inside a room, which is given a set of Chinese language inputs and produces a set of Chinese language outputs. Searle points out that a machine using only symbolic manipulation to match a list of pre- defined inputs with a list of pre-defined outputs may be capable of simulating conversation with a Chinese speaker well enough to pass the Turing test. However, Searle argues, such a machine cannot be said to understand Chinese in any sense, any more than a human who uses such a list to produce statements in Chinese can be said to understand Chinese. Searle ultimately concludes that meaning is in the head, not in the symbols, and furthermore that cognition cannot be just symbol manipulation, as it clearly requires some activity to take place at the semantic level.

Harnad puts an alternate spin on Searle’s Chinese Room scenario, asking the reader to imagine having to learn Chinese as a second language, where the only source of information available is a Chinese/Chinese dictionary. He observes that “The trip through the dictionary would amount to a merry-go-round, passing endlessly from one meaningless symbol or symbol-string (the definientes) to another (the definienda), never coming to a halt on what anything meant.” He then presents a second variant, where one has to learn Chinese as a first language, and again the only source of information available is a Chinese/Chinese dictionary. He argues that if the first variant is difficult, then the second must be impossible, relating it to the task faced by a purely symbolic model of the mind, and asking “How can you ever get off the symbol/symbol merry-go-round? How is symbol meaning to be grounded in something other than just more meaningless symbols? This is the symbol grounding problem.”

How indeed, are we to get off the Symbol/Symbol merry-go-round? Firstly though, let us consider in detail how the symbol grounding problem is relevant for the Semantic Web.

211

3. Why the Semantic Web needs Symbol Grounding

The ultimate vision of the Semantic Web is a web of data connected by meaning which is machine processable. The idea is to get meaning out of the technologists and domain expert’s heads, and into some explicit, machine processable representation which defines how to link it up appropriately, in real-time, without reference to human mediators. But the current Semantic Web building blocks are a long way from achieving this vision. Let’s take a look at why this is.

In building an ontology, the designer chooses terms for classes, instances and properties, and builds axioms/structure linking them. The terms are usually chosen for their meaning in some natural or domain-specific language. Additional annotations may explain the meaning of the term, using more natural and/or domain specific language. But natural language is notoriously ambiguous and slippery. Its symbols/semantic units are imperfectly grounded, as we will explain in the following section. And whilst domain specific-terminology may be unambiguous within the domain, it is not necessarily unambiguous when linking across domains.

If the basic terms used for ontologies are ambiguous, then having a well-defined structure that supports entailment is of dubious benefit. The structure by itself is not the meaning: as discussed, meaning requires both logical structure for the purposes of entailment, and grounding for the purpose of establishing correspondence between the domain and the logical structure. Only then can entailments made by virtue of the logical structure be guaranteed to be an accurate reflection of the real-world state. Garbage in, garbage out, as the old saying goes.

3.1. Meaningfulness

As an example of this principle, there is a considerable body of work e.g. [6] in Mathematical Psychology around determining which kinds of variables can be subjected to which kinds of mathematical operations, in order to produce only meaningful results and avoid meaningless conclusions. Considerable effort has gone into investigating “meaningfulness” to avoid the inappropriate use of statistics. In a classic example, the school football team are assigned numbers to wear on their football jerseys. This numerical assignment is simply to give each football player a unique label for the purpose of identification. However, this assignment does not support taking the average of those numbers, and asserting that this average reflects some meaningful property of the football team. This is because the numbers have no numerical properties attached to them - they are just labels, and could equally well be any other arbitrary symbol (letters of the alphabet, pictures of animals) - the only important factor is that each player is designated by a unique symbol. The underlying variable being represented (identity) is not a quantitative variable, so any mathematical inferences derived from the football jersey numbers are simply meaningless. This is not the case for a quantitative variable like the heights of the football players, where it is perfectly appropriate to represent heights as real numbers and calculate average and standard deviations in the height of this population. Chapter 7 Foundational Issues in Meaning

Note that the “meaningfulness” criteria is not necessary because of any problem to do with numbers themselves, or with mathematical reasoning. The problem is that the real-world dimension being represented does not have the same properties as the chosen representation. The representation is richer and more structured than what is being represented, and thus permits reasoning and inferences which have no correspondence with the real-world. Inferences which are perfectly valid inside the representation symbol are thus meaningless when we attempt to map them back to the domain of reference. As a formal logical system without appropriate grounding strategies to connect it to the real-world, the Semantic Web faces a similar problem.

3.2. Semantic Interoperability Problems

Pollock and Hodgson’s analysis of types of semantic conflicts [7] identified eleven kinds of semantic level clashes: DataType, Labeling, Aggregation, Generalization, Value Representation, Impedance Mismatch, Naming, Scaling & Unit, Confounding, Domain, and Integrity. This analysis has been adapted and re-organized to fit the Semantic Web and the focus of the current investigation.

PQ P Q P Q ! " & ' # $ % % % " ' % ( ) ' $ % ( *

" ) % - .S ( $ ! 0%1 . ! + ' ! ). , " ) " )2 0%1 ( ! % 0 # - $ * " 3 ) ( - ( 454 ' P Q

213

. + % 6 ! P'Q ', S ! ! % 7 % $ ' 8 % 9 8! : %

0 6 # % % ! , ! ! %

As semantic conflicts, every one of these problems requires reference to the meaning level of the information for its resolution. The following table looks at the resolution method for each class of semantic conflict and indicates how Symbol Grounding and Meaning is relevant in each case.

! " - ; % % ( % % 8 " < ! % % % = % 3 # < 9 + > 6 %+ > # % ! % % > Chapter 7 Foundational Issues in Meaning

! " . + -9 , - ', , % > , > 7 -9 6 8 , > > , 0 -9 6 , > , 1 >

3.3. Current Support for semantic interoperability conflict resolution

As the analysis of the previous section shows, symbol grounding and meaning is at the heart of these interoperability problems. Resolving these kinds of problems are common occurrences in mapping and aligning ontologies as they exist today. Whilst tools and heuristics are available to humans to assist the process, it is a problem essentially being addressed by human beings, rather than machines. Somehow, the underlying meaning needed to resolve these interoperability problems is not being explicitly represented in the Semantic Web, and/or there are not sufficient tools and techniques available for resolving it through automated processing. The remainder of this section makes an initial pass at identifying where the shortcomings are, starting with specifics of the OWL language and broadening out from there.

Mapping Constructs provided by OWL

OWL provides only very limited constructs for mapping ontologies, and essentially none for transformations. The OWL language has four constructs specifically for use in mapping ontologies for the purposes of merging. These constructs are intended for use when two or more ontologies have been built independently of each other and later need to be merged or linked, i.e. they are being mapped after the initial design phase has taken place. They are:

owl:equivalentClass asserts that two or more classes have exactly the same instances

owl:equivalentProperty asserts that two or more properties are equivalent

owl:SameIndividual asserts that two or more individuals are identical owl:DifferentIndividuals asserts that two or more individuals are identical 215

Whilst these constructs provide the means to specify that two or more classes, properties or individuals are equivalent / identical, or that two or more individuals are different, they are only useful once the equivalence, identity or difference is determined. This determination is outside of OWL, and the Semantic Web technologies do not provide any formal basis for a machine to determine this without human guidance (albeit supported by tools and heuristics). The use of class extensions and URIs is insufficient, as explained below.

Extensive vs Intensive Class Definitions

Some may argue that if two classes contain the same set of individuals, they must be the same, and machine processing can determine this. However, having the same extension does not necessarily prove that two classes are the same: they may happen to contain the same individuals, but have different intensive meanings: that is, the criteria for membership of the class is different. Philosophically, this is the denotation vs connotation distinction.

For example, the extension of the members of the school basketball team in a Sports Ontology and the extension of the school’s Grade A students in an Academic Ontology may conceivably, at some point in time, consist of exactly the same set of individuals, and machine processing may determine on the basis of these equal extensions that the two classes must be equivalent, and map them using owl:equivalentClass.

The possible ramifications of such an ill-advised mapping are obvious: if a student drops his grade, he will find he is no longer classified as a member of the basketball team. If she drops out of the basketball team, she will no longer be classified as a Grade A student. If a new student is added to the class of Grade A students, he will find himself automatically in the basketball team. A new member of the basketball team will find she is automatically classified as a Grade A student.

Clearly, the extensive approach to class definition and mapping is inadequate, especially when changes in the variables used for classification occur, or new, unclassified instances are encountered, simply because the classification criteria are not adequately captured. A complete specification of meaning needs to support a decision procedure which determines whether a previously unseen instance qualifies for membership of the class or not, by making the membership criteria explicit. We also note that it is the nature of some classes to have fuzzy boundaries [2] (e.g. the colour red), and support may be needed for graded class membership in such cases.

Representing Mappable Differences/Transformations

OWL also lacks the means to specify semantic differences and the transformations needed in a way that support interoperability. Programming code can get around this, but has to be written for each specific situation, straying from the Semantic Web ideal of explicit representation enabling automated processing at the meaning level. Chapter 7 Foundational Issues in Meaning

As a simple example, a temperature in Fahrenheit and a temperature in Celsius can easily be converted either way via a simple arithmetic equation. The underlying measurement scales are interoperable, but there is no support for representing such a relationship within an ontology. Assuming that both ontologies map temperature as a DataProperty to a real number value, the numerical values of the respective properties differ, but have a well-defined arithmetic conversion. However, OWL provides no mechanism for specifying the scale, the properties of the scale or for doing such a transformation. Meaning that is available to the designers, and could be made explicit in building the respective ontologies, subsequently supporting interoperability via machine processing, is currently not able to be represented within the ontology.

Independent and Dependent Ontologies

The mapping constructs in OWL for mapping independently designed ontologies were considered above. However, because OWL ontologies are built from URIs, the components can reside anywhere. The principle of composability in OWL means that when an ontology is being built or revised, the designer can freely use any constructs from any other available ontology. This results in one ontology having logical dependence on another. Interoperability is an issue for both dependent and independent ontologies, and some of the broader interoperability concerns are addressed below.

Dependent Ontologies: If a base ontology changes, other dependent ontologies are affected. This is potentially much more serious than just a “broken link” because it can potentially change the inferences made. In the World Wide Web, broken links and web page changes are not critical, they are simply a dead-end and one can always look for information elsewhere. However, when an external URI is an essential part of a logical structure, a change or deletion can have serious real-world consequences, such as an incorrect classification as an illegal alien for example.

Independent Ontologies: In the case where two independent ontologies need to be aligned or mapped and have no common constructs other than the Ontology language itself, currently the options for resolution are either heuristic approaches with varying success rates, or the human designers of the respective ontologies can communicate with each other or check other sources to establish the meanings of terms devise an appropriate mapping. The problem here is not only that this mapping problem is potentially in the order of N2 to achieve interoperability across the semantic web, but the problem is an N2 human-to-human problem. The resulting reward to effort ratio causes pause for reflection.

3.4. Why using URIs is not a sufficient grounding strategy

A commonly encountered argument is that Unique Resource Identifiers (URIs) can be used for any disambiguation needed for the semantic web. This is Sir Tim Berners- Lee’s own view (conversation with author, November 2006). After all, anything can 217 have a URI, why not simply use that as a unique identifier? If two concepts, individuals or properties have the same URI, they must therefore be the same. Problem solved!

Whilst this approach has its merits, it is not sufficient in itself to resolve all the kinds of semantic interoperability problems. Granted, it can identify cases where two ontologies are referring to the same thing, but it cannot identify what that same thing is that they both refer to, or that either of them are representing or processing it appropriately based on its meaning. This is because the URI does not have a grounding mechanism to connect it to anything outside the information system: A textual description residing at the URI or within the URI itself is natural language and thus subject to ambiguity and vagueness.

