D2.2.1 Methods for Selection and Integration of Reusable ... - CiteSeerX

NeOn-project.org NeOn: Lifecycle Support for Networked Ontologies Integrated Project (IST-2005-027595) Priority: IST-2004-2.4.7 — “Semantic-based knowledge and content systems”

D2.2.1 Methods for Selection and Integration of Reusable Components from Formal or Informal User Specifications

Deliverable Co-ordinator:

Marta Sabou

Deliverable Co-ordinating Institution:

The Open University (OU)

Other Authors: Sofia Angeletou(OU), Mathieu d’Aquin(OU), Jesús Barrasa(UPM), Klaas Dellschaft(UKO-LD), Aldo Gangemi(CNR), Jos Lehmann(CNR), Holger Lewen(UKARL), Diana Maynard(USFD), Dunja Mladenic(JSI), Malvina Nissim(CNR), Wim Peters(USFD), Valentina Presutti(CNR), Boris Villazón(UPM)

This deliverables describes a set of methods for data re-engineering, ontology evaluation and selection, that are necessary to support collaboration in the broadest sense defined in D2.1.1 (i.e., usage).

Document Identifier: Class Deliverable: Project start date Project duration:

NEON/2007/D2.2.1/0.5 NEON EU-IST-2005-027595 March 1, 2006 4 years

Date due: Submission date: Version: State: Distribution:

April 30, 2007 May 28, 2007 0.5 Final Public

c Copyright lies with the respective authors and their institutions. 2006–2007

Page 2 of 151

NeOn Integrated Project EU-IST-027595

NeOn Consortium This document is part of a research project funded by the IST Programme of the Commission of the European Communities, grant number IST-2005-027595. The following partners are involved in the project: Open University (OU) – Coordinator Knowledge Media Institute – KMi Berrill Building, Walton Hall Milton Keynes, MK7 6AA United Kingdom Contact person: Martin Dzbor, Enrico Motta E-mail address: {m.dzbor, e.motta}@open.ac.uk Universidad Politécnica de Madrid (UPM) Campus de Montegancedo 28660 Boadilla del Monte Spain Contact person: Asunción Gómez Pérez E-mail address: [email protected] Intelligent Software Components S.A. (ISOCO) Calle de Pedro de Valdivia 10 28006 Madrid Spain Contact person: Richard Benjamins E-mail address: [email protected] Institut National de Recherche en Informatique et en Automatique (INRIA) ZIRST – 655 avenue de l’Europe Montbonnot Saint Martin 38334 Saint-Ismier France Contact person: Jérôme Euzenat E-mail address: [email protected] Universität Kolenz-Landau (UKO-LD) Universitätsstrasse 1 56070 Koblenz Germany Contact person: Steffen Staab E-mail address: [email protected] Ontoprise GmbH. (ONTO) Amalienbadstr. 36 (Raumfabrik 29) 76227 Karlsruhe Germany Contact person: Jürgen Angele E-mail address: [email protected] Food and Agriculture Organization of the United Nations (FAO) Viale delle Terme di Caracalla 1 00100 Rome, Italy Contact person: Marta Iglesias E-mail address: [email protected]

Universität Karlsruhe – TH (UKARL) Institut für Angewandte Informatik und Formale Beschreibungsverfahren – AIFB Englerstrasse 28 D-76128 Karlsruhe, Germany Contact person: Peter Haase E-mail address: [email protected] Software AG (SAG) Uhlandstrasse 12 64297 Darmstadt Germany Contact person: Walter Waterfeld E-mail address: [email protected] Institut ‘Jožef Stefan’ (JSI) Jamova 39 SI–1000 Ljubljana Slovenia Contact person: Marko Grobelnik E-mail address: [email protected] University of Sheffield (USFD) Dept. of Computer Science Regent Court 211 Portobello street S14DP Sheffield United Kingdom Contact person: Hamish Cunningham E-mail address: [email protected] Consiglio Nazionale delle Ricerche (CNR) Institute of cognitive sciences and technologies Via S. Marino della Battaglia, 44 – 00185 Roma-Lazio, Italy Contact person: Aldo Gangemi E-mail address: [email protected] Asociación Española de Comercio Electrónico (AECE) C/Alcalde Barnils, Avenida Diagonal 437 08036 Barcelona Spain Contact person: Jose Luis Zimmerman E-mail address: [email protected] Atos Origin S.A. (ATOS) Calle de Albarracín, 25 28037 Madrid Spain Contact person: Tomás Pariente Lobo E-mail address: [email protected]

D2.2.1 Methods for Selection and Integration of Reusable Components

Page 3 of 151

Work package participants The following partners have taken an active part in the work leading to the elaboration of this document, even if they might not have directly contributed writing parts of this document:

• Open University (OU) • Universidad Politécnica de Madrid (UPM) • Universität Kolenz-Landau (UKO-LD) • Consiglio Nazionale delle Ricerche (CNR) • Universität Karlsruhe – TH (UKARL) • University of Sheffield (USFD) • Institut ‘Jožef Stefan’ (JSI) • Food and Agriculture Organization of the United Nations (FAO)

Change Log Version 0.1 0.2 0.3 0.4 0.5

Date 16-01-2007 13-01-2007 12-04-2007 17-04-2007 28-05-2007

Amended by Marta Sabou Marta Sabou and All Marta Sabou and All All Aldo Gangemi, Wim Peters, Marta Sabou and Boris Villazón

Changes First draft outline First draft contributions from partners Final contributions from partners Final additions and changes Updates based on comments from the internal reviewer


Page 4 of 151


Executive Summary This deliverables describes a set of methods that are necessary to support collaborative ontology design in the broadest sense defined in D2.1.1 (i.e., usage). A usage type collaboration is a collaboration process that takes place between two agents: the first agent reuses a knowledge component authored by the second agent (possibly at some different time and for some other purposes), while the second agent does not have to be aware of or involved in this process (Chapter 2). The material of the deliverable is structured in three parts dealing with the three major functionalities identified as important to support usage type collaborations: data re-engineering, ontology evaluation and ontology selection. In the first part of the deliverable we overview methods that allow re-engineering a wide range of resources to ontologies so that they can be reused and integrated into ontology based projects. These resources can be textual corpora (see Chapter 3), databases (Chapter 4), other conceptual structures such as Knowledge Organization Systems (KOS) or Lexica (Chapter 5), socially evolved structures such as folksonomies (Chapter 6) and data model patterns (Chapter 7). In the second part we describe ontology evaluation methods that are useful both for assessing the quality of the re-engineering methods and the quality of individual ontologies, thus making it possible that only suitable ontologies are reused. In Chapter 8 we give a principled overview of ontology evaluation techniques and classify the material from this part of the deliverable according to this framework. Then we detail three different gold standard based evaluation measures: (1) the OntoRand index for evaluating instantiated ontologies; (2) precision and recall based measures for the evaluation of the lexical and taxonomic layers of ontologies and (3) the balanced distance metric (BDM) for evaluating ontology population (Chapter 9). We end this part by describing how Open Rating Systems can be adapted to support collaborative ontology evaluation by combining evaluations done both by automatic methods and by a community of users (Chapter 10). In the final part, we describe work that investigates ontology selection in the context of automatic reuse. We define what we mean by ontology selection and overview existing work. After deriving a set of requirements imposed by two applications and experimentally investigating the characteristics of online ontology repositories, we present a selection algorithm that balances between obtaining a complete and precise coverage and offering a good performance. Finally, we make the first step from ontology selection towards selecting appropriate reusable knowledge components by combining the selection algorithm with ontology modularization techniques (Chapter 11). The material presented here was drawn from (1) pre-NeOn work which is currently applied on NeOn data, (2) novel approaches developed and published during the first year of the project and (3) ongoing efforts that will be fully reported in the next deliverable. As said in the introduction, it is worth noting the high level of integration of this work within the project: the majority of approaches described have been either already applied on case study data or have been integrated with methods developed in other WPs. Also, this active task lead to a broad area of future research which will be conducted on two separate foci. First, given their practical importance, data re-engineering methods will be explored in task T.2.2.a. Second, a closer integration between ontology selection and evaluation techniques will be the topic of T.2.2.b.


Page 5 of 151

Note on Sources and Original Contributions The NeOn consortium is an inter-disciplinary team, and in order to make deliverables self-contained and comprehensible to all partners, some deliverables thus necessarily include state-of-the-art surveys and associated critical assessment. Where there is no advantage to recreating such materials from first principles, partners follow standard scientific practice and occasionally make use of their own pre-existing intellectual property in such sections. In the interests of transparency, we here identify the main sources of such preexisting materials in this deliverable:

• Chapter 4 is partially based on [Bar]. • Section 5.3 is partially based on [vAGS06]. • Section 5.4 partly relies on material previously published in [GNV03]. • Section 5.6 partly relies on material previously published in [GFK+ 04]. • Section 6.1 presents material submitted for peer review in [ASSM07]. • Chapter 8 partly relies on material that will be published in [DS07]. • Section 9.2 reuses part of the material published in [BGM06]. • Section 9.4 includes material that has been previously published in [MPL06]. • Chapter 9 partly relies on material that has been previously published in [DS06]. • Chapter 10 partly relies on material that has been previously published in [LSNM06]. • Chapter 11 combines material that has been previously published in [SLMU06], [SLM06] and [dSM06].


Page 6 of 151


Contents 1 Introduction

13

1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.2 Integration with the Rest of the Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 1.3 A note on the source of this material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2 Collaboration

17

2.1 What is Collaboration? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.1.1 Formal Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.1.2 Types of Collaborations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.2 Functionalities that Support Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

I

Data ReEngineering

22

3 Ontology Learning

24

3.1 Ontology Learning Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3.1.1 Goal

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.1.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 3.2 OntoGen - a System for Semi-Automatic Ontology Construction . . . . . . . . . . . . . . . . . 25 4 Upgrading Database Content

28

4.1 Existing approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 4.2 R2 O and ODEMapster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 4.2.1 The R2 O Language

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.2.2 The ODEMapster Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 4.3 Case Study: Upgrading dabase content on FAO data

. . . . . . . . . . . . . . . . . . . . . . 32

4.4 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 5 Reengineering Thesauri and Lexica

33

5.1 Mapping Agrovoc onto WordNet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 5.1.1 Advantages of Mapping Agrovoc onto WordNet . . . . . . . . . . . . . . . . . . . . . . 35 5.1.2 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 5.1.3 Term Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 5.1.4 Lexical Match . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 5.1.5 Disambiguation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 5.1.6 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 5.1.7 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38


Page 7 of 151

5.2 Reengineering Lexica: the case of Princeton WordNet . . . . . . . . . . . . . . . . . . . . . . 38 5.3 Lexicon to Ontology ABox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.3.1 W3C WNET method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.3.2 Linking WordNet ABox to C-ODO ABox . . . . . . . . . . . . . . . . . . . . . . . . . . 43 5.4 Lexicon to Ontology TBox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 5.4.1 Assumptions made during Reengineering . . . . . . . . . . . . . . . . . . . . . . . . . 46 5.4.2 Applying Assumptions to WordNet Reengineering . . . . . . . . . . . . . . . . . . . . 47 5.4.3 Learning association links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 5.4.4 Learning conceptual relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 5.5 KOS to Ontology ABox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 5.6 KOS to Ontology TBox. An example in fishery . . . . . . . . . . . . . . . . . . . . . . . . . . 51 5.6.1 Formatting and Lifting in FOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 5.6.2 Formalizing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 5.6.3 Enriching the Conceptualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 6 Enriching and Exploring Folksonomies

58

6.1 Semantically Enriching Folksonomies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 6.1.1 Semantic Enrichment of Folksonomy Tag Space . . . . . . . . . . . . . . . . . . . . . 60 6.1.2 Experimental Results

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6.1.3 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 6.2 Learning from Folksonomies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 7 Deriving Ontology Design Patterns

70

7.1 Basic Description and Situation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 7.2 Plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 7.3 An example of re-engineering in business modeling . . . . . . . . . . . . . . . . . . . . . . . 76 7.4 Conclusion and future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

II

Ontology Evaluation

84

8 Ontology Evaluation for Collaboration

85

8.1 A Principled Overview of Ontology Evaluation

. . . . . . . . . . . . . . . . . . . . . . . . . . 85

8.2 Evaluation Scenarios and Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 8.2.1 Scenario 1: Quality Assurance During Ontology Engineering . . . . . . . . . . . . . . . 87 8.2.2 Scenario 2: Comparing Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . 89 8.2.3 Scenario 3: Evaluation for Ontology Selection 9 Gold Standard Based Evaluation

. . . . . . . . . . . . . . . . . . . . . . 91 92

9.1 Criteria for Good Evaluation Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 9.2 Instance Based Evaluation with OntoRand . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 9.3 Precision and Recall based Measures

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

9.3.1 Lexical Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 9.3.2 Taxonomic Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 9.3.3 Evaluation of the Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 9.4 The Balanced Distance Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102


Page 8 of 151


9.4.1 Learning Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 9.4.2 Evaluation of the BDM metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 9.4.3 Discussion and Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 10 User-based Evaluation of Ontologies

108

10.1 Evolution of Open Rating Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 10.2 Underlying Model and Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 10.2.1 Extended Open Rating System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 10.2.2 Different Trust Statements in the Extended Open Rating System . . . . . . . . . . . . . 111 10.2.3 Trust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 10.2.4 Propagation of Trust and Distrust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 10.2.5 Computing Trust Values for the Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . 113 10.2.6 Ranking Evaluations at the Property Level of an Ontology . . . . . . . . . . . . . . . . 114 10.2.7 Computing an Overall Evaluation of an Ontology . . . . . . . . . . . . . . . . . . . . . 115 10.2.8 Ranking Ontologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 10.3 Enabling Collaborative Ontology Evaluation with ORS . . . . . . . . . . . . . . . . . . . . . . 115 10.4 Combining Different Evaluation Techniques into the ORS Framework . . . . . . . . . . . . . . 116

III

Ontology Selection

118

11 Ontology Selection

119

11.1 What is Ontology Selection? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 11.2 Current approaches to ontology selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 11.2.1 Popularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 11.2.2 Richness of Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 11.2.3 Topic Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 11.2.4 Summary of Selection Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 11.3 Requirements for Supporting Automatic Knowledge Reuse Scenarios . . . . . . . . . . . . . . 123 11.3.1 Ontology based Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 11.3.2 Semantic Browsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 11.3.3 Requirements for Ontology Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 11.4 Preliminary Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 11.4.1 Experiment 1 - Obtaining Complete Coverage

. . . . . . . . . . . . . . . . . . . . . . 126

11.4.2 Experiment 2 - Dealing with Compound Labels . . . . . . . . . . . . . . . . . . . . . . 127 11.5 The Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 11.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 11.5.2 Step1: Query Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 11.5.3 Step2: Ontology Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 11.5.4 Step3: Identifying relevant ontology combinations. . . . . . . . . . . . . . . . . . . . . 131 11.5.5 Step4: Generality Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 11.5.6 Extending Semantic Match to Deal with Compound Labels . . . . . . . . . . . . . . . . 132 11.6 From Ontology Selection to Knowledge Selection

. . . . . . . . . . . . . . . . . . . . . . . . 133

11.7 Discussion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134


IV

Final Considerations

Page 9 of 151

136

12 Future Work

137

12.1 Future Work in Data Re-engineering (T2.2.a) . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 12.2 Future Work in Ontology Evaluation and Selection (T2.2.b)

. . . . . . . . . . . . . . . . . . . 138

A BNF Notation of the R2 O Language Grammar

140

Bibliography

143


Page 10 of 151


List of Tables 5.1 OWL property matrix obtained after the conversion. . . . . . . . . . . . . . . . . . . . . . . . 43 6.1 mushroom related tags that could not be connected semantically

. . . . . . . . . . . . . . . 64

6.2 fruit related tags that could not be connected semantically . . . . . . . . . . . . . . . . . . 65 6.3 beverage related tags that could not be connected semantically

. . . . . . . . . . . . . . . 66

6.4 mammal related tags that could not be connected semantically . . . . . . . . . . . . . . . . . 67 9.1 Semantic cotopies for the ontologies in Fig. 9.1. . . . . . . . . . . . . . . . . . . . . . . . . . 96 9.2 Common semantic cotopies for the ontologies in Fig. 9.1. . . . . . . . . . . . . . . . . . . . . 96 9.3 Evaluation of the ontologies in Fig. 9.3 with a semantic cotopy based measure . . . . . . . . . 99 9.4 Evaluation of the ontologies in Fig. 9.3 with a common semantic cotopy based measure . . . . 100 9.5 Evaluation of the ontologies in Fig. 9.4 with a semantic cotopy based measure . . . . . . . . . 100 9.6 Evaluation of the ontologies in Fig. 9.4 with a common semantic cotopy based measure . . . . 101 9.7 Structural evaluation of the reference ontology and the learned ontologies . . . . . . . . . . . 103 9.8 Comparison of the two learning algorithms Hieron and SVM with uneven margins for OBIE using three overall performance measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 9.9 Examples of entities misclassified by the Hieron based system . . . . . . . . . . . . . . . . . 106 10.1 Allowed Trust Statements and Their Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 10.2 Atomic Trust Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 11.1 Comparison of existing ontology selection approaches. * PowerAqua is still under development. 123 11.2 Number of ontologies retrieved for a set of queries. (X+ refines X.) . . . . . . . . . . . . . . . 127 11.3 Analysis of the appearances of some conjunctions and other terms. . . . . . . . . . . . . . . . 128


Page 11 of 151

List of Figures 3.1 Screen shot of the interactive system OntoGen for construction of topic ontologies. . . . . . . . 26 4.1 Schematic description of the R2 O and ODEMapster . . . . . . . . . . . . . . . . . . . . . . . 29 4.2 ODEMapster operation modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 5.1 Notions of ontology in the C-ODO library. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 5.2 Linking OWL WordNet classes to the C-ODO library. . . . . . . . . . . . . . . . . . . . . . . . 44 5.3 Common linking to C-ODO for OWL versions of WordNet, SKOS, FOAF, and DCMITYPE. . . . 44 5.4 A UML activity diagram for formatting and lifting activities in ONIONS@FOS. . . . . . . . . . . 54 5.5 Topic spaces (worldviews) in oneFish.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5.6 The activity diagram for metadata formalization and Core ontology building. . . . . . . . . . . . 55 6.1 Lion in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 6.2 Mushroom in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 6.3 Fruit in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 6.4 Beverage in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 6.5 Mammal in the Semantic Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 7.1 The DnS CODeP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 7.2 The Plan CODeP

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7.3 Generalization of plan elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 7.4 Conditions and Sequences in Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 7.5 Relations between Concepts of Plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 7.6 Plan Tasks Taxonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 7.7 Kinds of Contracts Data Model Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 7.8 Sales Order Use Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 7.9 Sales Order Process possible workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 7.10 Sales Order Process CODeP-based description . . . . . . . . . . . . . . . . . . . . . . . . . 80 7.11 Kind Of Contracts Ontology Classes

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

7.12 Kind Of Contracts Ontology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 7.13 Unified Model

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

9.1 Example reference ontology (OR1 , left) and computed ontology (OC1 , right) . . . . . . . . . . 95 9.2 Building blocks of the global taxonomic precision measure . . . . . . . . . . . . . . . . . . . . 97 9.3 Reference ontology (OR2 , left) and two learned ontologies (OC2 , middle; OC3 , right) . . . . . . 99 9.4 Reference ontology (OR3 , left) and two learned ontologies (OC4 , middle; OC5 , right) . . . . . . 100 9.5 Evaluation of the lexical layer depending on threshold θ . . . . . . . . . . . . . . . . . . . . . 101


Page 12 of 151


9.6 Evaluation of learned ontologies with T Pcsc depending on threshold θ

. . . . . . . . . . . . . 102

9.7 Evaluation of learned ontologies with T Psc depending on threshold θ . . . . . . . . . . . . . . 102 9.8 Subset of the Proton ontology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 10.1 This figure depicts how the 11 components of the extended ORS Model relate to each other. Note that agents can also simply rate other agents, they do not necessarily have to review ontologies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 10.2 This figure depicts how automated evaluation techniques can be used to automatically populate ontology reviews in ORS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 11.1 The main tasks and stages of the selection algorithm. . . . . . . . . . . . . . . . . . . . . . . 129 11.2 The knowledge selection process and its use for semantic browsing with Magpie. 11.3 A Prototype implementation of the selection and modularization methods.

. . . . . . . 133

. . . . . . . . . . . 134

11.4 The original ontology (a) and the resulting module (b). . . . . . . . . . . . . . . . . . . . . . . 135


Page 13 of 151

Chapter 1

Introduction The notion of collaboration is core to the vision of the NeOn project which assumes that: “large ontologies are built by teams, often distributed across time and space, and large semantic applications make use of knowledge generated from a variety of sources" [Con06]. Hence, collaboration covers a wide range of scenarios, related to collaboration between human users (expert teams) on one hand, and collaboration by reusing various sources of knowledge by integrating them in large scale applications. Workpackage 2, Collaborative Aspects for Networked Ontologies, investigates collaborative aspects of ontology development and reuse. The main objective of WP2 is to analyze and describe (by means of a suitable meta-model) the activities underlying collaborative design of networked ontologies, and to produce appropriate methods and tools to support the related workflow, by focusing on the collaborative aspects underlying the work of a knowledge community. This objective will be achieved through investigating the following five tasks:

• (T2.1) Formal specification of collaborative ontology design and characterization of design rationales for knowledge communities. The first deliverable of this task (D2.1.1) provides a formal framework to describe collaboration and design. Notions related to collaboration and collaborative design are explicitly formalized in the context of the C-ODO ontology.

• (T2.2) Methods and tools for evaluation, selection, and reengineering of ontologies and other knowledge sources. This task investigates the three major functionalities that are mandatory for supporting collaboration.

• (T2.3) Methods and tools for collaborative annotation and discussion of ontologies. This task is concerned with developing tools that support collaborative ontology design (e.g., argumentation, annotation).

• (T2.4) Multilingual and localization support for ontologies. This task focuses on the linguistic aspects of ontologies. It aims to overcome the linguistic barrier between users belonging to non-mutuallyaccessible linguistic communities, be them constituted of speakers of different natural languages, or of speakers belonging to different communities of practice, which use specialized lexica.

• (T2.5) Library creation and management support (registry) for ontology design patterns. This task focuses on ontology design patterns, i.e. modular components of ontologies that contain a clearly defined design rationale. T2.5 aims to support most degrees of collaboration, by providing a predefined set of choices, out of which new ontologies can be built, existing ontologies can be analyzed and possibly refactored, and that can be discussed during active collaboration. The goal of this deliverable is to give a comprehensive overview of methods developed by NeOn partners that can support the three main tasks needed to ensure collaboration at the level of reusing components that have not been designed for that purpose (re-engineering, ontology evaluation and ontology selection). As such,


Page 14 of 151


this deliverable complements the material presented in the other deliverables of this work package. First, we adopt and rely on the definition and formalization of collaboration as it was given in D2.1.1. According to that deliverable (see Chapter 2), the set of functionalities presented here are useful to support collaboration in general and virtual collaboration in particular (corresponding to the usage scenario). Second, we envision that the collaborative methods and tools designed in D2.3.1. will reuse and integrate methods and tools presented in this deliverable. For example, ontology selection techniques could be a core part of a collaborative environment. Third, issues related to multilinguality are important to provide language-independent ontology selection mechanisms (e.g., if the user requires an ontology which contains the concept Cat, the selection mechanism should also return ontologies containing the same concept but with a label in a different language, e.g., Gatto). Finally, re-use of components will be investigated at a much more fine-grained level in D2.5.1 where instead of reusing ontologies or relevant modules, we aim to support the reuse of design patterns. As other knowledge components, design patterns also require methods for re-engineering them (as proposed already in this deliverable), to evaluate them and to select them.

1.1

Overview

The material of this deliverable is structured as follows. We start by giving the required background on collaboration and justifying why the functionalities we discuss here are relevant for collaboration (Chapter 2). The rest of the material is divided in three main parts which describe a set of methods for data re-engineering, ontology evaluation and ontology selection respectively. Part I: Data ReEngineering In this part we overview a range of methods and tools that aim at obtaining ontologies by re-engineering various types of resources. In Chapter 3 we describe ontology learning methods that extract ontologies from textual corpora. We report on ongoing efforts in WP7 to evaluate how existing ontology learning tools can be used on actual data provided by FAO. We also present OntoGen, which specializes in building topic-hierarchy like ontologies from large document corpora. In Chapter 4 we describe a method to upgrade database content to instantiated ontologies and its use on FAO databases (WP7). In Chapter 5 we describe methods which instead of extracting an entirely new ontology from text or databases start out with conceptual structures such as Knowledge Organization Systems (KOS) or Lexica and evolve them into ontologies through a variety of methods. In Chapter 6 we present novel research on exploring socially built knowledge structures (folksonomies are types of KOSes) and try to align them to existing ontologies in order to enrich them semantically. In Chapter 7 we detail research on obtaining particular types of ontology elements (ontology design patterns) from data model patterns and business languages. Part II: Ontology Evaluation In this part we describe a variety of methods that address the issue of ontology evaluation. Note that, unlike in the case of the methods in Part I, these methods have not yet been applied to experimental data. In Chapter 8 we give a principled overview of ontology evaluation techniques and classify the material from this part of the deliverable according to this framework. We also detail the role of ontology evaluation both in the context of data re-engineering and ontology selection. In Chapter 9 we detail three different gold standard based evaluation measures: (1) the OntoRand index for evaluating instantiated ontologies; (2) precision and recall based measures for the evaluation of the lexical and taxonomic layers of ontologies and (3) the balanced distance metric (BDM) for evaluating ontology population. In Chapter 10 we describe how Open Rating Systems can be adapted to support collaborative ontology evaluation by combining evaluations done both by automatic methods and by a community of users. Note that this complex way of combining different evaluation measures can also be seen as one particular way


Page 15 of 151

to implement ontology selection, and thus makes a nice transition towards the final part of this deliverable concerned with issues on ontology selection. Part III: Ontology Selection In this final part we focus on the topic of ontology selection. Given the early stage of this work, the proposed method has not yet been validated within the case studies. In Chapter 11 we describe work that investigates ontology selection in the context of automatic reuse. We define what we mean by ontology selection and overview existing work. After deriving a set of requirements imposed by two applications and experimentally investigating the characteristics of online ontology repositories, we present a selection algorithm that balances between obtaining a complete and precise coverage and offering a good performance. Finally, we make the first step from ontology selection towards selecting appropriate reusable knowledge components by combining the selection algorithm with ontology modularization techniques. We conclude with the envisioned future work in Chapter 12.

1.2

Integration with the Rest of the Project

It is notable that the work reported in this deliverable has been tightly integrated with the efforts of other work packages in the project. In particular: WP1 As reported in Chapter 11, the ontology selection algorithm was extended with ontology modularization techniques developed in WP1 [dSM06]. WP3 The ontology selection technique developed in WP2 forms a core part of an innovative ontology matching approach which dynamically selects and combines knowledge from online available ontologies to provide alignments between a source and a target ontology [SdM06]. This technique was reported in WP3 and is now being integrated with the Alignment Server built within WP3. WP5 Expertise gathered during the work performed for this task helped in defining the glossary of NeOn functionalities. WP7 Already early in the project, several of the methods presented here have been successfully applied on the FAO data sets provided by WP7. First, preliminary experiments have been performed with using the ontology learning platform, OntoGen, in categorizing FAO documents into existing topic ontologies (Chapter 3). Then, R2 O [Bar06] and ODEMapster [Bar06] were used to populate three ontologies with data extracted from FAO databases (see Chapter 4). In Chapter 5, we describe a thesaurus alignment technique which was used to align AGROVOC to WordNet in order to enhance AGROVOC’s thesaural structure in terms of scope, coverage and detail. The same chapter reports on further work on enriching various KOSes provided by FAO. The work on ontology selection reported in Chapter 11 has lead to two concrete types of experiments with FAO data. First, the matching technique described in [SdM06] has been used to align AGROVOC with a similar thesaurus developed in the US, NALT. The same technique was used to align AGROVOC with two further FAO ontologies, ASFA and FICORE. Second, as a pre-requisite of understanding modularization techniques in order to integrate them with ontology selection, we are performing experiments with a variety of modularization techniques on top of FAO ontologies.

1.3

A note on the source of this material

The material collected in this deliverable has been selected and compiled from three main types of sources:

• Methods developed outside and before NeOn by NeOn partners which have the potential to be applied or useful for NeOn case studies or to support basic research performed in the project. We


Page 16 of 151


decided to include short summaries of these pieces of work as well as ways in which they are currently applied (or will be applied) within NeOn.

• Methods developed during NeOn. NeOn partners performed a significant amount of research related to the core topics of this deliverable within the duration of NeOn. We reproduce here an extended account of these novel achievements, mostly by reproducing relevant parts of already published material.

• Methods currently being developed in NeOn. Since this deliverable is only the first one in task T2.2., there is a lot of research currently going on in this direction. While this research is expected to reach maturity and to be published in the next deliverable, we decided to already give a short account of this novel and ground-breaking work.


Page 17 of 151

Chapter 2

Collaboration The role of this chapter is to explain why the functionalities that we investigate are necessary to support collaboration. For this, in the first part of this chapter we summarize important notions related to collaboration as it has been done in D2.1.1 (Section 2.1). Note that much of the content of Section 2.1 has been directly reused or slightly adapted from D2.1.1. After the basic background notions are provided, we motivate the content of this deliverable with respect the notion of collaboration (Section 2.2).

2.1

What is Collaboration?

The concept of collaboration has been formalized in D2.1.1. The authors observe that while there is a considerable literature concerned with generic collaboration issues, there is no common understanding of what this concept means in the context of ontology design. They also observe that this concept covers a wide range of situations depending on the (spatial-temporal) relationship and aims of the involved agents and the goal of the collaboration. For example, all the following situations (described in D2.1.1) can be intuitively perceived as some kind of collaboration: 1. Two prospective NeOn users both working at UN-FAO, Rome, speaking a fluent English, both expert in their respective fields, competent on using the same communication tools, and collaborating on the design of an ontology for fishery regulations. 2. Two prospective NeOn users working at UN-FAO, Rome, and CABI, Oxford, speaking a decent English, both experts in their respective fields, somehow competent on using at least one communication tool, and collaborating on the design of two ontologies for fishery regulations and for fishery techniques respectively. 3. Two users who work respectively for UN-FAO, and as an independent farmer in the Maluti Mountains, Lesotho. They have never met, have respectively inaccessible languages and, while being both experts in their respective fields, they are not necessarily competent on using the same communication tools. Moreover, they are not going to make any joint work on the design of an ontology for farming techniques; in principle, however, the knowledge owned by the independent farmer can be relevant for that task, and the UN-FAO user could involve the farmer in her project as a consultant. 4. A UN-FAO member reuses a distinction by Aristotle as a high level principle during the ontology modeling process, e.g. that between matter and attribute (substance and accident in Aristotle’s terms). Notice that in the above situations, the involved agents can be collocated in space (situation 1, or not, situations 2-4) and time (situations 1-3, or not, situation 4). Also, the involved agents can share a common goal (e.g., in situation 1-2 the building of an ontology) or not (situations 3-4). What is common to all situations, however, is that there is a knowledge transfer of some form between the involved agents.


Page 18 of 151


In the next section we provide a formal definition of this intuitive view on collaboration (taken from D2.1.1). Note that it is not necessary to fully understand all theoretical notions mentioned here to continue reading the deliverable.

2.1.1

Formal Definition

While, situations 1 and 2 are the most intuitive when we speak about collaboration, in order to cover a wide range of similar situations (such as 3 and 4), the authors of D2.1.1. define the broad notion of epistemic influence, which “is a relationship between rational agents that influences the knowledge of one or more agents in the relationship". Epistemic workflows further specialize the notion of epistemic influence by combining it with the notion of plan. As any plan, an epistemic workflow has:

• a goal - the goal that needs to be achieved, • a task - the instrument used for achieving the goal, • an agent-driven-role - the agent that takes part in the plan. D2.1.1 also defines three knowledgeRoles that are played by information objects involved in the workflow (all defined in D2.1.1, Section 3.2.1):

• a working-knowledge-item role - played by a working-knowledge-item which is an ontology or ontology elements when seen as evolving objects on which a knowledge creator is working (i.e., an input to the workflow that needs to be modified).

• a knowledge-resource role - played by an information object that is used for developing ontologies (e.g., ontology elements, design patterns). This is also an input to the workflow and is used to modify the working-knowledge-item.

• an knowledge-product role - played by ontology elements that are seen as the final result of the workflow (i.e., after the interaction between the previously described inputs). Finally, an epistemic workflow is also characterized by:

• an accountable Performer role - a performer that adopts the main goal of the task (i.e., it is accountable for the goal).

• an argumentation structure. The authors of D2.1.1. repeatedly re-use the descriptions and situations design pattern. According to this pattern, an epistemic workflow is a description which is satisfied by a situation modeled through the epistemic workflow enactment. Each enactment contains the following elements, each of them satisfying a role defined by the epistemic workflow. To help the understanding of these elements we will exemplify them in terms of the first collaboration situation described above:

• A1 = a rational agent involved in the workflow (can be a knowledge collective). In situation 1, one of the agents is one of the FAO employees.

• A2 = a rational agent involved in the workflow (can be a knowledge collective). In situation 1, one of the agents is one of the FAO employees.

• K1 = working knowledge item (expressing a viewpoint/rationale, e.g. a working hypothesis) of A1 . In situation 1, the ontology on fishery regulations is the working knowledge item (i.e., the item that is evolved and modified).


Page 19 of 151

• K2 = ontology-lifecycle resource (expressing a viewpoint/rationale) of A2 and playing the knowledgeresource-role. In situation 1, K2 is the input of A2 to the design process. In situation 4, K2 is the distinction of Aristotle which is reused to improve the ontology design.

• P = plan (e.g. ontology-design project), including at least one knowledge-production goal as main goal (G). The goal in situation 1 is to produce the fishery ontology, while the plan consists of a set of functionalities that need to be used to achieve this goal.

• T K = at least one task (e.g. functionality), of which the plan consists of. For example, using an editing tool to describe the ontology in situation 1.

• R = at least one accountable performer role, played by one or both rational agents, that is targeted at one or more tasks. In situation 1, both agents are involved in the workflow and they are both accountable for the outcome. On the contrary, in situation 4, only the ontology engineer is involved in and accountable for the process.

• AR = an argumentation (that has a design-making component) • Kf = a final information object playing the role of knowledge-product role (expressing an emerging viewpoint/rationale) resulting from epistemic influence. In situation 1, this is the final fishery ontology obtained as the product of the collaboration.

• T = at least one time interval, at which a rational agent interprets an ontology-lifecycle resource. For example, in situation 2, an expert gives a rationale for “purse-seigning fishing" technique interpretation at time T1, and then changes her mind at time T2.

