workshop semantic web mining

12th European Conference on Machine Learning (ECML'01) 5th European Conference on Principles and Practice of Knowledge

Discovery in Databases (PKDD'01)

Workshop Semantic Web Mining

Gerd Stumme, Andreas Hotho, Bettina Berendt

September 3, 2001 Freiburg Germany

Foreword

When we started the organization of the workshop, it was clear that the Se-mantic Web and Web Mining are two fast-developing research areas which havemany points of contact. However, there was not yet a precise idea what theintegration of the two areas might look like in detail. The workshop aims toadvance the convergence between Semantic Web and Web Mining research bybringing together researchers and practitioners from these two areas. Our aim isto improve, on the one hand, the results of Web Mining by exploiting the newsemantic structures in the web, and on the other hand to exploit Web Miningfor building the Semantic Web.

The Semantic Web is based on a vision of Tim Berners-Lee. He suggeststo enrich the web by machine processable information which is organized ondi�erent levels (see http://www.w3.org/DesignIssues/Semantic.html). For WebMining, the levels from XML and RDF to ontologies and logics are of particularinterest. Web Mining applies data mining techniques on the web. Three areascan be distinguished: Web usage mining analyzes user behavior, web structuremining utilizes the hyperlink structure, and web content mining exploits thecontents of the documents on the web.

In the workshop, we want to discuss the use of XML, RDF, ontologies, andlogics for the three web mining areas; and the support of web mining techniquesfor building XML and RDF schemes and ontologies. There are quite a numberof people from di�erent communities that approach the �eld of Semantic WebMining from di�erent, interesting angles. The goal of the workshop is to establishcommunication between these communities.

The contributions to this workshop represent di�erent approaches to Seman-tic Web Mining:

A number of methods are proposed for learning domain ontologies. Theserange from learning from natural-language data to exploiting existing meta-data representations. Engels, Bremdal, and Jones present a method that ex-tracts information from natural-language data and produces a graph visualiza-tion of relations between concepts, as well as XML/RDF ontologies. Kurematsu,Nakaya, and Yamaguchi describe an approach for constructing domain ontolo-gies from machine-readable dictionaries and text corpora. Clerkin, Cunningham,and Hayes cluster objects described by their attributes to generate class hierar-chies expressible as RDF schemas. Kavalec, Sv�atek, and Strossa use web direc-tories like Yahoo! as training data for automated meta data extraction.

Domain ontologies become more useful when they are related to the usersand their individual interests and preferences. Learning from a user's prior inter-action with the system can supply useful data for it. Kiss and Quinqueton dealwith the learning of user preferences. Their multi-agent system is dedicated to

corporate memory management in an intranet. Semantic structure on the Web,together with the mining of user navigation histories, can directly lead to betteradapted user interfaces. Mobasher describes an adaptive agent for informationretrieval which uses a concept hierarchy of terms found in documents, togetherwith clusters summarizing user search histories, to (semi-)automatically improvequeries.

Lastly, it is important to develop integrating architectures that range fromontology learning to the display of results for the user. Haustein describes anagent-based blackboard architecture that connects mining components, ontolo-gies, and applications, as for instance an interface for generating HTML. LeGrand and Soto cluster XML topic maps, online equivalents of printed indexes,by means of Formal Concept Analysis, to de�ne pro�les. Their aim is to supportnavigation and understanding of the set of documents.

With this collection of research papers, we aim to provide a starting pointfor the convergence of the Semantic Web and Web Mining. We wish to expressour appreciation to all the authors of submitted papers, to the members of theprogram committee, and to the additional reviewers for making the workshop avaluable contribution to Semantic Web Mining.

July 2001 Bettina BerendtAndreas HothoGerd Stumme

Organization

The Semantic Web Mining Workshop was organized as a workshop within the12th European Conference on Machine Learning (ECML'01) and the 5th Euro-pean Conference on Principles and Practice of Knowledge Discovery in Databases(PKDD'01). It was held on September 3, 2001, in Freiburg, Germany.

Workshop Chairs

Gerd StummeInstitut f�ur Angewandte Informatikund Formale Beschreibungsverfahren (AIFB)Universit�at KarlsruheD{76128 Karlsruhe, Germanyhttp://www.aifb.uni-karlsruhe.de/WBS/[email protected]

Andreas HothoInstitut f�ur Angewandte Informatikund Formale Beschreibungsverfahren (AIFB)Universit�at KarlsruheD{76128 Karlsruhe, Germanyhttp://www.aifb.uni-karlsruhe.de/WBS/[email protected]

Bettina BerendtAbteilung P�adagogik und InformatikHumboldt-Universit�at zu BerlinGeschwister-Scholl-Stra�e 7D{10099 Berlin, Germanyhttp://www.educat.hu-berlin.de/[email protected]

Program Committee

Soumen Chakrabarti(Indian Inst. of Technology, Bombay)

Rosine Cicchetti(Univ. de la M�editerran�ee, Marseille)

Stefan Decker(Stanford University, Palo Alto)

Ronen Feldman(Bar-Ilan University, Ramat Gan)

Klaus-Peter Huber(SAS, Heidelberg)

Alexander M�adche(Universit�at Karlsruhe)

Tom Mitchell(WhizBang! Labs, Pittsburgh)

Bamshad Mobasher(DePaul University, Chicago)

Katharina Morik(Universit�at Dortmund)

Claire Nedellec(Universit�e Paris Sud)

Myra Spiliopoulou(Handelshochschule Leipzig)

Rudi Studer(Universit�at Karlsruhe)

Further Reviewers

Martin Lacher(Stanford University, Palo Alto)

Lot� Lakhal(Univ. de la M�editerran�ee, Marseille)

Carsten Pohle(Handelshochschule Leipzig)

Karsten Winkler(Handelshochschule Leipzig)

Table of Contents

Learning Domain Ontologies

CORPORUM: a workbench for the Semantic Web : : : : : : : : : : : : : : : : : : : : : 1R.H.P. Engels, B.A. Bremdal, R. Jones

Acquiring Conceptual Relationships from a MRD and Text Corpus : : : : : : : 11M. Kurematsu, N. Nakaya, T. Yamaguchi

Ontology Discovery for the Semantic Web Using Hierarchical Clustering : : 27P. Clerkin, P. Cunningham, C. Hayes

Web Directories as Training Data for Automated Metadata Extraction : : : 39M. Kavalec, V. Sv�atek, P. Strossa

Integrating User Behavior and Domain Ontologies

Multiagent Cooperative Learning of User Preferences : : : : : : : : : : : : : : : : : : : 45A. Kiss, J. Quinqueton

Invited Talk:ARCH: An Adaptive Agent for Retrieval Based on Concept Hierarchies : : : 57

B. Mobasher

Architectures for Semantic Web Mining

Utilising an Ontology Based Repository to Connect Web Miners andApplication Agents : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 59

S. Haustein

XML Topic Maps and Semantic Web Mining : : : : : : : : : : : : : : : : : : : : : : : : : : 67B. Le Grand, M. Soto

�� "!�!$# %��&�'�()%�*,+.-0/1/32�465�7)8�9;:=<)>)8�?)5A@�9B8�CED�F�GIH�J1J�KLNMPORQSMPOUTWV XZYP[U\E]�^`_badcfehgji0\k]ml gZaPnoa�pqYrcIl stevu aw_

xzy|{by�}~y|��|�fy��by|� �1�)��$xzyw��)�3�� 6�P� y��

}ry��y�� A��h �¡h��¢¤£¥ �¦�§©¨{f��"�|¢ª��« �A¬

��®"y¯��)�°�)��®d�)�1�²±Ay¯®��1�)��B��³°´¥µ��©�1�"y ¶��)�)·©¸Z´��³�±�y��¹��¬»ºh ��|¼�¡�¡h