The exception, of course, is when the thing being referred to is, exactly, the information that resides at the URI. For instance, an identifier for a particular tax law can be grounded to a URI that contains the exact text of the tax law, and thus there is no need to go further. But if the URI is intended to reference a real world physical thing, like a person or a building, or something else outside the information system itself, it needs a symbol grounding strategy.

A further consideration when considering an information system on the intended scale of the Semantic Web, is who is going to check that every URI maps to one and only one real-world referent? No doubt many things will have more than one URI, in which case we still have the human problem discussed earlier, of determining that the URIs have the same referent and mapping them, which obviously cannot be resolved using the URI itself. And on the flip side, there will no doubt be many real-world things that have no corresponding URI, so the system is incomplete.

4. Next Steps and Conclusions

If Searle is right, cognition cannot be reduced to symbol manipulation. Semantic processing therefore requires an understanding of cognition in regard to meaning. The next steps underway in this line of research are the investigation of a wide range of grounded symbol systems, such as Musical Notation, Cartography, Chess Notation, Circuit Diagrams, Barcodes and even Knitting Patterns. These are being analyzed to determine the grounding strategies used, and how and why they are effective or ineffective. The analysis will identify the kinds of grounding strategies available, and determine appropriate criteria for assessing them, and it is hoped it will provide a theoretical basis for constructing Symbol Grounding strategies for the Semantic Web will be identified. Following this, the question of devising appropriate processing and procedures to produce meaningful results will be addressed.

In conclusion, this paper has put forward some of the hard questions the semantic Web needs to answer, examined some of the pitfalls that may occur if they are not addressed, and explained the relevance of the symbol grounding problem for the kinds Chapter 7 Foundational Issues in Meaning

of semantic interoperability issues commonly encountered. Some insights from measurement theory in Mathematical Psychology were briefly covered to illustrate how inappropriate correspondence between symbol and referent can result in logically valid but meaningless inference. Some of the shortcomings of the current Semantic Web technologies in dealing effectively at the level of meaning level were investigated. The arguments that set extensions and URIs can provide an appropriate basis for grounding the Semantic Web were considered and found wanting. Finally, next steps for identifying effective grounding strategies and doing meaning-level processing were briefly discussed.

Acknowledgments. To my infant nephew, Gaelen Raphael Bonenfant, for providing a beautiful example of how the grounding process gets started.

Research reported in this paper has been partially financed by NICTA (http://www.nicta.com.au). NICTA is funded by the Australian Government's Department of Communications, Information Technology and the Arts, and the Australian Research Council through Backing Australia's Ability and the ICT Centre of Excellence program. It is supported by its members the Australian National University, University of NSW, ACT Government, NSW Government and affiliate partner University of Sydney.

5. References

1. Berners-Lee, T., & Fischetti, M. (1999). Weaving the Web. SanFrancisco, Harper. 2. Gärdenfors, P. Conceptual Spaces. (2000) MIT Press, Cambridge MA. 3. Gruber, T. R. (1993) A translation approach to portable ontologies. Knowledge Acquisition, 5(2), 199-220. 4. Harnad, S. (1990). The Symbol Grounding Problem. Physica D, 42: 335-346. 5. Mill, J.S. (1843) A System of Logic, London. 6. Narens, L.E (2002) All you ever wanted to know about meaningfulness. Volume in the Scientific Psychology Series, Lawrence Erlbaum Associates, Mahwah, NJ. 7. Pollock, J.T. and Hodgson, R. (2004). Adaptive information: Improving business through semantic interoperability, grid computing, and enterprise integration. Wiley Series in Systems Engineering Management, John Wiley & Sons, Inc. 8. Rector, A., Drummond, N., Horridge, M., Rogers, J., Knublach, H, Stevens,R., Wang, H. and Wroe, C (2004). OWL Pizzas: Practical Experience of Teaching OWL-DL: Common Errors & Common Patterns. EKAW2006 Proceedings, (pp. 63-81). Available at www.co-ode.org/resources/papers/ekaw2004.pdf 9. Searle, J. (1980). Minds, brains and programs. Behavioral and Brain Sciences, 3: 417-457. 10.Smith, M.K., Welty, C., & McGuinness, D.L (eds) (2004). OWL Web Ontology Language Guide, W3C Recommendation, 10 February 2004. Available at http://www.w3.org/TR/2004/REC-owl-guide-20040210/. Latest version available at http://www.w3.org/TR/owl-guide/ 8 Conclusion

In this ﬁnal chapter, the conclusions of each publication are summarized, the contributions are reassessed with regard to the problems they addressed and the current state of development, and outstanding issues and future work required are highlighted.

8.1 Overview of Semantic Technologies

The “Overview of Semantic Technologies” chapter presented in Chapter 2 concluded that whilst there are still open issues and challenges, and work still in the development phase, semantic technologies are already sufficiently developed to be applied in real-world scenarios, to produce interoperability and other benefits. This was shown by the included FEA case study. Since publication, the Semantic Web stack of technologies has progressed, with SPARQL [132] becoming a W3C recommendation in January 2008, and work continuing via W3C working groups on OWL 2 [108] and the Rule Interchange Format [156]. The conclusion also pointed to the potential for Semantic Technologies to produce intelligent information services for finding, analyzing and using knowledge, and that ultimately this would provide a basis for effective, informed, automated action. This is borne out by the progress made by the examples covered at Section 1.3.3 [39, 52], albeit acknowledging further gains yet to be realized. A number of issues for Semantic Technologies were raised by the paper, and these are now reassessed in light of recent developments since the time of writing. The first issue raised the need for data owners to semantically markup their data, and the lack of automated and other techniques to ease the process. The W3C’s GRDDL ‘Gleaning Resource Descriptions from Dialects of Languages’ [33], eases the transition, by enabling RDF to be embedded in XML and XHTML documents, and to be extracted from them using XSLT. This makes it possible to add semantic content to existing web documents, and provides better integration with the existing web protocols and standards. GRDDL became a W3C Recommendation in September 2007. The potential to do automated semantic markup has taken quite a large step forward with the advent of Reuter’s Open Calais [134]. Open Calais provides a web service that automatically creates rich semantic metadata for user submitted content. Using natural language processing, machine learning and other methods, Calais recognizes and tags Named Entities, Facts and Events. Whilst it is unlikely to do markup quite as accurately as a human, it provides the ability to tag vast amounts of content with reasonable accuracy without human intervention, or alternately may be used as a first pass by a human conducting semantic markup and ontology building, to speed the process. The second issue raised the need for better tools and techniques for supporting large scale

219 Chapter 8 Conclusion semantic applications. In particular, there is a need to use traditional database technology in conjunction with ontologies, so that one may conduct semantic search and querying guided by ontologies but executed by database technology. However, the complexity of ontologies in relation to traditional database schemas requires a novel approach for fitting ontologies and databases together to achieve the best of both worlds. The work of Calvanese et al on query rewriting [128], and Franconi et al on the ICOM tool [40] are promising avenues for providing as much as possible of the expressivity of ontologies for conceptual modelling in combination with the performance achieved by traditional data warehouse technology for very large data sets. Enrico Franconi is currently visiting NICTA, and the author is discussing possible joint work on developing this approach with him. The third issue related to the need to better support non-logicians to participate in ontology building, particularly via the use of (controlled) natural language syntaxes. My own joint work on designing a Controlled Natural Language syntax for OWL addresses this need, and was covered at Chapter 5. The fourth issue was the need for better standards and methods for resolving meaning without recourse to human intervention. One possible solution identified was the use of jointly developed ontologies for use as common standards within a domain, to reduce the need for pairwise ontology mapping simply by commitment to a common ontology. Com- mitment to common ontologies may be encouraged via the approach described in Chapter 6 and the proposed n2mate tool. The second approach mentioned the need for a better underlying semantic theory that would encompass a broader view of semantics that would encompass cognitive aspects of meaning. The publications included at Chapter 7 address this, and further comments on the approach follow at Section 8.6. The fifth issue was the need for better methods for dealing with incomplete, uncertain and probabilistic data, for all those commonly encountered real-world situations where statements do not neatly map to a boolean truth value. In regards to this, we note that firstly that work on probabilistic extensions to OWL has been progressing, and Pellet [96] now offers an extension called “Pronto” that enables probabilistic knowledge representation and reasoning in OWL ontologies. Additionally, a W3C incubator group called ‘Uncertainty Reasoning for the World Wide Web XG’ (of which the author was a member), has been exploring representing and reasoning with uncertain information on the World Wide Web, and published its initial report at the end of March 2008 [155]. The final issue was the need for appropriate methods for dealing with proof, trust and security over semantically interlinked virtual data structure. Appropriate measures are needed to handle who can access what data and at what level, and for intelligent agents/reasoners to be able to account for and justify their actions. The W3C’s Policy Languages Interest Group (PLING) [153] has formed to address these kind of issues, as well as those of digital rights more broadly. The author’s NICTA colleague, Renato Iannella, is the PLING chair, and the author has been involved in some preliminary discussions with him about the creation of projects within NICTA which may assist in addressing these issues.

8.2 Competing Standards

One of the problems raised at Section 1.9 was that of competing standards for semantic representation, particularly those offered by the ISO under the Topic Map umbrella, as opposed to the W3C’s Semantic Web Recommendations. The publications in Chapter 3 presented a method for integrating the Topic Map standards into the Semantic Web by proposing an OWL-DL formalization of the ISO drafted 221 standard Topic Map Data Model [51], illustrating the fit between OWL and the TMDM, and the potential for the use of OWL-DL as a medium for representing the TMDM constructs. The proposed formalism provided the basis for end-users to author Topic Maps in OWL- DL by populating the constructed OWL DL ontology with their own instance data. An outstanding issue relating to the implementation of Type-Instance and SuperType-SubType relations as specified in the TMDM was identified, and three approaches for implementing the required functionality were outlined: firstly the use of OWL Full (not recommended), secondly by creating extra object properties to connect classes to classes to emulate the typing behaviour required, or thirdly by adding additional Java code. The conclusion stated that the use of OWL DL and its associated tools for constructing Topic Maps provides significant advantages over previous Topic Map representations in terms of explicit specification, formal semantics, constraint checking and querying capabilities. It was noted (see Section 1.10.2) that in May 2005 the relevant ISO Working Group including TMDM authors Lars Marius Garshol and Graham Moore accepted that the proposed approach was a faithful mapping of the TMDM at the object level. However, Garshol, who is very active in authoring the Topic Maps suite of standards and tools, has later noted this as a criticism of the approach, noting in a blog in 2007 [49], that the approach is essentially an ‘object mapping’ rather than a ‘semantic mapping’, and that this is in fact the wrong approach to take. The distinction between mapping at the object level versus mapping at the semantic level features in the W3C RDF/Topic Maps Interoperability Task Force’s Survey of RDF/Topic Maps Interoperability Proposals [125] and originates from Graham Moore’s 2001 proposal [106] where it is couched as “mapping the model” vs “modelling the model”, the key difference being that the former is “semantic” whilst the latter “uses each standard as a tool for describing other models”. Essentially, an object mapping uses the language of one standard to model the other, whilst a semantic mapping states equivalences between the constructs of the two languages. For example (mine not Moore’s), an object mapping might model the TM construct of ‘Topic’ as an OWL class named ‘Topic’ (as my TMDM OWL ontology did), whilst a semantic mapping might state that the TM construct ‘Topic’ is equivalent to the OWL construct ‘Class’. When considering a semantic mapping between Topic Maps and RDF, one requires a model to model mapping that identifies those entities and relationships that are in fact the same in both models. Quoting Moore, “The question we look to answer in this paper is ‘Is the nature of the relationships and the identified entities the same’. If we can show this to be true then we will be able to have a common model that can be accessed as a Topic Map or as a RDF Model. To make this clearer, if we have a knowledge tool X we would expect to be able to import some XTM syntax, some RDF syntax and then run either a RDF or TMQL query in the space of tool X and expect sensible results back across the harmonized model.” The Survey of RDF/Topic Maps Interoperability Proposals recognizes the relative merits of semantic mapping over object mapping and recommends that it be pursued as the basis of any recommended approach to translation. However, it also recognizes that for a semantic mapping there are a number of outstanding issues in deciding how to interpret a number of aspects of Topic Maps (identity, non-binary associations and roles are mentioned), and that there are also issues when the source model contains constructs that are not directly mappable into the target paradigm. This is clearly the case when attempting to map OWL into the Topic Map paradigm, as OWL is a far more expressive language, with many constructs that have no equivalents in the Topic Map language. Chapter 8 Conclusion