2.1.2

Types of Collaborations

D2.1.1 defines different types of collaboration scenarios, mostly based on the level of accountability of the involved agents. The weakest notion of collaboration is usage, characterized by the non-accountability of A2 . The notion of interaction is stronger than usage because A2 has some degree of accountability: she is aware that a knowledge resource authored by her (K2 ) is used to achieve the final goal, however she is not accountable for the goal of the project. Finally, the strongest notion of collaboration, called collaboration proper assumes that all the agents involved in the process have the maximum degree of involvement/accountability: they all subscribe to the goal of the project. We will now define and discuss each of these cases (again, all this has been taken from D2.1.1). Definition: Usage is a case of epistemic workflow that uses one additional role (with respect to epistemic workflow): non-accountable knowledge creator (for agents that have created the knowledge resource to be used in the workflow, but that are not actively involved in the current knowledge production plan). The example given in D2.1.1 is the scenario where an ontology expert needs to include a selection from an existing ontology into her own ontology. The participants to these scenario are:

• A1 = the ontology expert, who is accountable knowledge creator (i.e., accountable performer and knowledge creator) of K1 • A2 = the author of K2 (could be A1 at a previous time or any other ontology creator at any time), who, wrt P and Kf , is a non-accountable knowledge creator • K1 = the expert’s original ontology • K2 = the ontology to be selected, or take selection from • Kf = the final resulting ontology obtained after including the selection into the original ontology • P = how to build ontology Kf


Page 20 of 151


• G = having Kf built properly • T K = the task of selecting K2 , i.e., an ontology or an appropriate module from an ontology. Definition: Interaction is a case of epistemic workflow that prescribes the proactive involvement of two rational agents. A rational agent can also be a team (a collective), provided that the members perform actions expected by a shared interaction plan. It is crucial, however, that only some of the activated rational agents adopt the goal of the interaction (some of them are non-accountable performers). To exemplify this collaboration case, the authors reuse the same scenario as above with the difference that A1 asks A2 to send her the ontology she needs in order to achieve her goal, and A2 complies. This means that A2 is assigned a role in a task of the project, but it does not imply that she is accountable for anything related to it. Definition: Collaboration (proper) is a case of epistemic workflow that prescribes the proactive involvement of all rational agents, i.e. all the involved rational agents adopt the main goal of the collaboration plan. A rational agent can also be a team (a collective), provided that the members perform actions expected by a shared interaction plan. Continuing the scenario given as example previously, in this case A1 asks A2 to further develop her ontology in a specific direction, so that the final result will be best suited to the selection task A1 has planned to perform in order to achieve her goal. If A2 complies, this means that not only she is assigned a role in a task of the (new) ontology project, but also that she adopts the main goal of the project, thus becoming an accountable performer.

2.2

Functionalities that Support Usage

In this deliverable we concentrate on a set of functionalities typical for supporting the usage scenario (but not limited to this scenario). D2.1.1 distinguishes between a social and a system view of collaboration processes. Here, we adopt the system view and restrict ourselves to report on concrete methods that have been developed to support usage (and possibly other scenarios). The main characteristic of the usage scenario is that the agent which contributes the knowledge resource (K2 ) used to modify the working knowledge item K1 is un-accountable for the use of K2 and the overall goal of the workflow. She has authored K2 for the purposes of another application, in another context than that of the current integration process. Therefore usage scenarios cover all those cases when one wants to reuse a knowledge resource without involving its author. Such cases of “virtual collaboration" are numerous. For example:

• within organizations many diverse knowledge assets (e.g., data sources) exist, authored by different people, for different purposes that could be potentially useful in an ontology design process. For example, when designing an ontology for integrating the various data sources of an organization, a knowledge engineer can rely on existing ontologies if any, the schema of the internal databases, or on a repository of related documents to decide on the concepts that should be included (e.g., he could use semi-automatic methods to extract potential concepts from these sources). All these information objects have been authored by different people at different times and for different purposes, but nevertheless they can be reused in the ontology design process. In the sense of D2.1.1. whenever reusing such a resource, the ontology engineer engages in a collaboration activity of type usage, where the author of the reused asset must not be aware of or involved in this process.

• reusing knowledge assets is also an increasingly popular trend in Semantic Web research in general. As the Semantic Web is gaining momentum, more and more ontologies are made available online ready to be reused in diverse ontology projects. Besides ontologies, which provide a formal representation of knowledge, there are also other less-formal Web available resources that are potentially


Page 21 of 151

useful for ontology design. For example, folksonomies could be useful as they reflect knowledge about the world on which a wide number of people agree. There are three main issues that need to be solved in an usage scenario: Re-engineering data sources: Not all the resources that are useful in an ontology design process are directly reusable. Indeed, in many cases such resources need to be modified in order to be integrated with the original ontology. In the first part of this deliverable we describe a set of methods for reengineering a wide range of weakly formalized knowledge assets (e.g., text, databases schemas, folksonomies, thesauri) into ontologies. Evaluating the quality of the resource: Ontology evaluation is crucial both for assessing the quality of the re-engineering process (how strongly do the derived ontologies reflect the knowledge captured by the original data sources?) and for assessing the quality of individual ontologies so that they can be compared from this perspective (for example, in the context of ontology selection). In the second part of this deliverable we provide a set of methods for ontology evaluation. Locating the right resources: Both within an organization and (especially) on the Web, there can exist a large number of interesting resources for an ontology design task. A prerequisite for reusing a given resource is locating it either within an organization or on the Web. In the last part of the deliverable we propose methods for selecting ontologies from large, automatically populated ontology libraries (Chapter 11).


Page 22 of 151


Part I

Data ReEngineering


Page 23 of 151

As explained in the previous chapters, in this work-package and in NeOn, we adopt a broad notion of collaboration which spans from face to face argumentation between users involved in a joint project to virtual collaborations between an ontology engineer and the author of a knowledge resource where the ontology engineer reuses this asset without the necessary involvement of the author. The knowledge assets shared through a virtual collaboration (also termed usage) can be various ranging from entire ontologies that can be reused, to ontology modules/elements as well as assets that are not available as ontologies e.g., databases, explanatory texts, textual corpora etc. A pre-requisite for reusing the second type of weakly structured resources is that they are brought into an ontological form, i.e., they are re-engineered to formal, semantic resources. In this part of the deliverable we overview a range of methods and tools that aim at obtaining ontologies by re-engineering various types of resources or at adding a semantic layer on top of existing resources. We start by presenting a method that extract ontologies from textual corpora in Chapter 3. Then we describe a method to add a semantic layer on top of database schemas in Chapter 4. The rest of the material in this part focuses on methods that instead of extracting a whole new ontology, start out with some weakly structured assets and evolve them to ontologies. In Chapter 5 we describe a method for enriching thesauri with semantic information by relying on WordNet. Then, Chapter 6 presents novel research on exploring socially built knowledge structures (folksonomies which can be seen as vocabularies of terms) and try to align them to existing ontologies in order to enrich them semantically. Finally, in Chapter 7 we present research on obtaining particular types of ontology elements (ontology design patterns) from data model patterns and business languages. Note also, that many of the presented methods have been developed in close cooperation or tested on data provided by the case studies in the project.


Page 24 of 151


Chapter 3

Ontology Learning In the last decade, several tools have emerged that support ontology learning from structured (such as databases) as well as unstructured data (such as text). Ontology Learning is a knowledge acquisition activity that relies on (semi-) automatic methods to transform unstructured (e.g., corpora), semi-structured (e.g., folksonomies, html pages, etc.) and structured data sources (e.g., data bases) into conceptual structures (e.g., T-Box). The interested reader is referred to the following two surveys for an extensive overview of existing ontology learning methods: [MS01a] and [GPMM03]. In the context of NeOn, such tools come extremely handy in the tasks of assisting ontology editors and subject experts in keeping the information up to date and support the continuous growth of domain knowledge bases. Also, ontology learning methods allow re-engineering textual data into more formal representations and thus facilitate their reuse (and implicitly a usage type collaboration) in applications/projects that rely on the use of ontologies. In this section we report on ongoing NeOn work that is primarily focused on evaluating and adapting existing ontology learning methods to the needs of case study partners. In Section 3.1 we provide a brief overview of the experimental setup of task T7.3 which focuses on evaluating ontology learning methods on FAO data and in Section 3.2 we describe OntoGen, an ontology learning tool that is currently being tested on FAO data. Note that to the date of writing this deliverable ontology learning efforts focused only on data provided by WP7, but we expect that they will also be needed in WP8. We shall report on their use in WP8 in the next deliverable in this task.

3.1

Ontology Learning Experiments

Task T7.3. of WP7 aims at gaining a better understanding on which existing ontology learning methods can be used to support core knowledge acquisition and maintenance tasks within FAO. The FAO Fisheries domain lends itself as a perfect field for experimenting with ontology learning techniques since:

• FAO can provide both structured and unstructured knowledge sources to extract information from • some fishery ontologies already exist, which can be used for bootstrapping techniques as well as testbeds for evaluation

• fisheries experts can be called to judge and manually evaluate the performance of the various tools • tools can be efficiently integrated in the NeOn cycle as to optimise ontology creation in a collaborative fashion In order to obtain the best results from ontology learning experiments within a project that does not require specific innovative research in this field, we have used the following strategy. A survey of existing tools and their availability and functionalities was carried out in parallel with a survey of existing resources from which information could be extracted. Taking the combination of these two aspects, and the user requirements


Page 25 of 151

provided by FAO, in T7.3, we set out a battery of experiments that would exploit existing techniques on existing data to optimally match requirements of end users. This section provides a brief overview of the planned experiments.

3.1.1

Goal

The objective of the T7.3 experiments is to identify techniques, tools and algorithms that will assist ontology editors and subject experts in keeping the information up to date and support the continuous growth of the Fisheries knowledge base. Recommended (through D7.3.1) tools and algorithms will be considered to be included on the Semantic Web for fisheries ontology lifecycle, to ease the process of populating ontologies (with instances and relations/properties), keeping them up-to-date and managing mappings among them. Based on the requirements in D7.1.1 and further discussions with FAO, we have identified three major groups of experiments and candidates for each that can be carried out within this task. 1 The first two groups are focused on learning terminology as well as relations from text.

• Ontology enrichment from textual corpora – Enriching of AGROVOC ontology, by extracting terminology from corpora (Important) – Enriching RTMS-based ontologies, by extracting terminology from corpora (Critical) – Learning alternative (synonyms) names for fish (Important)

• Relation extraction from textual corpora – Refinement of AGROVOC relations (Interesting) – Finding relationships between instances of different ontologies (Critical)

• Ontology mapping

3.1.2

Evaluation

Pre-validation will be done by partners during the pilot experiments, possibly by adopting automatic techniques (e.g. cross-validation) during the pilot experiments phase. Manual subsequent evaluation (FAO) of the presented/demonstrated experiments and to provide feedback to partners. Crucially, the main evaluation criterion will be the reduction of time required to maintain the ontology due to the exploitation of the automatic processes. It will be estimated during the final experiments. The rationale behind this is that the trade-off between possibly correcting the output of a system and the benefit in terms of time economy that derives from semi-automating ontology learning must benefit performance. The outcomes of these experiments will be reported in the context of WP7, however, we envision that they will provide interesting material for our work on collaboration within WP2. For more details on the exact design of these experiments as well as their outcomes (which were not finalized at the time when this deliverable was written) the interested reader is referred to the deliverable of task T7.3.

3.2

OntoGen - a System for Semi-Automatic Ontology Construction

JSI has inside EU 6FP IP project SEKT developed OntoGen - a system for semi-automatic topic ontology construction from a collection of documents [FMG05]. Ontology construction is seen as a process where the user is constructing the ontology and taking all the decisions while the computer provides suggestions for the topics (ontology concepts), and assists by automatically assigning documents to the topics, naming the topics, etc. The user can use the suggestions for concepts and their names, further split or refine 1

Relevance for the case study is ranked as: Critical > Important > Interesting


Page 26 of 151


the concepts, move a concept to another place in the ontology, explore instances of the concepts (in this case documents), etc. The system supports also extreme case where the user can ignore suggestions and manually construct the ontology. All this functionality is available through an interactive GUI-based environment providing ontology visualization and the ability to save the final ontology as RDF. There are two main methodological novelties of the system: (i) suggesting concepts as subsets of documents and (ii) suggesting naming of the concepts. The system was extended as presented in [FGM06] so that it also enables automatic insertion of new documents into an existing topic ontology and proposes simple relations between the concepts. The main features of OntoGen include concept hierarchy visualization, management of a concept providing the concept details and suggestions for the concept naming, formation of a new concept based on unsupervised and supervised machine learning methods, concept visualization.

Figure 3.1: Screen shot of the interactive system OntoGen for construction of topic ontologies. Figure 3.1 shows the main window OntoGen as presented in [FMG05]. The system has three major parts that will be further discussed in following subsections. In the central part of the main window is a visualization of the current topic ontology (Ontology visualization). On the left side of the window is a list of all the topics from this ontology. Here the user can select the topic he wants to edit or further expand into subtopics. Further down is the list of suggested subtopics for the selected topic (Topic suggestion) and the list with all topics that are in relationship with the selected topic. At the bottom side of the window is the place where the user can fine-tune the selected topic (Topic management). While the user is constructing/changing topic ontology, the system visualizes it in real time as a graph with topics as nodes and relations between topics as edges. When the user selects a topic, the system automatically suggests what the subtopics of the selected topic could be. This is done by LSI or k-means algorithms applied only to the documents from the selected topic.


Page 27 of 151

The number of suggested topics is specified by the user. Then, the user selects the subtopics s/he finds reasonable and the system automatically adds them to the ontology with relation ‘subtopic-of’ to the selected topic. The user can also decide to replace the selected topic with the suggested subtopics. Figure 3.1 shows how this feature is implemented in OntoGen. The user can manually edit each of the topics changing its documents (one document can belong to more topics), changing the topic name and its relation to other topics. In NeOn OntoGen as it is can be useful when having different data sources including already populated topic ontologies and collections of documents to be inserted, as available in the NeOn FAO cases study. Preliminary experiments have been performed with OntoGen on the FAO data as a part of JSI activity on NeON WP7 study. Methods to support collaborative ontology design in the broadest sense are also machine learning methods and social network analysis methods. The idea is to analyse collaboration between the users in the networked ontology setting, where the user is working on ontology construction simultaneously with the other users and getting suggestions based on the actions of other users performed while constructing ontology on similar content/topic. In order to perform collaboration analysis, we assume that the data recording the users activities is available containing information on the provided suggestions and the user feedback. In NeOn, we will propose and implement methods based on machine learning and social network analysis and connect them to the existing tool for semi-automatic ontology construction OntoGen. Namely, inside the work performed in NeOn WP3 on providing context to the user based on the activity of the other users, OntoGen will be extended to support recording and exchange of information between different users on the system. In our case, mapping between the user activities and ontologies they are constructing is based on the assumption that each user has a collection of documents that s/he uses to construct ontology with support of OntoGen. Inside NeOn WP2 we will build on the top of that extensions to OntoGen.


Page 28 of 151


Chapter 4

Upgrading Database Content Databases are core components of all modern information systems holding large amounts of data sets. Providing methods that help upgrading the content available in databases into a semantically richer format (i.e., ontology instances) are crucial for providing a smooth transition from legacy databases to semantic infrastructures. This will also allow the reuse of database content in other projects/applications. Therefore, re-engineering methods that focus on databases should be an important part of the NeOn toolkit. Upgrading database content is a functionality which allows a NeOn user (or community of users) to "upgrade" information from an existing database to the Semantic Web by mapping its database schema to an ontology.

4.1

Existing approaches

Let us set the following scenario: we have a legacy DB and we want to generate semantic web content from it. Until now, the following approaches have been reported in literature. The first approach, described in [SSV02b, SSV02a], is based in the semi-automatic generation of an ontology from the database’s relational model by applying reverse engineering techniques supervised by the designer. Then mappings are defined between the database and the generated ontology. Because the level of similarity between both models is very high, mappings will be quite direct and complex mapping situations do not usually appear. A second approach, described in [HSV03], proposes the manual annotation of dynamic web pages which publish database content, with information about the underlying database and about how each content item in a page is extracted from the database. This approach does not deal with complex mapping situations and assumes we want to make our database schema public, which is not always the case. A third approach, the one described in [Biz03] proposes a language to define correspondences between ontology concepts and database schema (D2R) views with a processor that takes such descriptions and extracts massively the content of the database to generate a set of ontology instances out of it. This last approach is richer than the preceding ones but its expressiveness is limited to the definition of mappings between database views and ontology concepts and direct mappings between attributes/relations of the ontology and attributes of the database views. Conditions and transformations which are often needed to describe complex mapping situations can not be defined with this language, although they were added with some other features in the extended version of the language (D2Re), as described in [BCGP03]. A fourth approach described in [PC05], proposes an RDF based representation of both schema and instance data in the database. This RDF representation is based in the Relational.OWL ontology, which contains concepts such as table, column, foreign key, etc. In this approach there is actually no mapping with a pre-existing ontology. Instead, the Relational.OWL ontology is instantiated with the schema elements (tables, columns, keys, etc.) in the database’s relational model and with the instance data as well. This RDF


Page 29 of 151

based representation of the database is made available (published) makes possible the exploitation of the database content in the Semantic Web.

4.2

R2 O and ODEMapster

In this section we present R2 O [Bar06], an extensible and declarative language to describe mappings between relational database (DB) schemas and ontologies, and ODEMapster [Bar06] which is the processor in charge of carrying out the exploitation of the mappings defined using R2 O, performing both massive and query driven data upgrade. The first version of this technology has been developed in Esperonto Project (IST-2001-34373). This first version includes:

• First definition of the language. • Partial implementation of the language. • Simple queries to the system (e.g. instances for a particular class), without a query language. This technology proved extremely useful to address the needs of case study partners, within WP7 and WP8. As a result, this technique is being improved and extended within NeOn project. The modelling scenario contains three main components:

• A set of queries QS about a specific domain S that we want to be answered by a model • The model itself, M, a data source capable of answering certain queries QM described in terms of its elements.

• A correspondence enabling the transformation of QS queries into QM queries and an inverse correspondence enabling the translation of the answers provided by M into answers of S. Figure 4.1 shows a high level description of this situation following the R2 O approach. An ontology and a database schema with some semantic overlap in the domains they cover are to be related. Thus, the database schema can be queried in terms of the ontology elements in a transparent way.

Figure 4.1: Schematic description of the R2 O and ODEMapster


Page 30 of 151


As can be seen in Figure 4.1, an ontology (S in the description above) defines terms in a particular domain (Professor, University, PHDStudent, etc) and a database (M in the description above) does the same in another (Organization, Person). The level of overlap of these two domains allows the definition of correspondences between the terms of one and the other even if these terms model the intersection domain differently from a semantic point of view . A query QS like the one described in Figure 4.1 "Give me the names of all professors in UPM" would have its corresponding QM (probably SQL) over the DB and, similarly, the tuples returned by the DB would have their equivalent in a set of instances of the ontology answering the initial question. The question is expressed in ODEMQL query language [Bar], which is specificly designed for ODEMapster processor; the ontology must be represented in OWL or RDF(S); and the database must be stored in Oracle or MySQL database.

4.2.1

The R2 O Language

R2 O is a declarative, XML-based language that allows the description of arbitrarily complex mapping expressions between ontology elements (concepts, attributes and relations) and relational elements (relations and attributes). The strength of the R2 O language lies in its expressivity and in its DBMS independence. The elements of the language providing such qualities are conditions and operations and the rule-style mapping definition for attributes (see Appendix A for a BNF specification of the R2 O grammar). Conditions and operations Conditions and operations allow the description of "under which circumnstances a database individual (a relational tuple, a database record) can be upgraded to a Semantic Web individual (an instance of the target ontology)" and "what kind of transformations are needed to create a Semantic Web individual from a database individual" respectively. Both are defined in terms of an extendable set of primitives and are identified by their names and the set of named parameters they accept. The values of such parameters can be constant values (has-value), variables referring record fields from the database (has-column), or the result of the execution of other operations (has-transform). The first R2 O excerpts describe a condition based on the "match-regexp" primitive. The condition is verified if the content of column salaryRange of table jobs matches the regular expression. condition "match-regexp" arg-restriction on-param "string" has-column jobs.salaryRange arg-restriction on-param "regexp" has-value ([:digit:]*)-([:digit:]*)

The second fragment describes an operation based on the "concat" primitive. The operation concatenates two constant strings with the content of column id of table jobs. operation "concat" arg-restriction on-param "string1" has-value "http://net.testing.r2o/job-" arg-restriction on-param "string2" has-transform operation "concat"


Page 31 of 151

arg-restriction on-param "string1" has-columnjobs.id arg-restriction on-param "string2" has-column jobtypes.code

Other primitives defined in the first version of the language are: plus, minus, multiply, divide, apply-regexp, in-keyword, hi-tan, lo-than, equals, hieq-than, loeq-than, etc.

Attribute mapping definitions Mapping definitions for attributes are defined as sets of if-then rules that allow the conditional generation of attribute values as well as multivaluation. The structure of an attribute mapping definition is described by the following example. The value of the ontology attribute type is calculated based on the application of the set of rules (selector): If the condition part (applies-if) is verified, then the action part (aftertransform) is executed to generate a value. attributemap-def "http://net.testing.r2o/jobs#type" selector applies-if condition[ ... condition desc 1 ... ] aftertransform operation [ ... transformation desc 1 ... ] selector applies-if aftertransform ...

4.2.2

The ODEMapster Processor

The ODEMapster processor generates Semantic Web instances from relational instances based on the mapping description expressed in an R2 O document. ODEMapster offers two operation modes: query driven upgrade (on demand query transaltion) and massive upgrade batch process that generates all possible Semantic Web individuals from the data repository. Figure 4.2 shows the ODEMapster operation modes.

Figure 4.2: ODEMapster operation modes.


Page 32 of 151


The operations of ODEMapster are not limited by the expressivity of the DBMS. The set of primitives can be extended with delegable or non delegable primitive conditions and operations. The processor will delegate the execution of certain actions to the DBMS and execute the rest itself (post processing). The main steps of its executions are: Query and R2 O parsing, SQL generation, DBMS execution result grouping and finally post-processing.

4.3

Case Study: Upgrading dabase content on FAO data

In this section we present the preliminary experiments that we are performing with R2 O and ODEMapster on FAO data within WP7. So far we have worked with the reference data for biological species, FAO divisions of water areas and land; we have made the upgrade for the following ontologies:

• Water bodies Ontology This ontology was populated from the FIGIS data, this data is stored in 3 tables (fic_catch_area, fic_catch_area_agg_grp, and md_refobject). The DB table size is 137 Kb.

• Biological species Ontology This ontology was populated from the FIGIS data, this data is stored in 3 tables (fic_item, fic_item_grp and md_refobject). The DB table size is 3 Mb.

• Land areas Ontology This ontology was populated from the FIGIS data, this data is stored in 3 tables (area, area_yr_agg_grp, and md_refobject). The DB table size is 460 Kb. These ontologies are expressed in OWL format. We used the data sets from FIGIS (Fisheries Global Information System). The data sets are stored in a MySQL database. For each ontology we created a R2 O document for describing the mappings with their respective tables. Then we executed ODEMapster processor in batch mode exporting data from the database and storing the RDF results in an output file. The generated instances with the processor has been imported into the corresponding ontology using Protégé. Right now we are analyzing the results of these experiments, in the near future we will provide a complete description of them (this description will be part of the next deliverable in this task). Also we are planning to execute ODEMapster processor in query driven mode.

4.4

Future Work

Regarding the future trends of our work, the extension of the framework to include a semi automatic mapping discovery tool is under development. In addition, intensive testing with other DBs as well as the development of tools, middleware, APIs, etc, to generate and exploit R2 O mapping descriptions are carried out. A graphical user interface for both visualizing and writing R2 O mapping documents is under development. Also we are planning to follow the same philosophy with legacy XML data. We will create X2 O , a language that describes the mappings between XML schemas and ontologies; and XMapster, the processor in charge of carrying out the exploitation of the mappings defined using X2 O.


Page 33 of 151

Chapter 5

Reengineering Thesauri and Lexica Reengineering data sources for ontology design can take many forms, and there are two main dimensions, along which these forms can vary. The first dimension is the set of reuse and maintainance methods applied to data, while the second is the set of schema translation methods that are used to interpret the semantics of data. In the previous two chapters we have reported on methods that allow re-engineering unstructured data (i.e., texts), and weakly structured (i.e., databases) data into ontologies. Both chapters have concentrated mainly on (different) methods for reusing and maintaining those data. For example, ontology learning tends to privilege extraction methods, which have the main goal of drafting an ontology and its instance or fact population; no intention is expressed to continue maintaining any correspondence between texts and ontologies through time. On the contrary, the R2O approach shows an opposite method, which privileges access methods, which leave evolution and maintainance of data in their original database form (including also terminological databases). In this chapter, we concentrate on the second dimension, i.e. on the methods to interpret the semantics of data, and we address here the set of terminological databases, which, besides texts and generalized databases, contain a lot of useful, domain-oriented conceptual structures. Two kinds of terminological databases are distinguished: so-called Knowledge Organization Systems (KOS) and Lexica. Thesauri like Agrovoc ([SLL+ ]) and classification schemes like UMLS ([HL93]) are KOSes, while dictionaries, wordnets like Princeton WordNet ([Fel98]), and framenets like Berkeley FrameNet ([BFL98]) are lexica. KOSes, also called concept schemes in SKOS ([MB05]) are tailored to the needs of a certain community, and usually have a forest-like graph structure, without an explicit formal semantics. Typical relations that tag associations between nodes in those graphs include broader than/narrower than, related to, and used for. The nodes themselves are sometimes called descriptors, sometimes concepts, etc. Lexica also have a forest-like graph structure, without an explicit formal semantics, but their nodes denote lexical items used in a natural language, equivalence classes (“synsets”) of those items, or word senses. Their typical relations tag lexical associations: hyperonymy/hyponymy, synonymy, entailment, meronymy. When reengineering KOSes and lexica into ontologies for the semantic web, we need to understand what are the possible mappings between the (implicit semantics of) their graph structure, and the formal semantics underlying the graph structure of RDF and OWL models. Deciding on those mappings is relevant not only to reengineering (T2.2 in NeOn WP2), but also in order to provide multilingual natural language expressions to ontology elements (T2.4). For these two reasons, declaring the semantics assumed by KOSes and lexica is very relevant to NeOn WP2, and is being explicitly addressed in the C-ODO library of metamodels that allow a common conceptual framework for collaborative ontology design (T2.1, T2.3). The issue is clear from the part of C-ODO1 that represents ontology-related data (Fig. 5.1), where the notion of ontology from OMV (Ontology Metadata 1

http://www.loa-cnr.it/ontologies/OD/odData.owl


Page 34 of 151


Vocabulary2 ) admits KOSes as ontologies, while the formal notions of ontology, as those defined by the odData C-ODO module3 , or by the OWL metamodel4 , exclude KOSes as ontologies.

Figure 5.1: Notions of ontology in the C-ODO library.

The difficulty in establishing if a KOS or a lexicon is an ontology or not can be overcome within a reengineering perspective: what counts is not if they are ontologies or not, but rather how they can be reengineered and exploited by semantic technologies. In the first part of this chapter (Section 5.1) we report on the benefits of aligning thesauri to lexica. In particular, we show how the alignment of the AGROVOC thesaurus to the WordNet lexicon can enhance AGROVOC’s thesaural structure in terms of scope, coverage and detail. In the rest of the chapter (Sections 5.3, 5.4, 5.5, 5.6), we illustrate some sample practices to reengineer KOSes and lexica in OWL. We provide two alternatives for each type of resource: firstly, we show how we can reengineer a KOS or a lexicon as an OWL ABox (i.e. as a knowledge base), where the datamodel of the resource is reengineered as an OWL TBox. Secondly, we show how we can reengineer a KOS or a lexicon as an OWL TBox (i.e. as a full-fledged ontology), where the datamodel of the resource is reengineered as a fragment of the OWL metamodel. The ABox solutions leave the informal semantics of the reengineered resources mostly “untouched”, since it is assumed as a de facto situation; on the contrary, the TBox solutions try to enforce a formal semantics to them, even at the cost of changing their structure. This change can be cost-effective or not. Preliminary experiences as in FOS ([GFK+ 04]) indicate that when a large organization has a lot of conceptual structures already represented and maintained as KOSes, it is technically and socially difficult to impose a TBox reengineering. The work described in Sections 5.3, 5.4, 5.5, 5.6 has been partly done in past projects, but the way it is put together and harmonized here, as well as integrated within the C-ODO design approach, is novel (for example, in section 5.3.2), and specific to NeOn. In the next version of D2.2.1, a richer system of guidelines and computational support for KOS and lexicon reengineering will be presented.

5.1 Mapping Agrovoc onto WordNet The work described in this section concerns the alignment of the terminology/vocabulary of two resources: Agrovoc and WordNet. There are several reasons for choosing these two resources. First, both resources are KOS in the form of thesauri (in fact, WordNet is considered both a lexicon and a 2

http://omv.ontoware.org/2005/05/ontology http://www.loa-cnr.it/ontologies/OD/odData.owl 4 http://owlodm.ontoware.org/OWL1.0 3


Page 35 of 151

thesaurus by many), which means that they share the ontological commitments and axiomatizations that are typical for lightweight ontologies. This makes any matching more uniform. Second, the resources represent two different types of thesaurus. Agrovoc represents a terminological thesaurus restricted to a few domains. WordNet is considered to be a general language thesaurus. Third, the Agrovoc resource is a crucial resource for the fisheries case study in NeOn (WP7). The two thesauri are similar in many ways. WordNet was built on the basis of a number of existing resources, one of which is a biological taxonomy. Therefore, biological terminology constitutes part of the WordNet vocabulary, and, although WordNet is a general language resources, this causes an overlap in scientific terminology between the two resources. By means of simple techniques such as lexical matching and more complex mapping techniques in order to resolve ambiguity, one can obtain new terms and relations between new and existing concepts, which then constitute a richer structure. In this respect, we can speak of lightweight ontology enrichment, where the integration of new terminology and semantic relations leads to an increased conceptual coverage of the enriched resource. Previous work on the integration of domain-specific conceptual knowledge into the WordNet and EuroWordNet framework mostly involved manual, or semi-automatic techniques by means of the creation of equivalence relations, or the insertion of plugin relations that integrate the domain specific extension into the taxonomic structure of the resource that is to be enriched [PABC05]; [MS01b].

5.1.1

Advantages of Mapping Agrovoc onto WordNet

In general, it is foreseen that this alignment of these two lightweight ontologies will enhance Agrovoc’s thesaural structure in terms of scope, coverage and detail. More specifically, there are several advantages to the alignment of the conceptual vocabularies of Agrovoc and WordNet. In the first place, WordNet contains synonyms for the mainly scientific terms it shares with Agrovoc. These mostly non-scientific names for species are useful for detecting information from a variety of more informal sources that do not solely use the scienticic names, such as news items. The following example gives a rather extreme illustration of the potential synonym expansion for Agrovoc: Agrovoc termcode 46044: Lophius americanus. WordNet synset members: allmouth; angler; angler fish; anglerfish; goosefish; Lophius Americanus; lotte; monkfish. Secondly, Agrovoc does not contain any definitions or glosses for its terminological vocabulary. Links with WordNet will provide many glosses, which can be automatically added, and adapted by experts and developers with minimal overhead. The example above has the following description in WordNet: "fishes having large mouths with a wormlike filament attached for luring prey". A third advantage is that the available set of semantic relations into which Agrovoc terms are engaged can be extended with WordNet-derived semantic relations, e.g. meronymy relations (see section 5.1.4). Fourth, since this extension is an enlargement of semantic context, in the sense that it consists of a larger network of concepts and relations, resulting from an alignment between the two resources, it will also in the future facilitate the integration other thesauri. The increased vocabulary and semantic links will create a larger set for label matching and comparison in terms of semantic relations. In conclusion, by aligning the conceptual vocabulary of these two resources, we will have a larger base for the task of ontology population (see section 9.4). The result of the alignment is the superset of terms and relations in the two resources. This set has a greater vocabulary coverage than the original resources, which can be used for bootstrapping the automatic acquisition of new instances from text. Also, it has a larger ontological coverage in the form of all relations from both resources. For valid alignments, WordNet adds a substantial amount of semantic information to Agrovoc, especially with respect to synonymy and meronymy.


Page 36 of 151

5.1.2


Statistics

In order to put the alignment into context, we first describe the structural taxonomic properties of both resources in more detail. Agrovoc contains 39258 concepts. Each concept is lexicalized by one term only. In our analysis we have assumed that all these terms are nouns. All these nouns are monosemous, except one: “urban planning”, which has two senses. All in all, there are 27646 instances of a hypernymic relation. WordNet2.1 contains 117057 nouns, organized in 81426 concepts. 86.5% of its nouns are monosemous (101282), whereas 13.5% are polysemous (15775). The nature of the taxonomies of both resources differs to a great extent. WordNet has deep hierarchies (its average chain length is 7.4), with a small number of separate sub-hierarchies: it has 9 so-called unique beginners, each starting a sub-branch of the WordNet taxonomy. Agrovoc, on the other hand, has shallow hierarchies (the average chain length is 4.1), and there are 596 unique beginners. WordNet has an average branching factor of 5.2. For Agrovoc, it is 6.12. This means that the Agrovoc taxonomy fans out at a greater rate than WordNet’s taxonomy.

5.1.3

Term Matching

The conceptual matching of Agrovoc and WordNet2.1 has been performed on the basis of relational version of both resources. This has several reasons:

• the relational format allows straightforward data selection and manipulation in SQL, and scripting languages such as Perl and Awk.

• The owl versions of both resources are still in experimental format [FAO06]. • the owl versions, as they are at this moment, mirror the relational structure, and can therefore be regarded as derivationsn of the relational structure.

5.1.4

Lexical Match

The most straightforward alignment is a lexical match between two word forms [ES04]. An additional preprocessing step is the normalization of the labels for plurality. Matching between Agrovoc terms and WordNet nouns was performed at the lemma level, i.e. the plural forms from Agrovoc were normalised into singular headwords. The plural forms covered are: -s; -es for normal forms, and -ae; -i; -a for latinate forms. Examples:

• trees, bonuses • vertebrae; bronchi This process yielded 10602 matches, of which 8376 are monosemous in WordNet, and 2226 polysemous. The results of this match with respect to the enriching of Agrovoc are the following:

• a possible 10602 glosses; • 12682 additional synonyms for existing terms in Agrovoc; • 1087 additional part holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 1087 additional part meronyms between existing Agrovoc terms);

• 247 additional substance holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 247 additional substance meronyms between existing Agrovoc terms);


Page 37 of 151

• 1283 additional member holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 1283 additional member meronyms between existing Agrovoc terms). These are the numbers of matches for the whole of Agrovoc. Since this thesaurus covers a number of domains, amongst which “seed production”, “consumer economics” and “Protection of plants and stored products” (in total a set of 101 hierarchically organized domains), we selected the subset of domains that pertain to the fisheries domain. Inspection of the domains resulted in the selection of the following Agrovoc categories: 1. 1-94, amongst which:

• Miscellaneous demersal fishes • Herrings, sardines, anchovies • Tunas, bonitos, billfishes • Miscellaneous pelagic fishes • Sharks, rays, chimaeras • Marine fishes not identified • Crustaceans • Freshwater crustaceans • Crabs, sea-spiders • Lobsters, spiny-rock lobsters • King crabs, squat-lobsters • Shrimps, prawns • Krill, planktonic crustaceans 2. M11 Fisheries production 3. M12 Aquaculture production and management 4. M40 Aquatic ecology For these domains, there were 387 monosemous alignments with WN20, and 73 polysemous matches. The results of this match with respect to the enriching of the Agrovoc fisheries domain are the following:

• a possible 460 glosses; • 896 additional synonyms of existing Agrovoc terms; • 63 additional part holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 63 additional part meronyms between existing Agrovoc terms);

• 21 additional substance holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 21 additional substance meronyms between existing Agrovoc terms);

• 141 additional member holonym relations between existing Agrovoc terms (and because this relation is symmetrical, 141 additional member meronyms between existing Agrovoc terms).


Page 38 of 151

5.1.5


Disambiguation

In the cases where the Agrovoc term matches a polysemous WordNet lemma, we need to choose a disambiguation strategy, in order to proceed beyond the base line of label matching [ES04]. Within the fisheries domains, the Agrovoc terms listed in the table below were found to be polysemous in WordNet. The first column lists the concept identifier in Agrovoc. The second the English lexicalization, which is shared by both Agrovoc and WordNet. Column 3 indiactes the number of senses in WordNet, whereas column 4 lists the unique WordNet synset identifier, and column 5 contains teh WordNet gloss describing the synset.

5.1.6

Evaluation

FAO was requested to check the validity of the definitions of a sample of monosemous alignments, and all polysemous alignments. The purpose of this expert check is to give an indication of how reliably we can assume that a monosemous alignment involves conceptual equivalence, and the quality of the glosses associated with the WordNet synsets for inclusion into the Agrovoc datastructure. The overall result of the evaluation was that 19% of the WordNet glosses of monosemous alignments received an expert evaluation of 6 out of 10 or more. 100% of the 11 polysemous terms received score of more than 9 out of 10 for at least one of their senses.