½¿¾ZÀ�ÁhÂ�Ã�Ä�Á

ÅÆÈÇ�ÉmÊtÇ�Ë¤Ì¥ÍÏÎ°ÐÒÑtÊ�ÅÔÓÏÕ)ÖØ×.Ù0Ú$ÕUÛ�Ù0ÜÏÝßÞBà�á`âã�â�â�ÜZÕkÞ�â�ÚBáåä�à�Þ�Ó�Ù)æ�Þ�ákæÏç;ÓfçtÙ0Ü�ÞBâ�Ü�Þ�è|é"Ó�âä�â�ãmàÖzÙ0ÚBÝ)Õ0Ü�à�ê�à�Ü�Ý»à¯Þ;Öëâ�Û¯×�ì Õ�ÜÏíSà�ÞBÖbîÏÕ0Ý0â�ÖÕ�ÚBâ»Þ�ïAî�àç�Õ0Û�Û�ïðÕ)í�í�â1íUà�ÜñÕòÚBÕ0ÜÏí�Ù0áóÕ�Ü©íÌ0ôöõ�÷1Ño×ÒÕ0Ö�Ó�à�Ù)Üøã�ï�â�ù)â�ÚBï�ã�Ù�í�ïúä�ÓÏÙ�×.â�â�Û�ÖÛ�à�û)â,ç�Ù0Ü�Þ�ÚBà�ãÏæ�Þ�à�Ü�ÝÏèNédïAî�àç�Õ�Û�Û�ï0ìüÞBÓ�â�ÚBâ,ÓÏÕ0ÖÜ�Ù0Þbã�â�â�ÜýákæÏç;Óýç�Ù0ÜÏç�â�ÚBÜòÕ�ã�Ù0æ�Þ,Ó�Ù¥äPÞ�Ùî�ÚBâ�Ö�â�Ü�Þ ç�Ù0Ü�Þ�â�Ü�ÞBÖIà�ÜúÞ�Ó�âfã©â1Ö�Þ`ä"Õ3ï0ì�Ù�ÞBÓ�â�ÚÞ�ÓÏâ�Üoî�æÏÚ�âüÛÕ3ï�þ°Ù)æ�Þ|àÖ�Ö�æ�â�Ö�è�é"Ó�à�Ö|×ÒÕ0çtÞ�ìAçtÙ0á þã�à�Ü�â�ízä�à¯ÞBÓzÞ�ÓÏâE×ÒÕ0çtÞ~ÞBÓÏÕ¥ÞrÞ�Ó�â Ú�â�î�ÚBâ�Ö�â�Ü�ÞBÕ�þÞ�à�Ù0ÜkÛ�Õ0Ü�Ý0æÏÕ0Ý0â�æÏÖ�â�íkÕ¥ÞÿÞBÓ�âwä�Ù0ÚBÛ�í ä�àí�âdädâ�ãàÖ`á`Õ0à�ÜÏÛ�ï�×.Ù)Ú�á¤Õ�Þ`Ù)Ú�à�â�Ü�ÞBâ�í � à�è â0èøÜ�Ù0Þoí�âtþî�â�ÜÏí�à�Ü�ÝzÙ)Ü�ÕÈçtÙ0á`î�Û�â��b×.Ù0ÚBá¤Õ�Û|Û�Ù0Ý)à�ç ÚBâ�î�þÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0ÜÈá`â�ç;ÓÏÕ0Ü�àÖëá��6ì�á¤Õ�û)â�ÖüîÏæ�ã�Û�à�Ö�Ó�þà�Ü�ÝrÙ)ÜEÞBÓ�â�� â�Õ)Öëï)ì¥Ý0à�ù�à�Ü�Ýrà¯Þ�à�ÞBÖhâ�Ü�Ù)Úëþá`Ù0æÏÖ�â��î�Ú�â1Ö�Ö�à�ã�à�Û�à�Þ�ï0è�~Û¯ÞBÓ�Ù0æÏÝ0Ó ä�à�í�â�Û�ïEÕ)ç6þûAÜ�Ù¥ä�Û�â�í�Ý)â�íø×.Ù)Ú à�ÞBÖfÝ0â�Ü�â�Ú;Õ�Û Õ�Ü©í æ�Ü�à�ù0â�ÚëþÖBÕ�Û~Õ)í�ù¥Õ�Ü�ÞBÕ0Ý0â�Ö�ìdÞ�Ó�âzà�ÜÏç�Ú�â1Õ0Ö�à�Ü�Ý�î�Ù0îÏæ�Û�Õ0Úëþà�Þ�ï�Ù�×~ÞBÓ�âfädâ�ãØÕ0Û�Ö�ÙbÖ�Ó�Ù¥äüÖIæÏÖ`ÖëÙ)á âfá¤Õ¥þ� Ù0Ú�í�Ú;Õ3ä"þ�ãÏÕ0ç;û�Ö�èÿé"Ó�â�í�â�ù0â�Û�Ù)î�á`â�Ü�ÞBÖ Ù�×ÏÞ�Ó�âà�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0Ü,çtÙ)Ü)ÞBâ�Ü�ÞBÖrÙ0Ü Þ�Ó�â`ä�â�ã,í�æ�ÚBà�Ü�ÝÞ�ÓÏâ�ÛÕ0ÖëÞ ï0â1Õ�Ú,Õ�Û�Ù0Ü�â)ìIçtÛ�â�Õ0Ú�Û�ï á`Õ0Ú�û�ÖzÞ�Ó�âÜ�â�â�í,×.Ù0Ú ÖëÙ)á`âªç;ÓÏÕ�ÜÏÝ0â�Ö�è� �â�ÚBÓÏÕ�îÏÖEÙ0Ü�âoÙ0×Þ�ÓÏâ¤á Ù�Ö�Þ�×.â�Û¯ÞEîÏÚ�Ù)ã�Û�â�á¤Ö ä�à¯ÞBÓöÞ�Ó�â¤ä�â�ãöÕ0ÖÕ í�à�ÖëÞ�ÚBà�ãÏæ�Þ�â1í à�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0Ü�Öëï�ÖëÞ�â�á àÖkÞ�Ó�âí�à��oçtæ�Û�Þ�ï�Þ�Ù��ÏÜÏíØÕ�Ü©íúçtÙ)á`îÏÕ�ÚBâ à�Ü�×.Ù)Ú�á¤Õ¥þÞ�à�Ù0ÜZä�Ó�àç;ÓÈàÖ�î�ÚBÙ¥ù�àí�â1íªÙ)ÜÈà¯Þ1è��zÕ0Ü�ï î©â�Ù�þî�Û�âbÕ)í�íøî�ÚBà�ù¥Õ¥ÞBâ0ì�â1í�æÏç�Õ�Þ�à�Ù0ÜÏÕ0Û~Ù)ÚªÙ0ÚBÝ)Õ0Ü�à�þê�Õ�Þ�à�Ù0ÜÏÕ0Ûdç�Ù0Ü�Þ�â�Ü)ÞkÞ�ÙÈÞBÓ�âªä�â�ã ä�ÓÏà�ç;Ó�àÖkÙ0×�� ! "$#%"&��'�(*)%),+-�/.102�*34.$#%�5026�387:9<;

'�(*�*=*34=>+*.?�/@13BAC0EDCFG�*�H��IJ�*�BK<)%3B=��3BLM�1NO��;QPBR�R�RS;PBT�PBU�V�!<�W6*3X�/(�026��.2"�K<#%"?6Y0?�Z0?6-�/�*[\�/#%=*]^#%�*#%"102.2�S;0?��.J_�34=*34.?)`�/�*=ba:cd'e��.W0?6*3f(*"Q�/��3X�/'g026*34#%.Wh�#%"?(-�/) ;#%"Q�S0?#%��i0?�O��)%",(�"234=j02�khO#%"2(-�/)%#%"?3Z0?6*3Y"?0?.2(*AC0?(*.23Y�/'l ��(*.?3,mYK<# 0?6*#%�k026�#%"��/.102#%A4)%3�!,no)%3B�/"23,.?34'�34.G0?�Y0?6*3K��.?[O"?6*��+YK�34p*"?# 023J'��.�026�3 l �-�/)qh�34.2"?#%��q!