The question of disambiguating aspects of the models for intermapping has been tackled by the OMG in its Ontology Definition Metamodel [55], which includes RDF, OWL and Topic Maps, as well as other modelling languages including Common Logic, and uses meta- models underpinned by the meta object facility (MOF)to describe each. This is essentially a semantic mapping using the common language of MOF to express equivalences between models. In order to determine the appropriate meta-models, the ODM authors found it necessary to work closely with the designers of each language to capture their intentions accurately. To answer Garshol’s criticism, I would argue that whilst a semantic mapping approach is preferable in theory, given the actual situation at hand, there were definite merits in the object mapping approach taken by my contributions. The semantic mapping is certainly preferable given that one has adequate meta-level information about the modelling languages with which to do it (such as the meta-models crafted as part of the ODM effort), but at that time such information was unavailable. For a semantic mapping one must also assume that there are suitable utilities in place so that once mapped, the respective languages become interoperable: then one can for instance, as Moore requires, perform a TMQL query over an RDF model and get sensible results (assuming that this is at all feasible given the nature of the mapping; it certainly cannot be taken as given). Such utilities were not in place; furthermore, in doing a semantic mapping between OWL and Topic Maps, one must face the fact that OWL is far more expressive, and thus the mapping can only ever be partial. Garshol himself comments in the same blog post that TMCL is the appropriate language to compare with OWL for a semantic mapping, rather than the Topic Map language and data model. However, TMCL was in early development in 2005 and is still unfinalized as at August 2008. If one is in the situation of having no available meta-model, nor adequate meta-level information to perform a semantic mapping, nor suitable utilities for performing such tasks, and furthermore, one language is far more expressive than the other and has considerably more mature tools and capabilities, then an object mapping to the more expressive, mature language can provide considerable value by making its tools and capabilities available to the other, as my approach did. My approach compensated for the lack of having TMCL and TMQL standards and tools, by using OWL instead. With OWL being a far more expressive language, it was appropriate to map Topic Maps into OWL, but of limited value to attempt to map OWL into Topic Maps or to attempt a semantic mapping, as much as OWL’s expressivity would necessarily be lost. In concluding, it is worth noting that as at September 2008, both TMQL and TMCL specifications are still unfinalized, so the possibility of performing a TMQL query over an RDF model is still only theoretical, whereas the possibility of using OWL/RDF querying languages to query a topic map implemented as an OWL-DL ontology has been available since the time of my offering at the start of 2005. Therefore, we believe it is reasonable to argue that the contribution was of some utility.

8.3 Rules and the Limits of Ontology Languages

In Section 1.9, it was pointed out that many classification tasks require ontologies to be supplemented with rules, thus the fit between OWL and rule languages and tools is an important issue. Whilst understood at a logical level, there has been a need for practical testing and illustration of how rules could be used in tandem with OWL ontologies in practice, including exploration of the impact of OWL design features, such as the Open World Assumption, inspection of the extent to which tools can adequately support ontolo- 223 gies supplemented with rules, and encouragement of additional language features and tool development where necessary. The work described at Section 1.10.3 and included in Chapter 4 addressed these needs, using a simple example OWL DL ontology to show the difficulties and challenges of implementing some intuitively simple classification tasks using OWL, SWRL and the Protégé environment. The aim was to classify a number of Student groups as GoodGroup if they satisfied all of a number of intuitively simple criteria, and as BadGroup otherwise. Whilst we experienced implementation difficulties, ultimately we were able to successfully build an OWL DL ontology with SWRL rules that in theory would successfully classify all Good- GroupsandBadGroups. The publications showed the limitations of the expressivity of OWL, and illustrated several cases where it was necessary to supplement the ontology with SWRL rules in order to achieve the desired classification. The first paper was presented at OWLED in 2005, where possible extensions for OWL DL were first discussed, and it should be noted that some of the limits of OWL DL raised in the paper are being addressed in the design of OWL 2 extensions. In particular, OWL 2 supports qualified cardinality restrictions which would support capturing Condition 2 (groups should have at least one male member and at least one female member) more easily and intuitively. OWL 2 also includes property chain inclusion axioms which make it possible to implement more classification criteria in OWL without the need to use a rule language. Reasoning support for rules has also advanced substantially. Currently, most of the available OWL reasoners except FaCT / FaCT++ [114] offer support for rules, with SWRL being the usual rule language supported, although support for RuleML (Rule Markup Lan- guage) [79] is also now commonplace. (Note: SWRL combines sublanguages OWL and RuleML [76]). SWRL is supported by reasoners Hoolet [8], Pellet [96], KAON2 [42], and Racer, also known as RacerPro [60]. SweetRules [95] now provides an open source integrated set of tools for semantic web rules and ontologies, offering support for both SWRL and RuleML, whilst Bossam [105] is a RETE-based rule engine with native supports for reasoning over OWL, SWRL and RuleML. Additionally, it should be noted that the W3C’s RIF (Rule Interchange Format) working group [156] is chartered to produce a core rule language plus extensions which together allow rules to be translated between rule languages and thus transferred between rule systems. This will provide greater interoperability in the use of rules and reasoners. Although the languages and tools have evolved, the overall experience described in the papers remains pertinent. It illustrates that ontology building is not a straightforward process, and requires a deep understanding of the underlying logical structure. Without the input of our tutor, we might easily have fallen into the trap of believing we had captured all our conditions, when in actual fact we had not. In a real world situation this could have potentially disastrous consequences. We documented several experiences that are commonplace when working with ontologies: 1. the need to backtrack in the ontology design when one is unable to easily capture desired axioms 2. unexpected results in classification, due to inadequate understanding of the logical properties of the ontology 3. the need to think very carefully about the logical ramifications of the axioms and the OWL design principles, particularly the Open World Assumption. 4. the need to create classes and properties as intermediate steps in order to express the desired axioms. Chapter 8 Conclusion

5. diﬃculty in constructing the desired axioms using the tools and syntaxes available.

6. the need for extremely thorough testing to ensure that inferencing behaves as intended.

This illustrates that ontology building is best undertaken by qualified ontology builders who have a deep understanding of the ontology axioms and any rules, and their logical consequences. Ontology building is not well-suited for naive users or for casual collaboration: if one builds, modifies or adds to an existing ontology, unexpected results may ensue unless one has a deep understanding of the ontology as a logical structure. OWL DL (and subsequently OWL 2) has been designed with an Open World Assump- tion, in line with the general web policy that the web is an open environment, so there is always the potential for more material to be added and we should never assume we have complete knowledge. This means that negation as failure is not supported. Whilst this allows an open environment in which more material can always be added, in practice it can be rather unwieldy to work with. As the papers illustrate, one must capture one’s conditions positively rather than negatively - that is, in terms of what they are, rather than what they are not. In our experience this tended to require very complex axioms although the conditions themselves itself were intuitively quite simple. The absence of negation as failure can also be somewhat impractical for many use cases. When considering applications that use ontologies, one can imagine many situations where negation as failure is needed, and one would need to code it around the reasoning processes in order to provide useful results to the user. A common scenario would be semantic search, where a user sets a number of criteria and the application attempts to satisfy these criteria in returning search results. If none can be found, the user presumably would not want the application to hang indefinitely because of the possibility that at some time something that fits the criteria might appear. The application should be able to take the lack of a positive result from the reasoning processes as an indication that there is currently no useful solution, and might inform the user of this and request the user to relax their criteria, or perhaps relax the criteria by some pre-arranged method, embark on a revised search automatically and inform the user of the closest available alternatives. Also, the experiences highlight the importance of Proof and Trust, which are part of the overall plan for Semantic Web technologies as discussed at Section 1.3. Given that it is quite easy to obtain unanticipated results from inferencing processes even when one has built the ontology oneself, it seems paramount that users and applications that utilize an ontology/ies are able to be provided with some quality assurance as to their inferencing behaviour. Note that it is not enough that the results are logically sound: if the axioms are not set up correctly, then soundness is irrelevant and perhaps even misleading. What is needed is some assurance that the representation is appropriate, and the inferencing behaviour adequately captures or reflects the designers’ and users’ intentions and intuitions about how it should behave.

8.4 Readable Syntax

One of the issues identiﬁed for the Semantic Web in Section 1.9 was the need for a readable syntax so that domain experts and others could more easily and accurately author OWL ontologies. It was pointed out that it was important to ensure that logical precision was not compromised in the design of such a syntax: natural language comes with expectations and understandings inherent from its use as a natural language, and these may compete with the needs of a logical language, where precision is paramount. 225

The first publication included at Chapter 5 set out the scope, design goals, decisions and choices informing a proposed Controlled Natural Language syntax for OWL dubbed ‘Sydney OWL Syntax’. This syntax was presented at OWLED 2007, where it became apparent there were three parallel efforts on designing a CNL syntax for OWL, and a task force was set up to work towards agreeing on a common CNL syntax for OWL 2. The task force compared the three, and the comparison was published in the second paper included at Chapter 5. This comparison showed that although there are clearly differences between the three CNLs, there is considerable overlap between them and therefore much common ground to build on. The four principal areas of difference identified were: 1. Style. For example, ACE chooses to hyphenate noun phrases: river-stretch, whereas Rabbit and SOS allow River Stretch and river stretch (the capitalization Rabbit being another minor difference). However, these stylistic differences were considered easily resolvable and thus least important. 2. Phraseology. The three syntaxes take slightly different approaches in expressing certain constructs. For instance, when expressing part-whole relationships, ACE and Rabbit both opt for has-part/has part whereas SOS uses ‘has’ in conjunction with the phrase as a part at the end of the clause. 3. Mathematical Constraints. Probably the biggest area of difference is where the CNLs represent mathematical constraints such as transitivity. Rabbit’s approach pro- ceeds on the meta-level by talking about the construct explicitly e.g. ‘The relationship “is larger than" is asymmetric’ whereas SOS and ACE use variables in order to speak on the object level e.g. ‘If X is larger than Y then Y is not larger than X.’ In general, SOS has a less verbose form than ACE. 4. Design Principles Rabbit is designed to allow the domain expert to work in cooperation with a knowledge engineer, and thus does not attempt to remove notions such as ‘asymmetric’ from the syntax. ACE supports the user to express themselves in multiple ways, but sacrifices a one to one mapping between OWL and ACE: different ACE sentences can map to the same OWL axiom, and conversely, different OWL axioms can be represented by the same ACE statement. SOS enshrines the notion of one to one mapping between CNL and OWL syntax, and attempts to give as much naturalness of expression as possible without sacrificing logical precision. The paper concluded that there is sufficient commonality between the three CNLs described to provide a good base from which to proceed. Many of the differences are best resolved by proceeding empirically: that is, by conducting user testing which offers statements in different syntaxes and tests which ones give users the best comprehension. In conclusion, this exercise has some challenging territory to overcome on several fronts. One is that the precise meaning of an OWL axiom is defined by the underlying logical semantics, so it can be very challenging to provide comprehension for users who do not understand logic. All one can do is express the axiom in different ways in natural language and hope the idea comes across. However, users who are not familiar with formal logics may simply lack the mental architecture to fully understand the axiom, in whatever syntax it is written. In this case, whilst the syntax may not necessarily be at fault, the empirical results will not show user comprehension. However, at least we may expect that this factor would not disadvantage one syntax over another: they should all be equally affected. Also, the exercise walks a fine line between stipulative and lexical meaning: whilst logical semantics are specified stipulatively, the meaning of natural language arises lexically, from Chapter 8 Conclusion familiarity and use. The SOS approach attempts to define exactly one controlled natural language equivalent for every OWL statement, with the precise meaning of both being given by the OWL statement. Whilst the OWL statement thus constitutes the stipulative meaning, the SOS authors have attempted to delve into the English language and deliver an equivalent that both sounds natural and precisely corresponds to the OWL axiom, so that readers who do not understand the OWL axiom can use the English version instead to understand what the axiom means. However, any natural English statement derives its meaning primarily through lexical definition: that is, by how it and its constituents have been used in the readers’ experience. The natural meaning thus arises from a source other than the underlying logical semantics, and it can be difficult to balance these two aspects. In designing SOS, we attempted to use our own understanding of both OWL and natural English to come up with the closest possible match between the stipulative definition OWL provides, and a lexically- based English equivalent. However, lexical meaning, by its very nature, is neither fixed nor precise, so the match can only ever be approximate. Attempting to give readers who lack the ability to read a stipulative definition access to it through lexically defined language has obvious utility, but also necessarily comes with the lexical ‘baggage’ that the words of the controlled natural language carry. The difference between stipulative and lexical definition is treated in depth by the first paper included in Chapter 7.