5.1.7

Future Work

In the next phase of the work, we plan to evaluate the WordNet fish taxonomies with respect to their scientific value within the scope of possible re-use by FAO. Further analysis of the structure and content of the glosses will give insight into the dimensions along which the evaluations have taken place, and the reusability of the knowledge expressed in them for scientific purposes. Relevant additional conceptual material from WordNet will be offered to FAO for integration into the knowledge base, both in relational and in owl format. Also, the evaluated matches of the polysemous terms can serve as a gold standard for experimentation with word sense disambiguation strategies.

5.2

Reengineering Lexica: the case of Princeton WordNet

The number of applications where Princeton WordNet ([Fel98]) is being used as an ontology rather than as a mere lexical resource seems to be ever growing. Indeed, WordNet contains a good coverage of both the lexical and conceptual palettes of the English language. Moreover, many other languages have developed or are developing local wordnets, and multilingual wordnets, like EuroWordNet ([ewn98]) are available. A clean example of how wordnets can be used to suggest conceptualizations and lexicalizations in ontology design is provided by the OntoLing tool ([PS]), which implements a navigation system to wordnet databases (e.g. as a plugin to Protégé), and allows to choose individual lexical items or entire branches, and to promote them as ontology classes or subsumption hierarchies; alternatively, lexical items can be used to label existing classes. OntoLing implements an “on-the-fly reengineering method”, which is a good example of exploiting lexical resources on the reuse dimension alone, leaving the semantic interpretation to user choice, when a specific design need arises. On-the-fly reengineering is a clean method, well-suited to small projects and lightweight tools. On the other hand, the availability of completely reengineered resources, with a pre-built semantic interpretation, is an advantage within large projects or more complex tools. On the interpretation dimension, the two ABox and TBox approaches are exemplified in the next sections.


Term ID 2943

Termspell

Page 39 of 151

WN offset

WN gloss

Fish

Number of WN Senses 4

102489439

2943

Fish

4

107669669

2943

Fish

4

108566998

2943

Fish

4

109610846

6493

Regulations

6

100793467

6493 6493

Regulations Regulations

6 6

100795335 105773043

6493 6493

Regulations Regulations

6 6

106575164 113374899

6493 8488 8488 8488

Regulations Yields Yields Yields

6 4 4 4

114251720 100901487 104557797 113089677

8488

Yields

4

113577639

6400

Quality

5

104668449

6400 6400

Quality Quality

5 5

104672645 104929173

6400

Quality

5

105776237

6400 6839

Quality Scallops

5 4

113761936 101945963

6839

Scallops

4

107549875

6839

Scallops

4

107691816

6839

Scallops

4

113690322

any of various mostly cold-blooded aquatic vertebrates usually having scales and breathing through gills; "the shark is a large fish; "in the living room there was a tank of colorful fish" the flesh of fish used as food; "in Japan most fish is eaten raw"; the twelfth sign of the zodiac; the sun is in this sign from about February 19 to March 20 (astrology) a person who is born while the sun is in Pisces the act of controlling or directing according to rule; "fiscal regulations are in the hands of politicians" the act of bringing to uniformity; making regular a principle or condition that customarily governs behavior; "short haircuts were the regulation" an authoritative rule (embryology) the ability of an early embryo to continue normal development after its structure has been somehow damaged or altered the state of being controlled or governed production of a certain amount an amount of a product the income or profit arising from such transactions as the sale of land or other property the quantity of something (as a commodity) that is created (usually within a given period of time); "production was up in the second quarter" an essential and distinguishing attribute of something or someone. a degree or grade of excellence or worth (music) the distinctive property of a complex sound (a voice or noise or musical sound). a characteristic property that defines the apparent individual nature of something; "each town has a quality all its own". high social status; "a man of quality" edible marine bivalve having a fluted fan-shaped shell that swim by expelling water from the shell in a series of snapping motions thin slice of meat (especially veal) usually fried or broiled edible muscle of mollusks having fan-shaped shells; served broiled or poached or in salads or cream sauces one of a series of rounded projections (or the notches between them) formed by curves along an edge (as the edge of a leaf or piece of cloth or the margin of a shell or a shrivelled red blood cell observed in a hypertonic solution etc.)


Page 40 of 151

Term ID 7033

Termspell


WN offset

WN gloss

Shellfish

Number of WN Senses 2

101920999

7033

Shellfish

2

107677501

7321

Sponges

4

101887191

7321

Sponges

4

110099633

7321

Sponges

4

110344419

7321

Sponges

4

114403987

8408 8408

Winkles Winkles

2 2

101928577 107676959

868 868

Behaviour Behaviour

4 4

101205715 101205985

868

Behaviour

4

104842573

868

Behaviour

4

113822208

4317

Life cycle

2

111308334

4317

Life cycle

2

113333844

6507

Reproduction

5

100837424

6507 6507

Reproduction Reproduction

5 5

101005883 104030702

6507

Reproduction

5

105691746

6507

Reproduction

5

113376112

invertebrate having a soft unsegmented body usually enclosed in a shell meat of edible aquatic invertebrate with a shell (especially a mollusk or crustacean) primitive multicellular marine animal whose porous body is supported by a fibrous skeletal framework; usually occurs in sessile colonies a follower who hangs around a host (without benefit to the host) in hope of gain or advantage someone able to acquire new knowledge and skills rapidly and easily. a porous mass of interlacing fibers that forms the internal skeleton of various marine animals and usable to absorb water or any porous rubber or cellulose product similarly used edible marine gastropod small edible marine snail; steamed in wine or baked manner of acting or controlling yourself (psychology) the aggregate of the responses or reactions or movements made by an organism in any situation (behavioral attributes) the way a person behaves toward other people the action or reaction of something (as a machine or substance) under specified circumstances a series of stages through which an organism passes between recurrences of a primary stage the course of developmental changes in an organism from fertilized zygote to maturity when another zygote can be produced the sexual activity of conceiving and bearing offspring the act of making copies copy that is not the original; something that has been copied recall that is hypothesized to work by storing the original stimulus input and reproducing it during recall the process of generating offspring


5.3 5.3.1

Page 41 of 151

Lexicon to Ontology ABox W3C WNET method

Due to the relevance of WordNet for semantic applications, it is not surprising that W3C decided to start a task force (WNET) on “Porting WordNet to the Semantic Web”, within the Semantic Web Best Practices and Deployment Working Group ([SWB]). The way Princeton WordNet has been reengineered by WNET task force can be considered a good practice for wordnets in general, and we briefly summarize it here. We will also describe (5.3.2) how the resulting WordNet ABox can be linked to C-ODO ABox, in order to clarify the semantics of its reuse in the conceptualization and lexicalization tasks that are carried out in ontology design. WNET has produced a standard conversion of WordNet conceptual model and database to the RDF/OWL representation language in use in the Semantic Web community. It can be retrieved from: http: //wordnet.princeton.edu/wn. Such a standard representation is useful to provide application developers with a high-quality resource, as well as to promote interoperability. Important requirements in this conversion process are that it should be complete and should stay close to WordNetÕs conceptual model. The steps taken to produce the conversion included the following design decisions:

• the composition of class hierarchy and properties • the addition of a suitable OWL semantics • the chosen format of the URIs • a strategy to incorporate OWL and RDFS semantics in one schema such that both RDF(S) and OWL infrastructures can interpret the information correctly

• the description of the two versions that are provided (Basic and Full) to accommodate different usages of WordNet We mention here only the most relevant issues arisen for translating WordNet’s conceptual model to OWL semantics; the details of the work can be found in [vAGS06]. The three core notions in WordNet conceptual model are the synset, the word sense and the word. Words are the basic lexical units, e.g. “car”, while a sense is a specific sense in which a specific word is used (e.g. car as a motorcar or car as a railcar). Synsets group word senses with a synonymous meaning, such as car, auto, automobile, machine, motorcar or car, railcar, railway car, railroad car. There are four disjoint types of synset, containing exclusively nouns, verbs, adjectives or adverbs. There is one specific type of adjective, namely an adjective satellite. Furthermore, WordNet defines seventeen relations, of which ten between synsets (hyponymy, entailment, similarity, member meronymy, substance meronymy, part meronymy, classification, cause, verb grouping, attribute) and five between word senses (derivational relatedness, antonymy, see also, participle, pertains to). The remaining relations are gloss (between a synset and a sentence), and frame (between a synset and a verb construction pattern). The documentation defines characteristics for each relationship, such as (anti-)symmetry, inverseness and value restrictions on the lexical groups (e.g. nouns, verbs) that may appear in relations. Most of these informally stated requirements can be formalized in OWL and are present in the conversion. Investigation of the source files and its documentation reveals several conflicts. For each conflict we have proposed a resolution. We list a few typical examples:

• the order of synset arguments of the member meronym relation is the opposite from what the documentation asserts

• the documentation states that the hypernym relation has a reflexive relation (hyponym). The correct term is inverseness


Page 42 of 151


• the documentation states that derivational relatedness is reflexive, but here symmetry is meant • from the documentation it is not always clear if the symmetric relation is also present in the source file (e.g. if der(A,B) is in the file, is der(B,A) then also present?). It is not clear if the relation also holds when only one of the symmetrical facts is present in the source. The OWL class subsumption hierarchy obtained after the conversion is the following (see the OWL property matrix obtained after the conversion in Table 5.1): Synset AdjectiveSynset AdjectiveSatelliteSynset AdverbSynset NounSynset VerbSynset WordSense AdjectiveWordSense AdjectiveSatelliteWordSense AdverbWordSense NounWordSense VerbWordSense Word Collocation

Concerning the formatting of URIs and online querying, non-trivial choices have been required, because of the size of the ABox produced. We have chosen to introduce identifiers for the instances of classes Synset, WordSense and Word. We use a base URI + a locally unique ID. Three kinds of entities need a URI: instances of the classes Synset, WordSense and Word. Instead of generating any unique ID we have tried to use IDs derived from information in the source and also tried to make them human-readable. Since the IDs have distinct syntactic patterns, it is easy to identify the type of the resource (Synset, WordSense or Word), by examining the URI. The patterns are described below. Local IDs of Synset instances are composed of the synset ID, the lexical form of the first word sense in the synset and the lexical group symbol. Thus human readers can derive the lexical group of the word senses in the synset and get an idea about the kinds of words in the synset. For example:

http://wordnet.princeton.edu/wn/107909067-bank-n/ For WordSenses, the word + its lexical group + the sense number is used. Example:

http://wordnet.princeton.edu/wn/bank-noun-1/ For Words we use the lexical form, which is unique within English, plus the prefix “word-”. For example:

http://wordnet.princeton.edu/wn/word-bank/ There are two options in formatting the relationship between the namespace and the local part, usually termed “slash” URIs and “hash” URIs after the symbol used to connect the two parts (either or #). The disadvantage of hash URIs is that when a HTTP GET is done (e.g. for the first example above) the browser will return the whole document located at http://wordnet.princeton.edu/wn. The reason for this is that servers do not receive the fragment identifier. Since WordNet is very large, this is not a desirable option. The alternative is to use slash URIs. This choice implies that a decision needs to be made on which statements a server hosting WordNet should return when an HTTP GET is done for resources with a URI such as http://wordnet.princeton.edu/wn/107909067-bank-n/. Possible choices are:

• a graph that contains a pre-defined set of properties if the resource has values for them (e.g. rdf:type, rdfs:subClassOf);


Page 43 of 151

Property

Domain

Range

OriginalDBTable

synsetContainsWordSense word lexicalForm synsetId tagCount frame gloss hyponymOf entails similarTo memberMeronymOf substanceMeronymOf partMeronymOf classifiedByTopic classifiedByUsage classifiedByRegion causes sameVerbGroupAs attribute adjectivePertainsTo adverbPertainsTo derivationallyRelated antonymOf seeAlso participleOf classifiedBy meronymOf

Synset WordSense Word Synset Synset VerbWordSense Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset WordSense WordSense WordSense WordSense Synset Synset

WordSense Word xsd:string xsd:string xsd:integer xsd:string xsd:string Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset Synset WordSense WordSense WordSense WordSense Synset Synset

s s s s s fr g hyp ent sim mm ms mp cls cls cls cs vgp at per per der ant sa ppl cls mm,ms,mp

Table 5.1: OWL property matrix obtained after the conversion. • all statements connected to the resource with some offset, e.g. everything connected in at most two steps;

• the Concise Bounded Description of the URI ([Sti05]); • the Symmetric Concise Bounded Description of the URI ([Sti05]). The difference between the two last ones is that the Symmetric CBD not only includes statements for which the URI is the subject, but also those for which the URI is the object. We have chosen for the CBD of the URI because it “constitutes a reasonable default response to the request Õtell me about this resource”’ ([Sti05]).

5.3.2

Linking WordNet ABox to C-ODO ABox

The core OWL classes translated from WordNet conceptual model can be linked to C-ODO classes (Fig. 5.2), and the result enables semantically correct mappings between elements in lexically-derived ABoxes and other ABox elements, either from existing ontologies, or from ontologies derived from different resources, such as thesauri. Fig. 5.3 shows the closeness – within C-ODO –between the OWL translations of WordNet, SKOS (Simple Knowledge Organization Systems, [MB05]), FOAF (Friend of a Friend, [foa]), and DCMITYPE (Dublin Core vocabulary for resource types, [Dub]) classes. As an example, let’s consider the following ontology elements:


Page 44 of 151


Figure 5.2: Linking OWL WordNet classes to the C-ODO library.

Figure 5.3: Common linking to C-ODO for OWL versions of WordNet, SKOS, FOAF, and DCMITYPE.


the skos:Concept:

Fishes from Agrovoc thesaurus ([agr]), 02512053{fish1 }, as shown in section 5.1

Page 45 of 151

mapped to wn20:Synset:

an instance of wn20:Word: fish an instance of foaf:Image: TaiwanFish, depicting a fish encountered during snorkeling in Taiwan an instance of dcmitype:Sound: TaiwanFishOnCorals, which contains sounds produced by that fish knocking corals in a Taiwan coast. Now, given C-ODO linking, we can ensure that any mapping between concepts, and between concepts and information objects, will be semantically correct. For example, a consistency checker on the resulting OWL ABox will reveal a contradiction if we try to state an owl:sameAs between Fishes and fish, because edns:concept and edns:information-object are disjoint classes in C-ODO, so that skos:Concept and wn20:Word are disjoint by inheritance. On the other hand, an owl:sameAs between Fishes and 02512053{fish1 } would be semantically correct, because both are subclasses of edns:concept. More interestingly, we obtain new expressivity for free. For example, since we know that in C-ODO instances of edns:concept can be edns:expressedBy instances of edns:information-object, we can assert (and a runtime assistant can help doing that) that the skos:Concept: Fishes is expressed by a wn20:Word, a foaf:Image, and a dcmitype:Sound:

hFishes, fishi ∈ edns:expressedBy h02512053{fish1 }, fishi ∈ edns:expressedBy hFishes, TaiwanFishi ∈ edns:expressedBy hFishes, TaiwanFishOnCoralsi ∈ edns:expressedBy ...

5.4

(5.1)

Lexicon to Ontology TBox

The ABox translation of wordnets is very useful for ontology design, and even more if the ABox is linked to a common metamodel, as suggested in the previous section. However, if we want to use the conceptual structure of wordnets, so that they are serviceable as full-fledged ontologies (in the sense of a theory expressed in some logical language), we need to interpret lexical links according to a formal semantics that tells us something about the way we use a lexical item in some context for some purpose. In other words, we need a formal specification of the conceptualizations that are expressed by means of WordNet’s synsets5 . A formal specification requires a clear semantics for the primitives used to export WordNet information into an ontology, and a methodology that explains how WordNet information can be bootstrapped, mapped, refined, and modularized. In this section we describe the methodology and some results of the OntoWordNet project ([GNV03]), which aims to axiomatize conceptualizations derived from WordNet synsets. The methodology is hybrid because it employs both top-down techniques and tools from formal ontology, and bottom-up techniques from computational linguistics and machine learning. This section is taken from existing work done outside NeOn, but revisited here in order to produce a good example of a “full semantic interpretation” of a lexicon. In order to produce a formal specification of WordNet as an axiomatic theory (an ontology ), WordNet is reorganized and enriched in order to adhere to the following commitments: 5

Concept names in WordNet are called synsets, since the naming policy for a concept is a set of synonym words, e.g. for sense 1 of car: { car, auto, automobile, machine, motorcar}.


Page 46 of 151


• Logical commitment. WordNet synsets are transformed into logical types, e.g. owl:Classes, with a formal semantics for so-called lexical relations. The WordNet lexicon is also separated from the logical namespace. In other words, the taxonomy showing the alignment between WordNet ABox and C-ODO ABox does not hold for this second approach when synsets are interpreted, while it is also applicable here for words. I.e., synsets are not interpreted as individuals from the class wn20:Synset, but, depending on their meaning, they are interepreted as individuals from the (meta)classes owl:Class, owl:ObjectProperty, or owl:Individual.

• Ontological commitment. WordNet is transformed into a general-purpose ontology library, with explicit categorial criteria, based on formal ontological distinctions [GGMO01]. For example, we commit to a clear separation between (kinds of) concept-synsets, relation-synsets, meta-property-synsets, and enable the instantiation of individual-synsets.

• Contextual commitment. WordNet is modularized according to knowledge-oriented domains of interest ([MSPG02]). The modules should constitute a partial order.

• Semiotic commitment. WordNet lexicon is linked to text-oriented (or speech act-oriented) domains of interest, with lexical items ordered by preference, frequency, combinatorial relevance, etc. In [GGMO03] these commitments are discussed in detail and working hypotheses as well as first achievements are presented. Here we summarize the work originated by the ontological commitments, which deals with the creation of a TBox from WordNet structure.

5.4.1

Assumptions made during Reengineering

Substantial work has been done on the refinement of the hyponym/hyperonym relations, which have been investigated during 2001-2004 in the OntoWordNet project ([GNV03]). The hyperonymy relation is basically interpreted as formal subsumption, but hyperonymy for synsets referring to individuals (geographical names, characters, some techniques, etc.) has been interpreted as instantiation. This is called assumption A1 (“synset hyperonymy as class subsumption”). Examples are shown in 5.2. intAs

hypernym(09542339{devil2 }, 09541919{evil_spirit1 }) −→ 09542339{devil2 } v 09541919{evil_spirit1 } intAs

hypernym(095433531{devil1 }, 09504135{supernatural_being1 }) −→ 095433531{devil1 } ∈ 09504135{supernatural_being1 }

(5.2)

In the last version, Princeton WordNet has basically accepted A1, and has introduced instantiation relations for some synset classes, thus reducing the semantic ambiguity of hyperonymy. WordNet synonymy is a relation between word senses, therefore we assume that synsets are equivalence classes of words that share a same sense. This is called assumption A2 (“synset as meaning equivalence class”). However, we have no axiomatization of word senses in WordNet that allows us to create equivalence classes analytically, e.g. why devil2 is semantically equivalent to demon1 . The only reason for that equivalence is conveyed by textual definitions (called glosses), e.g. “chief spirit of evil and adversary of God” is the gloss for devil1 . A non-trivial assumption we make is that such glosses can be formalized in order to provide axioms that motivate semantic equivalences. This is called assumption A3 (“glosses as axiomatizations”). A related assumption that we make, required by the procedure we follow to axiomatize glosses (see below), is that word senses in glosses are consistent to WordNet synsets. E.g., we assume that “adversary” and “God” are used in a sense that is present in some synset. This is called assumption A4 (“glosses are synsetconsistent”).


Page 47 of 151

A4 lets us assume that the informal theory underlying synsets, hyperonymy relations, and glosses, can be formalized against a finite signature (the set of synsets and word senses), and a set of axioms derived from the associations (A-links) between any synset S and the synsets that can be assigned (by using appropriate learning techniques) to the words contained in the gloss of S. For example, an A-link between devil1 and God1 . This assumption is dependent on A3 and A4, and is called assumption A5 (“A-links are conceptual relations”).

5.4.2

Applying Assumptions to WordNet Reengineering

In this section, we summarize some techniques that have been used to produce a WordNet TBox. The task of axiomatizing WordNet, starting from assumptions A1-A5, requires that the informal definition in a synset gloss be transformed to a logical, axiomatic form. To this end, first, words in a gloss must be disambiguated, i.e. replaced by their appropriate synsets. This first step provides us with pairs of generic semantic associations (A-links) between a synset and the synsets present in its gloss. Secondly, A-links must be interpreted in terms of more precise, formally defined conceptual relations. The inventory of conceptual relations used here has been selected or specialized from DOLCE (http://dolce.semanticweb. org), since in WordNet only a limited set of relations are used, which are partly conceptual, partly lexical in nature. For example, part_of (meronymy in WordNet) and kind_of (hyponymy in WordNet) are typical conceptual relations, while antonymy (e.g. liberal and conservative) and pertonymy (e.g slow and slowly ) are lexical relations. Furthermore, WordNet relations are not axiomatized, nor are they used in a fully consistent way. The objective of the method summarized here is to:

• automatically extract a number of conceptual relations implicitly encoded in WordNet, i.e. the relations holding between a synset and the synsets in its gloss.

• (semi)-automatically interpret and axiomatize these relations. For example, sense 1 of the word driver has the following gloss “the operator of a motor vehicle”. The appropriate sense of operator is 2 , included in the synset operator, manipulator (“an agent that operates some apparatus or machine”), while motor vehicle sense is included in the synset: motor vehicle, automotive vehicle (“a self-propelled wheeled vehicle that does not run on rails”). After automatic sense disambiguation, we (hopefully) learn that there exists an A-link between driver1 and operator1 , and between driver1 and motor vehicle1 . Subsequently, given a set of axiomatized conceptual relations in DOLCE, we must select the relation that best fits the semantic restrictions on the relation universe (domain and range). For example, given an A-link between driver1 and motor vehicle1 , the best fitting relation is agentive-co-participation, whose definition is:

. agentiveCoParticipation = coParticipatesWith u Agent × FunctionalObject (5.3) The definition says that agentive co-participation is a relation of mutual participation (participation of two objects in the same event), with the domain restricted to “Agent”, and the range restricted to “FunctionalObject”. Domain and range in a conceptual relation definition are established in terms of the DOLCE ontology. Consequently, another necessary step of our method is to link at least some of the higher level nodes in WordNet to DOLCE, i.e. in the above example:

driver1 v Agent

(5.4)

motorVehicle1 v FunctionalObject

(5.5)

In the following section we summarize the procedures for gloss sense disambiguation and learning of association links, synset linking to DOLCE, and selection of conceptual relations.


Page 48 of 151

5.4.3


Learning association links

The first step is a bottom-up procedure that analyses the NL definitions (glosses) in WordNet and creates the A-links. For each gloss (i.e., linguistic concept definition), we perform the following automatic tasks: 1. POS-tagging of glosses and extraction of relevant words through appropriate tools 2. Disambiguation of glosses by the algorithm described in [GNV03]. The algorithm applies a set of heuristics grouped in two classes, Path and Context. Some of these heuristics have been suggested in [MM99] 3. Creation of explicit association links (A-links) from synsets found in glosses to synsets to which glosses belong. For example, the gloss of the synset for sense 1 of retrospective says: “an exhibition of a representative selection of an artist’s life work ”, while its hyperonym, art exhibition1 , is glossed as “an exhibition of art objects (paintings or statues)”. So initially we have (D is the given set of terms, P is the set of potentially related terms):

D = retrospective1 P ={work, object, exhibition, life, statue, artist, selection, representative, painting, art}

(5.6)

GlossSynsets(retrospective1 ) = { artist1 , exhibition2 , life12 }

(5.7)

Through several iterations, we finally obtain:

5.4.4

Learning conceptual relations

In the top-down phase, the A-links extracted in the bottom-up phase are refined. A-links provide just a clue of relatedness between a synset and another synset extracted from the gloss analysis, but this relatedness must be axiomatized, in order to understand if it is a hyperonymy relation, or some other conceptual relation (e.g. part, participation, location, etc.). First of all, we need a shared set of conceptual relations to be considered as candidates for A-links explicitation, otherwise the result is not easily reusable. Secondly, these relations must be formally defined. In fact, not only are A-links vague, but they also lack a formal semantics: for example, if we decide to represent associations as binary relations –like OWL “properties”– is an association symmetric? Does it hold for every instance, or only for some of the instances of the classes derived from the associated synsets? Is it just a constraint on the applicability of a relation to that pair of classes? Is the relation set a flat list, or there is a taxonomic ordering? To answer such questions, the shared set of relations should be defined in a logical language using a formal semantics. Since WordNet is a general-purpose resource, the formal shared set of relations should also be general enough, based on domain-independent principles, but still flexible, in order to be easily maintained and negotiated. For this reason, we have adopted DOLCE, a richly axiomatized, general purpose ontology, in its OWL encoding, i.e. the DOLCE-Lite+ library (http://www.loa-cnr-it/ontologies/DLP. owl). DOLCE assumes that its categories (top classes) constitute an extensionally closed set on any possible particular entity, i.e., entities that cannot be further instantiated within the assumptions of the theory (cf. Masolo et al. 2002, Gangemi et al. 2001). Of course, DOLCE does not assume an intensionally closed set,


Page 49 of 151

thus allowing for alternative ontologies to co-exist. Such assumptions will be referred to as A6_D (“extensional total coverage of DOLCE”). Consequently, we also assume that WordNet can be tentatively considered a (extensional) subset of DOLCE, after its formalization. A trivial formalization of WordNet might consist in declaring formal subsumptions for all unique beginners (top level synsets) under DOLCE categories, but this proved to be impossible, since the intension of unique beginners, once they are formalized as classes, is not consistent with the intension of DOLCE categories. Then we started (Gangemi et al. 2002) deepening our analysis of WordNet synsets, in order to find synsets that could be subsumed by a DOLCE category without being inconsistent. In our previous OntoWordNet work, WordNet 1.6 has been analyzed, and almost 1000 synsets have been linked to DOLCE-Lite+ library. A working hypothesis (A7_D) has been that the taxonomy branches of the relinked synsets are conceptually consistent with the DOLCE-Lite+ categories. Assumptions A4 and A5, together with A6_D, make it possible to exploit the axiomatized relations from DOLCE-Lite+. As a matter of fact, by looking at the A-links, a human expert can easily decide which relation from DOLCE-Lite+ is applicable in order to disambiguate the A-link, for example, from: (5.8)

A_link(car1 , engine1 ) we may be able to infer that cars have engines as components:

car1 v ∃hasComponent.

engine1

(5.9)

or that from

A_link(art_exhibition1 , painting1 )

(5.10)

we can infer that exhibitions as collections have paintings as members:

art_exhibition1 v ∃hasMember.

painting1

(5.11)

On the other hand, A-links are learnt from glosses, and the intention of the gloss writer is not necessarily that of imposing an existential dependence (implied by owl:someValuesFrom). For example, the A-link between art_exhibition#1 and painting#1 has been translated as in 5.11, but what if an art exhibition only includes sculptures? A more reasonable quantification would be:

art_exhibition1 v ≥ 0hasMember.painting1

(5.12)

An alternative design solution would be to generalize over paintings and sculptures:

art_exhibition1 v ∃hasMember.art1

(5.13)

This intellectual technique can be semi-automatized by using relation classification. Since relation equivalence is not supported in OWL1.0, we need rules to implement it. However, since OWL1.1 will support it, we assume it is available from generic OWL axioms. Our method consists in creating conceptual relations that have universes that partition the set of pairs of DOLCE categories. For example, the universe of hasMember (5.14) subsumes the universe created by the concepts in the A-link from 5.10, because art_exhibition#1 is subsumed by the domain of hasMember (Collection, 5.15), and painting#1 is subsumed by its range (Object, 5.16).

. hasMember = Collection × Object

(5.14)

art_exhibition1 v Collection

(5.15)

painting1 v Object

(5.16)


Page 50 of 151


Details on the creation and tuning of the set of the relation universes that, based on DOLCE-Lite+ and some cognitive principles, partition the category space, and how they have been applied to WordNet, are contained in [GNV03]. The results from the WordNet gloss axiomatization are encouraging (precision over 80%, recall about 40%, due to the desire of selecting only relevant A-links). An important aid to increase precision has been the adoption of “domains” (topics) that divide WordNet synsets into smaller sets, and provide a more stable association between words and synsets. Relation learning based on this method has been applied to several academic and practical projects [GFK+ 04, CGR+ 05], and can be applied to FAO testbed as well, as exemplified during early experiments with ASFA thesaurus “RT” relations (see section kostbox). Within T7.3 from WP7, some experiments will provide further hints on what and how to proceed with this kind of axiomatization for terminological database reengineering.

5.5

KOS to Ontology ABox

The methods to produce an ABox out of KOSes are varied, and depend on the specific conceptual model used for a KOS. Thesauri, classification schemes, folksonomies, etc. have each their own way to conceptualize a domain, without clarifying the related formal semantics. For space reasons, we do not provide here any detailed example of reengineering a KOS to an OWL ABox, but refer to several past and ongoing works. A more complete set of methods will be presented in the next deliverable of T2.2.

• SKOS (Simple Knowledge Organization Systems, [MB05]), is the most used metamodel to reengineering thesauri as OWL or RDF(S) models. It is currently aligned to C-ODO, and, as exemplified at the end of section 5.3, that linking provides a good semantic foundation for reusing KOSes in ontology design. SKOS has been thought primarily as an RDF(S) vocabulary, because it is intended to represent thesaurus nodes as (rdf) individuals (“concepts”), as well as (rdfs) classes. This ambiguity cannot be expressed in OWL-DL, thus preventing full compliance to the other metamodels used in C-ODO. The linking of SKOS to C-ODO has then requested firstly its reduction to OWL-DL

• The method presented in the next section 5.6 also envisages a KOS to ABox reengineering, by suggesting a preliminary translation of a KOS conceptual model to an OWL ontology, This approach is very similar to the one adopted for translating wordnets to OWL (see section 5.3)

• KOSes featuring a non-typical thesaurus structure are widely used in the biomedical domain, such as the Gene Ontology ([gen]), UMLS ([HL93]), etc. In the past, a lot of effort has been devoted to reengineering those KOSes to TBoxes ([GPS99, HS02]), but the size of the resources, and the vastity and dynamicity of the communities that maintain them make the task extremely difficult. On the other hand, in one case (SnoMed, [SCC97]) reengineering to TBox proved feasible and successful (the KOS is now maintained also as a TBox). In another case (NCI thesaurus, [nci]), the developers even wanted to switch its maintenance to a (huge) TBox

• Many different methods have been suggested for KOSes that express topics, or subject hierarchies, instead of concepts. DMOZ ([dmo]) is a good examples of a subject hierarchy, but many others exist in the business and industrial domains (one is also used in fishery, see next section). The ABox reengineering methods presented are varied, for examples: [Wel98] proposes a mereotopological representation of subject hierarchical relations, so that hierarchical relations are reengineered as either part or proximity relations; [SPR+ 06] suggests that subjects are just classes of documents, and the hierarchical relations can be interpreted as regular subsumptions; [Hep05] proposes to create a twofold reengineering, with a double semantics: the first ensures that each subject is translated to a class of documents, the second ensures that subjects are translated to regular OWL classes. Probably, the alignment of a metamodel for subject hierarchies to C-ODO would be a minimal solution for interoperability, because mappings between an ABox of subject hierarchies and other KOS or lexical ABoxes


Page 51 of 151

can be left to the user by choosing a particular semantics, and then entering the paradigm of networked ontologies.

5.6

KOS to Ontology TBox. An example in fishery

Large information systems often use metadata contained in Knowledge Organization Systems (KOSes), such as vocabularies, taxonomies and subject directories, in order to manage and organize information. KOSes support document tagging (thesaurus-based indexing) and information retrieval (thesaurus-based search), but their semantic informality and heterogeneity usually prevents a satisfactory integration of the supported documentary repositories and databases. Traditional integration techniques mainly consist of time-consuming, manual mappings that are made – each time a new source or a modification enter the lifecycle – by experts with idiosyncratic procedures. Informality and heterogeneity make them particularly hostile with reference to the semantic web. The different fishery information systems and portals that provide access to fishery information resources are one example of such scenario, and they have been targeted by the Fishery Ontology Service (FOS) Project (http://www.fao.org/aims/onto_domains.jsp). The following resources have been singled out from the fishery information systems considered in that project:

• OneFish [one] is a portal for fishery projects and a participatory resource gateway for the fisheries and aquatic research and development sector. It is organized as hierarchical topic trees (more than 1,800 topics, regularly increasing), topics have brief summaries, identity codes and attached knowledge objects (documents, web sites, various metadata)

• AGROVOC thesaurus [agr] has been developed by FAO and the Commission of the European Communities in the early 1980s and is used for document indexing and retrieval. AGROVOC contains approximately 2,000 fishery related descriptors out of about 16,000 descriptors

• ASFA thesaurus [asf] supports an abstracting and indexing service covering the world’s literature on the science, technology, management, and conservation of aquatic resources and environments, including their socio-economic and legal aspects. It consists of more than 6,000 descriptors

• FIGIS [fig] is a global network of integrated fisheries information. Presently its thematic sections (reference tables) are five: aquatic species, geographic objects, aquatic resources, marine fisheries, and fishing technologies. The tables consist of approximately 300 top-level concepts, with a max depth of 4, about 30,000 objects, with multilingual support. The sources to be integrated are rather variate under many perspectives (semantic, lexical and structural), then their interoperability requires reengineering based on a same framework of reference. An example of how that framework can be relevant for fishery information services is shown by the terminological knowledge related to aquaculture, provided by the legacy KOSes with different conceptual “textures”. For example, the AGROVOC thesaurus puts aquaculture in the context of different hierarchies, from the viewpoints of techniques and species. ASFA puts it differently, since its hierarchy focuses on the environment and disciplines related to aquaculture. FIGIS reference tables put aquaculture into the species context. Finally, oneFish directory returns a context related to economics and planning. The example is depicted in the following table, where NT means narrower than; indentation in FIGIS means subconcept of, rt means related term, Fr and Es point to the French and Spanish terms. AQUACULTURE (AGROVOC) NT1 fish culture NT2 fish feeding


Page 52 of 151


NT1 frog culture ...

rt agripisciculture rt aquaculture equipment ...

Fr aquaculture Es acuicultura AQUACULTURE (ASFA) NT Brackishwater aquaculture NT Freshwater aquaculture NT Marine aquaculture rt Aquaculture development rt Aquaculture economics rt Aquaculture engineering rt Aquaculture facilities Biological entity (FIGIS) Taxonomic entity Major group Order Family Genus Species Capture species (filter) Aquaculture species (filter) Production species (filter) Tuna atlas spec SUBJECT (OneFish) Aquaculture Aquaculture development Aquaculture economics @ Aquaculture planning

Notice that reengineering KOSes as TBoxes is not enough for interoperability, because different views will still be different after formalization. That’s why interoperability in FOS needed a common framework for KOS reengineering: a comprehensive set of reference ontologies that satisfy the following requirements:

• be (partly) domain-independent ontologies that are shared by the legacy KOSes • be flexible enough, so that different views are accommodated in a common context • be focused on the core reasoning schemata for the fishery domain, otherwise the common framework would be too abstract. In Figures 5.4 and 5.6 two UML activity diagrams are shown which summarize the main steps of the methods employed to create a Fishery Ontology Library in the FOS project. We refer to the global lifecycle as ONIONS@FOS, since it is an adaptation of the ONIONS methodology ([GPS99]). For the sake of readability, we have split the activity diagram into five pieces, as follows: 1. Terminological database (TDB) formatting and schema lifting


Page 53 of 151

2. TDB porting, formalization, and Core ontology building 3. Modularization, ontology library building, and alignment to reference ontologies 4. Annotation, refinement, and merging of the library 5. Measures for finalisation, maintenance, and exploitation but only two of them are discussed here, because directly relevant to reengineering.