à�á`á`â�ÜÏÖ�âÔí�à�ù)â�Ú;ÖëârÜÏÕ¥ÞBæ�Ú�â)èXrdÙ0Ü�Þ�â�Ü)Þ�Ù)ÜoÞ�ÓÏâä�â�ãøàÖoÝ)Ú�Ù¥ä�à�Ü�ÝöçtÛ�Ù)Ö�â�Ú¤ÞBÙ ÕöÚ�â1Õ�Ûüæ�ÜÏà�ù)â�Ú�þÖBÕ�ÛZs�Í�÷-t,u�Ç;ôOv)Ç,É;Ì3Ê�Çtìdä�Ó�â�ÚBâoÞBÓ�â�ÚBâfà�Ö Ù0ÜÏÛ�ïÙ)Ü�â î�ÚBÙ0ãÏÛ�â�á ÚBâ�ÛÕ¥ÞBà�ù)â�Û�ïxw æ�ÜÏí�â�Õ�Û�ÞOw`ä�à�Þ�ÓWyÞBÓ�â�î�ÚBÙ0ã�Û�â�áýÙ�×�Þ�Ó�âüà�Ü�Þ�â�Ú�î�ÚBâtÞ;Õ¥ÞBà�Ù)ÜIÙ0×�Ö�æÏç;Óç�Ù0Ü�Þ�â�Ü�ÞBÖ�è

z ÜøÞ�ÓÏà�Ö îÏÕ�î�â�Ú1ì�Þ�Ó�â,Õ�æ�ÞBÓ�Ù0Ú;Ö¤î�ÚBÙ¥ùAà�í�âzÕí�à�ÖBçtæÏÖBÖ�Ù0ÜöÙ0ÜúÕÈÞ�â�ç;ÓÏÜ�à�ç�Õ�Û"ÖëÙ)Û�æ�Þ�à�Ù0Ü�ä�Ó�àç;ÓàÖ Õ�à�á â1í Õ¥Þ`Ó�â�Û�î�à�Ü�Ý Þ�Ó�âfädâ�ã ÞBÙ ã�â�çtÙ)á`âá`Ù)Ú�â Ê�ÇtË¤Ì¥Í©Î�ÐÒÑ�è¤é"Ó�â{r^|\}Z ~|\},�� Þ�ÙAÙ0ÛÖ�âtÞ~Þ�ÓÏÕ�Þ~à�ÖÔí�â�ù0â�Û�Ù0î�â�í ×.Ù)ÚüÞ�ÓÏà�Ö~ÞBÕ)Öëû â��à�ÖëÞBÖÙ0×�ÕkÖ�âtÞ"Ù�×ÿî�ÚBÙ0Ý)ÚBÕ0á¤Ö�Þ�ÓÏÕ�Þ"ç�Õ�Ü¤×.æÏÛ��ÏÛ�ÛÿÕEù¥Õ�þÚBà�âtÞ�ïªÙ0×�ÞBÕ)Öëû�Ö�ìÏâ�à�Þ�ÓÏâ�ÚrÕ)Ö�w Ö�Þ;Õ�ÜÏí�þ�Õ0Û�Ù)Ü�â�w�ìÏÙ)ÚÕ0æ�Ý0á`â�Ü)ÞBà�ÜÏÝbâ1Õ0ç;ÓØÙ�Þ�ÓÏâ�Ú1è�rÖ Þ�Ó�âzÕ0à�á Ù�×ÞBÓ�â"Öëâ�á`Õ0Ü�Þ�àçwä�â�ãIà�Ö Þ�Ù â�Ü�ÓÏÕ�Ü©çtâdÞ�Ó�â��:�BÇBÑ�Ð��Ê6ÐÒ÷�Í Õ0ÜÏí��Ç;Ñ;Ì-u�u�Ù�×�Ö�â�Õ0ÚBç;Óÿì�ã�æ�Þ�Õ�ÛÖ�ÙEâ�ÜÏÕ0ã�Û�âÞBÓ�âfæÏÖëâfÙ�×$u�÷�v0ÐÒÑ;Ì-u^��Ç;Ì¥Êt÷�ÍÏÐ Í�vtÙ)Ü�ä�â�ãØçtÙ0Ü�þÞBâ�Ü�ÞBÖ"à�ÜZÙ0Ú;í�â�Ú"ÞBÙ¤Õ�ÜÏÖ�ädâ�Ú^�)æÏâ�ÚBà�â1Ö�è z á`î�Ù0Ú�þÞ;Õ�Ü�ÞÔÞBÕ0Ö�û�ÖüÞBÓÏÕ¥ÞEÕ0Ú�â í�â�Õ0Û¯Þ�ä�à�Þ�Ó,ã�ï�r^|\}�þ ,|\},�� Õ0Ú�âfÚBâ�ÛÕ¥Þ�â1í ÞBÙöà�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0ÜØÚ�â�þÞBÚ�à�â�ù¥Õ0Û � �ÏÜÏífÚBâ�Û�â�ù¥Õ�Ü�Þ�í�Ù�çtæ�á`â�Ü�Þ;Ö�ì�Ù)Ú"Ö�æ�î�þî�Ù0Ú�Þ,Þ�Ó�â»æÏÖ�â�Ú5�ÏÜ©í�à�ÜÏÝSÞ�Ó�â�áj�tì¤ã�æ�ÞúÕ0Û�Ö�Ùà�Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)Ü â��AÞBÚBÕ)ç6Þ�à�Ù0Ü � ç�Õ0Ü$ä�âbãÏæ�à�Û�ÞÈÕûAÜ�Ù¥ä�Û�â�í�Ý0âEãÏÕ)ÖëâI×.ÚBÙ0áWä�â�ãbí�ÙAç�æ�á`â�Ü�ÞBÖrÞBÙÕ0ÜÏÖ�ädâ�Ú��æ�â�Ú�à�â�Ö/�*�6ì à�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Üåí�àÖ�Ö�â�á`à�þÜ©Õ¥Þ�à�Ù0Ü � ÖëæÏá á¤Õ0Ú�à�ê�à�Ü�Ý�ÖëÞ�Ú;Õ¥ÞBâ�Ý)à�â1ÖoÕ�ÜÏí à�Ü�þ×.Ù)Ú�á¤Õ�Þ�à�Ù0ÜoùAàÖëæÏÕ0Û�àÖBÕ¥Þ�à�Ù0Ü:�6ìAÕ�ÜÏí Õ�æ�ÞBÙ0á¤Õ¥ÞBâ�íí�ÙAç�æ�á`â�Ü�Þ çtÛÕ0ÖBÖëà��©ç�Õ¥Þ�à�Ù0Ü Ö�ÞBÚBÕ�Þ�â�Ý0à�â�ÖØî©â�Úëþ×.Ù)Ú�á`â1í ã�ïWÖ�Ù�þ�ç�Õ0Û�Û�â�í à�Ü�Þ�â�Û�Û�à�Ý)â�Ü�Þ Õ0Ý0â�Ü�Þ;Öä�ÓÏà�ç;Ó¤Õ�ÚBâ�î�ÚBâ�Ö�â�Ü�Þ|Ù0Ü ÞBÓ�âüä�Ù0ÚBÛ�í ä�àí�â�ädâ�ãÙ)ÜªÕIî©â�ÚëÞBà�Ü�â�Ü�Þ"ã©Õ0Ö�à�Ö�èhé"ÓÏâÔç�æ�Ú�ÚBâ�Ü�Þ�Õ�Ú�Þ�àçtÛ�âí�à�ÖBçtæÏÖBÖ�â�Ö�Þ�ÓÏâ&r^|\}Z ~|\},�Y�¿ÞBÙAÙ0Û�Ö�âtÞ~Õ0ÜÏíÖ�Ó�Ù¥äüÖ Ó�Ù¥ä à�Þ|ç�Õ�Ü`Ö�æ�î�î�Ù0Ú�ÞhÝ)â�Ü�â�ÚBÕ�Þ�à�Ù0ÜkÕ0ÜÏíæ�Þ�à�Û�àÖ�Õ�Þ�à�Ù0ÜfÙ�×�Öëâ�á¤Õ�Ü�Þ�àç�Ö�Ù0ÜfÞBÓ�â�ädâ�ãÿè