8.5 Interoperability of Ontologies

Section 1.9 pointed out the need for semantic interoperability in order for application to use data from diﬀerent sources seamlessly. As explained at Section 1.10.5, the most straightforward way to go about this is by encouraging reuse of existing ontologies, rather than adding to the plethora of ontologies needing to be intermapped to achieve interoperability. The publication included at Chapter 6 proposed a tool called ‘n2mate’ to encourage vocabulary and ontology re-use. The unique aspect of this proposal was to leverage the otherwise hidden formal and informal knowledge networks created by existing business processes by marrying them with social networking models and tools. This provides a means to expose, organize and navigate the available information about existing knowledge artefacts, including pre-existing social and business agreements underlying them. The proposed n2Mate tool provides a central reference point to collect such information and user opinions, in order to provide potential vocabulary and ontology users with enough reassurance and conﬁdence to identify, choose between and commit to the use of existing knowledge artefacts instead of feeling compelled, through ignorance, to reinvent the wheel. Furthermore, it is expected that research into the strategy proposed would provide contributions to related projects, such as the development of:

• A lightweight mechanism revealing the state of interconnectedness in and between discourse communities.

• A bridging space between government, business, community, academia and science knowledge assets to enhance broadscale interoperability.

• A genetic algorithm to breed, select, and hybridize various standards artefacts such as ontologies, services, and trust authorities.

In short, we noted that there were significant inefficiencies in project scoping, the development of information products and online service provision, and that this is largely 227 attributable to the inadequacies of existing knowledge registers. In contrast, we believe that use of the proposed tool would enable consensus building and lead to the emergence of a relatively small number of high quality ontologies in each domain being endorsed by the community as the ones to use. For instance, factors such as the endorsement of one ontology by a particularly large or influential player could have a powerful flow on effect in brokering these de facto standards. We would then expect to see take-up of the most endorsed ontologies solidify and a community of practice coalesce around them. It is likely that the communities concerned would make the effort to intermap the most frequently used ontologies, producing a platform of semantic interoperability for their domain. Moving forward with tool building is dependent on obtaining funding to build such a tool, and unfortunately such funding has not yet been identified. A number of avenues are being considered, including partnership with large players in the domains and in search generally, and with bodies that have a mandate to produce vocabulary standards, such as the Australian Government Information Management Office (AGIMO) in the Australian Government space.

8.6 Making Meaning Machine Processable

Section 1.9 noted that in order for the Semantic Web to move closer to the vision of truly machine processable meaning, it needs to examine a number of questions to do with how one should go about defining one’s terms to make their meaning completely transparent, how one should determine the referent of a term and what kind of constructs might be needed to capture such notions. The papers included at Chapter 7, addressed these questions, albeit leaving much to be answered by future work. The first paper, which examined the potential for a science of definition concluded that such an enterprise would require an understanding of common semantic ground as it exists in the human conceptual landscape, and the ways in which divergences from it are made. The paper delineated some of the ways in which the conceptual landscape is naturally structured, and the means by which abstract constructs may be overlaid or built within it. Taking a cognitive viewpoint, it suggested some paths for investigating where common semantic ground may be located and how and where divergences from it arise. It also mentioned some avenues, including empirical studies, which may be of use in inferring the elements and structure of the conceptual landscape. It concluded that by putting together the various strands, it may well be possible to describe an underlying shared conceptual space which would provide a viable basis for robust methods of definition. The second paper put forward some of the hard questions the semantic Web needs to answer with regards to meaning, examined some of the pitfalls that may occur if they are not addressed, and explained the relevance of the symbol grounding problem for the kinds of semantic interoperability issues commonly encountered. Some insights from measurement theory in Mathematical Psychology were briefly covered to illustrate how inappropriate correspondence between symbol and referent can result in logically valid but meaningless inference. Some of the shortcomings of the current Semantic Web technologies in dealing effectively at the level of meaning were investigated. The arguments that set extensions and URIs can provide an appropriate basis for grounding the Semantic Web were considered and found wanting. Finally, next steps for identifying effective grounding strategies and doing meaning-level processing were briefly discussed. These included the investigation of a number of grounded symbol systems, such as Musical Notation, Cartography, Chess Notation, Circuit Diagrams, Chapter 8 Conclusion

Barcodes and even Knitting Patterns, with a view to analyzing these to determine the grounding strategies used, how and why they are eﬀective or ineﬀective, identifying the kinds of grounding strategies available, and determining appropriate criteria for assessing them. It was hoped that this would provide a theoretical basis for constructing Symbol Grounding strategies for the Semantic Web, and would inform the question of devising appropriate processing and procedures to produce meaningful results.

8.6.1 Further Analysis Whilst this analysis is far from complete, it is progressing and has produced the following initial insights:

1. Each of these representational systems represents a limited number of dimensions of human experience: for example, musical notation (excluding textual notes such as ‘Adagio’) represents the pitch and timing of sound. Chess notation represents chess pieces and their positions on the board in series as players take turns. Barcodes represent organizations and products and refer to registration schemes that record official identifiers for organizations and products. There is usually a very clear relationship between each dimension and some aspect of the representation: e.g. the vertical position of notes in musical notation has a one-to-one correspondence with tones of each pitch that are allowed by the musical system. Therefore the relationship between sign and signified is not purely arbitrary, but is in a clear correspondence with the referent (signified), possibly in a way that reflects the underlying structure of the system. 2. The Peircean distinction between modes of signs is informative ( [123], 1.564). Peirce outlined three fundamental semiotic modes reflecting the arbitrariness of the relationship of the sign to the signified: (a) In the symbol/symbolic mode, the signifier does not resemble the signified but is purely arbitrary or conventional e.g the use of red as the signal for stop. This is the mode discussed previously in relation to the Symbol Grounding problem. (b) In the icon/iconic mode, the signifier is perceived as resembling the signified in some way e.g. looking, sounding or feeling like it, so that it evokes some of the same qualities in a sensory way. This is the case with scale models or maps, cartoons, imitative gestures, sound effects, or onomatopoeia in speech. This mode is also commonly exploited in metaphor e.g. “a weighty argument”, “transparent motives”, as has been expounded by Lakoff and Johnson in their seminal work ”Metaphors We Live By” /citeLakoff relating the use of metaphor to embodiment as mediated by image schema. (c) In the index/indexical mode, the signifier is not arbitrary but is directly connected to signified either physically or causally, through a link which can either be observed or inferred. Examples include ‘natural signs’ such as smoke indicating a fire, measuring instruments such as weathercocks, thermometers and rain gauges, and recordings such as photographs and films. Whilst the symbol grounding problem relates to the symbolic mode, the iconic and indexical modes clearly have a more grounded relationship between sign and signified. 3. In the iconic mode, some systems map one or more dimensions into a spatial representation which is an analog of the physical one. The simplest example is standard cartography, where length of the area being represented is represented as length on the 229

map but is simply scaled down. However, length can be used to represent any quantitative dimension, such as temperature or time. Such representations usually preserve the inherent quantitative structure: For instance, if we are representing elapsed time spatially, and x is shown as twice as long as y in the spatial representation, this re- ﬂects that the corresponding elapsed time x took twice as long temporally speaking as elapsed time y. Spatial representation can also be logarithmic, and may preserve orderings or relations but lose quantitative structure, such as is the case with most subway maps.

4. Some representation systems describe a phenomena that has some underlying structure. For instance, Western musical notation assumes the use of a musical system predicated on octaves and tone/semitone scales that are themselves dependent on the use of harmonics in sound waves. This structure may be quite important in the interpretation of the representation.

5. Some representational systems are reliant on accompanying processes. For instance, retail store barcodes are reliant on two registration processes: the ﬁrst one sets up a correspondence between organizations and organization codes, whilst in the second, organizations catalogue their own products and assign an additional identiﬁer.

6. Different representational systems are set up for different purposes and should therefore be interpreted in different ways. For instance, they may be:

(a) Instructional. Examples include musical notation and knitting patterns, as musical notation may be interpreted as instructions for reproducing a piece of music physically, and knitting patterns as instructions for producing a physical garment. (b) Recordings. Notations can also be used to record events, for example, chess notation usually records the actual moves played in a game. (c) Pragmatic. In some instances, an exchange of notation reflects an activity in itself. For example, chess games can be conducted by correspondence by the exchange of moves in chess notation. This is possible because a game of chess is primarily a conceptual phenomenon rather than a physical one: whilst a board can be used to indicate each players intentions, it can be done equally well with representation that assumes the structure of the board and pieces. (d) Identificational. A barcode provides an entry point into an identification system, that is dependent on registration processes whereby organizations and products are put into one to one correspondence with registration codes that may then be encoded as barcodes. In this example the barcode is physically present as a label, connecting symbol directly to referent. Reading the barcode provides access to the mediating thought/reference, in this case the identificational system.

8.6.2 Building Situation Awareness from Sensor Input The three semiotic modes described by Peirce are illuminating in terms of identifying a way to relate symbols, or rather signs, more directly to the physical and causal world, and thereby bootstrap into the a fully symbolic mode. Recently I have worked full time with a team of researchers on a 12 month pilot project, concluding July 2008, that illustrates the principle and provides some insight into how this bootstrapping might occur. This project, entitled ‘Situation Awareness by Inference and Logic’ (SAIL) [4] was a joint project Chapter 8 Conclusion

CNL Query CNL Answer CNL Alert

CNL formalization CNL generation CNL generation

Alert generation CNL assertion handler - CNL query handler (LTL/BA)

Control program and reasoner interface CNL formalization

Semantic Analysis (Description Logics/RacerPro)

CNL background knowledge (scenario/intelligence) ...... ABoxj ABoxj+1 ABoxj+2

Data Aggregation (Rules/E-KRHyper)

GIS DB ... SD SDi SDi+1 i+2 ...