5.6.1

Formatting and Lifting in FOS

In the first phase of the methodology (Fig. 5.4), the original terminological databases are imported into a common database format. The conceptual models of the databases are lifted (either manually, or by using automatic reverse engineering components [11]). At the same time, a common Ontology Data Model (ODM) is chosen. This has been partly derived from the semantics of ontology representation languages (e.g. an OWL ODM, cf. http://owlodm.ontoware.org/OWL1.0), enhanced with criteria for distinguishing the different data types at the ontological level (e.g. individual, class, meta-property, relation, property name, lexicon, etc.). Based on the ODM, lifted schemata are translated and then integrated. In the case of FOS, the original TDBs resulted to be heterogeneous, specially FIGIS with respect to ASFA and AGROVOC. In fact, the first were controlled by means of a set of XML DTDs, while the seconds were implemented in relational databases. Semantically, the KOSes in the TDB schemata are even more heterogeneous:

• ASFA is a typical thesaurus, made up of so-called descriptors, equivalent to instances of skos:Concept, and relations among descriptors (BT, NT, RT, UF), which create a forest structure (an indirect acyclic graph, [pla]). Descriptors are directly encoded in the database via a preferred term

• AGROVOC is also a thesaurus, but contains multilingual terms that are assumed as equivalent across different languages; descriptors are encoded via alphanumeric codes, like synsets in WordNet

• FIGIS is a less typical thesaurus, composed as a collection of TDBs organised into modules encompassing different domains: vessels, organisms, techniques, institutions, etc. Each “concept” or “object” in a module is represented by means of an identifier, whose intuitive semantics is similar to Agrovoc’s descriptors: equivalence classes of multilingual terms. Each module has an own schema including local relations defined on classes of concepts or objects. E.g., a relation between institutions and countries, a relation between vessels and techniques, between organism species and genera, etc. These relations are more informative than generic RT thesaurus relations (see phase 2 about additional transformations to TDB).

• FIGIS DTDs used to encode heterogeneous metada for the management of the FIGIS database. These XML elements can refer to domain-specific information (e.g. Location), datatypes (e.g. Date), data about data (e.g. Available), foreign keys (e.g. AqSpecies_Text).

• OneFish is a tree structure of subjects (keywords used to classify documents, see previous subsectio), with multihierarchical links, similar to Web directories like DMOZ [dmo]. The top subjects in OneFish are depicted in Fig. 5.5. The integrated schema results to include all the elements from the TDBs. On the other hand, in order to make a KOS to TBox conversion, we need to interpret the original elements as ontology elements. Therefore, in FOS we created a mapping from each (domain-related) elements to an ODM element type, e.g. we have translated agrovoc:Descriptor and figis:Concept to owl:Class, asfa:RT to owl:ObjectProperty, onefish:Subject to owl:Thing, etc.


Page 54 of 151


Figure 5.4: A UML activity diagram for formatting and lifting activities in ONIONS@FOS.

Figure 5.5: Topic spaces (worldviews) in oneFish.


5.6.2

Page 55 of 151

Formalizing

After a common format and a mapping to an ODM have been obtained, the second phase (Fig. 5.6) starts by choosing an Ontology Representation Language (ORL), e.g. OWL. As shown in the diagram, two alternatives are available at this point. The first one goes back to the KOS-to-ABox reengineering approach, which preserves the KOS elements from the original schema, similarly to what we have seen for Lexicon-to-ABox translation. In this case, no mapping is performed from original schemata to ODM. The (dis)advantage is that no interpretation is performed on the KOS conceptual structure. In other words, there are pros and cons, as mentioned in past sections. An evaluation of the alternatives, in terms of project planning, sustainable workflows, cost-effectiveness, and other dimensions, will be made in the next T2.2 deliverable.

Figure 5.6: The activity diagram for metadata formalization and Core ontology building.

The second alternative is to map the KOS element types to ODM elements, and then to apply refinements in order to comply to the semantics of ODM. For example, AGROVOC makes no distinction between descriptors that can be interpreted as instances of owl:Class (e.g. agrovoc:River), and descriptors that can be interpreted as instances of owl:Individual (e.g. agrovoc:Amazon). Individuals are typically found in subdomains like geography and institutions. Another example concerns thesauri relations: some relations needed no refinement with reference to the ODM, e.g. RT (Related Term) in ASFA or Agrovoc is imported as an instance of owl:ObjectProperty. Other relations needed refinement, e.g. BT (Broader Term) can be usually interpreted as owl:subClassOf property, but sometimes it should be interpreted as an instance of owl:ObjectProperty. E.g. from:

hagrovoc:Blood_Cells, agrovoc:Bloodi ∈ agrovoc:BT

(5.17)

agrovoc:Blood_Cells v agrovoc:Blood

(5.18)

we can derive:

But this is false according to many design rationales (cf. NeOn deliverable D2.1.1, and the odRationales.owl module in C-ODO), e.g. because an instance of agrovoc:Blood_Cells is not an instance


Page 56 of 151


of agrovoc:Blood in all possible interpretations. For example, some biomedical ontologies would usually map the BT relation to some componency or constitution relation. Such inappropriate mappings can be easily discovered by linking reengineered KOSes to core or reference ontologies that contain restrictions and disjointness axioms, and running a consistency checker (see below, and [GFK+ 04] for details). In the KOS-to-TBox approach, RT relationships should be interpreted as associations between classes, and trasformations to ODM must clarify what is the intended semantics of those associations. In FOS, translation and refinement have been complemented by transforming the instances of ASFA/Agrovoc RT relation, and FIGIS relations into instances of owl:Restriction. The working hypotheses in making these transformations are that:

• RT is interpreted as a maximally generic owl:ObjectProperty, which means that any possible conceptual relation is a sub-property of it

• an application (triple) of RT to instances of e.g. owl:Restriction on instances of owl:Class

asfa:Descriptor is interpreted as an

• the resulting owl:Restriction instances are inherited to all subclasses of the instances of owl:Class that they pertain to • the quantification applicable to instances of owl:Restriction derived from RT application is owl:someValuesFrom • the intended meaning of RT applications can be extracted by following a technique similar to that used for learning the intended meaning of A-links (see section 5.4).

5.6.3

Enriching the Conceptualization

The soundness of the last two hypotheses is mostly empirical, but also based on the common sense of thesaurus builders. Here is an analogy with A-links extracted in the example of Lexicon-to-TBox reengineering approach. As observed in section 5.4, A-links are learnt from glosses, and the intention of the gloss writer is not necessarily that of imposing an existential dependence, as e.g. implied by owl:someValuesFrom. On the contrary, in the practice of thesaurus design, it seems that designers implicitly assume an existential dependence when RT relations are created, e.g. from the following RT application (notice that RT is considered symmetric in thesauri design):

hagrovoc:Rodenticides, agrovoc:Rodent_Controli ∈ agrovoc:RT

(5.19)

we can safely interpret the following restrictions:

agrovoc:Rodenticides v ∃agrovoc:RT. agrovoc:Rodent_Control agrovoc:Rodent_Control v ∃agrovoc:RT. agrovoc:Rodenticides

(5.20)

Of course, there remains to be seen what is the intended meaning of agrovoc:RT in that case, i.e. what is the intuitive relation between rodent control and rodenticides. As mentioned in section 5.4, an elegant technique is based on linking the top-level classes from the reengineered KOSes to “pivot” classes from core and reference ontologies, and applying rules to infer restrictions that apply to the pivot classes, to the linked classes. For example, from 5.21 (“reference:” is any well-axiomatized core or reference ontology to which a KOS can be linked):

agrovoc:Rodent_Control v reference:method agrovoc:Rodenticides v reference:functionalSubstance reference:functionalSubstance v ∃reference:resourceFor.reference:method reference:functionalSubstance v≥ 0 reference:usesResource.reference:method . reference:usesResource = reference:resourceFor−1 (5.21)


Page 57 of 151

we infer automatically that:

agrovoc:Rodenticides v ∃reference:resourceFor.reference:method

(5.22)

in addition, we can now propose to use the axiom at 5.20 as a good reason to infer the following:

agrovoc:Rodenticides v ∃reference:resourceFor.agrovoc:Rodent_Control agrovoc:Rodent_Control v ∃reference:usesResource.agrovoc:Rodenticides(5.23) Detailed explanations on the use of core and reference ontologies in KOS-to-TBox, with plenty of examples from FOS are contained in [GFK+ 04]. In particular, FOS used both a reengineered version of WordNet (OntoWordNet), and a dedicated core ontology of fishery (COF).


Page 58 of 151


Chapter 6

Enriching and Exploring Folksonomies In the previous chapter we described a range of methods to re-engineer and enrich a variety of Knowledge Organization Systems (KOS) such as thesauri (i.e., AGROVOC) and lexica (i.e., WordNet). In this chapter we are concerned with a particular type of KOS, folksonomies (see Fig. 5.1). Folksonomies [Wal07] are typical Web2.0 applications that allow users to upload, tag and share content such as photographs, bookmarks etc. The freedom of tagging largely contributed to the success of folksonomies: users need neither to have prior knowledge or specific skills to use the system ([HJSS06, WZY06]), nor to rely on a priori agreed structure or shared vocabulary. As a result, folksonomies capture socially engineered and evolved knowledge structures, thus being fundamentally different from thesauri and lexica which are typically built by a small group of experts. What is common, however, is that, semantically speaking, folksonomies are weak knowledge structures - in fact, they account for collections of tags as well as their inter-relations depending on the users who assigned them and the resources that they describe. In this chapter we report on two efforts to re-engineer folksonomy tag spaces into more formal ontologies so that this socially evolved, up-to-date knowledge can be reused in other ontology related projects.

6.1

Semantically Enriching Folksonomies

A distinctive feature of folksonomies is that they permit users to tag the same or similar resources with different tags depending on their social or cultural backgrounds, expertise and perception of the world ([GKF06, WZM06, Pet06, GH05]). For example, a zoologist can tag a photograph of a lion with {felidae, pantherinae, mammal}, while a non-zoology expert can use {lion, king, animal, jungle} for the same purpose. Unfortunately, the simplistic tag-based search used by folksonomies is agnostic to the way that tags used to describe different views about similar resources relate to each other. For example, a search for {mammal} ignores all resources that have not been tagged with this specific word, even if they are tagged with related concepts such as {lion, cow, cat}. As a result, content retrieval activities such as searching, subscription and exploration are limited ([GKF06]), they provide low-recall and hardly lend themselves to queryrefinement ([Sch06]). Therefore, to obtain satisfactory results, a searcher needs to build multiple complex queries that would cover all possible tags that could have been used by taggers ([WZM06, GH05, Pet06]). As searchers rely on their own view about what inter-related tags best describe the resource they are looking for, it follows that content retrieval could be enhanced if folksonomies were aware of the relations between their tags. Following this intuition, a variety of approaches have been proposed to identify tags that are inter-related based on the way they are used within the folksonomy. For example, [Sch06] uses a subsumption-based model, derived from the co-occurrence of tags, to find groups or related tags. [GKF06] organizes the tag space as an undirected graph, having frequently co-occurring tags as vertices, with the edges between them weighted according to their co-occurrence, and applying a spectral clustering algorithm to refine the resulting


Page 59 of 151

groups. [WZY06] uses a probabilistic model to generate groups of semantically related tags based on the co-occurrence of tags, resources, and users. These are represented as a multi-dimensional vector, where each dimension refers to a category of knowledge. Both the number of dimensions and the relationship values of entities to each dimension are determined using log-likelihood estimates. [Mik05] uses co-occurrence information to build graphs relating tags with users and tags with resources, and applies techniques of network analysis to discover sets of clusters of semantically related tags. [SM07] groups tags according to their co-occurrence by using a clustering algorithm similar to clustering by committee [Pan03]. Finally, most of the folksonomies provide facilities such as “clusters" and “related tags", which apparently also rely on co-occurrence information and clustering techniques. With the exception of [SM07], all the other approaches focus on finding groups of related tags rather than identifying the semantics of those relations. Specia et al. envisaged enriching tag spaces with semantic relations by exploring online ontologies. Their preliminary experiments on Flickr and Del.icio.us data confirmed that this is a promising strategy. Indeed, the recent growth of the Semantic Web has resulted in an increased amount of online available semantic data and has led to the first search engine to exploit this data, Swoogle [DFJ+ 04]. These facts made it possible to build applications that harvest the Semantic Web (i.e. dynamically select, combine and exploit online knowledge) to successfully solve a variety of tasks, such as: query disambiguation [GTEM06] and ontology matching [SdM06].

Figure 6.1: Lion in the Semantic Web

Applying this novel paradigm to folksonomies would make them explicitly aware of the inherent semantic relations between their tags. For example, subsumption relations such as the ones depicted in Figure 6.1 could be derived between the tags of the cluster {lion, animal, mammal, feline, tiger} by combining information from different online ontologies. The knowledge that Lions and Tigers are kind of Mammals would expand the potential of folksonomies. Users could make generic queries such as “Return all mammals" and obtain all resources that tagged with lion or tiger even if they are not explicitly tagged with mammal. While previous work has experimentally shown that harvesting online knowledge yields good results when applied to ontologies [SdM06], the folksonomy tag enrichment algorithm proposed in [SM07] was not fully automated. Therefore, an important research question is: Can we enrich folksonomies by automatically harvesting the Semantic Web? In particular, we are interested in finding out: What are the major characteristics of the Semantic Web and folksonomies that need to be taken into account to perform such enrichment? And if this enrichment is possible: What are its benefits? To answer these questions, we propose an approach to enrich the tag space of folksonomies which assumes the existence of previously defined groups of potentially related tags (these can be obtained by any of the above mentioned techniques) and which is entirely focused on the exploitation of the Semantic Web (Section 6.1.1). This approach is automated by using the algorithm described in [SdM06]. We present and discuss our experimental results which give an insight in the major characteristics of the Semantic Web and folksonomies that need to be considered when performing such enrichment (Section 6.1.2). We conclude and point out future work in Section 6.1.3.


Page 60 of 151

6.1.1


Semantic Enrichment of Folksonomy Tag Space

In this section we describe an approach to the semantic enrichment of folksonomy tag space which is a variation of the one proposed by [SM07]. Specia et. al describe a hybrid approach which combines harvesting the Semantic Web with using other Web resources such as Wikipedia and Google. As the goal of our work is to understand the potential and limitations of the Semantic Web when used to semantically enrich folksonomies, we have rephrased their algorithm so that it only relies on online ontologies. Our algorithm, presented next, takes as input a cluster of related tags and returns 1) a knowledge structure obtained by semantically relating these tags and 2) a set of tags which could not be semantically related to any other tag in their cluster.

Semantic Enrichment Algorithm The semantic enrichment of each cluster consists of two phases: (Phase 1) the concept definition of each tag (i.e., linking tags to ontology concepts) and (Phase 2) the relation discovery between all the possible pairs of tags. Phase 1. Concept Identification: The first step explicitly defines the meaning of each tag by extracting all Semantic Web Terms (SWT) whose label or localname are comparable with the tag. The comparison between the tag and the local name of the SWT can be achieved using anchoring techniques ranging from strict to flexible string matching as described in [SdM06]. Using the Semantic Web for extracting concepts is proposed in the work of [GTEM06] as a first step to query disambiguation. The authors search for candidate senses in online ontologies and then perform disambiguation based on the semantic similarity of the retrieved sense (e.g., bass can either refer to a fish or to musical notes depending on the context in which it is used). While we use the same technique for SWT identification we do not explicitly disambiguate between them. In our case, disambiguation is a side effect of relation discovery (Phase 2). The disambiguation of the tag sense (i.e., finding the right concept for a tag given its context) is approached differently in [SM07]. The authors rely on the heuristic that if pairs of tags from a cluster appear in the same ontology then this leads to an implicit disambiguation (i.e., searching for apple and fruit leads to ontologies about fruits, while when searching for apple and computer they identify ontologies about computers). While this intuition holds in the case of ontologies focused on a certain domain, it is problematic when the tags appear in broad, cross-domain ontologies such as WordNet or TAP1 . Also, by considering only ontologies that contain both tags, this approach potentially misses important information that might be declared in ontologies defining only one of the tags. This information can prove to be useful when combined with information from other ontologies. For example, an ontology containing Apple and Mac, can be combined with information from another ontology containing information about Mac and Computer. For these reasons, we retrieve all the potential SWTs for each tag and discover relations between them in Phase 2. Phase 2. Relation Discovery: This step identifies explicit semantic relations among all the pairs of SWTs (T1 and T2) discovered in the previous phase:

• Subsumption Relations: when one of the two SWTs is a subclass of the other, T1 subClassOf T2. This relation can be either declared in an ontology or inferred through transitivity. • Disjointness Relations: when T1 and T2 are disjoint, T1 disjointWith T2. Again this relation can be declared or inferred. We use the algorithm described in Section 6.1.1 to discover disjointness and subsumption relations.

• Generic Relations: when a generic relation holds between the two SWTs, e.g., Property1 hasDomain T1 and Property1 hasRange T2 or inversely. 1

http://tap.stanford.edu/data/


Page 61 of 151

• Sibling Relations: when the two SWTs share a common ancestor, which can be either a direct or an indirect parent. Note that our definition covers the three sibling definitions described in [SM07].

• Instance Of Relations: such as T1 instanceOf T2 or inversely. Unlike the previous relations, this relation is not considered by [SM07]. The identification of these relations can be made in two ways. First, a relation between SWT’s might be declared within a single ontology. Second, if no single ontology mentions both SWT’s, then a crossontology relation discovery can be performed by combining knowledge from several ontologies. In [SM07] the authors envision cross-ontology discovery as part of their future work thus strengthening our decision to perform such a search strategy. Cross-ontology relation discovery has been successfully implemented in the case of ontology matching [SdM06]. An important issue to be considered is how to deal with potentially contradictory relations, e.g., T1 subClassOf T2 and T1 disjointWith T2. This remains a future work topic. The semantically connected tags form the knowledge structures mentioned in the beginning of Section 6.1.1 and the tags not linked to SWTs or not related to other tags compose the set of uncovered tags. The study of the latter can provide hints about how to evolve the Semantic Web, as described in Section 6.1.2. Next we describe the current implementation of our approach which identifies only subsumption and disjointness relations found in single ontologies.

Subsumption/Disjointness Discovery Based on One Ontology The discovery of subsumption and disjointness relations between two terms within one ontology has been described and implemented on Swoogle’052 in [SdM06]. Given two candidate concept names (A and B) as an input, corresponding concepts are selected in online ontologies (A’ and B’) by using strict string based anchoring. The possible semantic relations occurring between concepts in an ontology are shown using description logic syntax, e.g., A’ v B’ means that A’ is a sub-concept of B’. The returned relations are v

expressed with arrows such as, e.g., A −→ B. The steps of this strategy in detail are: 1. Select ontologies containing concepts A’ and B’ corresponding to A and B; 2. For each resulting ontology: ≡

• if A’ ≡ B’ then derive A −→ B; v

• if A’ v B’ then derive A −→ B; w

• if A’ w B’ then derive A −→ B; ⊥

• if A’ ⊥ B’ then derive A −→ B; 3. If no ontology is found, no mapping is derived; In the simplest implementation, we can rely on direct and declared relations between A’ and B’ in the selected ontology. But, for better results, indirect and inferred relations should also be exploited (e.g., if A’ v C and C ⊥ B’, then A’ ⊥ B’). Different levels of inferences can be considered (no inference, basic transitivity, Description Logics reasoning), each of them representing a particular compromise between the performance of mapping and the completeness of the result. For our experiments, we used an implementation relying on basic transitivity reasoning (i.e., taking into account all parents of A’ and B’) and stopping as soon as a relation is found. 2

http://swoogle.umbc.edu/2005/index.php


Page 62 of 151

6.1.2


Experimental Results

The goal of our experiments is twofold. On the one hand we wish to reveal how much of the semantic enrichment of folksonomy tags can already be automated by using the software developed in [SdM06] which partially implements the current version of our envisioned algorithm (the part described in Section 6.1.1). On the other hand, we wish to understand any problematic issues so that they can be addressed in the design of the final, complete algorithm. At a higher level, these issues give an insight in how folksonomies and the Semantic Web relate. In a first experiment (Section 6.1.2) we applied the software developed in [SdM06] to Flickr and Del.icio.us clusters generated by [SM07]. This experiment lead to valuable insights into issues that hamper the enrichment and prompted us to repeat the experiments with another set of clusters selected directly from Flickr. We discuss the second set of experiments in Section 6.1.2. Experiment 1 The number of results obtained by running our algorithm with the clusters generated in [SM07] were surprisingly low. Two major reasons explain this. First, our implementation only searches for subClassOf and disjointWith relations. Unfortunately, the majority of tags in the clusters we work with are not related by these relations but by other, generic relations. The second major reason is that few of the tags in the analyzed clusters could be identified in ontologies in the Semantic Web. Taking a closer look to the tags that were not found we individuated the following cases: Novel terminology. Folksonomies are social artifacts, built by large masses of people, that dynamically change to reflect the latest terminology in several domains. As such, they greatly differ from ontologies which are developed by one person (or a small group of people) and evolve much slower. Therefore, it is not surprising that many of the tags used in folksonomies, e.g., {ajax, css}, have not yet been integrated into ontologies. Identifying frequent folksonomy tags that are missing from ontologies has a great potential for the Semantic Web as it can provide the first step towards enriching existing ontologies with these novel terms. Instances. When people tag resources, especially photographs, they more often tend to tag them with concrete names rather than more abstract concepts. In particular, we frequently find names of people {monica, luke, stephanie}, names of places {japan, california, italy} and particular dates {august2005, aug292005}. Unfortunately, the current version of our system only works at terminology level (it deals only with concepts and not with ontology instances), so we did not identify any of these instances in the experiments. Apart from that limitation it is unlikely that instances related to people and specific dates can be reliably identified in ontologies anyway. Photographic jargon. Given the scope of Flickr as a photo annotation and sharing site, many of the tags that are used reflect terms used in photography, such as, {nikon, canon, d50, cameraphone, closeup, macro}. Unfortunately, this domain is weakly covered in the Semantic Web. Multilingual tags. Both Flickr and Del.icio.us (but especially Flickr) contain tags from a variety of languages and not only English. These tags are usually hard to find on the Semantic Web because the language coverage of the existing ontologies is rather low. Indeed, statistics3 performed on a large collection of online ontologies (1177) in the context of the OntoSelect library indicate that 63% of these ontolgies contain English labels, while a much smaller percentage contains labels in other languages (German 13.25%, French 6.02%, Portuguese 3.61%, Spanish 3.01%). Concatenated tags such as {christmasornament, xmlhttprequest, librariesandlibrarians} appear frequently but obviously it is hard to identify concepts with the same spelling. 3

http://olp.dfki.de/OntoSelect/w/index.php?mode=stats


Page 63 of 151

Given the low coverage of the Semantic Web for the above mentioned categories of tags, we decided to repeat the experiments for clusters around tags that are well-covered in the Semantic Web. Also, since at this stage our system only discovers subsumption and disjoint relations, we decided that the experiments should consider significantly larger clusters than those provided by [SM07]. Experiment 2 In the second set of experiments we relied on the lessons learnt from the first experiment to identify clusters of tags that would be more appropriate for our goal. To address the first conclusion (i.e., that clusters should be potentially well covered in the Semantic Web), we relied on the results of previous work in the context of ontology matching [SdM06]. Follow up experiments revealed that domains related to food and animal species are well covered in the Semantic Web. Therefore, we selected a couple of tags from these domains, based on the concepts for which the most mappings were found during the matching experiments. We selected the tags: mushroom, fruit, beverage and mammal. The next step was to identify clusters of tags related to each tag above. As we said, we were looking for large clusters that would be more likely to accommodate subsumption relations and not just generic relations between tags. We chose the cluster facility provided by Flickr, since it returns much larger clusters of related tags than Del.icio.us and Technorati (moreover, since Del.icio.us and Technorati are mostly oriented towards news, business and web technologies, the clusters they provide for our tags in the food and animal domains are quite small). We used the Flickr API to retrieve the related clusters for each of the four tags4 . The same algorithm as in Experiment 1 was then applied to these clusters. As expected, we found several relations among tags as depicted in the figures bellow (directed arrows represent subClassOf relations, dotted lines depict disjointWith relations). Besides the tags between which we found relations, there were also sets of tags that could not be linked with any other tag in their cluster. We analyze these tag sets and describe a set of possible causes that lead to this failure. The case of Mushroom. The semantic relations identified between the tags related to mushroom by using online ontologies are depicted in Fig. 6.2. Mushroom was identified as a kind of Fungi and a kind of Plant. Also, we have learned that it is disjunct with Pizza, Pepper, Cheese and Tomato and so are these with each other. Mushroom also co occurs with Soup, Rice and Onion. As expected, there is no subsumption relation between these concepts and Mushroom. However, they are all subclasses of Food, as are Tomato and Cheese as well.

Figure 6.2: Mushroom in the Semantic Web

Table 6.1 shows some of the tags in the cluster of mushroom that could not be related semantically to any other tag, grouped according to the reason why they could not be linked. These are: 4

http://www.flickr.com/services/api/flickr.tags.getRelated.html


Page 64 of 151


Type Not covered by the SW Generic relation (location) Generic relation (seasons) Generic relation (usage) Colors Photo jargon

Tags {amanitamuscaria, toadstool, flyagaric} {nature, forest, garden, grass moss} {autumn, fall, herfst} {cooking, dinner, pasta, lunch} {green, white, yellow} {macro nikon closeup}

Table 6.1: mushroom related tags that could not be connected semantically Tags that are not covered by the Semantic Web. These tags refer to kinds of mushrooms or scientific names that are not described in the Semantic Web. Generally, our experience is that currently very few online ontologies cover scientific labels. Tags generically related to mushroom. The next three sets of tags are related to mushroom through other generic relations than subsumption or disjunction and describe locations, time and potential ways to use mushrooms. Tags about colors. This set of tags is not surprising taking into consideration that we retrieved the tag clusters from a photo-sharing system where users add color names to describe the image content of their photos. Note however, that these colors might be meant to describe the rest of the tags associated to a resource, e.g., {green pepper, white mushroom, yellow cheese}. Unfortunately, because the creation of compound tags such as these is not well handled by folksonomies, users have to add each tag separately, thus loosing the relationship between them. Photo jargon. The remaining group of tags are Flickr related tags, as we discussed in Experiment 1, and are not covered in the Semantic Web. Also, given the fact that they describe the photographs rather than their content, even if they were covered it is quite unlikely that they could be related to mushrooms or any other tag describing image content. The case of Fruit We obtained interesting results for the cluster of fruit (Fig.6.3). As fruits are wellcovered by the Semantic Web, the generated semantic structure contains much more information than just binary relations between the tags of the cluster. For example the multiple relations that exist between Fruit and Vegetable, and how this affects their common subclass, Tomato. In a biological context, a tomato is indeed the fruit of a tomato plant, however, normally one would classify tomatoes as types of vegetables. While such different views can co-exist, the fact that Fruit and Vegetable are disjoint makes this bit of knowledge inconsistent. Therefore, once such structures are derived from multiple ontologies, their consistency should be verified. Also, according to online ontologies, Fruit is disjoint with Dessert. The validity of this statement depends on the point of view we adopt: some would argue that fruits are desserts, while others might consider desserts generally far too unhealthy to contain fruits as well. Finally Strawberry and Watermelon were also found as subclasses of Fruit, but declaring them as subclasses of Berry and Melon, respectively, automatically infers their most generic parent, Fruit. The tags that could not be connected to Fruit fall into five categories (see Table 6.2), two of which are related to colors and photo jargons, as discussed before. A new set of interesting tags describes attributes generally related to fruits: {juicy, yummy, delicious, fresh, sweet}. Unfortunately, most concepts in ontologies model nouns. Attributes are often modeled as properties, as more geneneric relations. Finally, the other two sets of interesting tags refer to fruit cultivation methods and possibly best seasons for consumption of specific fruits, which again share generic relations with fruits, currently not in the scope of our software. The case of Beverage. The knowledge structure that emerged from the semantic enrichment of the cluster related to beverage is shown in Fig. 6.4. As in the case of fruit, the cluster for beverage contains


Page 65 of 151

Figure 6.3: Fruit in the Semantic Web

Type Attributes Generic relation (cultivation) Generic relation (seasons) Colors Photo jargon

Tags {juicy, yummy, delicious, fresh, sweet} {tree, nature, plant, seeds, leaves} {summer, autumn, fall, red, pink} {brown, green, white, red, pink} {closeup macro canon}

Table 6.2: fruit related tags that could not be connected semantically


Page 66 of 151


many concepts that were more specific than Beverage. Accordingly, these were identified to be in a subsumption relation with Beverage by our system. The tags that could not be related fall under the types of categories that we have already discussed in the previous cases and are presented in Table 6.3.

Figure 6.4: Beverage in the Semantic Web

Type Not covered by the SW Generic relation (container) Generic relation (event/place) Generic relation(ingredient) Attributes Colors Photo jargon

Tags {energy_drink, soda, martini, latte} {straw, mug, can, bottle, glass, cup} {breakfast, restaurant, party, starbucks} {lemon, fruit, cream, orange} {hot, delicious, refreshing} {brown, black, orange, green, red, pink} {closeup macro canon}

Table 6.3: beverage related tags that could not be connected semantically First, some types of beverages are not covered by the Semantic Web. It is interesting to note here that {latte} is not just an English word for a type of coffee, but also Italian for milk. The fact that it is not covered can be a side-effect of the low level of multilinguality in online ontologies, as we discussed in Experiment 1. Second, certain tags could be related to Beverage by generic relations, but these are not discovered by the current version of our system. These tags express types of containers, events and locations where beverages are served, as well as the ingredients of drinks. It is worth noticing that orange could belong both to the categories representing colors and ingredients. The final set of tags that could not be related refer to attributes which, as discussed before, have generally a weak coverage on the Semantic Web. The case of MammalThe last tag that was investigated is mammal. Fig. 6.5 shows the knowledge structure derived from its cluster. It is interesting to observe that the subclasses of Mammal do not represent the same level of abstraction. We note many common names of animals like Horse, Monkey, Rabbit, but also two subclasses of higher abstraction, Rodent and Feline. This is another evidence that users annotate their content with a variable level of generality: although Squirrel and Rabbit appear in the graph as subclasses of Mammal, their superclass, Rodent, appears as well. This confirms the hypothesis put forward by [GH05] according to which different users will settle at different “basic levels" depending on their level of expertise. The tags that could not be related are displayed in Table 6.4. Most of these categories have been discussed previously, along with a set of tags that could be related by generic relations indicating the location or habitat of mammals. Two tags were found to describe the state of the mammal when it was shot {eating,


Page 67 of 151

Figure 6.5: Mammal in the Semantic Web

sleeping}. Finally, an interesting set of tags is the one that depicts body parts which should be related to mammals through a part-of relation. Type Not covered by the SW Generic relation (location) Generic relation (action) Part-of Attributes Photo jargon

Tags {giraffe, seal, zebra} {zoo, nature, water, ocean, wild, farm, outdoors} {eating, sleeping} {fur, whiskers, eyes, face, nose} {cute, pet, funny, bunny} {portrait, closeup, macro, canon}

Table 6.4: mammal related tags that could not be connected semantically Finally, it is worth pointing out that in all of the above cases we identified certain tags, which were also found in Experiment 1, describing the places shown in the images, such as barcelona, japan, or the interests of the users, such as ilovenature, stilllife (we found 84.077 photographs annotated with ilovenature and 39.320 with stilllife).

6.1.3

Conclusions and Future Work

As an answer to our main research question, which is to explore whether folksonomies can be automatically enriched by harvesting the Semantic Web, based on the results of the preliminary experiments presented above, we can already conclude that it is indeed possible to automate the semantic enrichment of folksonomy tag spaces by harvesting online ontologies. By using these ontologies, we were able to automatically obtain semantic relations between the tags of several clusters of related tags. As an answer to our second research question, which is to identify the inherent characteristics of folksonomies and the Semantic Web and how they should be approached, the experiments also yielded relevant observations about the characteristics of folksonomies and the Semantic Web which have impact on the process of enriching folksonomies: 1. Folksonomy Characteristics. Our experiments show that many folksonomy tags fall in specific cat-


Page 68 of 151


egories that require special attention. First, by being dynamically updated by large masses of people, folksonomies reflect the newest terminology within several domains (novel terminology). Second, many folksonomy tags refer to specific instances (names of people, places, dates). Third, folksonomies contain tags representing words in a variety of languages (multilinguality). Fourth, some of the tags that are frequently used depend on the purpose of the folksonomy and usually describe the resource itself rather than its content (folksonomy jargon). Fifth, folksonomy tags often describe attributes of the content, for example, colors (especially in Flickr). Sixth, there are many concatenated tags which describe a large number of photographs and need to be exploited. Finally, there is a broad range of semantic relations that can exist between tags, including subsumption, disjointness, meronymy and many generic relations (e.g., location). 2. Semantic Web Characteristics. The most important observation regarding the Semantic Web, is that even if it is growing fast it still suffers from knowledge sparseness. Due to this limitation, we needed to restrict our experiments to domains that are well-covered (related to animals and food). Also, some of the categories of tags that appear frequently in folksonomies are difficult to find in online ontologies. First, novel terminology that emerges from folksonomies is often missing from ontologies. Second, the majority of specific instances that appear in folksonomies cannot be found (e.g., aug2004) or are difficult to reliably map to ontology instances (e.g., monica). Place names are an exception to this. Third, few of the online ontologies contain multilingual labels, therefore tags other than English are unlikely to be found in ontologies. Fourth, specific jargons, such as those related to photography are weakly covered as well. Fifth, online ontologies are rather poor in describing generic attributes such as color. One of the reason for this is that attributes are most often modeled as part of properties rather than concepts. We are confident, however, that surpassing some of the current limitations is a matter of time as many of them will be solved as more ontologies will appear online. For example, the AGROVOC5 ontology contains roughly 16000 concepts and their labels in 12 different languages. Making this single ontology available online will positively impact on the issue of anchoring multilingual tags. Nevertheless the appearance of more online ontologies can also be seen as a potential risk for this work as different ontologies reflect different views which often lead to contradictory bits of knowledge. Combining these bits then causes inconsistencies in the derived semantic structures, phenomenon which increases as the number of ontologies increases. However, existing reasoning techniques can be used to filter out and eliminate possible inconsistencies. Being aware of these characteristics help us to identify the current limitations of our software. Our software only implements a subset of the functionality envisioned for the enrichment algorithm. First, it is currently implemented on Swoogle’05 which lags behind in ontological content. Our final algorithm will be built on top of up-to-date semantic search engines. Second, the anchoring mechanism is based on strict string matching and therefore needs to be extended to more flexible anchoring. Third, from the broad range of semantic relations that can exist between tags, our software only identifies subsumption and disjointness. Obviously, extensions are needed that can discover the other types of relations as well. Finally, note that we have only experimented with finding relations within a single ontology and excluded cases when knowledge can be derived by combining facts from multiple ontologies. Another important future work will be to implement this cross-ontology relation derivation. The experimental work reported here indicates that the proposed enrichment process has the potential to benefit both folksonomies and the Semantic Web, thus answering our third research question. On the one hand, even if using a software with limited functionality we were able to derive explicit semantic relations between tags, thus going beyond existing methods that identify implicitly inter-related tags. We believe this could considerably enhance content retrieval in folksonomies. On the other hand, the differences between folksonomies and ontologies (such as novel terminologies emerging in several languages) can be used to evolve the Semantic Web. This valuable knowledge available in folksonomies could allow keeping online ontologies up to date, extending them with multi-lingual information and evolving them towards being truly shared conceptualizations of a much broader range of domains.