�1

AIFB AIFB

� �� ÀhÄ^��ZÃ�Âf� � ÀPÁ �� Á� � Â�� À�� Ã�� Á��tÄ�À � �Á�� ~¾

� â�ÜÏâ�Ú;Õ�Û�Û�ïkÖ�î�â�Õ�ûAà�Ü�ÝÏì�Þ�Ó�â�Ú�â�Õ�ÚBâ�Þ�ädÙÔ×.æ�Ü©í�Õ¥þá`â�Ü�ÞBÕ0Û�Û�ï¤í�à��â�Ú�â�Ü)Þ"Ö�ç�â�ÜÏÕ0Ú�à�Ù)Öwà�Üªä�Ó�à�ç;ÓoÞ�Ó�âä�Ù0ÚBÛ�íIä�à�í�âüädâ�ã¤çtÙ0æÏÛ�í`â�ù0Ù0Û�ù0â�×.æ�Ú�Þ�ÓÏâ�Ú1èG wà¯þÞ�ÓÏâ�Ú|ÞBÓ�â~ç�æ�Ú�ÚBâ�Ü�ÞBÛ�ïIâ��à�ÖëÞ�à�Ü�ÝEá¤Õ)Ö�Ö�Ù0×ÿí�Ù�çtæ�þá`â�Ü�ÞBÖ~Õ3ù3Õ0à�ÛÕ�ãÏÛ�â Ù0Ü Þ�ÓÏâEädâ�ãÈç�Õ�Üfã�âkÕ�ÜÏÕ0Û¯þï�Öëâ1ízà�Ü à�ÞBÖ�ç�æ�Ú�ÚBâ�Ü�Þ&w Õ)Ö~àÖw�×.Ù)Ú�ávÕ0ÜÏí ç�Ù0Ü�þÞ�â�Ü�ÞBÖ�ç�Õ�Ü ã�â|â��AÞ�Ú;Õ0çtÞ�â1í~×.ÚBÙ0á�à¯Þ1ì3Ù0Ú²ÞBÓ�â|ÚBâ�î�þÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0Ü�×.Ù)Ú�á¤Õ¥ÞhÙ�×�ÞBÓ�â"ädÙ)Ú�Ûí�ä�àí�â�ädâ�ãàÖoç;ÓÏÕ�ÜÏÝ0â�í�æ�îØ×.ÚBÙ0Ü�ÞoÖëÙ,Þ�Ó©Õ¥Þ í�Ù�ç�æ�á`â�Ü�ÞBÖÕ�ÚBâ¤Õ3ù¥Õ�à�Û�Õ0ã�Û�â à�ÜöÕf×.Ù)Ú�á¤Õ�ÞÔÞBÓÏÕ¥ÞEâ��î�Ú�â1Ö�Ö�â�ÖÖ�æÏç;Ó�w Ö�â�á¤Õ�Ü�Þ�àç�ÖOw�á`Ù0ÚBâ â��î�Û�à�ç�à¯Þ1è wÕ0ç;Ó�Ù�×IÞ�Ó�â1Öëâ�Õ�îÏî�Ú�Ù�Õ0ç;Ó�â1ÖªÓ©Õ3ù0âbÞ�Ó�â�à�Ú