Figure 8.1 SAIL Prototype System Architecture between NICTA and the Australian Defence Scientific and Technical Organization (DSTO). It involved processing data from sensors in order to build up a symbolic representation of a situation of interest, with the general objective of identifying threats or potential threats in a military scenario involving vehicles of various allegiance moving in time and space and exhibiting certain behaviours and intentions. Ultimately the system was able to issue alerts automatically when pre-defined situations or factors of interest arose, and could be queried by operators using a Controlled Natural Language interface. The pilot project resulted in a fully functional prototype SAIL system, with architecture as shown at Figure 8.1. The architecture includes a ‘Data Aggregation layer’ that processes input data streams providing time-stamped sensor data SDi about vehicles such as aircraft, ships, etc, moving in the environment, giving their location, speed, acceleration and so on. Using a First Order Logic Reasoner, E-KRHyper [124], this sensor data is processed using first order logic rules that aggregate the data, using numerical calculations where necessary, 231 and combining with background knowledge including GIS information and information about vehicles, facilities and capabilities to build primitive concepts that are then passed as primitives to the ‘Semantic Analysis layer’. This layer is powered by the RacerPro reasoner [60] and is able to issue alerts and be queried using a Controlled Natural Language (CNL) modules. The work conducted provides some interesting insights to the symbol grounding problem, as it deals with a case where instead of needing to deal with subjective human experience, as was attempted in Towards a Science of Definition, we were able with a more explicit analogue of it, provided in the objective form of sensor data streams. By virtue of this, instead of using mental processes to build concepts in a subjective, conceptual landscape, we were able to construct and execute the analogous explicit machine processable rules in FOL to convert sensor data into explicitly defined primitive concepts. These concepts then provided the base for a Semantic Analysis layer that defined non-primitive concepts in terms of these primitives. This layer uses description logic to define non-primitive concepts using the primitive concepts passed to it by the data aggregation processes. High-level concepts and roles (unary and binary relations, respectively) are introduced by means of DL axioms that define them in terms of the primitive concepts populated by the data aggregation layer and also other high-level concepts defined in the Semantic layer. For example, there is a first order logic rule in the Data Aggregation layer that identifies a ‘take off’ event when an object first appears in the sensor data, and its detected location is computed (by reference to GIS data) to be ‘in air’ (as opposed to ‘on land’ or ‘under water’). Firing the rule creates a new ‘take off’ event with a unique event ID, timestamped with the current time. This newly created instance of ‘take off’ event is then passed to the ABox of the Semantic Layer. The semantic layer uses the ‘take off’ event as a primitive concept that is a constituent in the construction of DL definitions for non-primitive concepts such as exhibiting ‘aggressive behaviour’ toward enemy targets. The primitive concepts used in the ontology in this project therefore have a direct relationship to sensor data streams by virtue of the Data Aggregation rules. Whilst the sensor data streams are presented to the data aggregation layer of the system in a symbolic form, they result from an indexical mode where they reflect measured values of an object on a number of physical dimensions. They may therefore be considered to be an explicit and machine processable analogue of perceptual data streams. We argue that since it is possible to show a direct, rule-based relationship to sensor- data, the primitives which are given by such rules are grounded in experience, in that they essentially ‘bottom-out’ to sensory experience, and we need go no further to determine what they mean. Given the sensor data streams and the rule-based definitions as executed by the FOL rules, the primitive concept is both explicit and machine processable: the primitive concept is stipulatively defined to be the rule that is executed at the data aggregation layer. It should be possible to resolve any disagreement about what the concept refers to by recourse to this definition. Furthermore, since it is an explicit rule, it should also be possible to negotiate the similarity or difference between it and other concepts by machine processing that is set up to reference and compare the rules underpinning the respective concepts.

8.6.3 A Strategy for Symbol Grounding It is suggested that this form of strategy is appropriate for use in symbol grounding, by the following method:

1. Identify an explicit, measurable analog for sensory data. It should make use of physical Chapter 8 Conclusion

dimensions. For example, colours might be captured by a data stream that gives wavelength values.

2. Deﬁne primitive concepts in terms of rules over these data streams. The rules must be explicit and complete. Rules may make use of both logical and numerical operators.

3. Each primitive concept should be assigned an associated symbol at Step 2, so that the deﬁnition of the symbol is then stipulatively the rule used.

Clearly this method is appropriate for those concepts that are directly related to sensory data such as ‘red’ or ‘hot’, where there are measurable physical variables that are objective analogues of subjective sensory experience. For example, colours perceived by humans have been related to measurable physical variables of wavelength, saturation and brightness, and empirical science has shown that varying these dimensions will result in a perceivable difference in colour [48]. A space defined by these dimensions may therefore be taken to be an explicit analogue for a subjective perceptual space such as those described by Gärdenfors[48]. This is not to argue that concepts reduce to the sensory inputs they generate - that would be far too simplistic. However, if a concept has meaning, it is argued that it is by virtue of some noticeable impact at some level of human experience, and therefore there is a path to connect it to the tangible. Applying this method becomes more challenging as one moves away from sensory-based concepts to more abstract concepts like ‘Tuesday’ or ‘justice’, in these cases there is need to include reference to functionality, social processes and other processes. There is also undeniable a need to perform logical and other operations over the raw inputs. In order to give an idea of how these kind of concepts might be grounded, let us consider the case of ‘Tuesday’. To ground the days of the week requires several steps. Firstly, one identifies the essential thing to which the symbol is grounded. Temporal concepts may be found to be grounded in physical events or processes, in this case, the relevant event is the rotation of the earth in its orbit around the sun and on its own axis over time. (For the sake of simplicity, let us leave aside for the moment the fact that the SI unit of time, the second, is now grounded to periods of the cesium-133 atom - whilst the method to be described can accommodate this, it makes the explication unnecessarily complex.) The arc of the earth’s movement provides the grounding for ‘year’ and ‘day’ respectively: a ‘year’ being a full rotation of the earth around the sun and a ‘day’ being a full rotation of the earth on its own axis. These physical processes are known to occur through observation of physical stimuli in combination with rational thought - we will not attempt to justify further here, except to note that such physical processes are empirically verifiable and exist independently of our own descriptions or definitions of them. The essential point for being appropriate as a grounding is to establish an objective and unambiguous means for resolving ambiguity. It is important to note that physical processes are not an appropriate form of grounding for all concepts; the appropriate grounding is simply the base level beyond which there is no need to clarify a notion any further in order to avoid ambiguity. More work is needed to fully understand the nature and forms of grounding in all their variety. Secondly, one relates the physical process to a symbol system: in this case the Gregorian calendar performs this function, as it is a symbolic system organised around the essential notions of ‘year’ and ‘day’, grounded to physical processes as described. The calendar itself has its own internal system of organisation, which relies on naming, sequence, numbering (which itself is a naming/sequence symbol system with particular properties), subdivision 233 and aggregation. By using such operations any aspect of the calendar (symbol system) can be reduced to one or both of the core grounded concepts, ‘year’ and ‘day’. Through empirical observation and counting, it is known that a year consists of some 365 days. These are aggregated into groups subdividing the year. Note that two subdivisions operate in parallel: the division of the year into months and the division into weeks. Months have a defined number of days each and are named ‘January’, ‘February’ and so on. The names are defined to occur in a given sequence, again corresponding to arcs in the sequential movement of the earth in its orbit around the sun - the movement of the earth in February 2008 is a continuation of the arc of movement of January 2008. The other system of subdividing years if weeks, which are defined to include precisely seven days. Unlike months, weeks within a year are not named, but may be numbered in sequence, again corresponding to the sequential movement of the earth in its solar orbit. The seven days within a week are named ‘Monday’, ‘Tuesday’ and so on, and the names are defined to occur in a given sequence - ‘Tuesday’ immediately follows ‘Monday’. One may move from the general level to specific years and dates by bedding down temporo-spatial points where a particular year or day starts and ends, for instance. Se- quential numbers are allocated to represent sequential years, ‘2008’ being assigned to the current year for instance, whilst of course specifying a date makes use of year, month and day. Thus in summary, to ground the concept ‘Tuesday’ requires a combination of factors:

1. identiﬁcation of the underlying physical process, itself established through empirical means;

2. a symbolic system - the Gregorian calendar - which has a clear correspondence between its own base symbols ‘day’ and ‘year’ and the key aspects of the physical process - this establishes grounded primitives, or what we might call ‘semantic primitives’. The system includes a full speciﬁcation of its own internal structure in terms of the primitives and

3. a set of operations to name and give internal structure to the various other elements of the representation system. For the Gregorian calendar, these include naming, sequencing, numbering, subdivision and aggregation, but others may be needed for other representational systems. using these operations ‘week’, ‘month’, ‘January’ and ‘Tuesday’ are deﬁned.

An interesting point to note is that the various ways in which symbols are assigned in this example, each level of deﬁnition building upon the previous one:

1. The symbols ‘year’ and ‘day’ each assign a symbol directly to an aspect of the physical process;

2. The symbols ‘month’ and ‘week’ are deﬁned in terms of subdivisions and aggregations of ‘year’ and ‘day’;

3. The symbols ‘January’ and ‘Tuesday’ are assigned by sequencing the months within a year and the days within a week, and assigning names accordingly.

4. Whilst the above may all be defined using relativities, a system of absolute reference is needed to specify specific years and dates. This may be done by giving a specific time/space referent where the arc of movement starts. Chapter 8 Conclusion

To complete the example, consider the case of another calendar system in addition to the Gregorian, which is also grounded to the movement of the earth, but has its own independent system of symbols with their own internal relationships. If the grounding and relationships within the representation system are made explicit, machine processing can feasibly be used to determine the semantic relationships between the symbols of the two systems, by reference to the relationships of the underlying arcs of movement the symbols refer to. The grounding provides a fully specified and unambiguous means to determine the referent of a symbol from either system, and thus the semantic relationship between symbols of the two systems. Without a known and specified grounding, the equivalence or relationship of symbols between the two systems must be determined by human negotiation. Note however, that it is sufficient to establish relationships between the primitives of the two systems, the other symbols then being described by the respective internal definitions. Examples of other grounded systems include latitude/longitude as a system for establishing locations on the surface of the earth; musical notation as a system for reproducing sound frequency and timing; and family trees as a representation of genetic relationships between individuals, the core primitive being parent/child inheritance.

8.7 Summation

In summary, this thesis has introduced the Semantic Web and presented it both as a vision and as the actual technologies produced and planned to date. The notion of meaning as it has so far been implemented in the Semantic Web languages OWL and RDF was explained as being based on model theoretic semantics, leaving the door open to consider other aspects of meaning and how they might be harnessed. A number of problems facing the Seman- tic Web were expounded, including accessibility of the concepts, competing standards for building semantic structures, adding rules to OWL ontologies, a readable syntax for OWL ontologies, interoperability between ontologies, and making meaning machine processable. The included publications offered a number of contributions addressing these problems. Chapter 2 introduced Semantic Technologies and offered an accessible introduction to the field and provided additional background material. In Chapter 3, a method for building topic maps in OWL-DL that were conformant to the ISO’s Topic Map data model was presented. Although the method has been criticized for being at the object level, it provides some utility in that it gives access to the OWL-DL semantics and tools. The need to add rules to OWL ontologies was explored in Chapter 4, illustrating first-hand the impact of some of OWL’s design choices, particularly the open world assumption. The additional expressivity of OWL 2 addresses a number of limitations that were identified by this work. The work also highlighted the need for ontology building to be undertaken with a deep level of understanding of the ontology as a logical structure. In Chapter 5, we identified the need for a more readable syntax for OWL, assessed the design choices for creating a CNL syntax for OWL, presented Sydney OWL Syntax and ultimately compared it with two other CNL efforts, ACE and Rabbit. More work remains to be done to agree on a common CNL syntax for OWL 2. The issue of semantic interoperability was considered in Chapter 6 and the n2mate tool was suggested as a mechanism to address the issue in a very simple fashion, by encouraging ontology re-use through tapping into contextual information and social networks. This work is unimplemented as it requires funding to proceed with implementation. Lastly, Chapter 7 considered some of the deeper issues concerning meaning and making it machine processable, and offered some insights into the nature of definition, and the Symbol Grounding problem and its importance for the Semantic Web. A method for tackling it was suggested at Section 8.6, involving the use 235 of rules operating over objective analogues for sensory experience. This approach was successfully executed by the author as part of a team working on the SAIL project at NICTA. Further extensions to the method for moving beyond sensory input were considered and an example expounded. Much work still remains to be done, especially on the vital question of how to provide a truly machine-processable notion of meaning. Resolving this question is crucial for providing a suitable foundation for what will ultimately be a ‘Pragmatic Web’, where machines can be trusted to take practical real-world actions based on the outcomes of mechanical inference, without fear of error due to semantic ambiguity and misinterpretation. In closing, it is hoped that the investigations undertaken in this thesis will provide some value and insights to those who continue to contribute to the Semantic Web technologies, and embark on the great journey towards implementing the modern vision of Leibniz’s dream.