5

http://www.fao.org/agrovoc


6.2

Page 69 of 151

Learning from Folksonomies

As it was previously shown, it is very promising to use tags for extending the lexical layer of an ontology. It was also shown that there exist problems with e.g. misspelled tags, different flexions of them or the ambiguity of tags. First ideas were implemented which address those problems. It is planned to further elaborate on the topic and to develop further techniques suitable for identifying merging and disambiguating tags. This may be used as a preprocessing step for improving the results of e.g. relation identification but it can also be used for defining a n:m mapping between the concepts contained in an ontology and the tags of a folksonomy and thus for extending an ontology with vocabulary coming from folksonomies. Furthermore, we will work on how the social dimension of folksonomies can be exploited for the enrichment of ontologies with vocabulary coming from folksonomies. For example, one may identify different user communities which are defined by their usage of tags. On the one hand, that may be language communities e.g. for German, English and French but also communities differing in their level of expertise may be distinguished. For example, a laymen may tag a photo only with “mushroom” while a biologist may also use the correct botanical name. The information about user communities and their different vocabularies will help to enrich ontologies with multi-lingual lexical representations of a concept or, in the case of a laymen and an expert community, to customize a given ontology to different expertise levels and requirements.


Page 70 of 151


Chapter 7

Deriving Ontology Design Patterns In the context of NeOn WP2 and in particular of task 2.5, ontology design patterns represent a main research topic. Specifically, in task 2.5 we are going to create a catalogue of ontology design patterns, and to design and prototype a registry supporting ontology design patterns storage and use. Ontology design patterns are of different types as it is preliminarly described in NeOn deliverable D5.1.1, and will be deeply studied in and reported in deliverables of task 2.5. In the context of this section we deal only with one type of ontology design patterns, i.e., Content Ontology Design Patterns (CODeP)[Gan05], and we focus to the domain of Business Interaction. We show how CODePs for business domain can be derived through the re-engineering of Data Model patterns [Hay96] and other formalisms (such as business modeling languages). The re-engineering is performed by specializing more general CODePs. Here we present two general patterns, namely Basic Description and Situation, and Plans (the latter is a specialization of the former). Then, we present an example business scenario which is modeled by using data model patterns and UML, and we show how to reengineer it in order to obtain an ontological description in terms of the early presented general CODePs. The presented approach is being adopted in the context of NeOn WP2 in order to define the CODeP catalogue.

Content Ontology Design Patterns (CODeP) A Content Ontology Design Pattern (CODeP) [Gan05] is a fragment of an existing ontology that is relevant because it is widely reusable (e.g. the Time interval relations pattern), or because it is well-suited or central to a specific domain of interest (e.g. an Invoicing pattern in a business transaction domain), or because it solves a recurring modelling problem (e.g. a domainindependent relation reification pattern). CODePs are a viable alternative to large and complex foundational or core ontologies, as well as to repositories of informal modelling patterns that cannot be easily composed or reused because of the difficulty of encoding them in a unique reasoning framework. Intuitively, CODePs are typically quite small (usually smaller than 10 classes), richly axiomatized (most classes have more than one association to the others), and close to the human expertise (they typically come out of reasoning patterns developed by experts, or are abstractions of them). Formally, CODePs are theories that include invariant structures against some specific transformations (“vocabulary morphisms") that can occur across a variety of domain patterns. Notably, many abstract patterns have similarities to socalled frames from cognitive linguistics [BFL98]. A formal theory of CODePs is still under development, but [CTP00] and [Gan05] have laid down some of its foundations.

7.1

Basic Description and Situation

Figure 7.1 depicts the general structure of Basic Description and Situation (DnS) CODeP. The classes involved in this pattern are Description, Concept, Entity, and Situation. A description is a social object that represents a sharable conceptualization, hence it is dependent on some agent and must be communicable (i.e. expressed by means of information objects). Descriptions can be also seen as viewpoints on some state of affairs or context. Descriptions typically define or use concepts, and the association Usage


Page 71 of 151

Figure 7.1: The DnS CODeP

captures this part of the pattern by relating concepts to viewpoint/s or description/s in which a concept is defined or used. Descriptions can be satisfied by situations as expressed by the association Satisfaction. A specific situation has to satisfy a description, and satisfaction is guaranteed by the fact that a situation is a setting for some specific entities, as described by the association Setting. Those entities are classified by at least one concept used by the description that the situation satisfies. The association Classification captures this part of the pattern. Finally, the association Specialization between concepts is used to specify concept taxonomies, e.g. reifications of UML generalizations. The DnS CODeP is a very general one. It should be read as a two-layer structure (imagine to split Figure 7.1 into two parts by drawing a vertical line exactly in the middle) that provides a vocabulary for reification. The first layer i.e., the left side of Figure 7.1, involves the classes Description and Concept, and provides a way to model a conceptualization. Classes Situation and Entity form the second layer i.e., the right side of Figure 7.1, and allow us to express occurrences of states of affairs that comply to descriptions. This CODeP is extracted from the ExtendedDnS ontology. For a comprehensive discussion and explanation, the reader can refer to [GM03],[MLB+ 04], and for the complete ontology [Thea]. ExtendedDnS takes into account also:

• social agents, • collectives or communities which these agents are members of, • the information object/s by which a description is expressed, • the time-spans characterizing the situations. For this reason ExtendedDnS can be widely used for extracting CODePs and use them in business modelling as well as in many other domains. However, an appropriate description and exemplification of CODePs that can be extracted from ExtendedDnS and used in business modelling will be subject for task 2.5 deliverables. For the sake of this deliverable, we provide anyway a simplified, first-order-logic explanation of the DnS Maximal Relation [GC06], a logical structure that is used as a generic design pattern to represent ontologies about social reality (the former UML diagram for ExtendedDnS is (logically) a projection1 of the DnS Maximal Relation. A toy use case is also provided, in order to strengthen the reader’s intuition. The DnS Maximal Relation has the following structure:

DnS(a, k, s, t, i, d, c1...n , e1...n ) → A(a) ∧ K(k) ∧ S(s) ∧ T (t) ∧ I(i) ∧ D(d) ∧ C(c1 ), ..., C(cn ) ∧ E(e1 ), ..., E(en )

(7.1)

1

A projection of a relation is a part of it, which keeps the relative position of arguments, and is implicitly dependent on the complete relation. E.g. given the relation eats(agent, f ood) can be a projection of the complete relation eats(agent, f ood, tool).


Page 72 of 151


A can be read as Social agent, K as Collection, S as Situation, T as Time interval, I as Information object, D as Description, C as Concept, and E as Entity. Intuitively, the DnS relation says e.g. that a social agent (a), as a member of a community (k ), singles out a situation (s) at a certain time (t), by using information (i) whose meaning is a descriptive relation (d) that assigns concepts (c1...n ) to entities (e1...n ) within that situation. The DnS Maximal Relation is complex due to its expressivity, which catches the social context surrounding the process of knowledge extraction and sharing. For example, a local situation of two employees exchanging emails about an order can be represented analytically as follows:

DnS(Supervisor#134, ACM E _Community, EmailT hread#13092008, T ime#13092008, LogM essage#13092008, ACM E _EmailW orkf low, {OrderM anager, ShippingCoordinator, emailM essage, Order}, {M ark, Sean, {email_texts}, EngineOrder#02062008})

(7.2)

The relationship formally states that a supervisor from the ACME company community has been able to verify an email thread on 13th September, 2008, as evidenced by a log message whose meaning is based on the ACME workflow for emails, which assigns the concepts: Order Manager, Shipping Coordinator, email Message, and Order, to the agents, information, and resources involved in the email thread situation, i.e. Mark (as order Manager), Sean (as Shipping Coordinator), some email texts (as email Message), and an order for an engine (as Order). This toy example illustrates the scope of the DnS pattern. The same pattern can be applied e.g. to represent team coordination situations, financial performance indicators, contract execution, social interaction data extracted from social network analysis, etc. Projections of the pattern can help focusing on specific aspects of social interaction, e.g. a social relationship situation can be represented by means of the following projection of the Maximal Relation:

SocRel(s, t, a1...n , k)

(7.3)

which says that a social relationship situation s at time t includes a set of agents (a1...n ) from a collective (k ). The projection only uses the s, t, and e1...n arguments of DnS, where entities are agents and collectives. In practice, DnS allows to talk about how agents and communities become aware of some situation, and at the same time, to talk about agents and communities within those situations. For example, in the previous example, the supervisor #134 is the agent that becomes aware of the email thread, while Mark and Sean are the agents within that thread situation. In the following, we will concentrate on smaller projections of DnS, in order to keep the complexity of UML diagrams under control.

7.2

Plans

In this section we are going to describe the Plan CODeP [Theb] (particularly useful for describing flows and processes). It is a specialization of DnS for a more specific domain, i.e., plans. This pattern is useful to describe procedures in contexts such as the enterprise business domain. Figure 7.2 depicts the general structure of the pattern. The Plan CODeP is more complex than that captured by such structure. Here we present a simplified version of it (although complete with main elements) and approach its description by adding elements step by step. A plan is characterized as a method for executing or performing a procedure or a stage of a procedure. It must use at least one role played by an agent, and at least one task. It might also use parameters, whose values are taken from appropriate regions (e.g. spatial, temporal, or physical values). Finally, a plan has always a goal as a proper part, and can also have regulations or other descriptions as proper parts. The above definition of plan characterizes its conceptual aspects. Following the two-layered approach described in section 7.1, we can say that such definition is captured by the first layer of the CODeP structure.


Page 73 of 151

Figure 7.2: The Plan CODeP

Figure 7.3: Generalization of plan elements

The second layer represents the possible settings that can comply to the first layer. Specifically, a plan can be satisfied by a plan execution, which provides the setting for entities like actions (classified by a given task), objects (classified by roles), and regions (classified by parameters). In order to clarify the link between Plan and DnS CODePs, Figure 7.3 depicts generalizations of plan classes with respect to DnS classes. Plans can be refined or expanded by other plans, and they can have pre- and post- conditions for their execution. Furthermore, a plan execution can be part of a sequence. These aspects provide further characterization of the Plan CODeP. Figure 7.4 shows relations precondition, postcondition, successor, predecessor, refines, and expands. Situations can be a pre-condition for the execution of a plan when they are assumed as required predecessors to its execution. Analogously, a situation is a post-condition for the execution of a plan when it is assumed as a required successor to that execution. Figure 7.5 shows that tasks are assigned to (i.e. thay are target of) some role played by an agent. Tasks can be complex, and ordered according to an abstract successor relation. Tasks are organized in a taxonomy. Figure 7.6 depicts some of them; for the full taxonomy see [Theb]. A task is complex when it has at least two other tasks as components, or elementary when they are atomic within a certain plan model. Elementary tasks can be either action task or control task. Action tasks represent the activities to be accomplished in the actual plan execution, while control tasks represent the activities


Page 74 of 151


Figure 7.4: Conditions and Sequences in Plan

Figure 7.5: Relations between Concepts of Plan


Page 75 of 151

Figure 7.6: Plan Tasks Taxonomy


Page 76 of 151


Figure 7.7: Kinds of Contracts Data Model Pattern

supporting the execution, i.e., they anticipate other “controlled” activities. In the following list we summarize some of control tasks that have been defined in [Theb]:

• Beginning task : a control task that is predecessor to all tasks defined in the plan; • Ending task : a control task that has no successor tasks defined in the plan; • Alternate task : a task branched to exactly 2 tasks, not executable in parallel. It is a specialization of the case task, which can be branched to a set of tasks not executable concurrently;

• Deliberation task : a task representing the decision taken after a case task execution; • Concurrency task : a task branched to a set of tasks executable concurrently; • Synchro task : is a merging aimed at waiting for the execution of all (except the optional ones) tasks that are direct successor to a concurrent task;

• Abstract merging task : a formal merging that is never executed and is used to indicate the closing of the branches of a case task.

7.3

An example of re-engineering in business modeling

In order to show an example application of our approach, we describe a possible use case scenario for a generic enterprise. In particular, from [Hay96], we take a data model pattern named Kinds of Contracts (KoC) and from [Sal], we take the description of the Sales Order Process (SOP). We use the UML graphical notation in order to model both, and in particular we use UML, class diagram for KoC and UML use case and activity diagrams for SOP. Furthermore, we highlight the workflow patterns used within this particular activity diagram by referring to [VDATHKB03]. After presenting the scenario, we describe our strategy for creating an ontology of that scenario. In practice, we describe how we can specialize a set of suitable CODePs for re-describing the scenario in a unified domain of discourse. Scenario

Figure 7.7 depicts the data model pattern Kinds of Contracts (KoC) as it is defined in [Hay96].

KoC says that a contract is set up between two parties that can be persons or organizations, while each possible party can be on one or the other “side” of a contract. Pragmatically, the subject of the contract can be some asset, asset type, service, or activity. All these elements are explicitly represented in the data model


Page 77 of 151

Figure 7.8: Sales Order Use Case

pattern. The goods that are involved in the contract are described by means of the entity line item, i.e., they can be bought via a certain line item. Actually, one or more line items compose the contract. The pattern considers also the case of dealing with services and their execution (i.e., activities). Entities of this data model pattern can then be associated to some workflows for a given enterprise. Since the information about the flows is not embedded within the data model pattern, it has to be expressed by means of a more suitable formalism for that specific issue (e.g., UML activity diagram). This approach can be enhanced by ontology-based model representation. It allows designers to capture all aspects by using only one formalism. An advantage of a unique formalism is the creation of a common domain of discourse for modeling, querying, mining, etc. In practice, we can compare different aspects of the enterprise business domain. As an example, from IBM WebSphere web site [Sal], we take the description of so called Sales Order Processing (SOP), which is, as it is stated there, “a core process in any business”. In [Sal] SOP is described as a set of tasks, performed by different actors in different possible sequences, i.e., workflows. The focus of SOP is on system requirements and the description of the intended meaning of the elements in SOP diagram is provided only as English comments. SOP also uses diagrams and other notations, which deal with system specifications based on that set of requirements. Although system specification is an interesting domain to be represented in ontologies, as it is shown in [Obe06, OMGS04, RU04], in this context we focus on models for enterprise business and production processes instead of software system models. We rephrase here the description found in [Sal] for SOP workflow. Figure 7.8 depicts the sales order use case. There are three actors involved in this use case, namely the customer, the Customer Service Representative (CSR), and the Back Office System (BOS). The order has to be requested, forwarded, and fulfilled. Between this main operations several others might be necessary in order to verify that certain conditions hold, e.g., check for reliability of customer, availability of goods, and so on. Figure 7.9 shows a possible workflow for sales order processing. A customer initiates a request for certain products that a company manufactures or distributes. The CSR records the request. In this particular scenario, before going on the CSR checks for customer reliability and if the check result is positive then the availability of goods is verified, otherwise the order is rejected and the process finishes. If the order is accepted and goods are not available, then it is needed to place an order to the appropriate supplier. On the other hand, if goods are already available the order is sent to the BOS, which is in charge of configuring products in order to meet customer’s needs, and to fulfill the order. This workflow is represented with UML activity diagram notation, as mentioned above. It is a simple example and it does not cover all possible workflow structures that might be necessary to represent. In [VDATHKB03]


Page 78 of 151


Figure 7.9: Sales Order Process possible workflow


Page 79 of 151

so called workflow patterns are listed, all structures needed for representing from basic to complex workflows. Each business model language has its own degree of coverage of these set of patterns. With CODePsbased modeling it is possible to have the desired expressiveness based on the specific enterprise domain needs. For example, in order to describe the workflow of Figure 7.9 we used the following workflow patterns [VDATHKB03]:

• Sequence: an activity in a workflow process is enabled after the completion of another activity in the same process;

• Exclusive choice: a point in the workflow process where, based on a decision or workflow control data, one of several branches is chosen;

• Simple merge: a point in the workflow process where two or more alternative branches come together without synchronization. It is assumed that none of the alternative branches is ever executed in parallel. In Figure 7.9 these are represented respectively with a directed arrow, a diamond, and a bar. The scenario we have built includes an enterprise, whose business consists in selling and/or buying products (either artifacts or services). The sales order process is managed by means of a defined workflow, and is regulated by a contract, one of many possible kinds of contracts. The scenario is split into parts observed from different perspectives, and each part is described in a different, most suitable way. As an example, in this section we have used natural language and UML class, use case, and activity diagram notation. In the rest of this section we show how to use suitable CODePs in order to describe the above described scenario as a whole within a unique domain of discourse and using only one formalism. The approach is to take general CODePs like those we have seen in sections 7.1 and 7.2, and to specialize them for the domain of interest i.e., the enterprise business.

Workflow ontology Figure 7.10 shows a UML object diagram, which describes the workflow of figure 7.9 as a specialization of the CODePs presented in sections 7.1 and 7.2. Each UML object maps to an OWL individual, and, according to the UML notation for object diagrams, the name of an individual is followed by the name of its class, and the two names are separated by a colon. For the sake of readability, Figure 7.10 contains only a few examples of uses relations, and target. The reader can infer all others by observing the types of individuals.

Sales Order Process is a plan, which uses three roles, namely Customer, CSR, and BOS, and a set of tasks e.g., Request For Order, Credit Check etc. Roles, according to the plans pattern (CODeP) of section 7.2, has some tasks as targets, and sequences of tasks are represented by means of the relation successor. In this specific workflow two AlternateTask s have been used: Credit Check and Inventory Availability Check. Each of them has its corresponding pair of DeliberationTask s : Reliable Customer and Not Reliable Customer for the former, and Goods Available and Goods Not Available for the latter. They also have their AbstractMergingTask, namely Credit Check Abstract Merge and Inventory Availability Check Abstract Merge, respectively. AlternateTask s (or in general CaseTask s) together with DeliberationTask s correspond to “exclusive choices”, while AbstractMergingTask s corresponds to “Simple merges” as defined in [VDATHKB03]. Finally, Sales Order Process Begin is a BeginningTask while the Sales Order Process End is an EndingTask, as they describe respectively the start and the end of the workflow.

Contract ontology In this section we show the re-engineering of the data model pattern Kind of Contracts (KoC) presented in Figure 7.7 by means of the CODePs presented in sections 7.1 and 7.2. This simple example shows that we can combine the same expressiveness of ad-hoc languages (e.g., Entity Relationship, UML) with the additional value of expressing all aspects in the same domain of discourse. As a first step we define KoC entities as ontology classes in terms of DnS and Plans CODePs, Figure 7.11 depicts that. In particular:


Page 80 of 151


Figure 7.10: Sales Order Process CODeP-based description


Page 81 of 151

A Contract is a Description while Service and Line Item are subclasses of Plan i.e., a special kind of Description. Party, and Asset, are defined as subclasses of class Role, emphAsset Type as subclass of class Concept, and Person, Organization, and Building are Objects.Finally, Activity is a subclass of Plan Execution. Notice that in Figure 7.7 the entity Activity is related to the entity Service, by means of an association labels as “execution of”. This is in line with the mapping of those entities to Plans CODeP classes and relations.

Figure 7.11: Kind Of Contracts Ontology Classes Figure 7.12 shows the relations that hold among the classes we have just defined.

Figure 7.12: Kind Of Contracts Ontology Specifically, a Contract has a Line Item as a proper part. The composition/aggregation associations are far too ambiguous, as Guizzardi shows [Gui05], then we propose here a specific properPartOf as defined in DOLCE [GGMO03, MGG+ 04]. Contract is stipulated from one Party to another; “from” is too generic as a reasonable relation name, then we propose its superrelation from the DnS codep, i.e. uses. The generalizations between organization/person and party, and between building (or other objects) and asset are “non-rigid” [GW02], because e.g. persons are not always parties, and buildings are not always assets. Therefore, based on the DnS CODeP, we transform those generalizations into classifies relations. Persons or Organizations can play the role of Party of a Contract i.e., Party classifies Person and Organization. The Line Item of a Contract may deal with one among Asset, Asset Type,Service, and Activity.

• an Asset is classified by an Asset Type e.g., a car which is of a specific brand and model;


Page 82 of 151


• an Activity is the execution of a Service e.g. the reparation of my car is the execution of a specific workflow (i.e., plan) for fixing that problem. Asset, Asset Type, Activity, and Service are all associated to Line Item through the “bought via” association. Since those classes are all disjoint in the reference ontology of our CODePs, and a best practice in ontology engineering is to provide different relations for disjoint domains, the association has therefore different meanings that, both legally and commercially, imply different practices and regulations. Consequently, we have numbered the four associations according to the class associated to Line Item. Each numbered association can then be formalized as a sub property of (i.e., owl:subPropertyOf) the involves property that composes the uses and classifies properties from the DnS CODeP. Unified model In this section we describe the example scenario above drawn as a CODeP-based unified model. Figure 7.10 contains the model of a SOP workflow based on the Plan CODeP, while Figures 7.11 and 7.12 contain that of KoC ontology after the reengineering of its data model pattern, based on the DnS CODeP. Now if we consider an enterprise that wants to endorse the SOP workflow to work out sales orders, and the KoC contract datamodel to regulate their sales order contracts, the merging is straightforward, and Figure 7.13 depicts the relationships that allow us to unify the model and to describe the scenario as a whole2 . The merging is simple because we can directly conceive the SOP workflow as an instance of the KoC: LineItem class, and the roles used by SOP as instances of the role class that includes the roles expected to be used by a Line Item instance.

Figure 7.13: Unified Model

7.4 Conclusion and future work In this chapter we have shown how CODePs for business domain can be derived through re-engineering of existing components, which were not initially thought for ontology modeling e.g., Data Model patterns [Hay96]. To this aim, we have presented an example: a possible business scenario, which has been modeled by means of data model patterns and UML activity diagrams. We have re-engineered these models in order to obtain an ontology model, which specializes two general CODePs we have previously described i.e. DnS, and Plans. We are applying the described approach in order to build an ontology design pattern library in the context of NeOn T2.5. However, as we are in an early stage of development of such library, the definition of generalized methods for re-engineering existing components to ontology design patterns is untimely. In the course of the NeOn project, and specifically in the context of T2.5 we will further experience practical 2

arrows with dashed line represent the rdf:type relation that holds between an individual and its class


Page 83 of 151

application of the above presented method and possibly new ones. Next versions of this deliverable will contain results from this experience. Specifically, we will identify general principles, and define re-engineering methods and techniques for deriving ontology design patterns from existing components, which were not designed for ontology modeling.


Page 84 of 151


Part II

Ontology Evaluation


Page 85 of 151

Chapter 8

Ontology Evaluation for Collaboration Ontology evaluation is an important topic that has attracted a considerable amount of work in the field of the Semantic Web. The goal of this deliverable is to report on evaluation work performed within the NeOn consortium, to understand how this work can support collaboration and to identify further research issues derived from the context of collaboration. While we reference a wide range of existing work, our primary aim is not to provide a thorough state of the art deliverable. Readers interested in state of the art overviews on ontology evaluation are referred to two recent surveys in [BGM05] and [HSG+ 05]. As argued earlier in this deliverable, re-engineering of resources as well as their selection play an important role in supporting collaboration. Ontology evaluation is an important issue that needs to be considered both during re-engineering (to assess the quality of the obtained ontology) and during ontology selection (to compare the quality of a set of candidate ontologies and to rank them accordingly). In the next subsection we give a principled overview of ontology evaluation techniques and classify the material from this part of the deliverable according to this framework. Then we detail the role of ontology evaluation both in the context of data re-engineering (Section 8.2.1 and Section 8.2.2) and ontology selection (Section 8.2.3).

8.1

A Principled Overview of Ontology Evaluation

In [GCCL06] a multi-layered approach to ontology evaluation is presented which we wish to use here as a framework for categorizing the work on ontology evaluation performed within NeOn. In this approach there are three layers that are directly related to evaluation: O 2 , a meta-ontology that allows to treat an ontology as a semiotic object; oQual (for Ontology Quality), a pattern based on O2 , which models ontology evaluation as a diagnostic task; qood (for Quality-Oriented Ontology Description), the component of oQual which describes the desired evaluation criteria. Brief descriptions of each of these layers are provided below. O2 meta-ontology provides a semiotic meta-model for ontologies in terms of the following classes of entities: Rational Agent, Ontology Profile, Ontology Description, Ontology Graph, Conceptualization, Ontology Element and Semantic Space. The central idea is that an ontology element has an intended conceptualization as well as a (formal) semantic space admitted by that conceptualization. The element and the conceptualization are related to one another by a rational agent who, on the one hand, encodes/interprets the element and, on the other hand, internally represents the conceptualization. In addition, the agent may encode a profile of the ontology element, which contains the metadata to express a description of the ontology. Relevant examples of such descriptions are those describing methods to measure and evaluate an ontology element. oQual is a pattern based on O2 ’s ontology descriptions, which models ontology evaluation as a diagnostic task. oQual1 is based on two main classes: Quality-Oriented Ontology Description (qood) and Ontology Intended Use Situation. The main idea here is that instances of the latter class may be related to 1

oQual is an application of the reification paradigm Description and Situation introduced in chapter 7 of this deliverable.


Page 86 of 151


instances of the former by a relation of satisfaction. This relation holds when the considered ontology satisfies (in terms of its intended use) the quality criteria defined in the qood. Note that O2 and oQual will be better integrated with C-ODO in the future deliverable D2.1.2 (and therefore also in D2.2.2). qood is the component of oQual which describes the desired evaluation criteria for ontologies. A qood is based on the following classes: Ontology Element Role, Ontology Element Task, Quality Parameter2 . The instances of these three classes classify the instances of the classes Ontology Elements, Ontology-driven process, respectively, Values Space - which form the basis of the instances of the class Ontology Intended Use Situation3 . The central idea here is that a qood provides parameters i.e. range of values - for the values of the attributes of the ontology elements (described in the qood as roles) and for the values of the attributes of the ontology-driven processes (described in the qood as tasks). In order to deploy the measures for evaluating an ontology via qoods, a number of issues must be addressed, as follows. What to measure in an ontology and how? The quality of an ontology may be measured relative to three main groups of dimensions (structural, functional, usability-related), pertaining to distinct attributes of an ontology and of its elements. Structural dimensions (like breadth, depth, tangledness, fan-outness, differentia specifica, density, modularity, consistency, complexity, various ratios and various qualified types of these dimensions) are related to the topological and logical attributes of an ontology. Structural dimensions are measured against (the elements of an) ontology, represented as a graph - i.e. by contextindependent metrics. In other words, structurally speaking, an ontology should be considered no more than an information object. Functional dimensions (precision, recall, accuracy, adequacy and various qualified types of these dimensions) are related to the attributes of the intended use of an ontology and of its elements i.e. their function in a context. Functional dimensions are measured against the agreement, the task, the topic, the design of an ontology’s conceptualization. Functionally speaking, an ontology is a language (i.e. an information object plus its intended conceptualization). Usability-related dimensions (presence, amount, completeness, and reliability) are related to the attributes of the ease of use of an ontology (how easy it is for users to recognise its properties, how easy is to find out which one is more suitable for given tasks, etc.). Usability-related dimensions are measured against the level of annotation of a given ontology. From the viewpoint of usability, an ontology is a meta-language (i.e. an information object plus its intended conceptualization plus its semiotic context). Which parameters for the quality of an ontology? Each measure can have more than one quality parameter, depending on other parameters/measures, and the overall composition for a given ontology project implies a non-linear procedure to quality assessment. For example, an ontology project may combine measures like logical complexity and presence of dense areas (e.g. of design patterns). If high density is chosen as a quality parameter, then the parameter associated with high complexity is chosen too, because usually dense areas involve a lot of restrictions, sometimes with indirect cycles; in other words, the high-density parameter depends on the high-complexity parameter. On the other hand, if the quality parameter is low complexity, then the parameter associated with low density is chosen too, because the first depends on the second. Hence, different trade-offs denote good/bad quality according to which criterion is preferred. oQual formalizes the observation that quality parameters are defined according to some principle. In the example, “high parameters” could be associated to a 2

In Description and Situation terms, a qood is said to define such classes. In Description and Situation terms, the instances of the classes Ontology Elements, Ontology-driven process, Values Space are said to be in the setting of the instances of Ontology Intended Use Situation. 3


Page 87 of 151

“transparency” principle, while the “low” parameters could be associated to a “computational efficiency principle”. When combining principles, the need for a trade-off typically arises, producing either a preference ordering function, or a relaxation of parameters. Again in the example, either high density can be relaxed to e.g. medium-high density, or low complexity can be increased to e.g. medium-low complexity. Which examples? There are typical examples and patterns (Gold Standards) of good/bad quality for each measure. The material presented in this part of the deliverable does not provide insight on measuring structural dimensions - as these are not directly relevant to the evaluation of the knowledge encoded in an ontology. The material we describe mainly proposes ways of measuring functional dimensions. The methods based on Gold Standards for topic assessment (presented in chapter 9.2) either measure how close a given ontology is to a Gold Standard ontology (which in turn defines and/or gives an example of the general boundaries of the considered knowledge) or compare a set of extracted instances with a Gold Standard set of instances. Moreover, the notion of user-satisfaction in Open Rating Systems (presented in chapter 10) may too be seen as relating to the functional dimension, in terms of the adequacy of the evaluated ontology4 . Finally, concerning the above mentioned trade-offs between (set of) parameters, no matter whether a given set of parameters is dictated by a Gold Standard or by a (set of) principle(s), the choice between alternative parameters probably depends on the purpose for which the evaluation is performed (i.e. selection or reengineering or etc.).

8.2

Evaluation Scenarios and Approaches

When it comes to the evaluation of ontologies, one has to distinguish between three scenarios which have different requirements with regard to the evaluation of the structural, functional and usability-related dimension of ontologies. Before we proceed in the following chapters with describing concrete evaluation measures we think it is important to give an overview of the three scenarios and the already available evaluation approaches.

8.2.1

Scenario 1: Quality Assurance During Ontology Engineering

In this scenario ontologies are evaluated during the ontology engineering process as part of the quality assurance. Typical questions for evaluating an ontology are whether it is consistent, complete, concise and expandable (see [GP03]). For this purpose, we previously proposed in 8.1 we proposed to measure the structural and functional dimension of ontologies and their usability profile. The requirements which should be fulfilled by an ontology with regard to the dimensions will be usually defined during the start phase of an ontology engineering project. For example, one may check whether the target domain of the ontology is sufficiently modeled to fulfill the functional requirements and/or whether the ontology helps to improve the performance in the task for which it is designed. As a consequence of such an evaluation one may e.g. decide to further extend some aspects of the ontology. During the structural evaluation it is checked whether certain criteria are fulfilled which are related to the design principles of good ontologies and which help to improve an ontology’s overall quality. In this scenario, the evaluation of ontologies is seen as an important part of the quality assurance process. In many cases the ontologies will be manually engineered. But also data re-engineering techniques like ontology learning or population (see the previous chapter) may be applied during the ontology engineering process. Thus, also the output of such re-engineering techniques can be evaluated during the quality assurance process. 4

Even tough it is not unreasonable to presume that user-satisfaction somehow reflects/depends on an evaluation of the usability of the ontology.


Page 88 of 151


The following approaches to a functional and structural evaluation can be identified in this scenario: Task-based Approaches The task-based approaches try to measure how far an ontology helps to improve the results of a certain task. For example, if one designs an ontology for improving the performance of a web search engine (cf. [WKCC03]) one may collect several example queries and compare whether the search results contain more relevant documents if a certain ontology is used. A task-based evaluation is influenced by many aspects which have to be kept constant during all evaluations so that changes in the results can be put down to the changes in the used ontologies. The choice of concrete measures for such an evaluation is dependent on the task, e.g. for the web search engine from above one may adapt measures known from information retrieval but also other success criteria may be defined. Because every task-based evaluation is individual, no finite set of well-suited measures can be defined. Nevertheless, some principles can be identified: Usually, it is not enough to know whether an ontology is better or worse than another but one wants to conclude on concrete shortcomings in its conceptualization. Thus, in [PM04] it is required that a task-based evaluation allows for concluding on insertion, deletion and substitution errors in the ontology, i.e. whether there are superfluous, missing or off-target concepts and/or relations. But again, there is no universally valid way about how these principles can be realized in a concrete task-based evaluation. In [PM04] it is only demonstrated for a single example task. Corpus-based Approaches Corpus-based approaches are used for checking how well (sufficiently) an ontology covers a given domain. For this purpose, the ontology is compared with the content of a text corpus which is representative for the domain. The content of the corpus is analyzed with natural language techniques, e.g. in [BADW04] Latent Semantic Analysis and a clustering method were applied for identifying terms in the corpus. The list of identified terms was then compared with the terms in the evaluated ontology. Similar approaches for evaluating the lexical layer of an ontology are described in [DR04] and [SR05] while [Spy05] contains a preliminary method applicable for evaluating triples in ontologies. All the corpus-based approaches have in common that they involve information extraction and/or ontology learning techniques in the evaluation. Thus they are not suitable for evaluating learned ontologies but only for evaluating manually engineered ontologies. Indeed, in [DR04] it is proposed to evaluate and extend ontologies at the same time with such an approach, e.g. by suggesting terms which are currently missing in the ontology and which would improve the evaluation results. Criteria-based Approaches In this category fall a wide variety of evaluation measures which all have in common that they measure how well an ontology or taxonomy adheres to certain desirable criteria. One can distinguish between measures related to the structure of an ontology, e.g. if it is represented as a graph, and more sophisticated measures which e.g. evaluate a taxonomy based on philosophical notions. Structural measures are quite straight forward and easy to understand: For example, one may measure the average depth of paths from root to leave nodes in a directed graph, how many nodes have more than one ingoing arc (i.e. multi-hierarchical nodes) or whether there are cycles in the directed graph (cf. [GCCL06] and [GCCL05]). But also for frame- or description logic based ontologies one may define structural measures, e.g. for detecting potential inconsistencies in the partitioning of a taxonomy (cf. [GP03]). Such a partitioning error measure may for example find instances belonging to more than one class and where two or more of the classes are defined as disjoint. For the structural measures it is usually no problem to have a fully automatic evaluation. This is not the case for the more sophisticated measures like OntoClean [GW03] which evaluates taxonomies based on philosophical notions like the essence, identity and unity which should be taken into account during modeling an ontology in order to avoid common pitfalls. For example, a property is essential for an entity if it holds for that entity in every possible world. Furthermore, a property is rigid if it is essential for all possible instances. In [GW03], this is explained with the example relations having a brain and


Page 89 of 151

being a student. In this example, the having a brain relation is essential for all human beings thus it is a rigid property. In contrast the being a student relation, which is not essential for any human being as everyone can become a student or cease to be a student at any time. Thus it would be a anti-rigid property. (For more examples see [GW03].) Because of this high complexity, OntoClean is thought for manually analyzing ontologies although an approach for partially automating this process was recently proposed (see [VVS05]). The most important success criterion for an ontology engineering project is whether the final ontology helps to improve the task for which it was engineered. Thus, improving the results during a task-based evaluation can be seen as the most important goal. Corpus-based and criteria-based evaluation approaches only help to pinpoint the remaining problems which should be addressed in an improved version of the ontology. The main assumption behind corpus-based and criteria-based evaluation measures is that an improvement with regard to the measures correlates with an improvement in the task-based evaluation (see [WKCC03] where the correlation was shown for OntoClean [GW03]).

8.2.2

Scenario 2: Comparing Learning Algorithms

Ontology learning and ontology population are two frequently used re-engineering methods within the Semantic Web in general, and for supporting collaboration in particular. In this scenario one tries to directly assess and compare different learning algorithms with each other. It can be used by researchers to improve existing algorithms or to find out how changing the values of input parameters affects the results. Thus the quality of a learning algorithm itself should be evaluated. This is usually done by looking at the output (i.e. the derived ontology or the instantiated ontology) and comparing it with the input (i.e. the data source). With regard to the functional dimension of the evaluation one is interested in measuring in how far the learning algorithm is able to conceptualize the information from a given corpus (e.g. whether an ontology learning technique extracts isA-relations between relevant concepts) or, in the case of ontology population, whether it correctly links ontology concepts and their corresponding instances in the textual document (or other data source). Furthermore, one wants to know which fraction of the information available in the data source is found. This corresponds to measuring a kind of precision and recall. But also from evaluating the structural dimension of the learned ontology or ontology instances one may draw interesting conclusions on the qualities of a learning algorithm. In the following, two approaches to measuring the functional dimension will be presented which are specific for the needs of this scenario and which are different to the approaches from the first scenario. In contrast, it is possible to re-use a subset of the structural measures described in 8.2.1. Thus, we will concentrate here on the evaluation of the functional dimension. In the previous scenario, a task-based evaluation was considered as ideal for evaluating the functional dimension of an ontology. This is not the case, if a more general evaluation and comparison of learning algorithms should be done. The reason is that the results of a task-based evaluation are influenced by many other factors like the choice of the corpus, the task itself or the algorithm used for performing the task. Even if all of these aspects are kept constant during the evaluation, it is difficult to transfer the evaluation results to other tasks. Taking the task-based evaluation seriously would mean to re-do the evaluation for several tasks and algorithms, which are used for performing the task. Another aspect, why a task-based evaluation can not be considered as ideal in this scenario is that it is very difficult to conclude from the results of a task-based evaluation on the concrete precision and recall values achieved by the learning algorithm. Instead, it would be valuable to have a more direct approach to measuring those dimensions of interest. All in all, the following list of criteria should be fulfilled by an evaluation in this scenario:

• The evaluation should be task neutral and allow developers to easily pinpoint the advantages and disadvantages of a learning algorithm. Weighing the different advantages and disadvantages is then up to the ontology engineer who has a concrete task in mind. This weighing can be based on his


Page 90 of 151


experience or even on a task-based evaluation where it was shown that certain aspects are more important than other.