Ù¥ä�Üoí�Ú;Õ3ä�ãÏÕ0ç;û�Ö�ì0Õ ×ÒÕ)ç6ÞwÞBÓÏÕ¥Þ�á`à�Ý)Ó)Þdã©â~Þ�Ó�âÚBâ�Õ0Ö�Ù0ÜzÞ�ÓÏÕ�ÞÔÞBÓ�â�ÚBâ à�Ö ÖëÞ�à�Û�Û|Õ0Ü Ù0ÜÏÝ0Ù0à�Ü�Ýfí�âtþãÏÕ�Þ�âkÙ)ÜÈä�ÓÏÕ¥Þ~Þ�Ó�âIÜ�â��AÞrÝ)â�Ü�â�ÚBÕ�Þ�à�Ù0Ü z Ü�Þ�â�ÚëþÜ�â�Þ Ö�Ó�Ù0æ�Ûí�Û�ÙAÙ0û,Û�à�û0â)è é"Ó�â í�à�ÖBÕ0í�ù¥Õ0Ü)Þ;Õ�Ý)âÙ�×$w ¡©Õ�ÞOw�ì|á¤Õ�à�Ü�Û�ïö×.Ù0ÚBá`Õ�ÞIãÏÕ)Öëâ1í�ÚBâ�î�ÚBâ�Ö�â�Ü�þÞBÕ�Þ�à�Ù0ÜmÛÕ�Ü�Ý)æÏÕ�Ý)â�Ö � ç6×�è£¢~é^�¥¤wì\¤ Õ0é â¦\�àÖ�Þ�Ó©Õ¥Þ~ÞBÓ�â�ïZá`à��Zà�Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)ÜÈÙ0Üzç�Ù0Ü�Þ�â�Ü)Þ� ÞBÓ�â�ÞBâ��AÞrÕ`ä�ÚBà¯ÞBâ�Úüä"Õ�Ü�ÞBÖ�Þ�Ùªí�àÖBÖëâ�á à�ÜÏÕ�Þ�â��Õ�Ü©íkÞBÓ�â�×.Ù)Ú�á¤Õ¥Þ|à�Ü¤ä�Ó�àç;Ó`Þ�Ó�àÖ�àÖwí�Ù)Ü�â � ÛÕ3ïÙ0æ�Þwà�ÖBÖ�æ�â�Ö/�6èf§AæÏç;Ó ÕEí�àÖ�Õ)í�ù¥Õ�Ü�ÞBÕ0Ý0â"à�ÖwÞ�Ùkã©âÙ0îÏî©Ù�Öëâ1íØÞ�Ù�ÕúÚBÕ�Þ�Ó�â�Úªâ1Õ0Ö�ïúÞBÙúÛ�â�Õ�ÚBÜ»ÛÕ�Ü�þÝ0æ©Õ�Ý0â)ì�ÖëÙØÞ�Ó©Õ¥ÞzùAà�Ú�Þ�æ©Õ�Û�Û�ï Õ�ÜAïAã©Ù�í�ï ä�à�Þ�Óä�â�ã�Õ)ç�ç�â�ÖBÖªç�Õ0Ü â�Õ)Öëà�Û�ïØîÏæ�ã�Û�à�Ö�Ó�à�Ü�×.Ù)Ú�á¤Õ¥þÞ�à�Ù0Ü�ì�ûAÜ�Ù¥ä�Û�â1í�Ý0â Õ�Ü©íªÙ)î�à�ÜÏà�Ù)ÜÏÖ�èüÜ�Ù0Þ�Ó�â�ÚÔÚBâ�Õ)ÖëÙ)ÜZ×.Ù0ÚÔÞ�ÓÏâIÜ�â�â�íz×.Ù0ÚÔá Ù)Ú�âtwÇ1É`Ê�ÇtË¤Ì¥Í©Î�ÐÒÑtÊüàÖÔÞ�ÓÏÕ�Þ Õ0Û¯ÞBÓ�Ù0æ�Ý)ÓzÞBÓ�â`ädâ�ãàÖ,Õ$á`â�í�àÕ ×.Ù)Úöî�æ�ã�Û�à�Ö�Ó�à�Ü�ÝSçtÙ0Ü�ÞBâ�Ü�Þ�ìI×ÒÕ�Ú×.ÚBÙ0á Õ�Û�Ûrà�ÞBÖfçtÙ)Ü)ÞBâ�Ü�ÞBÖ Õ�ÚBâzçtÚBâ�Õ�Þ�â�íøÙ)Ü»Ù0Ú×.Ù0Úrà¯Þ�¨ z ÜÈá`Ù)ÖëÞrç�Õ)Öëâ1Öüí�Ù�çtæ�á`â�Ü)Þ;Õ¥ÞBà�Ù)ÜÈÓÏÕ0ÖÞ�ÙØã�âbÚBât×.Ù)Ú�á¤Õ¥Þ�Þ�â1í Õ�ÜÏí�Õ�ÜÏÕ0Û�ï�Ö�â�í»ã�ât×.Ù)Ú�âà�Þ�ç�Õ0Üðã�â î�æÏã�Û�àÖ�Ó�â�íåÙ)ÜðÞ�Ó�â ä�â�ã�ì Õ�Ü©íâ��AÞBÚBÕ)ç6Þ�à�Ü�ÝöÖëâ�á¤Õ�Ü�Þ�àç ç�Ù0Ü�Þ�â�Ü�ÞBÖI×.Ú�Ù)á Öëæ©ç;Óæ�Ü � ädâ�ão��ÖëÞ�ÚBæÏçtÞ�æ�ÚBâ�í¤í�Ù�ç�æ�á`â�Ü�ÞBÖdá`à�Ý)Ó)Þ�Õ�î�þî�â�Õ�Ú�ÜÏÙ�Þ�Þ�Ù`ã�â�â�Õ)ÖëïªÕ�ÞüÕ�Û�Û�è�~Ö�à�Ü�Ý Õ0Ü â��î�Û�àçtà�ÞUÚ�â�î�Ú�â1Öëâ�Ü�ÞBÕ¥ÞBà�Ù)Ü ÛÕ�Ü�þÝ0æ©Õ�Ý0âªä�à¯ÞBÓØç�Û�â1Õ�ÚIÖ�â�á¤Õ0Ü)ÞBà�ç�Ö�ì�ä�ÓÏâ�ÚBâöÑ;÷¥Ío�Î�ÇtÍ©Î�àÖhÚ�â�î�ÚBâ�Ö�â�Ü�Þ�â1íEâ��î�Û�àçtà�Þ�Û�ï0ì�æÏÖ�æÏÕ0Û�Û�ïEÖëâ�ÞBÖÕ ÓÏÕ0Û¯Þ�ÞBÙ Þ�Ó�â�â1Õ0Ö�â�Ù�×�æÏÖëâ"×.Ù)Ú�á`Ù)ÖëÞwÕ3ù0â�Ú;Õ�Ý)âæÏÖ�â�Ú;Ö�èf�rÖëà�Ü�Ý�w Ó�à�Ý0Ó�â�ÚOw�ÚBâ�îÏÚ�â1Öëâ�Ü)Þ;Õ¥ÞBà�Ù)ÜªÛÕ�Ü�þÝ0æ©Õ�Ý0â1Ö � çt×�èM¦Y�d¤G©-}�ªY«�ì|Ù)Ú¤×.Ù0ÚBá¤Õ�Û�ÛÕ�Ü�þÝ0æ©Õ�Ý0â1ÖS�¤à�ÜSÕúÖ�à�á`à�Û�Õ0Ú á¤Õ0Ü�Ü�â�ÚZÕ0ÖªÞ�Ù�í�Õ3ï�Öä�â�ã¤î�æ�ãÏÛ�àÖëÓÏà�Ü�ÝEÞ�ÙAÙ0ÛÖ�á`à�Ý)Ó�Þ|Þ�Ó�â�Ú�â�×.Ù0ÚBâ�Ü�Ù�Þã�âIÞBÓ�â`ã©â1Ö�ÞEä"Õ3ïfÞ�ÙZÝ0Ù©ì²ã©â1ç�Õ0æÏÖëâ à�Þ àÖ â��Aþî�â�ç6ÞBâ�í Þ�Ù ÞBÓ�ä"Õ�Ú�Þ"î�æ�ã�Û�à�Ö�Ó�à�Ü�Ý¤Õ�ÜÏífÖ�ÓÏÕ�ÚBà�ÜÏÝÓ�àÖ Ù0ÚkÓÏâ�Ú`ûAÜ�Ù¥ä�Û�â1í�Ý0â Õ�ÜÏíöÞBÓ�Ù0æÏÝ0Ó�ÞBÖ í�æ�âÞ�Ùúà¯Þ;ÖªÓ�à�Ý0ÓÏâ�Ú ç�Ù0á`î�Û�â��à�Þ�ï0è¬~íÏí�à¯ÞBà�Ù)ÜÏÕ�Û�Û�ï0ìÉ;Ì0Ñ/s-twÌ��ôöÑ;÷¥Ë��Ì�Î�ÐÒÉ�Ð�u=Ð Î2®¤Ù�×�ÕZÜ�â�ä Ê�ÇtË¤Ì¥Ío�

Î°ÐÒÑ5twÇ1ÉfÚ�â�î�Ú�â1Öëâ�Ü�ÞBÕ¥ÞBà�Ù)Ü ÛÕ�Ü�Ý)æÏÕ�Ý)â ÖëÓÏÙ0æ�Ûíã�â�Ý0æ©Õ�Ú;Õ�Ü�Þ�â�â�í²è¢~Õ3ùAà�Ü�Ý ÓÏÕ0íIÞBÓ�à�Ödí�â�ã©Õ¥Þ�â"×.Ù)Ú|ÕÔ×.â�äøï0â1Õ�Ú;Ö

ÜÏÙ¥ä�ìÿà�Ü,ÞBÓ�â¤á`â�Õ�ÜAä�Ó�à�Û�â`ç�Ù0ÜÏÖ�â�Ü©ÖëæÏÖ�Ö�â�â�á¤ÖÞBÙoã�â�Þ�ÓÏÕ�ÞÔÕ¤ç�Ù0ákãÏà�ÜÏÕ�Þ�à�Ù0ÜÈÙ0×hÞBÓ�â�Þ�ä�ÙªÕ�î�þîÏÚ�Ù�Õ0ç;Ó�â1ÖoçtÙ)æ�Û�í$ÖëÙ)Û�ù)âÈá Ù�Ö�ÞfÙ�×Eà�ÞBÖZí�Ú;Õ3ä"þã©Õ0ç;û�Ö�è ~Öøî�Ù)ÖBÖ�à�ã�Û�âmÖëÙ)Û�æ�Þ�à�Ù0Ü Ù0Ü�âmç�Õ0Üà�á¤Õ�Ý)à�Ü�âªÞ�ÙAÙ0Û�Ö�æ�î�î�Ù0Ú�ÞIà�Ü�Ù)ÚBí�â�ÚEÞBÙbâ�à¯ÞBÓ�â�ÚÕ0ÜÏÕ�Û�ï�ÖëâÈî©Õ�Ý0â1Ö`Þ�Ó©Õ¥ÞfÕ0Ú�âÈÜÏÙ�ÞªÚBâ�î�ÚBâ�Ö�â�Ü�ÞBâ�íà�ÜZÕ{w Öëâ�á¤Õ�Ü�Þ�àç�Õ0Û�Û�ï¤Ú�àç;ÓWwAá¤Õ�Ü�ÜÏâ�Ú1ì�Ù)Ú"Ù��â�Ú�þà�Ü�Ý Ý0Ú;Õ�î�ÓÏà�ç�Õ�Û�à�Ü�ÞBâ�Ú�×ÒÕ0çtâ1Ö � â�í�à¯ÞBÙ0Ú;ÖS�wÞ�Ù î�â�Ù�þîÏÛ�âkÞ�ÓÏÕ�Þ Ö�æ�î�î�Ù0Ú�Þ ç�Ú�â1Õ¥ÞBà�Ü�Ý ÖëæÏç;Ó,Öëâ�á¤Õ�Ü�Þ�àçÚBâ�îÏÚ�â1Öëâ�Ü)Þ;Õ¥ÞBà�Ù)ÜÏÖ � Öëâ�á`à¯þ4��Õ0æ�Þ�Ù)á`Õ�Þ�àç�Õ0Û�Û�ï0èé"Ó�â1Öëâ`Þ�ä�ÙzÖ�ç�â�ÜÏÕ0Ú�à�Ù)Ö�Õ0Ú�â¤ã�Ù�ÞBÓ Öëâ�â�Ü Õ)Ö