Bibliography

[1] Aduna, NLnet Foundation, and Ontotext. Sesame - open source framework for stor- age, inferencing and querying of RDF data. [Software]. See http://www.openrdf.org/. [2] R. Aggarwal. Semantic web services languages and technologies: Comparison and discussion. LSDIS Lab, University of Georgia, 2004. citeseer.ist.psu.edu/700682.html. [3] D. Allemang, R. Hodgson, and I. Polikoff. FEA Reference Model Ontol- ogy (FEA RMO). [Online]. Available: http://www.topquadrant.com/documents/ TQFEARMO.pdf, 2005. [4] Franz Baader, Andreas Bauer, Peter Baumgartner, Anne Cregan, Alfredo Gabaldon, Krystian Ji, Kevin Lee, and Rolf Schwitter. A Novel Architecture for Situation Aware- ness Systems. [Online] Available: http://nicta.com.au/ data/assets/pdf file/0019/ 15472/AI-08-submission-long-version.pdf; revised version accepted for publication in Proceedings of TABLEAUX 2009 http://tableaux09.ifi.uio.no/, 2008. [5] Franz Baader, Diego Calvanese, Deborah L. McGuinness, Daniele Nardi, and Peter F. Patel-Schneider, editors. The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge University Press, 2003. [6] S. Bechhofer, I. Horrocks, P. F. Patel-Schneider, and S. Tessaris. A proposal for a description logic interface. In Proc. of the 1999 Description Logic Workshop (DL’99), pages 33–36, 1999. CEUR Workshop Proceedings http://ceur-ws.org/Vol-22/. [7] S. Bechhofer, T. Liebig, M. Luther, O. Noppens, P.F. Patel-Schneider, B. Suntisri- varaporn, A.-Y. Turhan, and T. Weithöner.DIG 2.0 – towards a flexible interface for description logic reasoners. In B. Cuenca Grau, P. Hitzler, C. Shankey, and E. Wal- lace, editors, Proceedings of the 2nd Workshop OWL Experiences and Directions 2006, 2006. [8] Sean Bechhofer. Hoolet. Department of Computer Science, University of Manchester [Online] Available: http://owl.man.ac.uk/hoolet/, 2004. [9] D. Beckett. New syntaxes for rdf. Technical Report, 2004. Institute for Learning and Research Technology, Bristol. [10] D. Beckett. Turtle - terse rdf triple language. Technical Report [Online] Available: urlhttp://www.dajobe.org/2004/01/turtle/, 2007. [11] Freie UniversitätBerlin. Disco - Hyperdata Browser. [Software]. See http://www4. wiwiss.fu-berlin.de/bizer/ng4j/disco/. [12] Raffaella Bernardi, Diego Calvanese, and Camilo Thorne. Lite Natural Language. In IWCS-7, 2007. [13] Tim Berners-Lee. Realising the Full Potential of the Web. [Online] Available: www. w3.org/1998/02/Potential.html, 1998. [14] Tim Berners-Lee. Semantic Web Roadmap. [Online] Available: www.w3.org/ DesignIssues/Semantic.html, 1998. [15] Tim Berners-Lee and Mark Fischetti. Weaving the Web: The Original Design and Ultimate Destiny of the World Wide Web by Its Inventor. Harper, San Francisco, 1999.

237 BIBLIOGRAPHY

[16] Tim Berners-Lee, James Hendler, and Ora Lassila. The Semantic Web. Scientific American, 284:34–43, May 2001. [17] Tim Berners-Lee and Ralph Swick. Semantic Web Development Final Tech- nical Report (Tech. Rep. No. AFRL-IF-RS-TR-2006-294). DTIC Research Report ADA458366. [Online] Available: http://www.scribd.com/doc/122258/ semantic-webdevelopment-2006, 2006. [18] Paul V. Biron and Ashok Malhotra (Eds). XML Schema Part 2: Datatypes Second Edition. World-Wide Web Consortium (W3C) Recommendation 28 October 2004. [Online] Available: http://www.w3.org/TR/xmlschema-2/. [19] C. Bizer, T. Heath, D. Ayers, and Y. Raimond. Interlinking Open Data on the Web (Poster). Proceedings of the 4th European Semantic Web Conference (ESWC2007), pages 802–815, 2007. [20] Tim Bray, Jean Paoli, C. M. Sperberg-McQueen, Eve Maler, and Franois Yergeau (Eds). Extensible Markup Language (XML) 1.0 (Fourth Edition). World-Wide Web Consortium (W3C) Recommendation 16 August 2006, edited in place 29 September 2006. [Online] Available: http://www.w3.org/TR/REC-xml/, September 2006. [21] D. Brickley and R. V. Guha. RDF Vocabulary Description Language 1.0: RDF Schema. World Wide Web W3C Recommendation 10 February 2004 [Online] Avail- able: http://www.w3.org/TR/rdf-schema/, February 2004. [22] Mark Burstein, Jerry Hobbs, Ora Lassila, Drew Mcdermott, Sheila McIlraith, Srini Narayanan, Massimo Paolucci, Bijan Parsia, Terry Payne, Evren Sirin, Naveen Srini- vasan, and Katia Sycara. OWL-S: Semantic Markup for Web Services. [Online] Available: http://www.w3.org/Submission/2004/SUBM-OWL-S-20041122/, Novem- ber 2004. [23] Peter Clark, Philip Harrison, Thomas Jenkins, John Thompson, and Richard H. Wo- jcik. Acquiring and Using World Knowledge Using a Restricted Subset of English. In FLAIRS 2005, pages 506–511, 2005. [24] Confluence project. Confluence project, 2007. http://www.comp.leeds.ac.uk/ confluence/. [25] Oracle Corporation. Oracle 11g RDF Database. [Software]. See http://www.oracle. com/technology/tech/semantic technologies/index.html. [26] Louis Couturat. La Logique de Leibniz. Felix Alcan, Paris, 1901. [27] A. Cregan. Symbol Grounding for the Semantic Web. Proceedings of the 4th European Semantic Web Conference, pages 429–442, 2007. [28] A. Cregan, R. Schwitter, and T. Meyer. Sydney OWL Syntax - towards a Con- trolled Natural Language Syntax for OWL 1.1. Proceedings of the third OWLED 2007 Workshop on OWL: Experiences and Directions, Innsbruck, Austria, 6-7, June 2007, (OWLED 2007), CEUR-WS Vol 258, 2007. [29] Anne Cregan, Rolf Schwitter, and Thomas Meyer. Sydney OWL Syntax — towards a Controlled Natural Language Syntax for OWL 1.1. In Christine Golbreich, Aditya Kalyanpur, and Bijan Parsia, editors, 3rd OWL Experiences and Directions Workshop (OWLED 2007), volume 258. CEUR Proceedings, 2007. 239

[30] Hamish Cunningham, Diana Maynard, Kalina Bontcheva, and Valentin Tablan. GATE: A framework and graphical development environment for robust NLP tools and applications. In Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics, Morristown, NJ, USA, 2002. Association for Compu- tational Linguistics.

[31] Mike Dean and Guus Schreiber (Eds). OWL Web Ontology Language Reference. World Wide Web Consortium (W3C) Recommendation 10 February 2004. [Online] Available: http://www.w3.org/TR/owl-ref/, February 2004.

[32] Patrick Durasau and Steve Newcomb. Topic maps - reference model. ISO 13250 Committee Draft 12 February 2005. [Online] Available: http://www.isotopicmaps. org/TMRM/TMRM-5.0/TMRM-5.0.html, February 2005.

[33] Dan Connolly (Ed). Gleaning Resource Descriptions from Dialects of Languages (GRDDL). World Wide Web Consortium(W3C) Recommendation 11 September 2007. [Online] Available: http://www.w3.org/TR/grddl/, September 2008.

[34] Dave Beckett (Ed). RDF/XML Syntax Speciﬁcation (Revised). World Wide Web Consortium (W3C) Recommendation 10 February 2004. [Online] Available: http: //www.w3.org/TR/rdf-syntax-grammar/, February 2004.

[35] Patrick Hayes (Ed). Resource Description Framework (RDF) Semantics. World Wide Web Consortium (W3C) Recommendation 10 February 2004. [Online] Available: http: //www.w3.org/TR/rdf-mt/, February 2004.

[36] World Wide Web Consortium (W3C) Semantic Web Education and Outreach (SWEO) Interest Group. Linking Open Data - a SWEO Community Project. [On- line] Available: http://esw.w3.org/topic/SweoIG/TaskForces/CommunityProjects/ LinkingOpenData, October 2007.

[37] Tim Berners-Lee et al. Tabulator generic data browser and editor. [Software]. See http://www.w3.org/2005/ajar/tab.

[38] J. Euzenat, A. Isaac, C. Meilicke, P. Shvaiko, H. Stuckenschmidt, O. Sv´ab,V.ˇ Sv´atek, W.R. van Hage, and M. Yatskevich. Results of the ontology alignment evaluation initiative 2007. Proceedings of the International Workshop on Ontology Matching (OM) at the International Semantic Web Conference (ISWC)+ Asian Semantic Web Conference (ASWC), pages 96–132, 2007.

[39] Lee Feigenbaum, Ivan Herman, Tonya Hongsermeier, Eric Neumann, and Susie Stephens. The Semantic Web in Action. Scientiﬁc American, 297:90–97, Decem- ber 2007.

[40] P. Fillottrani, E. Franconi, and S. Tessaris. The New ICom. 2006 International Workshop on Description Logics DL06, page 259, 2006.

[41] Stanford Center for Biomedical Informatics Research. The Prot´eg´eOntology Editor and Knowledge Acquisition System.