• All influencing factors of the evaluation have to be sufficiently described so that its results can be reproduced at another time and place. This is important for having a proper scientific evaluation.

• It should be possible to do additional evaluation runs at low cost because frequent and large-scale evaluations are required during developing learning algorithms. It has to be ensured that all evaluation runs are performed under the same conditions in order to have comparable results. By looking at the literature, one can identify the following two approaches for measuring the functional dimension in this scenario: Manual Evaluation by Human Experts This evaluation approach can be found in several papers about learning algorithms like in [BC99] and [GM02] where a learned ontology is presented to one or more human experts which have to judge the correctness of the extracted information (i.e. the precision is measured). But the approach has several drawbacks: First of all, the extracted information is not compared with the information found in the corpus but with the knowledge of the human expert. While this is not so problematic for measuring the precision of the learning algorithm it makes a reliable measurement of the recall nearly impossible. Furthermore, the most important influencing factor of the evaluation is the choice of the human experts. Because they may not be available at another time and place the last two criteria from above are not fulfilled. This problem can only be avoided by asking a sufficiently large number of experts. Additionally, every evaluation run comes with the same high costs as the first run thus making frequent and large-scale evaluations unfeasible. Gold Standard Based Approaches Gold standard based approaches compare the output of the learning algorithm with a previously created gold standard which represents an idealized outcome. A learning algorithm is considered to be better when its output has a high similarity with the gold standard. Examples for this kind of evaluation can be found in papers like [CPSTS05], [SR05] and [SWGS05]. The gold standard based evaluation fulfills all the criteria from above: It can be used for directly measuring the precision and recall with regard to the gold standard. Furthermore, the evaluation results can be reproduced and are comparable if the same corpus, learning algorithm and gold standard are used. Additionally, only for the first run of the evaluation the high costs of creating the gold standard exist. Subsequent runs of the evaluation are then fully automatic. Although the gold standard based evaluation seems to be ideal in this scenario there remains one big issue: Where to get or how to create such a gold standard? On the one hand, one may ask a human expert to create a gold standard based on the information in the used corpus. Depending on the size of the corpus, this can constitute a very work intensive approach. Another approach might be to take an already existing ontology and choose the corpus accordingly so that it can be assumed that most of the information of the gold standard is available in the corpus. An example of the latter approach is available in [CHS05]. Independent from this decision, the term “gold standard” may be misleading as there exists not only one gold standard but, depending on who is asked, one may get several gold standards which differ in their details. This is due to the different conceptualization humans may have of a domain (cf. [GCCL06]). The same problem exists for the manual evaluation by human experts. There it is typically addressed by measuring the consensus between several experts (cf. [MS02]). A similar way may be used for the creation of the gold standard. For example, one may involve several experts in the creation of the gold standard and measure their consensus or one may compare with several gold standards (and measuring the agreement between those gold standards). But regardless from this decision, the main advantage of gold standard based evaluation remains that it makes the expected output explicitly available in form of the gold standard. This ensures that every learning algorithm is compared against the same standard and that everyone can control how thorough the gold standard was created.


Page 91 of 151

In this deliverable, we present in more detail three different gold standard based evaluation measures which were developed or should be further developed in the context of NeOn. In Section 9.2 we describe the OntoRand index, a method for evaluating ontologies which have instances assigned to their concepts. It may be used for evaluating how well a concept hierarchy was learned as well as for evaluating the arrangement of the assigned instances, i.e. the way the ontology was populated. In Section 9.3, precision and recall based measures for the evaluation of the lexical layer of ontologies and the evaluation of concept hierarchies are described. Finally, we present in Section 9.4 the balanced distance metric (BDM) and how it can be used for evaluating ontology population.

8.2.3

Scenario 3: Evaluation for Ontology Selection

We define ontology selection as the process that allows identifying one or more ontologies or ontology modules that satisfy certain criteria. The actual process of checking whether an ontology satisfies certain criteria is, in essence, an ontology evaluation task. Therefore ontology evaluation is central to the process of ontology selection. One important requirement from the perspective of ontology selection is that it should be possible to automatically assess the important criteria. Current approaches to ontology selection evaluate ontologies depending on (a) the coverage of a set of keywords, (b) their popularity (i.e., how often they are used by other ontologies) and (c) by the richness of the conceptualized knowledge (see [SLMU06] for an overview). All these criteria can be assessed automatically. However, since this line of work is very new, the current metrics are rather arbitrary, they do not take into account the semantic nature of the ontologies and they have not been extensively tested. One of the goals of this deliverable is to focus on defining a set of automatic evaluation metrics that could be used within ontology selection and consequently use them in our own selection mechanism. In future work, we will rely on the set of metrics proposed by CNR [GCCL06], select those that can be evaluated automatically, implement them on top of the WATSON Semantic Web gateway [dSD+ 07] and test their behavior on a large number of ontologies (thousands). While automatic metrics can be used to evaluate several aspects of a given ontology, there exist a number of quality criteria that cannot be automatically assessed. In these cases users can be asked to provide their own ratings of ontologies that they tried to use/re-use. Such a mechanism is complementary with that of traditional, automatic selection mechanisms, however, it is more user and community oriented. Users rely on other user’s ratings and in time a trust relation develops between the members of the community that use a given repository. In Section 10 we describe a user-based rating system employed for evaluating ontologies.


Page 92 of 151


Chapter 9

Gold Standard Based Evaluation In this chapter, we detail three different gold standard based evaluation measures which were developed or should be further developed in the context of NeOn. In Section 9.2 we describe the OntoRand index, a method for evaluating ontologies which have instances assigned to their concepts. In Section 9.3, precision and recall based measures for the evaluation of the lexical layer of ontologies and the evaluation of concept hierarchies are described. Finally, we present in Section 9.4 the balanced distance metric (BDM) and how it can be used for evaluating ontology population.

9.1

Criteria for Good Evaluation Measures

Before we start describing concrete evaluation measures for doing a gold standard based evaluation, it is important to establish what how to compare different measures. For this we need to understand what we expect from a “good” measure and which are thee criteria according to which the different measures can be evaluated. Measures fulfilling the following criteria will help avoiding the misinterpretation of evaluation results and ease drawing the right conclusions about how to improve the evaluated ontology or ontology learning algorithm. The most important criterion is that a measure allows for evaluating an ontology along multiple dimensions. This criterion is formulated in several papers like [MPL06] and [HSM+ 04]. Thus a user can weight different kinds of errors based on his own preferences. This enables to better analyze the strengths and weaknesses of a learned ontology. If a multi dimensional evaluation is performed, each measure should be influenced just by one dimension, i.e. by one type of error only. For example, if one uses measures for evaluating the lexical layer of an ontology (e.g the lexical precision and recall) and one also wants to evaluate the quality of the learned concept hierarchy (e.g. with the taxonomic overlap), then a dependency between those measures should be avoided. The second criterion is that the effect of an error onto the measure should be proportional to the distance between the correct and the given result, i.e. a graded correctness score should be given. For example, an error near the root of a concept hierarchy should have a stronger effect on the evaluation measure than an error nearer to the leafs (see also [HSM+ 04]). The third criterion is closely related to the previous one. For measures with a closed scale interval (e.g.

[0..1]), a gradual increase in the error rate should also lead to a gradual decrease in the evaluation results. For example, if a measure has the interval [0..1] as its scale but already slight errors lead to a decrease of the returned results from 1 to 0.2 then it is difficult to distinguish between slight and severe errors (see [BGM06]).

9.2

Instance Based Evaluation with OntoRand

JSI has previously developed an approach to ontology evaluation that may be adopted for NeOn, if we have a gold standard ontology to be evaluated against. The approach is primarily geared to enable automatic


Page 93 of 151

evaluation of an ontology that includes instances of the ontology concepts. The approach is based on the gold standard paradigm and its main focus is to compare how well the given ontology resembles the gold standard in the arrangement of instances into concepts and the hierarchical arrangement of the concepts themselves. While it bases the evaluation on instances assigned to the ontology concepts, it does not rely on natural language descriptions of the concepts and instances (unlike e.g. the string edit distance approaches of Maedche and Staab [MS02]). No assumptions are made regarding the representation of instances, only that we can distinguish one instance from another (and that the ontology is based on the same set of instances as the gold standard). The approach is based on the analogies between ontology learning task as defined in OntoGen (see Chapter 3) and traditional unsupervised clustering. In clustering, the task is to partition a set of instances into a family of disjoint subsets. Here, the topic ontology can be seen as a hierarchical way of partitioning the set of instances. There are different techniques for comparing two partitions of the same set of instances, which can be used to compare the output of an automated clustering method to a golden-standard partition. These measures are usually defined for traditional “flat” partitions but can be extended to hierarchical partitions and can then be used to compare a learned ontology to the golden-standard ontology. OntoRand index, as defined in [BGM06] measures similarity over ontologies as follows. Let us denote by U (o) the cluster of U that contains the instance o ∈ O, and similarly by V (o) the cluster of V that contains the instance o ∈ O. Let δX (Xi , Xj ) be some distance measure between clusters Xi and Xj of a partition X . Then we define the OntoRand index by the following formula:

P OntoRandIdx(U, V ) = 1 −

1≤i

(10.1)

where α = (α1 , α2 , α3 , α4 ) is a vector representing weights for combining the four atomic propagation schemes, B is the belief matrix and B > is the transposed belief matrix. Entries in CB,α indicate how trust can be propagated within the Web of Trust. To propagate the trust, it is necessary to apply CB,α on the initial trust information available. Let P (k) be a propagation matrix where each entry describes how strong the trust is between users after k propagation steps: P (k) = CB,α · (T − D) (10.2) Using a combination of P (k) with different propagation depths, a final propagation matrix F can be computed


Table 10.2: Atomic Trust Propagation Operator Description

Propagation Direct Propagation

B

B> · B

Co-Citation

Transpose Trust

Trust Coupling

Page 113 of 151

B>

B · B>

If A trusts B , someone trusted by B should also be trusted by A

If A trusts B and C , someone trusting C should also trust B

If A trusts B , someone trusting B should also trust A

If A and B trust C , someone trusting A should also trust B

using a weighted linear combination:

F =

K X

γ k · P (k)

(10.3)

k=1

where K is a suitably chosen integer and γ is a constant that is smaller than the largest eigenvalue of CB,α . K represents the maximal depth of trust propagation in the Web of Trust, γ is a parameter basically determining the rate of decay of trust as propagated within the Web of Trust (the further trust is propagated, the weaker it becomes). While it is mathematically sound to perform the computation of trust values on a continuous scale, at some point, those values have to be interpreted as trust or distrust. The most successful method presented by Guha and colleagues [GKRT04] is called “majority rounding”. The basic idea is to use information from the original belief matrix B to make assumptions about whether an inferred trust value should be interpreted as trust or distrust. Suppose a user i expresses trust and distrust for n people (entries in the trust matrix T or distrust matrix D ), and we need to infer a trust relationship towards a user j . Using the final propagation matrix F , all inferred trust values linking i to the n users initially trusted or distrusted are sorted in the ascending order, including the entry fij . Then, depending on the local neighborhood of trust statements in the ordered set, fij is interpreted as trust or distrust, based on the majority of trust statements in the neighborhood. Based on the inferred trust information, a ranking of reviews of other users can be made. We will go into details on how we perform such a ranking in the following paragraph.

10.2.5

Computing Trust Values for the Ranking

Now that all trust statements are in the form W , trust ranks can be computed. Note that in contrast to the traditional model, we do compute trust relationships for every property of every ontology (every On Xk combination) specifically. We have to distinguish two possibilities when a ranking of evaluations of a property of an ontology has to be computed for a user querying the system: Either that user has made a specific trust statements for any available evaluation covering that On Xk combination or not. If a user has not made any trust statement, no local trust information can be inferred. In that case, the ranking has to be based on all trust statements (made by all other users) affecting that On Xk combination. This is done by using a modified version of the TrustRank (see equation 10.4) and DistrustRank (see equation 10.5) algorithms introduced in [Guh03]:

X T rustRankN +1 (Au ) = (1 − d) + d · (

v∈Tv

T rustRankN (v) ) Nv

(10.4)


Page 114 of 151


where Au is the agent whose TrustRank is computed, v ∈ Tv is the agent trusting Au , Nv is the total number of agents agent v trusts, d is a damping factor between 0 and 1 (usually set to 0.85), and N is the number of iterations.

T rustRank(v) (10.5) v∈Bv Nv where Au is the agent whose DistrustRank is computed, v ∈ Bv is the agents distrusting Au , Nv is the total number of agents the agent v distrusts. Intuitively speaking, TrustRank assigns trust to agents based on how DistrustRank(Au ) =

X

many other agents trust them and how important the opinion of those agents is. The same holds true for DistrustRank, it is taking into account who distrusts an agent and how important the distrusting agents are. As it is evident, TrustRank is basically just a PageRank [PBMW98]. In contrast to TrustRank, DistrustRank can be computed with only one iteration of the algorithm. If local trust information is available, propagation of trust along a user’s web of trust can be performed. Guha and colleagues performed an extensive evaluation, testing their algorithm using real world data and provided valuable insights towards the best choice of parameters [GKRT04]. Because the algorithm was proven to produce good results on real-world data, we use the same parameters in our computation of the final propagation matrices F . We compute a propagation matrix F for every On Xk combination featuring evaluations. We perform the calculation of FOn Xk using single-step distrust propagation and majority rounding (see section “Propagation of Trust and Distrust” on page 112) as follows: 1. BOn Xk = TOn Xk 2. CBO

n Xk

,α

(k0 )

> > > = 0.4 · BOn Xk + 0.4 · BO BOn Xk + 0.1 · BO + 0.1 · BOn Xk BO n Xk n Xk n Xk 0

k 3. POn Xk = CB

4. FOn Xk =

On Xk

,α

· (TOn Xk − DOn Xk )

k0 k0 =1 0.9

P7

(k0 )

· POn Xk

5. Interpret values using “Majority Rounding” where TOn Xk and DOn Xk are Trust and Distrust matrices (as defined in section “Propagation of Trust and Distrust” on page 112) specific to the On Xk combination, and K is set to 7 (since it is not sensible to propagate trust further than 7 steps). Explanations of the different matrices and operations can be found in section “Propagation of Trust and Distrust” on page 112.

10.2.6

Ranking Evaluations at the Property Level of an Ontology

Evaluations that exist for an On Xk combination are linked to their author. The quality of an evaluation is determined by feedback from the user community on how helpful it was. The TrustRank, DistrustRank and F values provide the information about the global ranking of authors (TrustRank and DistrustRank) and about the ranking of authors as perceived by each single user (F ). The first choice a user has to make when querying the system is how to combine TrustRank and DistrustRank. Some may tend to put a great emphasis on TrustRank values while others rely on the significance of DistrustRanks. Each user therefore has to choose a parameter α ∈ [0, 1] that can be stored in a profile and is used to compute:

CombinedRank(Ai ) = T rustRank(Ai ) − (α · DistrustRank(Ai ))

(10.6)

When evaluations have to be ranked, the system first looks if any local trust information exists for the user – On Xk combination. If yes, using FOn Xk , the information who is trusted and who is distrusted is retrieved and then both trusted and distrusted users are ordered using their inferred local trust rank. In case two users share the same local trust value, the order of those reviewers is determined by their CombinedRank . The results of the ranking will start with the users that are trusted locally. When the ranks of all the locally trusted


Page 115 of 151

users have been determined, the following ranks are filled with all reviewers that no local trust information is available for, using their CombinedRank . Lastly, the locally distrusted reviewers are ranked starting with the least distrusted reviewer ending with the most distrusted reviewer. If no local trust information is available, ranking is solely based on CombinedRank . At the end of the ranking of the reviewers, their reviews are presented to the user in that order. Note that local ranks always override global ranks, so that a user having a very subversive view will have reviews by his favorite reviewers ranked first instead of reviews that the majority of users like.

10.2.7

Computing an Overall Evaluation of an Ontology

Each ontology to be ranked has several properties it can be evaluated on, such as degree of formality, maturity, quality of content or reusability. Combining the ratings provided in the context of evaluation of its properties, an overall rating can be inferred for an ontology. It is important to note that there is no single right way to combine the evaluations of an ontology’s properties. Depending on the intended application, different aspects may be important to the user. Therefore, for every query, weights µk (that are normalized P to ensure that µk = 1) have to be assigned to all ontology properties Xk that the system should take into account for computing the overall rating. A user who wants to find ontologies that have a high maturity and reusability might choose to assign µ1 = 0.5 to property “maturity” and µ2 = 0.5 to property “reusability”. While we assume most users searching for an ontology will know exactly which ontology properties are important for their intended use of the ontology, µk -weight-presets will be offered for the rest. Since every ontology property Xk will have one top-ranked evaluation (specific to the user querying) featuring a rating P On Xk . Since the Dm associated by R, an overall rating for an ontology can be computed as DOn = µk · Dm system uses parameters specific to each user, DOn can not be preprocessed. In contrast to the traditional model, it is now possible to compose an overall rating using evaluations of different reviewers.

10.2.8

Ranking Ontologies

The main tasks the Open Rating System has to perform in an ontology repository is ranking ontologies that show up as result of a query or when browsing categories in the domain hierarchy. The first step in getting a ranking is finding the objects that should be ranked. In the case of a query, a simple pattern-based comparison of the search term and metadata annotation in the system should provide a subset of ontologies that have to be ranked. If the task is to rank all ontologies belonging to a certain domain concept, the concept hierarchy HC is traversed down adding all ontologies that are defined to have that domain by L at each concept Ci to the result space. An example would be a user browsing the science domain. First all ontologies being science ontologies would be added to the result space, then the domain hierarchy would be traversed down adding all biological, chemical, computer science a.s.o. ontologies until all ontologies covering science or any subcategory are added. Once all ontologies that have to be ranked have been found, they are ordered using the inferred overall rating DOn (see section “Computing an Overall Evaluation of an Ontology” on page 115). The ranking results are highly user-specific, because they are based on a user’s trust statements, the parameter α, and the weights µk assigned to the different ontology properties.

10.3

Enabling Collaborative Ontology Evaluation with ORS

As has been shown in the previous section, ORS can be used to combine a number of different opinions/reviews. One of the key aspects here is the fact that no reviewer has to review the whole ontology or all ontologies, but it is sufficient to just review aspects of an ontology or an ontology that one is familiar with and qualified to judge. So in a way, ORS can be seen as a giant collaboration platform. Every user adds to the utility of every other user by simply stating an opinion. Of course, collaboration can also be planned, for example by assigning certain aspects to review to reviewers who are experts on these aspects. So a team of ontology experts could collaboratively evaluate a number of ontologies, each focussing on their own


Page 116 of 151


Prerequisites: - Knowing what property of an ontology and what kind of ontologies can be evaluated using your automated evaluation technique - Knowing the mapping (interpretation) from evaluation technique results to 1-5 star ratings

Interpret evaluation technique output and write reviews back into ORS

Extracting relevant ontologies to be evaluated

Ontology 1 Output for Ontology 1 ORS with Ontologies

Ontology 2

Review by Reviewer "Eval" for Ontology 1

Ontology 3 1 Output for Ontology 2

ORS with Ontologies

Applying eval. technique to relevant ontology properties / ontologies Review by Reviewer "Eval" for Ontology 2

Automated Evaluation Technique "Eval"

Ontology 1

Output for Ontology 1

Ontology 2


Ontology 3


Review by Reviewer "Eval" for Ontology 3

Output for Ontology 3 2

3

Figure 10.2: This figure depicts how automated evaluation techniques can be used to automatically populate ontology reviews in ORS.

individual strength. Another important aspect to point out is that even if a user does not agree with the expert, there might be a review of a different reviewer that he or she can relate to and use.

10.4 Combining Different Evaluation Techniques into the ORS Framework We have encountered several different evaluation techniques in this part of the deliverable. They each have their own strength and weaknesses and might be more appropriate for one application but not for another. Since the idea of ORS is to stay generic in your ability to express thoughts (you can put any explanation you like in the free text part of the rating), and the reviews are written by conscious human beings and not a computer, ORS have the potential to accommodate all challenges and requirements posed to ontology evaluation. For example, let us assume a user wants to have an ontology to plug into his application. Given that another user has already faced a similar problem and wrote reviews for different ontologies he considered to be adequate for this task, finding the right ontology comes down to finding the one with the best reviews. An automated technique will most likely not be adaptable for each potential use case and will thus likely produce results that are not directly usable. So how can one combine different evaluation techniques into a meta-evaluation framework using an ORS? Let us assume that each evaluation technique is assigned a fixed author id (e.g. “Gold1” for gold standard evaluation 1). Given that the results of the evaluation techniques can be interpreted in terms of quality


Page 117 of 151

(e.g. high similarity with gold standard is good) every ontology can be assessed by the (potentially automated) evaluation technique and the review be added into the ORS automatically. So imagine the gold standard evaluation technique was run on all ontologies in a certain domain (or whatever a sensible scope of the evaluation technique might be taking into account the ontologies in the system) and the results added as reviews of these ontologies by user “Gold1” already interpreted in terms of 1–5 star ratings with the actual results as the review. Then the ORS could be easily filled automatically with automated reviews provided by different evaluation techniques. What would be restrictions on evaluation techniques that can be added to ORS? The most important restriction would be that the results of the evaluation technique has to be interpretable in terms of 1–5 star ratings. However that should be the case. What good is an evaluation technique if the results cannot be interpreted in terms of quality? Given that the mapping from evaluation results to 1–5 star ratings exists and the reviews of the different evaluation techniques can be automatically added into the ORS, we have an easy way to fill up the reviews of ontologies by these automated methods. Ideally, it can be decided in advance which property of the ontology is assessed by which evaluation technique. Examples could be allowing evaluation techniques that rate the domain coverage, or formal complexity automatically generating reviews for these ontology properties. For a depiction of this idea see Figure 10.2. To tie the ideas of ORS together, what could one do with an ORS containing both human reviews and automated reviews? Given one understood how the automated techniques work, one might assign trust to a certain evaluation technique for a certain property of an ontology or a certain domain. Let us assume reviewer “FormalQuality” was an automated evaluation technique for assessing the formal quality of an ontology and formal quality would be a property of the ontology that is ratable, then we should trust reviewer “FormalQuality” to rate that property of ontologies. So in the end the final ranking of ontologies as determined in the system could rely on a mixture of “subjective” (user) reviews and “objective” (automated) reviews. While this framework looks similar to the CORE Framework proposed by Fernandez et al [FCC06], it follows a different approach under the surface. Not only is it more generic (potentially any new kind of evaluation technique can be accommodated) but also the way the ranking works is quite different (as can be seen by comparing the actual ranking algorithms). Obviously, without reviews and ontologies an ORS is worthless. Adding ontologies in the system is the easy part, since those can be crawled from the web or existing repositories. Getting users to review ontologies is considerably more difficult. First, there are not so many experts and second, these expert normally do not want to spend their time rating ontologies. However, once the interest in ontologies and the semantic web increases to a level where ontologies become more mainstream, this hopefully will change. Up till then, using automated evaluation techniques together with human reviews will be a promising way to sell the benefit of these systems and get people to use it. So even without human reviews it can be used as a evaluation framework for existing (automated) evaluation techniques, enabling the user to easily combine different approaches potentially using different weights for each query.


Page 118 of 151


Part III

Ontology Selection


Page 119 of 151

Chapter 11

Ontology Selection According to the view taken on collaboration within WP2 (described in D2.1.1), a frequently occurring collaboration process is that of usage. Usage denotes situations of “virtual collaboration", when somebody reuses knowledge without the implicit collaboration of the author of this asset. A pre-requisite to establish such usage type collaborations is the availability of methods that allow locating the relevant ontologies for a given task. More generally, robust mechanisms for selecting ontologies are crucial to support knowledge reuse in large scale and open environments as the Semantic Web. The context of reuse has a major influence on the requirements for the selection algorithm and should be taken into account when developing such algorithms. In particular, the context of the reuse highly depends on the agents that reuse this knowledge. We distinguish between human agents (that have the capability to interpret the results of the selection before reuse) and software agents (which automatically integrate the returned ontologies). All existing work on ontology selection has focused on human mediated reuse (Section 11.2). Also, the ORS model presented in the previous chapter is mostly focused on scenarios where a human user is involved in the ontology selection process (by providing ratings and interpreting the ratings of other fellow users). We consider that, while ambitious, the context of automatic reuse complements that of human mediated reuse and raises novel challenges that can lead to further development of existing selection algorithms. In this section we describe work that investigates ontology selection in the context of automatic reuse. We start by defining what we mean by ontology selection and by giving an intuitive example about how the requirements for ontology selection differ in human mediated and automated scenarios (Section 11.1). Then we overview existing work in Section 11.2. In Section 11.3 we derive a set of requirements imposed by two applications that are extended to perform automatic knowledge reuse. We also consider characteristics of online ontologies explored through a set of indicative experiments (Section 11.4). We then present an initial design of a selection algorithm that balances between obtaining a complete and precise coverage and offering a good performance (Section 11.5). Finally, we describe our work (performed jointly with WP1) in combining this selection algorithm with ontology modularization techniques (Section 11.6). Through this work, we go one step beyond ontology selection towards selecting appropriate reusable knowledge components.

11.1

What is Ontology Selection?

We define ontology selection as the process that allows identifying one or more ontologies or ontology modules that satisfy certain criteria. The actual process of checking whether an ontology satisfies ceratin criteria is, in essence, an ontology evaluation task. For example, when one needs to select an ontology that has the best coverage for a given corpus, a prerequisite of the selection lies in evaluating all considered ontologies on this criterium. We distinguish the following elements that characterize the ontology selection process: The information need. The aim of the selection process is to identify an ontology structure that satisfies a certain information need. The information need can be expressed differently. For example, it could


Page 120 of 151


be expressed as a set of keywords, a logical query, it could be represented by a corpus or by an ontology. Obviously, the way the information need is expressed is influenced by the requirements of the application that will use the results of the selection (an application defines a usage scenario for the selection task). The selection criteria. The core task of the ontology selection algorithm is to evaluate a set of ontologies in order to identify the ones that fulfill the selection criteria. These criteria can be related to topic coverage, ontology structure or ontology popularity. The ontology library. Ontology selection is performed on top of a collection of ontologies, i.e., an ontology library. The output. The ontology selection process could have different outputs. For example, selected ontologies can be presented as a ranked list of ontologies. In other cases selection might return possible combinations of ontologies that jointly satisfy a certain information need. Often, consumers are only interested in a part of the ontology - so the relevant module should be presented from the perspective of the users and using the right level of granularity. In [N.F04], N. Noy points out that objective evaluations do not often support the ontology users to their best and that particular care should be taken to help naive users find ontologies and evaluate their suitability for the user’s tasks. From the perspective of ontology selection this translates in providing a friendly output for ontology selection facilities. For example, a summarization of the selected ontologies. Before understanding the requirements for an ontology selection algorithm, it is important to distinguish between the differen scenarios in which selection is used. We further discuss and contrast the requirements imposed by the contexts of human mediated and automatic reuse. As background for our discussion, consider the following news snippet: The Queen will be 80 on 21 April and she is celebrating her birthday with a family dinner hosted by Prince Charles at Windsor Castle.1 Human mediated tasks. Imagine a person wishing to annotate this news snippet and in search of an ontology containing the Queen, birthday and dinner concepts. When queried for these terms, a selection mechanism is expected to return an ontology that best covers them. It is not a problem if the returned ontology contains only a subset of the terms (partial coverage) as the user can extend the ontology according to his needs. It is also admissible for the system to make mistakes when mapping between the query terms and ontology concepts as the user can filter out such errors (imprecise coverage). For example, ontologies containing the concept Queen as a subclass of Bee or dinner_fork as an approximation for dinner will be rejected as irrelevant for this user’s context. Finally, users are willing to wait some minutes for reusable ontologies, since this time is negligible compared to that needed to build an ontology from scratch. Automatic knowledge reuse. As opposed to the previous scenario, imagine that the output of the selection is automatically processed. For example, a semantic browser such as Magpie [DDM03] which identifies and highlights entities of a certain type in Web pages needs to find an ontology according to which to describe the page above. The requirements are much stricter than before. First, a complete coverage of the query terms is needed to fully support the sense making activity offered by the browser. If no completely covering ontology is found, a set of ontologies that jointly cover the query could be returned. Or, alternatively, an ontology with more generic concepts such as woman, event and meal could be useful, provided that a machine interpretable explanation of the relation between the query terms and the returned concepts is available (e.g., a dinner is a kind of meal). Indeed, another requirement relates to the quality of mappings between terms and concepts. Errors such as those described in the context of human mediated tasks are not admissible. Finally, a quick response becomes more important when the selection is used at run time as in this case. Existing selection approaches function well in the context of human mediated tasks, but are insufficient when it comes to automatic knowledge reuse. Regarding the level of coverage, none of the existing approaches 1

http://news.billinge.com/1/hi/entertainment/4820796.stm


Page 121 of 151

enforces complete coverage. Further, the quality of the mapping between query terms and concept labels is quite low, as all these approaches rely only on syntactic matches. For example, ActiveRank, currently the most advanced algorithm, uses a fuzzy match between terms and concept names (i.e., project is mapped to projectile) but makes no provision to filter out obviously irrelevant hits. The meaning of the concepts given by their position in the hierarchy is not considered by any of the approaches. Finally, our only indication about performance is that ActiveRank needs 2 minutes to evaluate each ontology - a baseline that needs improvement in the case of automated tasks. This definition of the ontology selection task illustrates our perspective that evaluation is core to selection. At the same time it is clear that this evaluation is influenced by the characteristics of the underlying libraries and the requirements of the scenarios that use selection (these influence the way the information need is expressed, the selection criteria to be used and the way output should be provided). In this deliverable we report on ontology selection from online ontology libraries. We discuss some characteristics of these online libraries in Section 11.4. We also discuss typical characteristics of automated selection scenarios in Section 11.3.3.

11.2

Current approaches to ontology selection

Several approaches have already been proposed for the problem of selecting (ranking) ontologies. A distinguishing feature between these approaches is the selection criteria that they rely on (i.e., the kind of evaluation that is performed). Based on this feature, we identified three categories of approaches that select ontologies according to their popularity (Section 11.2.1), the richness of semantic data that is provided (Section 11.2.2) and topic coverage (Section 11.2.3).

11.2.1

Popularity

Approaches from this category select the “most popular" (i.e., well established) ontologies from an ontology collection. They rely on the assumption that ontologies that are referenced (i.e., imported, extended, instantiated) by many ontologies are the most popular (a higher weight is given to ontologies that themselves are referenced by other popular ontologies). These approaches rely on metrics that take into account solely the links between different ontologies. In fact, these approaches use the same principle as current Web search engines (the importance of a Web page is proportional to the number of pages that reference it) and they often use a modified version of the PageRank algorithm. To our knowledge there are three approaches that consider ontology popularity. OntoKhoj [PSLP03] is an ontology portal that crawls, classifies, ranks and searches ontologies. For ranking they use the OntoRank algorithm which is in spirit similar to the PageRank algorithm but instead of relying on HTML links it considers the semantic links between ontologies. Semantic links are denoted by instantiation and subsumption. Swoogle [DPF+ 05] is a search engine that crawls and indexes online semantic Web documents. Swoogle allows querying its large base of semantic data and provides also some metrics for ranking ontologies. They rely on a similar principle as OntoRank and use a PageRank-like algorithm on semantic relations between ontologies (i.e., using terms of an ontology to define new terms, populating ontology terms, importing ontologies). OntoSelect [BED04] is one of the first comprehensive ontology libraries that offers a complex ontology selection algorithm relying, among others, on selecting the most “well established" ontologies. The authors name this as the “connectedness" criteria since they look at how well an ontology is connected to other ontologies in order to determine its popularity. Unlike Swoogle and OntoKhoj, they use a less complex metric which only considers ontology imports as denoting semantic links between ontologies.


Page 122 of 151

11.2.2


Richness of Knowledge

Another way to rank ontologies is to estimate the richness of knowledge that they express. When approximating this aspect, most approaches investigate the structure of the ontology. The ActiveRank [AB05] algorithm is the only selection algorithm that has been developed independently from an ontology library. ActiveRank combines a set of ontology structure based metrics when ranking ontologies. To determine the richness of the conceptualization offered by the ontology they use the Density Measure(DEM) metric. This measure indicates how well a given concept is defined in the ontology by summing up the number of its subclasses, superclasses, siblings, instances and relations. ActiveRank introduces two other measures that rely on the ontology structure and aim to evaluate the quality of the ontological knowledge. First, the Centrality Measure(CEM) metric relies on the observation that concepts which are in the “middle" of the ontology are the most representative and have the right level of generality. CEM is computed by taking into account the longest path from the root through the branch that contains a concept C to its node and the path from the root to the concept C. Second, the Semantic Similarity Measure (SSM) measures how close the concepts that correspond to the query are placed in the ontology by relying on the links between these concepts. The assumption is that an ontology that contains all queried concepts close enough to be treated as a module is better than an ontology in which these concepts are spread in different parts of the hierarchy. In OntoSelect a similar metric, called Structure, is used. The value of the Structure measure is simply the number of properties relative to the number of classes in the ontology. The rationale behind this metric is that “more advanced ontologies have a large number of properties".

11.2.3

Topic Coverage

Finally, ontologies can be ranked based on the level to which they cover a certain topic. To determine this, most approaches consider the labels of ontology concepts and compare them to a set of query terms that represent the domain. The Class Match Measure (CMM) of ActiveRank denotes how well an ontology covers a set of query terms. It is computed as the number of concepts in each ontology whose label either exactly or partially matches the query terms. Note that the matching is purely syntactic and no attention is paid to discovering synonyms or indeed to make sure that the concept is used in the same sense as intended by the query term. The OntoSelect algorithm allows to specify the information need by supplying a whole corpus. The concepts that are the most relevant for a corpus are determined by statistical processing of the corpus. Then, coverage is measured by comparing the number of concept/property labels of the ontology with the query terms extracted from the corpus. This selection algorithm relies on the evaluation approach proposed in [BADW04]. In OntoKhoj they have considered word senses when ranking ontologies to cover a topic. In their algorithm they accommodate a manual sense disambiguation process, then, according to the sense chosen by the user, hypernyms and synonyms are selected from WordNet. The algorithm first tries to determine ontologies that contain the supplied keyword. If no matches are found, the algorithm queries for the synonyms of the term and then for its hypernyms. The algorithm was designed for a single word and it does not take into account relations. Swoogle also offers a limited search facility that can be interpreted as topic coverage. Given a search keyword Swoogle can retrieve ontologies that contain a concept (or a relation) matching the given keyword. The matches are lexical and one can select between different levels of matches (exact, when the keyword matches exactly the concept label, prefix, when the keyword appears at the beginning of the concept label; suffix, when the keyword appears at the end of the concept label and fuzzy, when the keyword appears at any position in the concept label). While still under development, the ontology selection algorithm which is part of the PowerAqua [LMU06] question answering tool should be mentioned here. This algorithm aims to find the ontologies that cover a set of triples derived from a question. The minimum requirement is that any of the triples submitted as a query


Information need Selection Criteria * Popularity * Knowledge Richness * Topic Coverage Ontology Library Output

OntoKhoj [PSLP03] one keyword Yes No Yes OntoKhoj

Swoogle [DPF+ 05] one keyword

OntoSelect [BED04] corpus

Yes Yes No Yes Yes Yes Swoogle OntoSelect ranked list of ontologies

Page 123 of 151

ActiveRank [AB05] set of keywords

*PowerAqua [LMU06] triples

No Yes Yes Swoogle

No No Yes Any combinations of ontologies

Table 11.1: Comparison of existing ontology selection approaches. * PowerAqua is still under development. should be completely covered by an ontology. If elements of a triple are discovered in distinct ontologies than the triple is broken down in two more specific triples and the selection is reiterated. The output can contain more than one ontologies if different triples are covered by different ontologies. The selection itself is more semantic than existing approaches because it relies on WordNet senses, it checks for coverage of relations as well as concepts and considers the position of concepts within an ontology hierarchy to perform the selection.