à�á`î©Ù)ÚëÞ;Õ�Ü�Þ�ìIÕ0ÖbÛ�Ù)Ü�Ý$Õ0ÖzÞ�Ó�â�ïmçtÙAâ��àÖ�Þ Ù)ÜÞBÓ�â$ä�â�ã�è¯ ù¥Õ�ÚBà�âtÞ�ïðÙ�×zî�ÚBÙ � â1ç6ÞBÖúä�à¯ÞBÓÖ�â�á¤Õ0Ü)ÞBà�ç ÚBâ�î�ÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0Ü�Û�Õ0Ü�Ý0æÏÕ0Ý0â1ÖkÓÏÕ3ù)âÖ�Ó�Ù¥ä�ÜñÞ�ÓÏÕ�ÞøÚBâ�îÏÚ�â1Öëâ�Ü)ÞBà�ÜÏÝßÕ�Û�ÛÈä�â�ã çtÙ0Ü�þÞBâ�Ü�ÞBÖ`à�Ü>w ÓÏà�Ý)Ó�â�Ú¤Ù)ÚBí�â�Ú`Û�Õ0Ü�Ý0æÏÕ0Ý0â1Öw�á`à�Ý)Ó�ÞÜÏÙ�Þ ã�âb×.â�Õ0Ö�à�ã�Û�âbæ�Ü�Û�â�ÖBÖªá`Ù)Ú�â,Õ�æ�ÞBÙ0á¤Õ¥ÞBâ�íÕ0î�î�ÚBÙ)Õ)ç;Ó�â�ÖÔã�â�çtÙ)á`âoÕ3ù3Õ0à�ÛÕ�ãÏÛ�â)è8¢üÙ¥ä�â�ù)â�Ú1ìÕ0ÜWà�Ü©çtÚBâ�Õ0Ö�à�Ü�ÝNÕ0ç�çtâ�î�ÞBÕ�Ü©çtâòÙ0× á Ù)Ú�âßÖëâ�þá¤Õ0Ü)ÞBà�çöÚBâ�îÏÚ�â1Öëâ�Ü)Þ;Õ¥ÞBà�Ù)Ü�ÛÕ�Ü�Ý)æÏÕ�Ý)âöÙ0ÜSÞ�ÓÏâä�â�ã ç�Õ0Ü$ã�â,ÜÏÙ�Þ�àçtâ1í$Õ0ÜÏí�Ö�â�ù)â�Ú;Õ�Ûrà�Ü�à�Þ�àÕ¥þÞBà�ù)â�Ö Õ0à�á Õ¥Þ Öëæ�îÏî©Ù)ÚëÞBà�Ü�ÝUÞBÓ�â�áZè §Aâ�ù)þâ�ÚBÕ0ÛdîÏÚ�Ù � â�ç6ÞoÕ0Ú�â à�Ü�à�Þ�àÕ¥ÞBâ�íØä�Ù0ÚBÛ�íAþ�ä�àí�âoÞBÙÖ�æ�î�î�Ù0Ú�ÞEÞBÓ�âzÊ�Ç�Ë`Ì�ÍÏÎ°ÐÒÑ{twÇ1É �B° « ùH¢Y±k�²�²³°ì° ¢üâ�Üµ´�´-³°ì�âtÞ;ç�è%�6ì�ÛÕ�Ü�Ý)æÏÕ�Ý)â�Ö�ì�â��ÞBâ�ÜÏÖ�à�Ù0ÜÏÖ�Ù)ÜÛÕ�ÜÏÝ0æÏÕ0Ý0â�Ö¤Õ0ÜÏí��æ�â�ÚBïØÛ�Õ0Ü�Ý0æÏÕ0Ý0â1ÖoÕ�ÚBâzí�âtþ�©Ü�â�í ��° ¶ � ´·´-³�ì ° «J¢�¢�¸X´�´*³¹�ñÕ0ÜÏí ÞBÙAÙ0ÛÖ×.Ù)Ú � Öëâ�á à�þ4��Õ�æ�Þ�Ù0á¤Õ�Þ�àçªçtÙ)Ü�Þ�â�Ü�Þ â��AÞ�Ú;Õ0çtÞ�à�Ù0ÜÕ0Ú�âWà�á`î�Û�â�á`â�Ü�Þ�â1í �B° ¶,º ²·²*³�ì ° Ù·�¥ � ´·´*³°ì° Õ�Y´�´�³¹�6è

» â�ù)â�Ú�Þ�Ó�â�Û�â1Ö�Ö�ì�ÛÕ0ÖëÞwï0â1Õ�Ú;Ö�Ù�×²Ú�â1Öëâ1Õ�Ú;ç;Óÿì¥ã�â~à�Þç�Ù0ÜÏí�æÏç6ÞBâ�í¤Ý)Ù¥ù0â�ÚBÜ�á`â�Ü�ÞdÖ�æ�î�î�Ù0Ú�Þ�â1íoÙ)Údî�ÚBà�þù¥Õ�Þ�â0ì ÓÏÕ3ù0â¤Ö�Ó�Ù¥ä�Ü,â�á`â�ÚBÝ0à�Ü�ÝfÞ�â1ç;Ó�Ü�Ù0Û�Ù0Ý)à�â1ÖÞBÓÏÕ¥Þ1ìEÕ�Û�Þ�Ó�Ù)æ�Ý0Ó Ù�× ÞBâ�Ümî�ÚBâ�í�à�çtÞ�â�íSÕ0ÜÏíSÕ�Û�þÚBâ�Õ)í�ï¤à�Ü�à¯ÞBà�Õ�Þ�â1íf×.Ù0Ú�á¤Õ�ÜAïªï)â�Õ0ÚBÖ"Õ�Ý)ÙÏìAÙ0ÜÏÛ�ïÚBâ�ç�â�Ü�Þ�Û�ïòæ�ÜÏÛ�â1Õ0Ö�Ó�â�íðÖëÙ)á`âØÙ0× ÞBÓ�â»î�Ù¥ädâ�ÚÞBÓÏÕ¥ÞÈÛ�à�â1Öfà�ÜSÕøçtÙ0áIã�à�ÜÏÕ¥ÞBà�Ù)Ü$Ù�×`Öëâ�á¤Õ�Ü�Þ�àçÕ0ÜÏÕ�Û�ï�ÖëàÖ�Õ0ÜÏíkí�à�ÖëÞ�ÚBà�ãÏæ�Þ�â1íEà�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0ÜIÖëï�ÖëþÞBâ�á¤Ö�è,|rÜ�â�Ù0× ÞBÓ�â�Ö�â Þ�ÙAÙ0ÛÖ�àÖüí�â�ù0â�Û�Ù0î�â�ífà�ÜÞBÓ�âÈî�ÚBà�ù¥Õ�Þ�âzÖ�â�ç6ÞBÙ0Úoí�æÏÚ�à�Ü�ÝöÞ�ÓÏâzîÏÕ0ÖëÞ`×.Ù0æ�Úï)â�Õ0ÚBÖ�ì©Õ0ÜÏízÓÏÕ)Ö~Ý)Ú�Ù¥ä�Ü á`Õ�Þ�æ�ÚBâkâ�Ü�Ù0æÏÝ0ÓÈÞBÙÖ�â�ÚBù0â Õ0Ö ã�Ù�Þ�Þ�Ù0ávÞ�â�ç;ÓÏÜ�Ù0Û�Ù0Ý)ïÈà�ÜöÕfù3Õ0Ú�à�âtÞ�ïÙ0×�î�Ú�Ù�í�æ©ç6ÞBÖ�ì~Õ)Öªä�â�Û�ÛrÕ)Öªà�Ü$çtÙ0þ°Ù)î©â�ÚBÕ�Þ�à�Ù0ÜîÏÚ�Ù � â�ç6Þ;ÖÿÙ)ÜkÕ� wæ�Ú�Ù)î©â1Õ�ÜEÛ�â�ù)â�Û � çt×�è�ÞBÓ�â^|rÜ�þé Ù·± ÜÏÙ¥ä�Û�â1í�Ý0ârî�Ú�Ù � â�çtÞ ° «hùH¢�±k�²·²*³¹�6è