[42] Research Center for Information Technologies, University of Karlsruhe, and Uni- versity of Manchester. KAON2 OWL-DL reasoner. [Software]. See http://kaon2. semanticweb.org/. BIBLIOGRAPHY

[43] Gottlob Frege. Begriffsschrift, eine der Arithmetischen Nachgebildete Formelsprache des Reinen Denkens. Verlag von Louis Nebert, Halle, 1879. [44] Norbert E. Fuchs, Kaarel Kaljurand, and Gerold Schneider. Attempto Controlled En- glish Meets the Challenges of Knowledge Representation, Reasoning, Interoperability and User Interfaces. In Geoff Sutcliffe and Randy Goebel, editors, Nineteenth Inter- national Florida Artificial Intelligence Research Society Conference (FLAIRS 2006), pages 664–669, Melbourne Beach, Florida, May 11–13th 2006. The AAAI Press, Menlo Park, California. [45] Norbert E. Fuchs, Uta Schwertel, and Rolf Schwitter. Attempto Controlled English — Not Just Another Logic Specification Language. In Pierre Flener, editor, Logic-Based Program Synthesis and Transformation, number 1559 in Lecture Notes in Computer Science, Manchester, UK, June 1999. Eighth International Workshop LOPSTR’98, Springer. [46] Adam Funk, Valentin Tablan, Kalina Bontcheva, Hamish Cunningham, Brian Davis, and Siegfried Handschuh. CLOnE: Controlled Language for Ontology Editing. In Proceedings of the Sixth International Semantic Web Conference (ISWC), Busan, Korea, 2007. [47] A. Gangemi et al. Sweetening ontologies with DOLCE. In A. Gómez-Pérezand V. R. Benjamins, editors, Proc. of the 13th EKAW, volume 2473 of LNAI, pages 166–181, Siguenza, Spain, 2002. Springer. [48] Peter Gärdenfors. Conceptual Spaces: The Geometry of Thought. MIT Press, Cam- bridge Massachussetts, 2000. [49] Lars Marius Garshol. Comparing Topic Maps and RDF. Blog Entry 06 January 2007. [Online] Available: http://www.garshol.priv.no/blog/92.html, January 2007. [50] Lars Marius Garshol and Robert Barta. Topic map query language (tmql) requirements. ISO 13250 Draft 7 November 2003. [Online] Available: http://www1.y12.doe. gov/capabilities/sgml/sc34/document/0448.htm, November 2003. [51] Lars Marius Garshol and Graham Moore. Topic Maps Part 2: DataModel. Draft ISO/IEC International Standard (ISO/IEC JTC 1/SC34 ISO/IEC FCD 13250-2) 10 January 2005. [Online] Available: http://www.isotopicmaps.org/sam/sam-model/, January 2005. [52] Jennifer Golbeck and Peter Mika (Chairs). Semantic Web Challenge. [Online] Avail- able http://challenge.semanticweb.org/. [53] C. Golbreich and I. Atsutoshi. Combining SWRL rules and OWL ontologies with ProtégéPOWL plugin, Jess, and Racer. In R. Fergerson and N. Noy, editors, Proceed- ings of the 7th International ProtégéConference, 2004. [54] James Gosling, Bill Joy, Guy Steele, and Gilad Bracha. The Java language specification. Addison-Wesley, third edition, 2005. [55] Object Management Group. Ontology Definition Metamodel. Object Manage- ment Group Adopted Specification pending finalization. OMG document Num- ber: ptc/2007-09-09[Online] Available http://www.omg.org/docs/ptc/07-09-09.pdf, November 2007. 241

[56] Talis Group. Talis Semantic Web Application Platform. [Software]. See http://www. talis.com/platform/.

[57] T. R. Gruber. A translation approach to portable ontology speciﬁcations. Knowledge Acquisition, 5(2):199–220, 1993.

[58] Thomas R. Gruber. A Translation Approach to Portable Ontology Speciﬁcations. Knowledge Acquisition, 5(2):199–220, 1993.

[59] V. Haarslev and R. M¨oller.Racer: An owl reasoning agent for the semantic web. In Proc. of the International Workshop on Applications, Products and Services of Web- based Support Systems, in conjunction with 2003 IEEE/WIC International Confer- ence on Web Intelligence, Halifax Canada, Oct 13, pages 91–95, 2003.

[60] Volker Haarslev, Ralf Moller, and Michael Wessel. RACER, Semantic Middleware for Industrial Projects based on RDF/OWL. [Online] Available: http://www.sts. tu-harburg.de/∼r.f.moeller/racer/.

[61] Christian Halaschek-Wiener, Jennifer Golbeck, Bijan Parsia, Vladimir Kolovski, and Jim Hendler. Image browsing and natural language paraphrases of semantic web annotations. In First International Workshop on Semantic Web Annotations for Multimedia (SWAMM), Edinburgh, Scotland, May 22nd 2006.

[62] S. Harnad. The Symbol Grounding Problem. Physica, 42:335–346, 1990.

[63] Glen Hart, Catherine Dolbear, and John Goodwin. Lege Feliciter: Using Structured English to represent a Topographic Hydrology Ontology. In Christine Golbreich, Aditya Kalyanpur, and Bijan Parsia, editors, 3rd OWL Experiences and Directions Workshop (OWLED 2007), volume 258. CEUR Proceedings, 2007.

[64] Glen Hart, Catherine Dolbear, John Goodwin, and Katalin Kovacs. Do- main Ontology Development. Technical Report I-0003, Ordnance Survey Re- search, 2007. http://www.ordnancesurvey.co.uk/oswebsite/partnerships/research/ publications/docs/2007/Domain Ontology Development V1.pdf.

[65] M. Hausenblas, W. Halb, Y. Raimond, and T. Heath. What is the Size of the Semantic Web? - Metrics for Measuring the Giant Global Graph. International Conference on Semantic Systems (I-Semantics08) at TRIPLE-I. To appear., 2008.

[66] J. Hendler, T. Berners-Lee, and E. Miller. Integrating applications on the semantic web. Journal of the Institute of Electronic Engineers of Japan, pages 676–680, 2002.

[67] H. Herre, B. Heller, P. Burek, R. Hoehndorf, F. Loebe, and H. Michalek. General formal ontology (gfo): A foundational ontology integrating objects and processes. part i: Basic principles. Research Group Ontologies in Medicine (Onto-Med), University of Leipzig. [Online] Available: http://www.onto-med.de/en/theories/gfo/, 2007.

[68] Daniel Hewlett, Aditya Kalyanpur, Vladimir Kolovski, and Chris Halaschek-Wiener. Eﬀective Natural Language Paraphrasing of Ontologies on the Semantic Web. In End User Semantic Web Interaction Workshop (ISWC 2005), 2005.

[69] Jaako Hintikka. Lingua Universalis vs. Calculus Ratiocinator. An ultimate presup- position of Twentieth-century philosophy. Kluwer Academic Publishers, Dordrecht, 1997. BIBLIOGRAPHY

[70] Jan Hladik. Reasoning about nominals with FaCT and RACER. In Proc. of the 2003 International Workshop on Description Logics (DL2003), CEUR-WS, 2003.

[71] R. Hodgson and D. Allemang. Semantic technologies for e-government (SiCoP Working document). [Online] Available: http://colab.cim3.net/ﬁle/work/SICoP/ 2006-02-09/Proposals/RHodgson.pdf, (accessed 6 June 2007).

[72] Matthew Horridge, Nick Drummond, John Goodwin, Alan Rector, Robert Stevens, and Hai H Wang. The manchester owl syntax. In Proc. of the 2006 OWL Experiences and Directions Workshop (OWL-ED2006), 2006. [Online] Available: http://owl- workshop.man.ac.uk/acceptedLong/.

[73] Matthew Horridge, Holger Knublach, Alan Rector, Robert Stevens, and Chris Wroe. A practical guide to building OWL ontologies using the Prot´eg´e-OWL plugin and CO-ODE tools, 2004. University of Manchester.

[74] I. Horrocks and P. Patel-Schneider. Reducing OWL entailment to description logic satisﬁability. Web Semantics: Science, Services and Agents on the World Wide Web, 1(4):345–357, 2004.

[75] I. Horrocks, P. F. Patel-Schneider, S. Bechhofer, and D. Tsarkov. Owl rules: A proposal and prototype implementation. Conditionally accepted for publication. Journal of Web Semantics, 2005.

[76] I. Horrocks, P.F. Patel-Schneider, H. Boley, S. Tabet, B. Grosof, and M.Dean. SWRL: A semantic web rule language combining OWL and RuleML. DARPA DAML Program [Online] Available: http://www.w3.org/Submission/2004/SUBM-SWRL-20040521/, 2003.

[77] Ian Horrocks, Peter F. Patel-Schneider, and Frank van Harmelen. From SHIQ and RDF to OWL: The making of a web ontology language. Journal of Web Semantics, 1(1):7–26, 2003.

[78] U. Hustadt, B. Motik, and U. Sattler. Reducing SHIQ− description logic to disjunctive datalog programs. In D. Dubois, C. Welty, and M.-A. Williams, editors, Proceedings of the KR2004, pages 152–162. AAAI Press, 2004.

[79] Rule Markup Initiative. RuleML: Rule markup language. [Online] Available: http: //www.ruleml.org, 2008.

[80] Mustafa Jarrar, Maria Keet, and Paolo Dongilli. Multilingual verbalization of ORM conceptual models and axiomatized ontologies. Technical report, Vrije Universiteit Brussel, February 2006.

[81] M. Johnson. The Body in the Mind: The Bodily Basis of Meaning, Imagination and Reason. University of Chicago Press, Chicago, 1987.

[82] N. Jolley, editor. The Cambridge Companion to Leibniz. Cambridge University Press, 1995.

[83] Kaarel Kaljurand. Attempto Controlled English as a Semantic Web Language. PhD thesis, Faculty of Mathematics and Computer Science, University of Tartu, 2007. 243

[84] Kaarel Kaljurand and Norbert Fuchs. Verbalizing OWL in Attempto Controlled English. In Christine Golbreich, Aditya Kalyanpur, and Bijan Parsia, editors, 3rd OWL Experiences and Directions Workshop (OWLED 2007), volume 258. CEUR Proceedings, 2007. 10 pages. [85] Kaarel Kaljurand and Norbert E. Fuchs. Bidirectional mapping between OWL DL and Attempto Controlled English. In LNCS 4187, pages 179–189, 2006. [86] Adhitya Kalyanpur, Bijan Parsia, Bernardo Cuenca-Grau, and Jim Hendler. Swoop: A ‘web’ ontology editing browser. Journal of Web Semantics, (4(2)), 2005. [87] H. Kamp and U. Reyle. From Discourse to Logic. Dordrecht: Kluwer, 1993. [88] Esther Kaufmann, Abraham Bernstein, and Lorenz Fischer. NLP-Reduce: A “na¨ıve” but Domain-independent Natural Language Interface for Querying Ontologies. In 4th European Semantic Web Conference (ESWC 2007), pages 1–2, November 2007. [89] G.A. Kelly. The Psychology of Personal Constructs. Norton, New York, 1955. [90] R. I. Kittredge. Sublanguages and controlled languages. Oxford University Press, 2003. [91] Graham Klyne and Jeremy Carroll (Eds). Resource Description Framework (RDF): Concepts and Abstract Syntax. World Wide Web Consortium (W3C) Recommen- dation 10 February 2004. [Online] Available: http://www.w3.org/TR/rdf-concepts/, February 2004. [92] Tobias Kuhn. AceWiki: A Natural and Expressive Semantic Wiki. In Semantic Web User Interaction at CHI 2008: Exploring HCI Challenges, 2008. To appear. [93] G. Lakoﬀ and M. Johnson. Metaphors We Live By. University of Chicago Press, Chicago, 1980. [94] Ora Lassila and Ralf Swick (Eds). Resource Description Framework (RDF): Model and Syntax Speciﬁcation. World Wide Web Consortium (W3C) Recommendation 22 February 1999, REC-rdf-syntax-19990222. [Online] Available: http://www.w3.org/ TR/1999/REC-rdf-syntax-19990222/, February 1999. [95] Benjamin Grosof (lead). SweetRules: Tools for Semantic Web Rules and Ontologies. [Software]. See http://sweetrules.projects.semwebcentral.org/. [96] Clark & Parsia LLC. Pellet OWL-DL reasoner. [Software]. See http://pellet.owldl. com/. [97] Image Matters LLC. Gnizr - open source application for social bookmarking and web mashup. [Software]. See http://www.gnizr.com. [98] Leroy Loemker, editor. Leibniz: Philosophical Papers and Letters. D. Reidel, Dor- drecht, 1969. [99] P. Lyman and H.R. Varian. How much information? 2003. [Online]. Available: http://www.sims.berkeley.edu/how-much-info-2003, (accessed 6 June 2007), 2003. [100] S. Newcomb (Eds) M. Biezunski, M. Bryan. Topic Maps Second Edition. ISO/IEC International Standard (ISO/IEC 13250) 19 May 2002. [Online] Available: http:// www1.y12.doe.gov/capabilities/sgml/sc34/document/0322.htm, May 2002. BIBLIOGRAPHY

[101] F. Manola and E. Miller. Resource Description Framework (RDF) Primer. World Wide Web W3C Recommendation 10 February 2004 [Online] Available: http://www. w3.org/TR/rdf-primer/, February 2004.