11.2.4

Summary of Selection Approaches

To summarize Section 11.2, we provide a comparative overview of the ontology selection methods described above in Table 11.1. Our first obvious observation is that all the existing methods (except PowerAqua) rely on rather simple ways to specify an information need (a keyword, a set of keywords or a corpus from which a set of keywords are distilled) and use the same format to provide the output (i.e., a list of ranked ontologies). We also note that all approaches are designed for large-scale, automatically built ontology libraries. Several interesting conclusions can be drawn with respect to the selection criteria used by these approaches. It is interesting to see that all methods offer some functionality for estimating topic coverage. This functionality is complemented with support for other selection criteria such as popularity or knowledge richness. In our discussion we have seen that there is a correlation between the selection criteria and the aspect of ontology that is evaluated. Namely, popularity based evaluation takes into considerations the links between ontologies, when knowledge richness is estimated then the structure of the ontologies is considered. Finally, topic based approaches consider the labels of the ontology elements when comparing them to the query terms. It is surprising that none of these approaches take advantage of ontological knowledge but rather treat ontologies as interconnected objects, graphs or bags of labels.

11.3 Requirements for Supporting Automatic Knowledge Reuse Scenarios While current ontology selection tools primarily target human users, we are working on two Semantic Web tools (Sections 11.3.1 and 11.3.2) that are evolving from using a single, rich and manually crafted ontology to exploring and combining ontologies available on the Web. These tools rely on automatic ontology selection on which they pose a set of requirements (Section 11.3.3). Awareness of such practical scenarios is essential to understand and develop ontology selection.

11.3.1

Ontology based Question Answering

AquaLog [LPM05] is an ontology based question answering system which relies on the knowledge encoded in an underlying ontology to disambiguate the meaning of questions asked using natural language and to


Page 124 of 151


provide answers. To shortly give an impression about how the system operates, consider that it is aware of an ontology about academic life2 which has been populated to describe KMi related knowledge3 . Also, suppose that the following question is asked4 : Which projects are related to researchers working with ontologies? In a first stage the system interprets the natural language question and translates it to triple-like data structures. Then, these triples are compared to the underlying ontology centered knowledge base using a set of string comparison methods and WordNet. For example, the term projects is identified to refer to the ontology concept Project and ontologies is assumed equivalent to the ontologies instance of the Research-Area concept. After the modifier attachment is resolved by using domain knowledge, two triples are identified: (projects, related to, researchers) and (researchers, working, ontologies) The relations of the triples are also mapped to the ontology. For example, for the second triple, there is only one known relation in the ontology between a Researcher and a Research-area, namely has-researchinterest. This relation is assumed to be the relevant one for the question. However, when disambiguating the relation that is referred to by related to, the system cannot find any syntactically similar relation between a Project and a Researcher (or between all more generic and more specific classes of the two concepts). Nevertheless, there are four, alternative relations between these two concepts: has-contact-person, hasproject-member, has-project-leader, uses-resource. The user is asked to choose the relation that is closest to his interest. Once a choice is made, the question is entirely mapped to the underlying ontological structure and the corresponding instances can be retrieved as an answer. While the current version of AquaLog is portable from one domain to another, the scope of the system is limited by the amount of knowledge encoded in the ontology used at that time. The new implementation of AquaLog, PowerAqua [LMU06], overcomes this limitation by extending the system in the direction of open question answering, i.e., allowing it to benefit from and combine knowledge from the wide range of ontologies that exist on the Web. One of the challenges is the selection of the right ontology for a given query from the Web.

11.3.2

Semantic Browsing

The goal of semantic browsing is to exploit the richness of semantic information in order to facilitate Web browsing. The Magpie [DDM03] Semantic Web browser provides new mechanisms for browsing and making sense of information on the Semantic Web. This tool makes use of the semantic annotation associated with a Web page to help the user get a quicker and better understanding of the information on that page. Magpie is portable from one domain to another as it allows the user to choose the appropriate ontology from a list of ontologies that are known to the tool. However, similarly to AquaLog, the current version relies on a single ontology being active at any moment in time. This limits the scope of the sense making support to the content of the current ontology. Our current research focuses on extending Magpie towards open browsing. This means that the tool should be able to bring to the user the appropriate semantic information relevant for his browsing context from any ontology on the Web. This extension relies on a component that can select, at run time, the appropriate ontologies for the given browsing context. In the case of Magpie, the query for the ontology selection is more complex than for AquaLog as it is defined by the current browsing context. This includes the content of the currently accessed Web pages and, optionally, the browsing history and the profile of the user. Web pages typically span several different topics. For example, the following short news story5 is both about trips to exotic locations and talks. Therefore, the query sent to the selection mechanism is likely to contain terms drawn from different domains. 2

http://kmi.open.ac.uk/projects/akt/ref-onto/. See the populated ontology at http://semanticweb.kmi.open.ac.uk. 4 See the AquaLog demo at http://kmi.open.ac.uk/technologies/aqualog/.

3

5

http://stadium.open.ac.uk/stadia/preview.php?s=29&whichevent=657


Page 125 of 151

“For April and May 2005, adventurer Lorenzo Gariano was part of a ten-man collaborative expedition between 7summits.com and the 7summits club from Russia, led by Alex Abramov and Harry Kikstra, to the North Face of Everest. This evening he will present a talk on his experiences, together with some of the fantastic photos he took."

11.3.3

Requirements for Ontology Selection

Hereby we formulate the requirements imposed by our applications on ontology selection and discuss to which extent they are addressed by current approaches. These requirements drove the design of our selection algorithm (Section 11.5). 1. Complete coverage. A complete coverage is probably the most important requirement for our applications (though it might not be so important for other tools). Because in these applications the retrieved knowledge is automatically processed, they require that all the needed knowledge should be retrieved. While existing approaches rank ontologies that cover most terms as best, they do not enforce a complete coverage. 2. Precise coverage. Automatic knowledge reuse requires a rigorous mapping between query terms and ontology concepts as well as a formal representation of the mapping relation (e.g., more generic). Assuming that a human user would filter out (and eventually enrich) the returned ontologies, current tools treat the comparison between query terms and ontology concepts rather superficially, relying only on (often approximate) lexical comparisons. 3. Returning ontology combinations. Our preliminary experiments indicate that the sparseness of knowledge on the Web often makes it impossible to find a single ontology that covers all terms (Section 11.4). However, it is more likely to find ontology combinations that jointly cover the query terms. Existing tools return lists of single ontologies rather than their collections. 4. Performance. Our applications rely on the results of selection at run time and therefore require a good performance. While simple selection tools perform rather well, the more complex ActiveRank needs 2 minutes per ontology to compute all its metrics. This is acceptable for supporting ontology building, but needs to be improved in an automatic scenario. 5. Dealing with relations. Our applications, especially PowerAqua, illustrate a need for considering relations and not just concepts when selecting an ontology. Currently, only OntoSelect considers relations. 6. Dealing with instances. Our applications help users in their information gathering activities. Most often, people are interested in finding out things about certain entities rather than generic concepts. This requires that selection should consider instances as well (i.e., match between instances in a query and those in online ontologies). Matching instances is a difficult problem in itself given the large number and high level of ambiguity when dealing with instances (e.g., many instances can share the same or similar names). 7. Modularization. Knowledge reuse is closely related to ontology modularization. Indeed, our tools would require selection mechanisms to return a relevant ontology module rather than a whole ontology. Note that the work in [AB05] has already considered this issue when introducing a metric to measure how close the hits in an ontology are (assuming that this indicates the existence of a module). As with instance mappings, ontology modularization is a difficult and as yet unsolved issue, though a large amount of work in this area [Spa05] could be reused to some extent. In Section 11.6 we describe our approach to combine ontology selection with ontology modularization techniques.


Page 126 of 151

11.4


Preliminary Experiments

To better design our algorithm, we wanted to get an insight in the characteristics of the ontological data available online. Since the requirement of complete and precise coverage of the query terms was identified as the most important one in the context of automatic knowledge reuse, our experiments are centered towards 1) exploring the factors that hamper obtaining a complete coverage and 2) getting an insight in the nature of compound concept labels in preparation to provide a more precise mapping to query terms. We performed both experiments on top of Swoogle6 because it is currently the largest ontology repository. It is important to note that our experiments have an exploratory role rather than trying to rigourously test our hypotheses.

11.4.1

Experiment 1 - Obtaining Complete Coverage

The goal of this experiment is to get an indication about how difficult it is to find a completely covering ontology when using Swoogle. One of the motivations for this experiment was that, while important, complete coverage has not been investigated in any previous work (although best covering ontologies are rated best). In fact, with the exception of OntoSelect, all selection algorithms are tested for the rather trivial case of one or two query terms. On the contrary, our tools require ontologies that cover at least three query terms (e.g., AquaLog translates each question in one or more triples). Our intuition was that the number, topic relatedness and type of the query terms will influence the success of obtaining an all covering ontology. Namely, a single, all covering ontology is difficult to find if 1) there are many query terms, 2) if query terms are drawn from different topic domains or 3) relations are considered. According to these considerations, we devised four sets of queries. The first three queries represent an optimal scenario where few concepts are drawn from the same domain (we chose a well covered domain in terms of online ontologies, the academic domain). The second set of queries (4 - 6) have terms drawn from different (and less covered) topic domains. They were inspired by the actual text snippets in Section 11.1 and Section 11.3.2, therefore being representative for real life scenarios encountered with Magpie. The queries in set three (7 - 10) have terms drawn from the same domain but, unlike the first set, contain a relation as well (these are typical AquaLog queries). Our final queries (11 - 14) explore overcoming failure of finding a completely covering ontology by replacing query terms in queries 4, 6, 9 and 10 with their hypernyms. The experimental software queries Swoogle for ontologies that contain concepts/properties that exactly match the query terms (variations in capitalization are allowed)7 . For each query, the software outputs the number of ontologies that cover each term, their pairwise combinations and all terms. The results are summarized in Table 11.2. Notice that as the number of terms increases less completely covering ontologies are found. The drop in the number of returned ontologies is significant when adding even one extra term. This phenomena is evident throughout the table even in our optimal scenario where terms were chosen from the same, well covered domain. Our second set of queries containing terms drawn from different topic domains return less ontologies than previously (mostly zero). At a closer look, however, one might argue that the null results are caused by the fact that the domains from which the terms were drawn are weakly covered in Swoogle in the first place (indicated by the low number of ontologies returned for individual terms). While this observation does not necessarily undermine the intuition that topic heterogeneity has negative effects, it indicates that the knowledge currently available online is sparse, as many domains are weakly covered (or not at all). Therefore, null results can be expected even when query terms are topically related but refer to a weakly covered topic. The third set of experiments indicates that the presence of relations seriously hampers retrieving an all covering ontology even when the query terms are chosen from the same, well represented domain. In the last four queries, by refining query terms through hypernym replacement, better results were obtained. An obvious worry is that if the refinement uses too generic terms (e.g., Entity) the returned ontologies will be 6

We use Swoogle 2005 as our software was written before Swoogle 2006 was released. Exact matching is an extreme case (e.g., hasAuthor, authorOf, authored all mean the same thing) and as it will be evident from the results, it is too limiting. 7


Page 127 of 151

Query 1 2 3

(t1 , t2 , t3 )

(t1 )

(t2 )

(t3 )

(t1 , t2 )

(t1 , t3 )

(t2 , t3 )

(t1 , t2 , t3 )

(project, article, researcher) (researcher, student, university) (research, publication, author)

84 24 15

90 101 77

24 64 138

19 16 8

13 15 5

9 38 36

8 13 4

4 5 6

(adventurer, expedition, photo) (mountain, team, talk) (queen, birthday, dinner)

1 12 0

0 25 9

32 9 2

0 2 0

1 1 0

0 1 1

0 1 0

7 8 9 10

(project, relatedTo, researcher) (researcher, worksWith, Ontology) (academic, memberOf, project) (article, hasAuthor, person)

84 24 21 90

11 9 36 14

24 52 84 371

0 0 0 8

13 3 3 32

0 0 5 2

0 0 0 0

(person, trip, photo) (woman, birthday, dinner) (person, memberOf, project) (publication, hasAuthor, person)

371 32 371 77

7 9 36 14

32 2 84 371

1 1 16 2

20 1 46 52

1 1 5 2

1 1 5 2

11 (4+) 12 (6+) 13 (9+) 14 (10+)

Table 11.2: Number of ontologies retrieved for a set of queries. (X+ refines X.)

too generic to be of any use for the concrete knowledge reuse task at hand. While only preliminary, our experiments do indicate that query size, topic heterogeneity and type might influence the chance to find an all covering ontology. They have also revealed the sparseness of the online knowledge. As a bottom line, independently of having verified our intuitions, we can observe that the chance to find an all covering ontology is rather low, especially in scenarios such as those provided by Magpie (many terms, drawn from different, possibly weakly represented domains) and AquaLog (properties as query terms).

11.4.2

Experiment 2 - Dealing with Compound Labels

Considering the results of the previous experiment, some mechanisms might be needed to expand the search for potentially relevant ontologies. Besides the synonym/hypernym extension, the more lexical oriented strategy of selecting concepts whose labels partially match the query terms can be explored. For example, Swoogle’s fuzzy search functionality returns concept labels that contain the query term as a substring. This mechanism is rather brittle, and, while it returns several important hits (e.g., GraduateStudent when searching for Student), it also generates clearly invalid hits (e.g., update when searching for date). To ensure our second requirement referring to precise coverage, all the compound labels returned by fuzzy search need to be interpreted in order to understand their relation with the query term. A special case of compound labels are those containing conjunctions (e.g., BlackAndWhite). Some researchers have proposed a set of rules to interpret such labels [MSS03]. Naturally, reading, splitting and interpreting all these labels can seriously hamper the time performance, thus questioning the usefulness of performing a fuzzy search at all. In this experiment we explore the feasibility of performing fuzzy search. We illustrate some cases when it pays off and some when it does not. We also evaluate how frequently conjunctions are used in compound labels. To support our experiments we implemented a program (LabelSplitter) that splits compound labels according to the most common naming conventions and checks if a given term is a well formed part of that label (i.e., its base form is the same as the base form of one of the components of the label). For example, TeachingCourse is a relevant compound label (CL) for the term teach, but an irrelevant one for the term tea. In Table 11.3 we summarize the results obtained when querying some random terms and then some conjunctions showing the total number of hits returned by Swoogle, which is broken down into the number of exact matches, relevant and irrelevant CLs.


Page 128 of 151


Word project (clarifying example) project student tea mountain university and or of not except but not

Total -

Exact project

644 190 492 36 86 2363 18178 6353 840 45 0

90 84 3 12 64 37 11 4 23 0 0

Relevant CLs PastProject ProjectPartner 413 97 23 21 22 444 184 4743 77 0 0

Irrelevant CLs Projectile Projector 141 9 466 4 0 1882 17983 1606 740 45 0

Table 11.3: Analysis of the appearances of some conjunctions and other terms. As expected, fuzzy search is a good mechanism to broaden the search space as it can return a lot of broader hits that contain the term. In general, in the case of longer words (less likely to be substrings of other words) more relevant than irrelevant compound labels are found. This is not true in the case of shorter words such as tea where an overwhelming number of irrelevant hits are returned. Therefore, taking into account that fuzzy search is rather expensive, it should be used only when all other alternatives fail. Regarding the frequency of conjunctions, in current online ontologies “or" appears the most frequently but in the large majority of cases as a substring and not a well formed part. While the “of" conjunction appears less often than “or" it is the most frequently used as a proper part of the compounds (mostly as part of property labels). “And" appears quite frequently as well in its role of well formed part (444). Surprisingly, negation and disjunction indicators appear infrequently or at all in the collection that we have queried. We conclude that interpretation rules for some conjunctions have to be written.

11.5 The Algorithm In this section we present the design of an algorithm which aims to address some of the requirements stated in Section 11.3.3 and also draws on our conclusions regarding the nature of online ontologies detailed in the previous section. We first give an overview of the method in which we motivate our main design choices and then explore each major step of the algorithm in detail. The algorithm has been entirely specified and partially implemented (with the exception of the ideas reported in Sections 11.5.4 and 11.5.6) .

11.5.1

Overview

For our first implementation we wish to satisfy the first five requirements: we aim to identify ontologies (or combinations of ontologies - R3) that completely and precisely cover our query (R1 and R2). The query can contain both concepts and relations (R5). The performance of the algorithm should be such that it can be used by other applications at run time (R4). The final two requirements, related to instances and modularization, are not addressed yet. From our experiments we have learned that in cases when query terms are drawn from different domains or when they represent relations it is challenging to find ontologies that would cover all terms (therefore R1 is not so easy to fulfill). We have also seen that in such cases the search space can be expanded either 1) by query expansion with semantically related terms or 2) by searching for labels that incorporate the query term. However, our second experiment indicates that fuzzy search should be used only when absolutely needed.


Page 129 of 151

Given these considerations, we have designed an algorithm that adapts itself to the particular context and can employ increasingly complex methods in order to achieve a complete coverage. The algorithm in Figure 11.1 executes increasingly complex combinations of a couple of main steps until complete coverage is achieved. We will first explain the role of each step and then describe how they are combined in increasingly complex stages.

Figure 11.1: The main tasks and stages of the selection algorithm.

Step1: Query Expansion. This step supplements the query terms with their semantically related terms such as synonyms and hypernyms. The reason for this is that we aim for a semantic rather than a lexical match to our query terms. Also, the chosen query terms might not be the most representative for a certain concept and this might lead to empty result sets. Step2: Ontology identification. In this step we identify ontologies that cover to some extent the query terms. After an initial syntactic mapping between query terms (either exact or fuzzy) and ontology concepts, we perform a more in depth analysis of these mappings and define their semantic type (i.e., exact, generic or more specific). We call this task semantic match. Step3: Identify ontology combinations. Using the output of the previous step, here we decide on the ontology combinations that provide a complete coverage of the query terms. Step4: Generality Ranking. The ontologies that are returned contain hits that can be more generic or more specific than the query terms. In this step we evaluate the ontology combinations according to their level of generality and choose those with the appropriate level of abstraction.


Page 130 of 151


These basic steps are combined in the following increasingly complex and expensive stages. The algorithm enters in a new stage only if the previous stage has failed: Stage I relies on the simplest combination of the main steps. It uses an exact match to identify relevant ontologies thus circumventing complex semantic matching and the generality ranking step. This stage is likely to succeed only if the query terms are few or drawn from the same, well covered domain. Stage II is used only if Stage I fails (no ontology was found for at least one term) and some kind of broadening of the search space is needed. Query expansion is used for the problematic terms and then the same ontology identification and combination steps as in stage I are performed. Notice that at this stage we can already use the generality ranking step because query broadening is likely to identify hypernyms for some of the query terms. Stage III is the most complex one, as besides query expansion, it also relies on more flexible syntactic matching to identify even more concepts potentially related to the query terms. This fuzzy match complicates the semantic matching step as the retrieved compound labels need to be split and interpreted. After the semantic match has identified the semantic relations between query terms and ontology concepts we apply the ontology combination and generality ranking steps.

11.5.2

Step1: Query Expansion

Query expansion is needed in order to broaden the search space for ontologies in cases when no or few ontologies are returned for a term. Our experiments indicate that such cases will be often encountered given the knowledge sparseness of online ontology collections. Term expansion allows searching not just for the term but for all its semantically related terms (e.g., synonyms, hypernyms). This can be allowed because we aim to perform a semantic rather than a syntactic selection and therefore synonyms that denote the same concept as the query term are relevant. Currently, we use WordNet to augment each query term with their synonyms and hypernyms. The only system that uses a similar expansion approach is OntoKhoj [PSLP03].

11.5.3

Step2: Ontology Identification

In this step we identify ontologies that contain the concepts specified in our query. This is in essence a mapping stage between the query terms and the concepts of the ontologies. We distinguish two substages: Step 2.1. Syntactic Match. The syntactic match identifies lexically similar concept labels. It can be either exact (the query term is exactly the same as the concept label) or fuzzy (the query term is a substring of the concept label, e.g., the term Student is part of GraduateStudent). In the case when a fuzzy match is performed, this step is also responsible for splitting the compound labels and returning only the compound labels that are relevant for the given term (as done by the LabelSplitter module described in Section 11.4.2). Current ontology selection techniques only use syntactic matches when identifying relevant ontologies. Step 2.2 Semantic Match. Semantic matching goes beyond the state of the art in ontology selection as it checks the soundness and the semantic nature of the previously identified syntactic mappings. Concretely, the input to this step is a term and a concept in an ontology that is lexically related to the term. The task is to find out the semantic relation between the term and the concept. This can be equivalence, more specific or more general. An obviously relevant body of work is that on mapping techniques. However, according to a recent survey of mapping techniques [SE05] most matchers return a probability coefficient to describe the significance of a mapping rather than its semantic interpretation. A notable exception is the S-Match algorithm which returns the semantic category of each mapping in terms of (among others) the exact, more generic or more specific operators [GSY04]. Following the general model of the S-Match algorithm, we distinguish two steps to obtain a semantic matching:


Page 131 of 151

A. Acquiring the sense of the concept label also taking into account its position in the hierarchy (i.e., parent and children nodes). B. Deriving the semantic relations between the term and the concept.

A. Acquiring the sense of the concept label. We use information about the position of a concept in the ontology to determine its sense according to a method originally presented in [MSS03]. In a nutshell, given a concept c and either one of its ancestors or descendants r all WordNet synsets for both labels are retrieved. Then, if any of the senses for c is related to any of the senses of r either by being a synonym, hypernym, holonym, hyponym or a meronym, then that sense of c is considered the right one. For example, if Apple (which can have two senses: tree and fruit) has Food as its ancestor, then there exists a hyponym relation between apple#1 (fruit) and food#1, so we retain this sense and discard the one referring to apple being a tree.

B. Deriving Semantic Relations. After identifying the sense of the concept, we derive semantic relations between the terms and the concepts such as equivalence, more generic or more specific. We use a WordNet based comparison between the senses of the term and that of the concept label. Therefore, equivalence is established when two terms share a synset, and more general/more specific relations are indicated when hyponym/holonym (or even meronym/holonym) relations exist between their senses. In cases when none of these relations hold we investigate whether there is any similarity at all between the terms (and return a weaker “related" relationship). For this we investigate whether there exists an allowable “is-a" path in WordNet connecting their synsets by relying on the depth and common parent index (C.P.I) measures described in [LPM05]. Matching relations. Our previous experience in AquaLog [LPM05] was that mapping relations is more difficult than mapping concepts. One of the reasons is that many relations are vaguely defined (a classical example is relatedTo which can have a variety of meanings) and therefore can have a large number, hard to automatically predict lexical variations. Also, the meaning of a relation is given by the type of its domain and its range so the precondition of a mapping between two relations is that their domain and range classes match to some extent. Inspired by our previous work [LPM05], we treat relations as “second class citizens" and concentrate on finding matches for the classes that denote their domain and range first. Then, if only one relation exists between these classes we adopt it as such. If more relations exist we attempt a lexical based disambiguation of the one that is closest to the relation that we seek. An interesting case is when some relations are present in other ontologies as concepts (e.g., hasAuthor can be modeled as a concept Author in another ontology). This case is also explored.

11.5.4

Step3: Identifying relevant ontology combinations.

Ideally, one would expect that the selection mechanism finds a single ontology which contains all the query terms. However, in practice this is seldom the case. Most often query terms are spread over two or more ontologies. Unfortunately, previous approaches provide a set of ontologies ranked by the coverage of each individual ontology. Our task therefore is to identify the best combinations of ontologies that cover the query. There are two criteria to rank ontology combinations. On one hand, the number of ontologies should be minimal. On the other hand, the number of terms that they cover should be maximal. The ultimate best is one ontology covering all terms, and the worst is a collection of ontologies each covering a single term. We are currently working on an optimal implementation of this multiple criteria optimization problem.


Page 132 of 151

11.5.5


Step4: Generality Ranking

Due to our semantic matching, the returned concepts can be more generic or more specific than the query terms. In this step we identify the ontology combinations that are closest in terms of abstraction level to the query terms. We are not aware of any work that directly addresses the issue of measuring the generality of an ontology or evaluating the generality of an ontology with respect to a set of terms. A recent publication investigates evaluating the generality of a document with respect to a query [YLS06]. After concluding that most of the generality related work in the field of IR is based on statistical measures, they propose a method to compute the generality of a document with respect to a domain ontology (in that case, Mesh) by relying on the depth and proximity of the concepts in the domain ontology (i.e., the deeper and closer the concepts are in the ontology the more specific the document/query is). Generality is computed both for the query and the document and then the obtained scores are compared. The major drawbacks of this approach are that (1) it is time consuming because all terms need to be looked up in the oracle and their positions have to be computed and (2) it depends on the coverage of the used oracle. We agree with [YLS06] that generality computation should be based on the meaning of the terms rather than on statistical measures. Instead of computing generality both for the query and an ontology and then comparing them, we assume that the query provides the baseline and we only compute the generality deviation of the ontology from this baseline. Another optimization is that we circumvent the use of an external oracle by reusing the generality relation between terms and concepts as established by the semantic mapping step (we consider a function genRel between a term and its hit returning -1 if the concept is more specific, 0 if it is equivalent and 1 if it is more generic than the query term).

RD(T, O) =

P

n i=1

|genRel(ti ,ci )| ; GS(T, O) n

P = σ( ni=1 genRel(ti , ci ))

Given a set of n query terms (t1,n ) and their semantically related concepts (c1,n ) we compute the relative generality (RD(T, O)) of the ontology/ontologies containing these concepts with respect to the query as the mean of the absolute value of the genRel function. We also compute the sign of the generality deviation as the sign of the sum of all the values of the genRel function.

11.5.6

Extending Semantic Match to Deal with Compound Labels

Compound labels derived in Stage III complicate semantic matching. Hereby we describe some of the problems and the solutions that we are investigating. A. Acquiring the sense of a compound concept label. Establishing the sense of compound labels by using WordNet is difficult as WordNet does not have an extensive coverage of compound words. We are currently investigating the strategy of interpreting the meaning of compound labels in terms of logical relations that hold between the senses of their constituents (similarly to work in [MSS03] and [GSY04]). According to this previous work, compound labels can be interpreted as the conjunction of their constituents and according to these rules: Rule1. Commas and coordinate conjunctions are interpreted as a disjunction; Rule2. Prepositions like in and of are interpreted as a conjunction; Rule3. Exclusion expressions (e.g., except, but not) translate into negation. However, we are not convinced that all these rules are useful in the context of online ontologies. For example, only five labels returned by Swoogle contain commas, so this is just an isolated case. Also, we found that no labels contain “except" and “but not", thus making the third rule redundant.


Page 133 of 151

B. Deriving Semantic Relations between compound terms. The limited multi-word coverage of WordNet also prohibits using it to derive semantic relations between compound labels. We investigate a solution along the lines of that presented in [GSY04] where compound labels, after being interpreted as logical formulas, are compared with the help of a reasoner.

11.6

From Ontology Selection to Knowledge Selection

Selection algorithms tend to run into two major problems. First, if the selection returns a large ontology this is virtually useless for a tool such as Magpie which only visualises a relatively small number of concepts at a time. Unfortunately, in the experiments we have performed large ontologies are often returned (especially WordNet). What is needed instead is that the selection process returns a part (module) of the ontology that defines the relevant set of terms. A second problem relates to the fact that in many cases it is difficult to find a single ontology that covers all terms (we observed this knowledge sparseness phenomenon in Section 11.4 and our current systematic experiments in this direction confirm our initial intuitions). However, a combination of one or more ontologies could cover all the query terms. This problem is related to modularization in the sense that it is easier to combine small and focused knowledge modules than ontologies of large size and coverage.

Figure 11.2: The knowledge selection process and its use for semantic browsing with Magpie. These considerations justify the need to extend selection techniques with modularization capabilities. In Figure 11.2 we depict the three major generic steps of the knowledge selection process that integrates ontology selection, modularization and merging. We envisage three major steps: 1. Selection of relevant ontologies. In a first step, given a set of terms for which an ontology is required, the selection technique identifies online ontologies or sets of ontologies that cover the given query terms. By coverage we mean that the identified ontologies contain concepts, properties or instances that are semantically related to the query terms. We use the selection algorithm described above. 2. Modularization of the selected ontologies. Given the ontologies discovered in the previous step, a modularization technique is applied on each ontology to identify a module in the ontology that contains relevant knowledge for the query terms that were mapped into that ontology. A modularization technique has been designed and implemented as described in [dSM06]. 3. Merging of the relevant ontology modules. Finally, in the case when the query terms are covered by several different ontologies, the separate modules extracted from each individual ontology need to be merged in a meaningful way to form the final result of the selection. The merging algorithm will be built after the selection and modularization parts are finalized.


Page 134 of 151


Figure 11.3: A Prototype implementation of the selection and modularization methods. The first two steps of this knowledge selection algorithm have been implemented and integrated in a prototype tool. A screenshot of this tool is shown in Figure 11.3. The user can enter the terms for which he wishes to retrieve relevant knowledge in the text box designated to Search Terms. In the example, the user wishes to retrieve ontology modules that cover the terms: cat, dog, duck, old_lady. When OK is pressed, the tool selects all the ontologies that contain these terms, displaying the URI of each retrieved ontology and for each ontology the URI’s of the concepts that match each search term. Then, for each ontology, the user can visualize either the entire ontology (by selecting the “Ontology" button) or he can extract and visualize the ontology module that is relevant for the search terms (by selecting the “Module" button). Figure 11.4 shows the results of both visualizations for the first ontology. On the left hand side (part a), the entire ontology is visualized. Even if this ontology is a medium-sized ontology, visualizing the entire hierarchy is not very useful as the searched terms cannot be identified. On the right hand side of the figure (part b), we only visualize the ontology module that is relevant for the search terms. Obviously, this visualization is much easier to understand by a human user. Also, if the output of this knowledge selection process is directly used by a software tool, we believe that it will be much easier to integrate the smaller bit of knowledge tailored in such a way that it contains the complete semantic description of the query terms (all the knowledge that is relevant for these terms, see details of the algorithm in [dSM06]) rather than making sense of a much larger ontology which contains a lot of useless knowledge for the particular task of the tool.

11.7 Discussion and Future Work The key contribution of this work is that of exploring ontology selection in the context of automatic knowledge reuse. Indeed, as discussed in the introduction, this complements current selection techniques which have focused on human mediated tasks so far. While both contexts are equally important, we think that exploring the automatic context can lead to novel challenges and improvements of this technology. We have analyzed the requirements of two Semantic Web tools, a question answering tool and a semantic browser, and concluded that current approaches only marginally address them. This is not a surprise in itself because these requirements raise hard to address research issues. In fact, our proposed algorithm limits


(a)

Page 135 of 151

(b)

Figure 11.4: The original ontology (a) and the resulting module (b).

itself to tackle only five of the seven requirements. These requirements indicate that selection will need to adapt techniques from currently developing research directions such as ontology evaluation, mapping and modularization. Ontology mapping has been the focus of the proposed algorithm which balances between providing a complete, precise coverage and an acceptable performance. Our strategy is to use a self-adaptation metaphor, the algorithm adapts its complexity to the case of each query by invoking increasingly complex stages as necessary. As such, the simplest stage is just a bit more complicated than state of the art techniques, while the most complex stage raises yet unsolved research issues. The major difference from existing approaches is the emphasis on the correctness of the mapping between query terms and ontology concepts. We go beyond current techniques which exclusively rely on lexical matches by performing a semantic match. Naturally, establishing a semantic mapping at run-time without interpreting the entire ontology structure is a challenging issue by itself. Our work on combining ontology selection with ontology modularization techniques has proved promising and, as a result, we are continuing this line of work. One of the major efforts is put into better understanding existing modularization techniques by using them on data sets provided by case study partners (e.g., the AGROVOC thesaurus). While there are several state of the arts of existing techniques, no comprehensive study exists yet that actually runs all these techniques on real life data sets and compares them based on their results. By doing this, we will get a better understanding of which of these different techniques can be better combined with ontology selection. After gaining a better understanding of various modularization approaches, we wish to finalize the entire knowledge selection process by implementing techniques to merge multiple ontology modules resulting from the execution of the selection and modularization steps (see Figure 11.3). Another major line of work is centered in bringing the ontology selection work closer to ontology evaluation. In particular, we wish to integrate a variety of ontology selection techniques into the ontology selection technique. Note that our current algorithm focuses mostly on the coverage of the query terms but does not take into considerations any other ontology quality criteria. To overcome this limitation, our future work will focus in selecting a set of suitable evaluation metrics and combining them with the selection mechanism. Finally, our ultimate goal is to make the complete selection mechanism (extended with evaluation metrics, with ontology modularization techniques) available as a core part of the WATSON Semantic Web gateway [dSD+ 07].


Page 136 of 151


Part IV

Final Considerations


Page 137 of 151

Chapter 12

Future Work

During performing the work that is reported in this deliverable, the involved partners have drawn the following major conclusions which will influence the further evolution of this task. First, data re-engineering methods proved to be of crucial importance for our case study partners. Indeed, as it has been reported here, most of the re-engineering techniques have already been successfully applied to data provided by FAO. We believe that, generally speaking, such methods will be also important to other organizations that aim to adopt Semantic Web technology in order to semantically enrich their legacy data sources. Therefore, we should pay special attention to further developing the existing methods, to investigate novel methods, to test these methods on realistic data sets (provided by WP7 and WP8) and to provide these methods as part of the NeOn toolkit (in close collaboration with WP6) and methodology (in close collaboration with WP5). Given the importance of data re-engineering, it was decided that, from now on, a special subtask (T2.2.a) will be devoted to these methods. Second, ontology evaluation and selection methods are strongly related. We will continue to investigate them together in the context of task T2.2.b. In this deliverable we present an overview of both types of methods and make the first step towards combining them. A tighter integration between ontology evaluation and selection techniques will be the core topic of our next deliverable. Meanwhile, individual selection and evaluation techniques will be further developed also individually, evaluated on case study partner data sets and integrated in the NeOn toolkit and methodology (thus, working in strong collaboration with WP5, WP6, WP7 and WP8). Besides these generic considerations about future work, each of these lines of research also has a variety of more concrete research foci as we detail next.

12.1

Future Work in Data Re-engineering (T2.2.a)

Data re-engineering will focus on evolving existing methods and on investigating novel ones as follows: Learning support for semi-automatic ontology construction. Methods to support collaborative ontology design in the broadest sense are also machine learning and social network analysis methods. The idea is to analyse collaboration between the users in the networked ontology setting, where the user is working on ontology construction simultaneously with the other users and getting suggestions based on the actions of other users performed while constructing ontology on similar content/topic. In order to perform collaboration analysis, we assume that the data recording the users activities is available containing information on the provided suggestions and the user feedback. We will propose and implement methods based on machine learning and social network analysis and connect them to the OntoGen, an existing tool for semi-automatic ontology construction. Namely, inside the work performed in WP3 on providing context to the user based on the activity of the other users, OntoGen will be extended to support recording and exchange of information between different users on the system. In our case,


Page 138 of 151


mapping between the user activities and ontologies they are constructing is based on the assumption that each user has a collection of documents that s/he uses to construct ontology with support of OntoGen. Inside NeOn WP2 we will build on the top of that extensions to OntoGen. Upgrading database and XML content to semantic formats. The ODEMapster framework described in Chapter 4 is currently extended with a semi automatic mapping discovery tool and a graphical user interface for visualizing and writing R2 O mapping documents. In addition, intensive testing with other DBs as well as the development of tools, middleware, APIs, etc, to generate and exploit R2 O mapping descriptions are carried out. Also, we are planning to follow the same philosophy with legacy XML data. We will create X2 O, a language that describes the mappings between XML schemas and ontologies; and XMapster, the processor in charge of carrying out the exploitation of the mappings defined using X2 O . Extracting ontologies from folksonomies. Our initial work on semantically enriching folksonomies has lead to promising results which prompt us to further explore this line of work. In particular, we will implement an semantic enrichment algorithm that will rely on ontologies from WATSON to derive a broad range of semantic relations between folksonomy tags. We will also investigate how to derive synonyms, homonyms, multilingual information and domain/community specific sub-languages (jargons) from folksonomies. Extracting ontologies from lexicons and thesauri. The work reported in Chapter 5 will continue with evaluation and further investigation of disambiguation procedures. Relevant additional conceptual material from WordNet will be offered to FAO for integration into the knowledge base, both in relational and in owl format. Further, in the next deliverable, a richer system of guidelines and computational support for KOS and lexicon reengineering will be presented.

12.2

Future Work in Ontology Evaluation and Selection (T2.2.b)

Within this task we aim both at the individual development of evaluation and selection techniques as well as their tighter integration with each other an the rest of the project. Quality measures. We will continue our work on similarity measures between ontologies. Also, we will further evaluate the use of the BDM measure within the fisheries domain, and investigate effectiveness of the BDM score compared to other similarity metrics. Finally, in order to understand which evaluation measures make sense to be integrated in ontology selection, we want to perform a data-driven evaluation of existing ontology measures. While we have a theoretical hypothesis about how certain ontology measures work, we have no understanding about how they behave on real life data sets. The usefulness of these metrics is hard to decide theoretically especially when applied to heterogeneous ontologies such as those on the Semantic Web. Therefore we think it would be useful to test the theoretical hypothesis by applying each selected metric on a large collection of ontologies and study the produced results. We can then conclude in which cases a certain metric works and in which it does not. Those metrics that are found useful will be implemented on top of WATSON and then used as parts of the selection mechanism. Selection methods. The selection mechanism proposed here will be further extended (with evaluation metrics, modularization and mapping techniques), refined, implemented on top of WATSON and evaluated. Identifying sub-components. We have performed the first steps of providing the selection of ontology modules instead of selecting entire ontologies. We believe this is a very important line of work if the results of selection are to be easily understood by humans and efficiently reused in automatic scenarios. As a result, we will integrate more modularization techniques with the selection algorithm.


Page 139 of 151

Integrating the identified reusable components. To provide a complete knowledge selection process, it is not enough to identify reusable components, but they also need to be integrated (1) with each other and (2) in the context where they are reused. We will investigate ontology matching based methods that will allow integrating these components. Supporting context and user-specifications, open rating systems. We will explore the use of Open Rating Systems (ORS) as an alternative ontology selection mechanism that combines both automatic evaluation measures and ratings provided by a community of users. While supporting collaboration by allowing the selection of appropriate components, an ORS base system would also be a fine example of collaborative ontology evaluation.


Page 140 of 151


Appendix A

BNF Notation of the R2O Language Grammar In this section we provide the BNF notation of the R2 O language grammar, grouped according to the different types of transformations that can be performed with the language.

BNF for R2 O mapping descriptions (1) r2o::= r2o import-element? dbschema-description onto-description+ conceptmapping-definition+ (2) import-element::= import literal (3) literal::= ’’

BNF for R2 O DB schema descriptions (4) dbschema-description::= dbschema-desc name documentation? (has-table table-desc)+ (5) name::= name literal (6) documentation::= documentation literal (7) table-desc::= name tabletype? documentation? (has-column column-description)+ (8) tabletype::= tableType literal (9) column-description::= (keycol-desc | forkeycol-desc | bothkeycol-desc | nonkeycol-desc) name columnType col-reference? documentation? (10) columnType::= columnType datatype (11) col-reference::= refers-to literal (12) datatype::= string | boolean | decimal | float | double | date | integer ... (XML Schema Datatypes)


Page 141 of 151

BNF for R2 O ontology description (13) onto-description::= onto-desc name documentation? (has-concept concept-desc)+ (14) concept-desc::= name documentation? (hasAttribute att-description)+ (hasRelation rel-description)+ (15) att-description::= name attType datatype? (16) rel-description::= name range (17) range::= has-range literal

BNF for concept mapping definitions in R2 O (18) conceptmapping-definition::= conceptmap-def name identified-by+ (uri-as selector)? (applies-if cond-expr)? (joins-via concept-join-expr)? documentation? (described-by propertymap-def)* (19) identified-by::= identified-by literal (20) concept-join-expr::= (join-expr conceptJoinOpers cond-expr)? (21) conceptJoinOpers::= join | union | difference

BNF for condition expressions in R2 O (22) cond-expr::= orcond-expr | AND andcond-expr orcond-expr (23) orcond-expr::= condition | OR orcond-expr condition (24) condition::= primitive-condition (arg-restriction arg-restrict)* (25) primitive-condition::= lo_than | loorequal_than | hi_than | hiorequal_than| equals | not_equals | in_keyword | between | match_regexp (26) arg-restrict::= parameter-selector restriction (27) parameter-selector::= on-param literal (28) restriction::= has-value constant-value | has-column literal | has-transform transformation (29) constant-value::= datatype literal


Page 142 of 151


BNF for transformations in R2 O (30) transformation::= primitive-transf (arg-restriction arg-restriction)* (31) primitive-transf::= get _delimited | concat | add | subtract | Multiply | divide | square | constant| apply_regexp

BNF for property mappings in R2 O (32) propertymap-def::= attributemap-def | relfromatt-def | relationmap-def (33) attributemap-def::= attributemap-def name selector* use-dbcol* documentation? (34) relfromatt-def::= attributemap-def name selector* use-dbcol* newobj-type? documentation? (35) relationmap-def::= relationmap-def to-concept (applies-if cond-expr)? (joins-via relation-join-expr)?

(36) relation-join-expr::= join (join-expr cond-expr)? (37) use-dbcol::= use-dbcol literal (38) selector::= selector (applies-if cond-expr)? (aftertransform transformation)? (39) newobj-type::= newobject-type literal (40) to-concept::= to-concept literal


Page 143 of 151

Bibliography [AB05]

H. Alani and C. Brewster. Ontology Ranking based on the Analysis of Concept Structures. In Proceedings of the Third International Conference on Knowledge Capture( K-CAP 05), Banff, Canada, 2005. ACM.

[agr]

The AGROVOC Ontology. http://www.fao.org/agrovoc.

[asf]

The ASFA Ontology. http://www.fao.org/fi/asfa/asfa.asp.

[ASSM07]

S. Angeletou, M. Sabou, L. Specia, and E. Motta. Bridging the Gap Between Folksonomies and the Semantic Web: An Experience Report. Submitted for peer review, 2007.

[BADW04]

C. Brewster, H. Alani, S. Dasmahapatra, and Y. Wilks. Data-driven ontology evaluation. In Proceedings of International Conference on Language Resources and Evaluation, 2004.

[Bar]

J. Barrasa. PhD thesis, Facultad de Informatica, Universidad Politecnica de Madrid, Madrid, Spain, Title = Modelo para la definición automática de correspondencias semánticas entre ontologías y modelos relacionales, Year = 2007, March.

[Bar06]

J. Barrasa. Semantic upgrade and publication of legacy data. In Ontologies for Software Engineering and Software Technology, 2006.

[BC99]

M. Berland and E. Charniak. Finding parts in very large corpora. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics, 1999.

[BCGP03]

J. Barrasa, O. Corcho, and A. Gomez-Perez. Fundfinder – a case study of database-toontology mapping. In In Proc. ISWC Semantic integration workshop, 2003.

[BED04]

P. Buitelaar, T. Eigner, and T. Declerck. OntoSelect: A Dynamic Ontology Library with Support for Ontology Selection. In Proceedings of the Demo Session at the International Semantic Web Conference, Hiroshima, Japan, 2004.

[BFL98]

C.F. Baker, C.J. Fillmore, and J.B. Lowe. The Berkeley FrameNet project. In C. Boitet and P. Whitelock, editors, Proceedings of the Thirty-Sixth Annual Meeting of the Association for Computational Linguistics and Seventeenth International Conference on Computational Linguistics, pages 86–90, San Francisco, California, 1998. Morgan Kaufmann Publishers.

[BGM05]

J. Brank, M. Grobelnik, and D. Mladenic. A survey of ontology evaluation techniques. In In Proceedings of the Conference on Data Mining and Data Warehouses (SiKDD 2005), Ljubljana, Slovenia, 2005.

[BGM06]

J. Brank, M. Grobelnik, and D. Mladenic. Automatic Evaluation of Ontologies. In A. Kao and S.R. Poteet, editors, Natural Language Processing and Text Mining. Springer Verlag, 2006.

[Biz03]

C. Bizer. D2R MAP - A Database to RDF Mapping Language. In 12th International World Wide Web Conference, 2003.


Page 144 of 151


[CGR+ 05]

M. Ciaramita, A. Gangemi, E. Ratsch, J. Saric, and I. Rojas. Unsupervised Learning of Semantic Relations between Concepts of a Molecular Biology Ontology. In IJCAI, pages 659–664, 2005.

[CHS05]

P. Cimiano, A. Hotho, and S. Staab. Learning concept hierarchies from text corpora using formal concept analysis. JAIR – Journal of AI Research, 24:305–339, 2005.

[Con06]

NeOn Consortium. NeOn: Lifecycle Support for Networked Ontologies. NeOn Technical Annex, 2006.

[CPSTS05]

P. Cimiano, A. Pivk, L. Schmidt-Thieme, and S. Staab. Learning taxonomic relations from heterogenous sources of evidence. In Ontology Learning from Text: Methods, Applications, Evaluation. IOS Press, Amsterdam, 2005.

[CST00]

N. Cristianini and J. Shawe-Taylor. An introduction to Support Vector Machines and other kernel-based learning methods. Cambridge University Press, 2000.

[CST03]

P. Cimiano, S.Staab, and J. Tane. Automatic Acquisition of Taxonomies from Text: FCA meets NLP. In Proceedings of the ECML/PKDD Workshop on Adaptive Text Extraction and Mining, pages 10–17, Cavtat-Dubrovnik, Croatia, 2003.

[CTP00]

P. Clark, J. Thompson, and B. Porter. Knowledge Patterns. In A.G. Cohn, F. Giunchiglia, and B. Selman, editors, KR2000: Principles of Knowledge Representation and Reasoning, pages 591–600, San Francisco, 2000. Morgan Kaufmann.

[DDM03]

M. Dzbor, J. Domingue, and E. Motta. Magpie - towards a semantic web browser. In Proceedings of the Second International Semantic Web Conference, 2003.

[Del05]

K. Dellschaft. Measuring the Similiarity of Concept Hierarchies and its Influence on the Evaluation of Learning Procedures. Diploma thesis, Universitat Koblenz-Landau, December 2005. http://www.uni-koblenz.de/FB4/Institutes/IFI/AGStaab/Theses/2005/DADellschaft.pdf.

[DFJ+ 04]

L. Ding, T. Finin, A. Joshi, R. Pan, R. S. Cost, Y. Peng, P. Reddivari, V.C. Doshi, and J. Sachs. Swoogle: A Search and Metadata Engine for the Semantic Web. In Proc. of the 13th ACM Conf. on Information and Knowledge Management, 2004.

[DKS04]

O. Dekel, J. Keshet, and Y. Singer. Large Margin Hierarchical Classification. In Proceedings of the 21st International Conference on Machine Learning (ICML-2004), Canada, 2004.

[dmo]

The DMOZ Ontology. http://www.dmoz.org.

[DPF+ 05]

L. Ding, R. Pan, T. Finin, A. Joshi, Y. Peng, and P. Kolari. Finding and Ranking Knowledge on the Semantic Web. In Y. Gil, E. Motta, V.R. Benjamins, and M.A. Musen, editors, Proceedings of the 4th International Semantic Web Conference, volume 3729 of LNCS, pages 156 – 170. Springer-Verlag GmbH, 2005.

[DR04]

W. Daelemans and M.-L. Reinberger. Shallow text understanding for ontology content evaluation. IEEE Intelligent Systems, 19(4):74–81, 2004.

[DS06]

K. Dellschaft and S. Staab. On How to Perform a Gold Standard Based Evaluation of Ontology Learning. In Proceedings of the 5th International Semantic Web Conference (ISWC), 2006.

[DS07]

K. Dellschaft and S. Staab. Strategies for the Evaluation of Ontology Learning. In P. Buitelaar and P. Cimiano, editors, Bridging the Gap between Text and Knowledge – Selected Contributions to Ontology Learning and Population from Text. IOS Press, December 2007.


Page 145 of 151

[dSD+ 07]

M. d’Aquin, M. Sabou, M. Dzbor, C. Baldassarre, L. Gridinoc, S. Angeletou, and E. Motta. WATSON: A Gateway for the Semantic Web. Poster at ESWC, 2007.

[dSM06]

M. d’Aquin, M. Sabou, and E. Motta. Modularization: a Key for the Dynamic Selection of Relevant Knowledge Components. In Proc. of the ISWC 2006 Workshop on Modular Ontologies., 2006.

[Dub]

DCMI Metadata Terms. http://www.dublincore.org/documents/dcmi-terms/.

[ES04]

M. Ehrig and Y. Sure. Ontology mapping - an integrated approach. In Proceedings of the First European Semantic Web Symposium, number 3053 in Lecture Notes in Computer Science, pages 76–91. Springer Verlag, 2004.

[ewn98]

EuroWordNet: A Multilingual Database with Lexical Semantic Networks. Springer, October 1998.

[FAO06]

FAO. AGROVOC and the OWL Web Ontology Language. The Agriculture Ontology Service/Concept Server OWL model. Technical report, FAO, Rome, IT, 2006.

[FCC06]

M. Fernández, I. Cantador, and P. Castells. CORE A Tool for Collaborative Ontology Reuse and Evaluation. In Proc. of the 4th International Workshop on Evaluation of Ontologies for the Web (EON2006) at the 15th International World Wide Web Conference (WWW 2006), Edinburgh, UK, MAY 2006.

[Fel98]

C. (ed.) Fellbaum. WordNet - An electronic lexical database. MIT Press, Cambridge, Mass., 1998.

[FGM06]

B. Fortuna, M. Grobelnik, and D. Mladenic. Semi-automatic data-driven ontology construction system. In Proc. of the 9th International multi-conference Information Society IS-2006, pages 223–226. Ljubljana: Institut Jozef Stefan, 2006.

[fig]

The FIGIS Ontology.

[FMG05]

B. Fortuna, D. Mladenic, and M. Grobelnik. Semi-automatic construction of topic ontologies. In Semantics, web and mining : joint international workshops, EWMF 2005 and KDO 2005, volume 4289 of LNAI, pages 121–131. Berlin; Heidelberg; New York: Springer, 2005.

[foa]

The FOAF Ontology. http://www.foaf-project.org/.

[Gan05]

A. Gangemi. Ontology Design Patterns for Semantic Web Content. In M. Musen et al. (eds.): Proceedings of the Fourth International Semantic Web Conference, Galway, Ireland, 2005. Springer.

[GC06]

A. Gangemi and C. Catenacci. A constructive ontology of descriptions and situations. Technical report, ISTC-CNR, 2006. http://www.loa-cnr.it/TR/ConstructiveDnS.pdf.

[GCCL05]

A. Gangemi, C. Catenacci, M. Ciaramita, and J. Lehmann. Ontology evaluation and validation: An integrated formal model for the quality diagnostic task. Technical report, Laboratory for Applied Ontology, ISTC-CNR, 2005.

[GCCL06]

A. Gangemi, C. Catenacci, M. Ciaramita, and J. Lehmann. Modelling Ontology Evaluation and Validation. In Proceedings of the Third European Semantic Web Conference, volume 4011 of LNCS. Springer, 2006.

[gen]

The Gene Ontology. http://www.geneontology.org.


Page 146 of 151


[GFK+ 04]

A. Gangemi, F. Fisseha, J. Keizer, I. Pettman, and M. Taconet. A Core Ontology of Fishery and its use in the Fishery Ontology Service Project. In First International Workshop on Core Ontologies, EKAW Conference, CEUR-WS, volume 118, 2004.

[GGMO01]

A. Gangemi, N. Guarino, C. Masolo, and A. Oltramari. Understanding Top-Level Ontological Distinctions. In H. Stuckenschmidt (ed), editor, Proceedings of the IJCAI Workshop on Ontologies and Information Sharing, 2001, 2001.

[GGMO03]

A. Gangemi, N. Guarino, C. Masolo, and A. Oltramari. Sweetening WORDNET with DOLCE. AI Mag., 24(3):13–24, 2003.

[GH05]

S. Golder and B.A. Huberman. The Structure of Collaborative Tagging Systems. HPL Technical Report, 2005.

[GKF06]

G.Begelman, P. Keller, and F.Smadja. Automated Tag Clustering: Improving search and exploration in the tag space. In Proc. of the Collaborative Web Tagging Workshop at WWW’06, 2006.

[GKRT04]

R. Guha, R. Kumar, P. Raghavan, and A. Tomkins. Propagation of Trust and Distrust. In Proc. of the Thirteenth International World Wide Web Conference, pages 403–412, New York, NY, MAY 2004. ACM Press.

[GM02]

R. Girju and D. Moldovan. Text mining for causal relations. In Proceedings of the FLAIRS Conference, 2002.

[GM03]

A. Gangemi and P. Mika. Understanding the Semantic Web through Descriptions and Situations. In Proceedings of ODBASE03 Conference, pages 689–706. Springer, 2003.

[GNV03]

A. Gangemi, R. Navigli, and P. Velardi. The OntoWordNet Project: Extension and Axiomatization of Conceptual Relations in WordNet. In CoopIS/DOA/ODBASE, 2003.

[Gol05]

J. Golbeck. Computing and Applying Trust in Web-based Social Networks. PhD thesis, University of Maryland. College Park, MD, 2005.

[GP03]

A. Gomez-Perez. Ontology Evaluation. In S. Staab and R. Studer, editors, Handbook on Ontologies. Springer Verlag, Heidelberg, 2003.

[GPMM03]

A. Gomez-Perez and D. Manzano-Mancho. A Survey of Ontology Learning Methods and Techniques. OntoWeb Delieverable 1.5, May 2003.

[GPS99]

A. Gangemi, D. M. Pisanelli, and G. Steve. An Overview of the ONIONS Project: Applying Ontologies to the Integration of Medical Terminologies. Data Knowl. Eng., 31(2):183–220, 1999.

[GSY04]

F. Giunchiglia, P. Shvaiko, and M. Yatskevich. S-Match: An Algorithm and Implementation of Semantic Matching. In The Semantic Web: Research and Applications. Proceedings of the First European Semantic Web Conference, volume 3053 of LNCS, pages 61 – 75. SpringerVerlag, 2004.

[GTEM06]

J. Gracia, R. Trillo, M. Espinoza, and E. Mena. Querying the Web: A Multiontology Disambiguation Method. In Proc. of ICWE’06, 2006.

[Guh03]

R. Guha. Open Rating Systems. Technical report, Stanford University, CA, USA, 2003.

[Gui05]

G. Guizzardi. Ontological foundations for structural conceptual models. PhD thesis, University of Twente, Enschede, The Netherlands, Enschede, October 2005.


Page 147 of 151

[GW02]

N. Guarino and C.A. Welty. Evaluating ontological decisions with OntoClean. Communications of the ACM, 45(2):61–65, 2002.

[GW03]

N. Guarino and C.A. Welty. An overview of OntoClean. In S. Staab and R. Studer, editors, Handbook on Ontologies. Springer Verlag, Heidelberg, 2003.

[Hay96]

D.C. Hay. Data Model Patterns. Dorset House Publishing, 1996.

[Hea92]

M. Hearst. Automatic aquisition of hyponyms from large text corpora. In Proc. of the 14th International Conference on Computational Linguistics, 1992.

[Hep05]

M. Hepp. Representing the hierarchy of industrial taxonomies in owl: The gen/tax approach. In Proc. of the ISWC Workshop on Semantic Web Case Studies and Best Practices for eBusiness (SWCASE), pages 49–56, Galway, Ireland, November 2005.

[HJSS06]

A. Hotho, R. Jaschke, C. Schmitz, and G. Stumme. Information Retrieval in Folksonomies: Search and Ranking. In Proc. of ESWC’06, 2006.

[HL93]

B. L. Humphreys and D. A. Lindberg. The UMLS project: making the conceptual connection between users and the information they need. Bull Med Libr Assoc, 81(2):170–177, April 1993.

[HS02]

U. Hahn and S. Schulz. Turning Lead into Gold? Feeding a Formal Knowledge Base with Informal Conceptual Knowledge. In Proceedings of 13th EKAW Conference. Springer, 2002.

[HSG+ 05]

J. Hartmann, Y. Sure, A. Giboin, D. Maynard, M. C. Suarez-Figueroa, and R. Cuel. Methods for ontology evaluation. Knowledge Web Deliverable D1.2.3, 2005.

[HSM+ 04]

J. Hartmann, P. Spyns, D. Maynard, R. Cuel, M. Carmen Suarez de Figueroa, and Y. Sure. Methods for ontology evaluation. Deliverable D1.2.3, Knowledge Web, 2004.

[HSV03]

S. Handschuh, S. Staab, and R. Volz. On deep annotation. In 12th International World Wide Web Conference, 2003.

[KPO+ 04]

A. Kiryakov, B. Popov, D. Ognyanoff, D. Manov, A. Kirilov, and M. Goranov. Semantic annotation, indexing and retrieval. Journal of Web Semantics, ISWC 2003 Special Issue, 1(2):671– 680, 2004.

[LBC05]

Y. Li, K. Bontcheva, and H. Cunningham. SVM Based Learning System For Information Extraction. In Proceedings of Sheffield Machine Learning Workshop, Lecture Notes in Computer Science. Springer Verlag, 2005.

[LBC06]

Y. Li, K. Bontcheva, and H. Cunningham. Perceptron-like learning for ontology based information extraction. Technical report, University of Sheffield, Sheffield, UK, 2006.

[LMU06]

V. Lopez, E. Motta, and V. Uren. PowerAqua: Fishing the Semantic Web. In Proceedings of the Third European Semantic Web Conference, 2006.

[LPM05]

V. Lopez, M. Pasin, and E. Motta. AquaLog: An Ontology-portable Question Answering System for the Semantic Web. In Proceedings of the European Semantic Web Conference, 2005.

[LSNM06]

H. Lewen, K. Supekar, N.F. Noy, and M.A. Musen. Topic-Specific Trust and Open Rating Systems: An Approach for Ontology Evaluation. In Proc. of the 4th International Workshop on Evaluation of Ontologies for the Web (EON2006) at the 15th International World Wide Web Conference (WWW 2006), Edinburgh, UK, MAY 2006.


Page 148 of 151


[Mae02]

A. Maedche. Ontology Learning for the Semantic Web. Kluwer, Boston, 2002.

[May05]

D. Maynard. Benchmarking ontology-based annotation tools for the Semantic Web. In UK e-Science Programme All Hands Meeting (AHM2005) Workshop "Text Mining, e-Research and Grid-enabled Language Technology", Nottingham, UK, 2005.

[MB05]

A. Miles and D. Brickley. SKOS Core Vocabulary Specification. Technical report, World Wide Web Consortium (W3C), November 2005. http://www.w3.org/TR/2005/WD-swbp-skos-corespec-20051102/.

[MGG+ 04]

C. Masolo, A. Gangemi, N. Guarino, A. Oltramari, and L. Schneider. Wonderweb eu project deliverable d18: The wonderweb library of foundational ontologies. http://wonderweb.semanticweb.org/deliverables/documents/D18.pdf, 2004.

[Mik05]

P. Mika. Flink: Semantic Web Technology for the Extraction and Analysis of Social Networks. Journal of Web Semantics, 3(2), 2005.

[Mla98]

D. Mladenic. Machine Learning on non-homogeneous, distributed text data. PhD thesis, University of Ljubljana, Slovenia, Ljubljana, Slovenia, October 1998.

[MLB+ 04]

C. Masolo, L.Vieu, E. Bottazzi, C. Catenacci, R. Ferrario, A. Gangemi, and N. Guarino. Social Roles and their Descriptions. In Didier Dubois, Christopher A. Welty, and Mary-Anne Williams, editors, KR, pages 267–277. AAAI Press, 2004.

[MM99]

R. Mihalcea and D. Moldovan. An automatic method for generating sense tagged corpora. In Proceedings of AAAI ’99, pages 461–466, 1999.

[MPL06]

D. Maynard, W. Peters, and Y. Li. Metrics for Evaluation of Ontology-based Information Extraction. In WWW 2006 Workshop on "Evaluation of Ontologies for the Web" (EON), Edinburgh, Scotland, 2006.

[MS01a]

A. Maedche and S. Staab. Ontology Learning for the Semantic Web. IEEE Intelligent Systems, 16(2):72 – 79, March/April 2001.

[MS01b]

B. Magnini and M. Speranza. Integrating Generic and Specialized WordNets. In Proceedings of Recent Advances in Natural Language Processing, RANLP-2001, pages 149–153. Wolf Legal Publishers, 2001.

[MS02]

A. Maedche and S. Staab. Measuring Similarity between Ontologies. In Proc. of the European Conference on Knowledge Acquisition and Management (EKAW-2002), 2002.

[MSPG02]

B. Magnini, C. Strapparava, G. Pezzulo, and A. Gliozzo. The role of domain information in Word Sense Disambiguation. Nat. Lang. Eng., 8(4):359–373, December 2002.

[MSS03]

B. Magnini, L. Serafini, and M. Speranza. Making explicit the semantics hiden in schema models. In Proceedings of the Human Language Technology for the Semantic Web workshop at ISWC’03, 2003.

[nci]

The NCI Thesaurus.

[N.F04]

N.F.Noy. Evaluation by Ontology Consumers. IEEE Intelligent Systems, 19(4):74 – 81, July/August 2004.

[NGM05]

N.F. Noy, R. Guha, and M. A. Musen. User ratings of ontologies: Who will rate the raters? In Proc. of the AAAI 2005 Spring Symposium on Knowledge Collection from Volunteer Contributors, Stanford, CA, 2005.


Page 149 of 151

[Obe06]

D. Oberle. Semantic Management of Middleware, volume I of The Semantic Web and Beyond. Springer, New York, FEB 2006.

[OMGS04]

D. Oberle, P. Mika, A. Gangemi, and M. Sabou. Foundations for service ontologies: Aligning OWL-S to DOLCE. In Proceedings of the World Wide Web Conference (WWW2004), volume Semantic Web Track, 2004.

[one]

The OneFish Portal. http://www.onefish.org/global/index.jsp.

[PABC05]

W. Peters, N. Aswani, K. Bontcheva, and H. Cunningham. Quantitative Evaluation Tools and Corpora v1. Technical report, SEKT project deliverable D2.5.1, 2005.

[Pan03]

P.A. Pantel. Clustering by Committee. PhD thesis, 2003.

[PBMW98]

L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank Citation Ranking: Bringing Order to the Web. Technical report, Stanford University, CA, USA, 1998.

[PC05]

C. Perez and S. Conrad. Relational.OWL - A Data and Schema Representation Format Based on OWL. In APCCM 2005., 2005.

[Pet06]

E. Peterson. Beneath the Metadata: Some Philosophical Problems with Folksonomy. D-Lib Magazine, 12(11), November 2006.

[pla]

Forest.

[PM04]

R. Porzel and R. Malaka. A task-based approach for ontology evaluation. In Proc. of the ECAI Workshop on Ontology Learning and Population, 2004.

[PS]

M.T. Pazienza and A. Stellato. Linguistically motivated Ontology Mapping for the Semantic Web. citeseer.ist.psu.edu/745894.html.

[PSLP03]

C. Patel, K. Supekar, Y. Lee, and E. K. Park. OntoKhoj: A Semantic Web Portal for Ontology Searching, Ranking and Classification. In Proceeding of the Workshop On Web Information And Data Management. ACM, 2003.

[RU04]

J.-M. Rosengard and M. Ursu. Ontological Representations of Software Patterns. LNCS, 3215:31, 2004.

[Sal]

Sales Order Processing Collaboration Template. http://publib.boulder.ibm.com/infocenter/wbihelp/v6rxmx/ind topic=/com.ibm.wbi_order_management.doc/doc/collaboration_templates/ct_salesorderprocessing.htm.

[SCC97]

K.A. Spackman, K.E. Campbell, and R.A. Coté. SNOMED RT: A reference terminology for health care. In Proceedings of AMIA 97 Conference, 1997.

[Sch06]

P. Schmitz. Inducing Ontology from Flickr Tags. In Proc. of the Collaborative Web Tagging Workshop at WWW’06, 2006.

[SdM06]

M. Sabou, M. d’Aquin, and E. Motta. Using the Semantic Web as Background Knowledge for Ontology Mapping. In Proc. of the Int. Workshop on Ontology Matching (OM-2006), 2006.

[Se99]

S. Staab and et al. Getess - searching the web exploiting german texts. In Proc. of the 3rd Workshop on Cooperative Information Agents, 1999.

[SE05]

P. Shvaiko and J. Euzenat. A Survey of Schema-Based Matching Approaches. Journal of Data Semantics, IV:146 – 171, 2005.


Page 150 of 151


[SLL+ ]

D. Soergel, B. Lauser, A. Liang, F. Fisseha, J. Keizer, and S. Katz. Reengineering Thesauri for New Applications: AGROVOC Example. Journal Of Digital Information, 4.

[SLM06]

M. Sabou, V. Lopez, and E. Motta. Ontology Selection on the Real Semantic Web: How to Cover the Queens Birthday Dinner? In Proc. of the European Knowledge Acquisition Workshop (EKAW), Podebrady, Czech Republic, 2nd-6th October 2006.

[SLMU06]

M. Sabou, V. Lopez, E. Motta, and V. Uren. Ontology Selection: Ontology Evaluation on the Real Semantic Web. In Proceedings of the Evaluation of Ontologies on the Web Workshop, held in conjunction with WWW’2006, 2006.

[SM07]

L. Specia and E. Motta. Integrating Folksonomies with the Semantic Web. In Proc. of ESWC’07, 2007.

[Spa05]

S. Spaccapietra. Report on modularization of ontologies. D2.1.3.1, July 2005.

[SPR+ 06]

H. Stoermer, I. Palmisano, D. Redavid, L. Iannone, P. Bouquet, and G. Semeraro. Contextualization of a RDF Knowledge Base in the VIKEF Project. In ICADL, pages 101–110, 2006.

[Spy05]

P. Spyns. EvaLexon: Assessing triples mined from texts. Technical Report STAR-2005-09, STAR Lab, 2005.

[SR05]

P. Spyns and M.-L. Reinberger. Lexically evaluating ontology triples generated automatically from texts. In Proc. of the second European Conference on the Semantic Web, 2005.

[SSV02a]

L. Stojanovic, N. Stojanovic, and R. Volz. A Reverse Engineering Approach for Migrating Data-intensive Web Sites to the Semantic Web. In Intelligent Information Processing, 2002.

[SSV02b]

L. Stojanovic, N. Stojanovic, and R. Volz. Migrating data-intensive Web Sites into the Semantic Web. In Proc. of the Symposium on Applied Computing, 2002.

[Sti05]

P. Stickler. CBD - Concise Bounded Description. W3c member submission, World Wide Web Consortium, 2005.

[SWB]

Semantic Web Best Practices http://www.w3.org/2001/sw/BestPractices/.

[SWGM05]

M. Sabou, C. Wroe, C. Goble, and G. Mishne. Learning domain ontologies for web service descriptions: an experiment in bioinformatics. In Proc. of WWW05, 2005.

[SWGS05]

M. Sabou, C. Wroe, C. Goble, and H. Stuckenschmidt. Learning domain ontologies for semantic web service descriptions. Journal of Web Semantics, 3(4), 2005.

[Thea]

The Description and Situation ontology. http://www.loa-cnr.it/ontologies/ExtendedDnS.owl.

[Theb]

The Plan ontology. http://www.loa-cnr.it/ontologies/Plans.owl.

[vAGS06]

M. van Assem, A. Gangemi, and G. Schreiber. Conversion of WordNet to a standard RDF/OWL representation. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy, May 2006.

and

Knowledge Web Deliverable

Deployment

Working

Group.

[VDATHKB03] W.M.P. Van Der Aalst, A.H.M. Ter Hofstede, B. Kiepuszewski, and A.P. Barros. Workflow Patterns. Distributed and Parallel Databases, 14:5–51, 2003. [vR79]

C. van Rijsbergen. Information Retrieval. Butterworths, London, 1979.


Page 151 of 151

[VVS05]

J. Völker, D. Vrandecic, and Y. Sure. Automatic Evaluation of Ontologies (AEON). In Proceedings of the 4th International Semantic Web Conference (ISWC2005), 2005.

[Wal07]

T. Vander Wal. Folksonomy coinage and definition. 2007.

[Wel98]

C. Welty. The ontological nature of subject taxonomies, 1998.

[WKCC03]

C. Welty, R. Kalra, and J. Chu-Carrol. Evaluating ontological analysis. In Proceedings of the ISWC-03 Workshop on Semantic Integration, 2003.

[WZM06]

H. Wu, M. Zubair, and K. Maly. Harvesting Social Knowledge from Folksonomies. In In Proc. of HYPERTEXT ’06, 2006.

[WZY06]

X. Wu, L. Zhang, and Y. Yu. Exploring Social Annotations for the Semantic Web. In Proc. of WWW’06, 2006.

[YLS06]

X. Yan, X. Li, and D. Song. Document Generality: Its Computation and Ranking. In Proceedings of the Seventeenth Australasian Database Conference, 2006.


D2.2.1 Methods for Selection and Integration of Reusable ... - CiteSeerX

D2.2.1 Methods for Selection and Integration of Reusable ... - CiteSeerX

Suggest Documents

Integration of FAHP and TOPSIS Methods for the Selection of ...

Comparison of Selection Methods for Evolutionary ... - CiteSeerX

Comparison of Selection Methods for Evolutionary ... - CiteSeerX

Formal Methods Integration for Software Development ... - CiteSeerX

Relay Selection Methods for Wireless Cooperative ... - CiteSeerX

Integration of Statistical Methods and Judgment for Time ... - CiteSeerX

Design Principles for Reusable, Composable and ... - CiteSeerX

Dynamic Integration of Data Mining Methods Using Selection in a ...

Dynamic Integration of Data Mining Methods Using Selection in a ...

Formal Methods for Integration of Automotive Software - CiteSeerX

A Methodology for integration of formal methods in a ... - CiteSeerX

Comparing Efficiency of Integration Methods for Cloth ... - CiteSeerX

Selection of Reverse Engineering Methods for Relational ... - CiteSeerX

Methods for Selection of Aptamers to Protein Targets - CiteSeerX

Integration of Reusable Software Components and ... - Semantic Scholar

A comparison of feature selection methods for an evolving ... - CiteSeerX

New Feature Selection Methods for Qualification of the ... - CiteSeerX

Projective Integration Methods for Distributions

methods for studying morphological integration and modularity

methods for studying morphological integration and modularity

methods for studying morphological integration and modularity

Model Selection Methods for Unidimensional and

Assignment of Reusable and Non-Reusable

An empirical study of representation methods for reusable software ...