[102] D.L. McGuinness and F. van Harmelan (Eds). OWL Web Ontology Language Overview. World Wide Web Consortium (W3C) Recommendation 10 February 2004. [Online] Available: http://www.w3.org/TR/owl-features/, February 2004.

[103] Chris Mellish and Xiantang Sun. Natural Language Directed Inference in the Presen- tation of Ontologies. In 10th European Workshop on Natural Language Generation, Aberdeen, Scotland, August 8–10th 2005.

[104] John Stuart Mill. A System of Logic. Harper & Brothers, New York, 1874.

[105] Joo-Chan Sohn Minsu Jang. Bossam: An extended rule engine for owl inferencing. volume 3323/2004, pages 128–138, Hiroshima, 08/11/2004 2004. Springer Berlin / Heidelberg.

[106] Graham Moore. Rdf and topic maps: An exercise in convergence. [Online] Available: http://xml.coverpages.org/moore-topicmapsrdf200105.pdf, 2001.

[107] Graham Moore, Mary Nishikawa, and Dmitry Bogachev. Topic map constraint language (tmcl) requirements and use cases. ISO 13250 Editors Draft 16 October 2004. [Online] Available: http://www.jtc1sc34.org/repository/0548.htm, October 2004.

[108] Boris Motik, Peter Patel-Schneider, and Ian Horrocks. OWL 2 Web Ontology Lan- guage: Structural Speciﬁcation and Functional Syntax. World Wide Web Consortium (W3C) Working Draft 11 April 2008. [Online] Available: http://www.w3.org/TR/ owl2-syntax/, April 2008.

[109] L.E. Narens. Theories of Meaningfulness. Scientiﬁc Psychology Series. Lawrence Erlbaum Associates, New Jersey, 2002.

[110] B. Niemann, R.F. Morris, H.J. Riofrio, and E. Carnes (Eds). Introducing semantic technologies and the semantic Web (SICoP White Paper Series Module 1). [Online] Available: http://colab.cim3.net/ﬁle/work/SICoP/WhitePaper/, (accessed 6 June 2007), 2005.

[111] Ian Niles and Adam Pease. Towards a standard upper ontology. In In Proceedings of the 2nd International Conference on Formal Ontology in Information Systems (FOIS- 2001, pages 2–9. ACM Press, 2001.

[112] N. Noy, R. Fergerson, and M. Musen. The knowledge model of Protégé-2000:Com- bining interoperability and flexibility. In R. Dieng and O. Corby, editors, Proc. of the 12th EKAW, volume 1937 of LNAI, pages 17–32. Springer, 2000.

[113] Natalya F. Noy and Deborah L. McGuinness. Ontology Development 101: A Guide to Creating Your First Ontology. [Online] Available: http://protege.stanford.edu/ publications/ontology development/ontology101.html, 2001.

[114] University of Manchester. Fact++ OWL-DL reasoner. [Software]. See http://owl. man.ac.uk/factplusplus/. 245

[115] C. K. Ogden and I. A. Richards. The Meaning of Meaning: A Study of the Inﬂuence of Language Upon Thought and of the Science of Symbolism. Routledge and Kegan Paul, London, 1923.

[116] C.K. Ogden. Basic English: A General Introduction with Rules and Grammar. Paul Treber & Co., Ltd, London, 1930.

[117] Ontopia. Omnigator Topic Map Browser. [Software]. See http://www.ontopia.net/ solutions/omnigator.html.

[118] Ordnance Survey. Buildings and Places, 2008. http://www.ordnancesurvey.co.uk/ ontology/v1/BuildingsAndPlaces.owl.

[119] G. H. R Parkinson and Mary Morris, editors. Leibniz: Philosophical Writings. J. M. Dent, London, 1973.

[120] Bijan Parsia and Peter F. Patel-Schneider. Meaning and the semantic web. In Stuart I. Feldman, Mike Uretsky, Marc Najork, and Craig E. Wills, editors, WWW (Alternate Track Papers & Posters), pages 306–307. ACM, 2004.

[121] Peter Patel-Schneider and Ian Horrocks. OWL 1.1 Web Ontology Language Overview. World Wide Web Consortium (W3C) Member Submission 19 December 2006. [Online] Available: http://www.w3.org/Submission/2006/SUBM-owl11-overview-20061219/, December 2006.

[122] Peter F. Patel-Schneider, Patrick Hayes, and Ian Horrocks (Eds). OWL Web Ontology Language Semantics and Abstract Syntax.

[123] Charles Sanders Peirce. The Collected Papers of Charles Sanders Peirce(8 Vols.). Harvard University Press, Cambridge, 1931-58.

[124] Bj¨ornPelzer and Christoph Wernhard. System description: E-KRHyper. In Frank Pfenning, editor, Automated Deduction - CADE-21, 21st International Conference on Automated Deduction, Bremen, Germany, July 17-20, 2007, Proceedings, volume 4603 of Lecture Notes in Computer Science, pages 508–513. Springer, 2007.

[125] Steve Pepper and Fabio Vitali (Eds). Rdftm: Survey of interoperability proposals. W3C Editor’s Draft 24 February 2005. [Online] Available: http://www.w3.org/2001/ sw/BestPractices/RDFTM/survey-2005-02-24, February 2005.

[126] Steve Pepper and Graham Moore (Eds). XML Topic Maps (XTM) 1.0. Top- icMaps.Org Speciﬁcation 06 August 2001. [Online] Available: http://www.topicmaps. org/xtm/1.0/xtm1-20010806.html, August 2001.

[127] J. Piaget. The Principles of Genetic Epistemology. Basic Books, New York, 1972.

[128] Antonella Poggi, Domenico Lembo, Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Riccardo Rosati. Linking data to ontologies. J. on Data Semantics, X:133–173, 2008.

[129] J.T. Pollock and R. Hodgson. Adaptive information: Improving business through semantic interoperability, grid computing, and enterprise integration. Wiley Series in Systems Engineering Management. John Wiley & Sons, Inc., 2004. BIBLIOGRAPHY

[130] HP Labs Semantic Web Programme. Jena An open source Semantic Web Framework for Java. [Software]. See http://jena.sourceforge.net/. [131] Apache Lucene project. Solr - open source enterprise search server. [Software]. See http://lucene.apache.org/solr/.

[132] Eric Prudhommeaux´ and Andy Seaborne(Eds). SPARQL Query Language for RDF. World-Wide Web Consortium (W3C) Recommendation 15 January 2008. [Online] Available: http://www.w3.org/TR/rdf-sparql-query/, January 2008. [133] A. Rector, N. Drummond, M. Horridge, J. Rogers, H. Knublauch, R. Stevens, H. Wang, and C. Wroe. OWL pizzas: Practical experience of teaching OWL-DL: Common errors & common patterns. In E. Motta, N. R. Shadbolt, and A. Stutt, editors, Proc. of the 14th EKAW, volume 3257 of LNCS, pages 63–81. Springer, 2004. [134] Reuters. Open Calais. [Online] Available: http://www.opencalais.com/, 2008. [135] R. Robinson. Deﬁnition. Oxford University Press, London, 1950. [136] E. Rosch, C.B. Mervis, W.D. Gray, D.M. Johnson, and P. Boyes-Braem. Basic Objects in Natural Categories. Cognitive Psychology, 8:382–439, 1976. [137] R. Schwitter. English as a formal speciﬁcation language. In Proceedings of the Thir- teenth International Workshop on Database and Expert Systems Applications (DEXA 2002), pages 228–232, 2002. [138] Rolf Schwitter, Anna Ljungberg, and David Hood. ECOLE — A Look-ahead Ed- itor for a Controlled Language. In Controlled Translation, Proceedings of EAMT- CLAW03, Joint Conference combining the 8th International Workshop of the Euro- pean Association for Machine Translation and the 4th Controlled Language Applica- tion Workshop, pages 141–150, Dublin City University, Ireland, May 15–17th 2003. [139] Rolf Schwitter and Marc Tilbrook. Let’s Talk in Description Logic via Controlled Natural Language. In Logic and Engineering of Natural Language Semantics 2006, (LENLS2006), Tokyo, Japan, June 5–6th 2006. [140] Rolf Schwitter and Mark Tilbrook. Annotating websites with machine-processable information in controlled natural language. In Advances in Ontologies 2006, Pro- ceedings of the Second Australasian Ontology Workshop (AOW 2006), pages 75–84, 2006. [141] J. Searle. Minds, brains and programs. Behavioral and Brain Sciences, 3:417–457, 1980. [142] Clay Shirky. The Semantic Web, Syllogism, and Worldview. First published November 7, 2003 on the “Networks, Economics, and Culture” mailing list. On-line at http: //www.shirky.com/writings/semantic syllogism.html, 2003. [143] Nuno Silva, Zoltan Varady, Frank Westerhausen, Oliver Fodor, Pedro Vieira Silva, and Paulo Maio. MAFRA - Ontology MApping FRAmework Toolkit. [Software]. See http://mafra-toolkit.sourceforge.net/. [144] M.K. Smith, C. Welty, and D.L. McGuinness (Eds). OWL Web Ontology Language Guide. World Wide Web Consortium (W3C) Recommendation 10 February 2004. [Online] Available: http://www.w3.org/TR/owl-guide/, February 2004. 247

[145] D. Snowden. Complex Acts of Knowing: Paradox and Descriptive Self Awareness. Journal of Knowledge Management, 6(2):100–111, May 2002. [146] John F. Sowa. Common Logic Controlled English. Technical report, 2007. Draft, 15 March 2007, http://www.jfsowa.com/clce/clce07.htm. [147] John F. Sowa. Current semantic web layer cake. Email sent to ontolog- forum [Online]. Available: http://ontolog.cim3.net/forum/ontolog-forum/2007-07/ msg0025.html, July 2007. [148] Susan Thomas. The importance of being earnest about deﬁnitions. In Business, Technologie und Web (BTW 2005), pages 560–577, 2005. [149] TopQuadrant. TopBraid Composer. Ontology Editor [Software]. Available http:// www.topquadrant.com/topbraid/composer/index.html. [150] A. Wierzbicka. Semantics: Primes and Universals. Oxford University Press, Oxford, 1996. [151] World Wide Web Consortium (W3C). Protocol for Web Description Resources (POW- DER) Working Group Charter. World Wide Web Consortium (W3C) webpage [On- line] Available: http://www.w3.org/2007/02/powder charter. [152] World Wide Web Consortium (W3C). Simple Knowledge Organization System (SKOS) Home Page. World Wide Web Consortium (W3C) webpage [Online] Avail- able: http://www.w3.org/2004/02/skos/. [153] World Wide Web Consortium (W3C). W3C Policy Languages Interest Group. [On- line] Available: http://www.w3.org/Policy/pling/, year = 2008. [154] World Wide Web Consortium (W3C). W3C Semantic Web Activity Statement. [On- line] Available: www.w3.org/2001/sw/Activity, accessed 1 July 2008. [155] World Wide Web Consortium (W3C). Uncertainty Reasoning for the World Wide Web Incubator Group Initial Report. [Online] Available: http://www.w3.org/2005/ Incubator/urw3/XGR-urw3-20080331/, April 2008. [156] World Wide Web Consortium (W3C). W3C RIF Working Group. [Online] Available: http://www.w3.org/2005/rules/wiki/RIF Working Group,accessed1August2008, 2008. [157] World Wide Web Consortium (W3C). W3C Semantic Web Home Page. [Online] Available: www.w3.org/2001/sw/, accessed 1 July 2008, 2008.