¼2

AIFB AIFB

«hà�Ý)æ�ÚBâ �·½ rdÙ0ÚBâ\~ÜÏÕ�Û�ï�Öëâ�ÚüçtÙ)á î�Ù0ÜÏâ�Ü�ÞBÖOwg�Èà¾�Èà�Úü×.æ�ÜÏç6ÞBà�Ù)ÜÏÕ�Û�à�Þ�ï0è

¿ À �üÀhÄ�Âf� � Á�� ÂÁ ÁJ��ÃxÄxÅÇÆÈÄÉÅËÊÍÌ À�ÎªÀ�ÁG� �

«�Ù)ÚrãÏæ�à�Ûí�à�Ü�Ýfæ�îÿì�æ�Þ�à�Û�à�Ö�à�ÜÏÝfÕ0ÜÏíÈá¤Õ�à�Ü�ÞBÕ�à�Ü�þà�Ü�Ý�Þ�ÓÏâ~Öëâ�á¤Õ�Ü�Þ�àç"ädâ�ãÿì)Þ�Ó�â�Ú�âüÕ0Ú�âüÕ ù¥Õ�ÚBà�âtÞ�ïÙ�×~ÞBÕ0Ö�û�ÖEÞBÓÏÕ¥Þ¤Õ�ÚBâoÞBÙbã�âZí�â�Õ0Û¯Þ`ä�à¯ÞBÓÿèÏüÛ�ÛÙ�×wÞ�Ó�â1ÖëâIÞBÕ)Öëû�ÖY�ÏÜÏízÞ�Ó�â�à�Új��Ì�ÐÊ�÷¥ÍúôÿÅ Ç�Î2��ÇÔà�ÜÞ�ÓÏâ`×ÒÕ0ç6Þ�ÞBÓÏÕ¥ÞIî©â�Ù0î�Û�â`Ü�â�â1í ÞBÙÈÝ0â�ÞEÙ0Ü,Þ�Ù)îÙ�×dÞ�Ó�â¤à�Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)ÜbÙ¥ù)â�Ú�¡ÏÙ¥ä Þ�Ó�â�ï Ý)âtÞ�Ù�× þ×.â�ÚBâ�íØÞBÙ�Þ�ÓÏâ�áZèýé"Ó�àÖªÓÏÙ0Ûí�Ö¤×.Ù)Úoà�Ü©í�à�ùAàíAþæÏÕ0Û�Ö,Û�â�Õ�ÚBÜ�à�Ü�ÝÏì Ù0ÚBÝ)Õ0Ü�àÖëà�Ü�Ý$Õ0ÜÏíýà�Ü�Þ�â�ÚBÕ)ç6Þëþà�Ü�ÝoÙ0ÜZÞBÓ�âkä�â�ãzÕ)Ö�ákæ©ç;ÓzÕ0Ö�×.Ù)ÚüÙ0ÚBÝ)Õ0Ü�àÖ�Õ�þÞ�à�Ù0Ü©Ö�Þ�Ó©Õ¥ÞIä"Õ�Ü�ÞEÞBÓ�â�à�Úkâ�á`î�Û�Ù¥ï)â�â1Ö Þ�Ùbákæ�þÞ�æ©Õ�Û�Û�ïöã�â�Ü�â��Þ¤Ù�×~Õzã�âtÞ�Þ�â�Ú`í�à�ÚBâ�çtÞ�â�íÿì�ã©â�ÞëþÞ�â�Ú æ�Ü©í�â�Ú;Ö�Þ;Õ�ÜÏíÏÕ�ã�Û�â Õ�ÜÏíSá`Ù)Ú�âúç�Û�â1Õ�Úzà�Ü�þ×.Ù0ÚBá¤Õ¥ÞBà�Ù)ÜIÕ�Ü©íkûAÜ�Ù¥ä�Û�â�í�Ý)âdÖ�ÓÏÕ�ÚBà�Ü�Ý~×ÒÕ)çtà�Û�à�Þ�ï�B° ¶,º §g¸�²�²*³¹�6è|rÜ�ÞBÓ�âÈÞBÓ�â�Ù)Ú�â�Þ�àç�Õ�ÛüÖ�àí�âZÞ�Ó�â ÊtÇ�Ë¤Ì¥ÍÏÎ°ÐÒÑtwÇ1ÉZà�Ö í�â��ÏÜÏâ�íSÕ0ÖfÞBÓ�â�á`â�Õ�Ü©ÖZã�ï�ä�Ó�àç;ÓÞ�ÓÏà�ÖfçtÙ)æ�Û�í ã©â,Ú�â1Õ0ç;Ó�â1í²èÐ�Þ Þ�Ó�â Þ�â�ç;ÓÏÜ�Ù�þÛ�Ù0Ý0àç�Õ0ÛÿÖëàí�â�ÞBÓ�â$r^|\}Z ,|\}~�Y� Þ�ÙAÙ0Û�Ö�âtÞrà�Öí�â�ÏÜ�â�ífÕ0ÖwÞ�ÓÏâ Ö�â�ÚBù0â�Ú|×.Ù)Ú"Ö�â�á¤Õ0Ü)ÞBà�çÔÕ�Ü©Õ�Û�ï)þÖ�â�Ö �B° ¶,º ²·²*³�ì ° ¶ ´�´�³¹�6èwé"Ó�â1Öëâ�Õ0ÜÏÕ�Û�ï�Öëâ1Ö�Õ�ÚBâî�â�Ú�×.Ù0ÚBá â1íZãAï�r^|\}Z ,|\},��zÖOw©ç�Ù0ÚBâEçtÙ0á þî�Ù0Ü�â�Ü)Þ1ìhÕZÖ�â�á¤Õ�Ü�ÞBà�ç¤Õ0ÜÏÕ�Û�ï�Öëâ�Ú ç�Ù0á`î�Ù0Ü�â�Ü)Þç�Õ0Û�Û�â�íÂ�Èàe�zà�Ú1è<�»Ó�â�Ú�â1Õ0Ö��Èàe�zà�Úhà�Ö�Þ�ÓÏâ�ç�Ù0ÚBâÕ�Ü©Õ�Û�ïAÖ�àÖhç�Ù0á`î�Ù0Ü�â�Ü)Þ�à�ÜkÞBÓ�âYr^|\}Z ,|\}~�Y�

ÞBÙAÙ0Û0Ö�âtÞ1ì�ÞBÓ�àÖÿù0â�Ú�ïÔçtÙ)á î�Ù0ÜÏâ�Ü�Þ ç�Õ0Ü�ã©âwæÏÖëâ1íà�ÜzÕ¤ù¥Õ�ÚBà�âtÞ�ïªÙ0×|Öëâ�ÞëÞ�à�Ü�Ý�Ö~í�æ�âkÞ�Ùoà�ÞBÖrÕ�ã�à�Û�à¯Þ�ïÞBÙ¤â��AÞ�Ú;Õ0çtÞüçtÙ0Ü�ÞBâ�Ü�ÞBÖ�ì�Ý0â�Ü�â�Ú;Õ¥ÞBâEÕ¤Öëâ�á¤Õ�Ü�Þ�àçÚBâ�îÏÚ�â1Öëâ�Ü)Þ;Õ¥ÞBà�Ù)Ü,Ù0×üÞ�ÓÏâfç�Ù0ÜÏç�â�î�Þ;Ö � à�á`î�Û�à�ç�à¯ÞÕ)Ö|ä�â�Û�Û�Õ)Ö|â�AîÏÛ�àçtà�Þ�Û�ï`î�Ú�â1Öëâ�Ü�Þdà�Ü¤Þ�â��Þ;ÖS�tìAÚ�â�þÛÕ¥ÞBà�Ù)ÜÏÖ�Ó�à�î©Ö~Õ�Ü©íÈÚ�Ù)Û�â1Ö�èY�Èàe�zà�Ú;Öw©×.æ�ÜÏçtÞ�à�Ù0Ü�þÕ0Û�à�Þ�ï à�ÖbãÏÕ0Ö�â�íSÙ)Ümá`Ù0ÚBâ�×.Ù0ÚBá¤Õ�ÛIç�Ù0á`î�æ�þÞ;Õ¥ÞBà�Ù)ÜÏÕ�ÛdÛ�à�Ü�Ý0æÏà�ÖëÞ�àç�Ö�èbé"Ó�âfçtÙ)á îÏæ�ÞBÕ�Þ�à�Ù0ÜÏÕ0ÛÛ�à�Ü�Ý0æ�àÖëÞ�àçøîÏÕ�Ú;Õ0í�à�Ý0á � ç6×�èÑ�ÏÝ)æ�Ú�â ¼ Õ0ÜÏí° ¶ ´·´�³��hÞ;Õ�û)â�Ö�î�ÛÕ0ç�â�Ù0Ü`ÞBÓ�Ú�â�â�á¤Õ�à�Ü¤Û�â�ù0â�ÛÖ ½ÞBÓ�â5��õÏ÷¥Í�÷-u�÷Sv�ÐÒÑ;Ìu&u�Ç�Ò¥Ç�u¯ì�tw÷-��ôÓu�ÇÒ3Çu=ì¤Ê�Ç�Íµ�Î�Ç�Í�Ñ;Çbu�ÇÒ3Çu�Õ�ÜÏíªÞ�ÓÏâIÊ�Ô��Õ��Ì-��Ê�Ç�ÍÏÎ�ÇtÍ©Î�ÐÒÌu�u�Ç�Ò¥Çu� ù0â�Ú�ïoÖ�à�á`à�Û�Õ0Ú"Þ�Ù`ÞBÓ�âoô¥ÐÊ�Ñ;÷-ÔH�;Ê�Ç$u�Ç�Ò¥Çu �6è|rÜfÞ�Ù)îÈÙ�×hÞ�Ó�àÖüãÏÕ)Öëàç ×.æ�Ü©ç6Þ�à�Ù0Ü©Õ�Û�à¯Þ�ï)ì�Þ�ÓÏâ

Õ0ÜÏÕ�Û�ï�Öëâ�ÚzçtÙ)á î�Ù0ÜÏâ�Ü�ÞbÓÏÕ)ÖÈÞBÓ�â�Õ0ã�à�Û�à�Þ�ï$ÞBÙç�Ù0á`îÏÕ0Ú�âzÖ�æÏç;Ó»ÚBâ�î�ÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0ÜÏÖ`Ù0×Eá`â�Õ�Ü�þà�Ü�Ý©ì�à�Ü`Ù)ÚBí�â�Ú�Þ�Ù\�ÏÜÏí`Ù)æ�Þ|Ó�Ù¥äØÖëà�á`à�ÛÕ�Ú�ÞBÓ�â�ïÕ0Ú�â)è ¶ Õ)Öëâ1íoÙ0ÜªÞ�Ó�âÔÚBâ�Ö�æ�Û¯Þ;Ö�Ù�×�ÞBÓ�àÖ�Ö�à�á à�ÛÕ�Ú�þà�Þ�ïIÕ0ÜÏÕ�Û�ï�ÖëàÖ�ì¥Þ�Ó�âY�Èàe�zà�Ú�çtÙ0á`î�Ù0Ü�â�Ü�Þ|Ù��â�Ú;ÖÕ)í�ùAà�ç�â�Ù0Ü ä�Ó�àç;ÓEí�Ù�çtæ�á`â�Ü)Þ;ÖÿÕ�ÚBâ�á`Ù)ÖëÞÿî©â�ÚëþÞBà�ÜÏâ�Ü�Þ�Þ�Ù�Õ Öëî�â�ç�à��©ç"Õ0ÜÏÕ�Û�ï�Öëâ1í�Þ�â�AÞ�ì)Õ0ÜÏíIÚ�â�þÞBæ�ÚBÜbÞBÓ�Ù)Ö�âoîÏÕ0ÚëÞ;Ö�à�ÜöÞBÕ�ÚBÝ0â�Þ�â1í,í�ÙAç�æ�á`â�Ü�ÞBÖá`Ù�Ö�ÞoÖ�à�á`à�Û�Õ0Ú`Þ�Ù Õ,îÏÕ�Ú�Þ�àçtæÏÛ�Õ0Ú¤à�Ü�îÏæ�ÞoÞ�â�AÞ� ç6×�èJ�ÏÝ)æ�ÚBâ � �6è~ÖöÖ�ÙAÙ0ÜßÕ)Ö,â�ákã�â�íÏí�â�íßà�ÜßÞ�ÓÏâÓr^|\}�þ ,|\},�� ÞBÙ�Ù)ÛÈÖëâ�Þ�ìÈÞBÓ�âÖ�zàe�Èà�Ú çtÙ)á`î©Ù0þÜÏâ�Ü�Þ�à�ÖüÕ0ã�Û�ârÞBÙ¤æ�Ü�Û�â�Õ)ÖëÓfà¯Þ;Ö�ÚBâ�Õ�ÛÿÖëÞ�ÚBâ�ÜÏÝ�Þ�Ó©ÖÕ0ÜÏíPÖëâ�Ú�ù)â�ÖòÕ)ÖSÞ�ÓÏâ×w ãÏÚBÕ0à�ÜÏÖOwzÙ0×øà�Ü�Þ�â�Û�Û�à¯þ

Ø3

AIFB AIFB

« à�Ý0æÏÚ�â ¼H½ � Ú;Õ�á`á¤Õ¥ÞBà�ç�Õ�ÛÿÕ�Ü©Õ�Û�ïAÖ�àÖdÙ0×�ÞBâ��AÞBÖ�è

Ý0â�Ü�Þ Õ�Ý)â�Ü�ÞBÖkÝ)Õ�Þ�Ó�â�Ú�à�Ü�ÝÈà�Ü)ÞBâ�Û�Û�à�Ý0â�ÜÏçtâ Ù�×~Õ0Û�ÛÖ�Ù0Ú�ÞBÖ Ù0Ü�Þ�Ó�âZä�â�ãÿì�ã�âfà�Þoá`â�í�àç�Õ0Û�ûAÜ�Ù¥ä�Û¯þâ�í�Ý0â�×.Ù)ÚwÕEÖ�î©â1çtà��©ç�Ü�â�äØîÏÚ�Ù0þ°Û�â�æ�ûAà�Ü�â�ãÏÕ)Öëâ1íá`â�í�àçtà�Ü�â)ì ÕøÖëÞ�æÏí�â�Ü�Þzä�Õ0Ü�Þ�à�Ü�ÝúÞBÙ»çtÙ0Û�Û�â�ç6Þá¤Õ¥ÞBâ�ÚBà�Õ0Û�×.Ù0ÚªÕúçtÙ0æÏÚBÖ�âZÙ0Úoã�æ©Öëà�Ü�â�ÖBÖoà�Ü�Þ�â�Û¯þÛ�à�Ý)â�ÜÏç�â Õ0ã©Ù)æ�Þüî�Ù�Þ�â�Ü�Þ�àÕ�Ûÿá¤Õ�ÚBû0â�Þ�Ù)î�î�Ù0Ú�Þ�æ�þÜ�à�Þ�à�â�Ö�Ù0ÚdÞ�ÚBâ�Õ�ÞBÖ�è�«�Ù0ÚdÞ�ÓÏâÔà�Ü�Þ�â�Û�Û�à�Ý)â�Ü�Þ�Õ0Ý0â�Ü)ÞÖBçtâ�Ü©Õ�ÚBà�Ù©ì,Õåä�â�ã Öëâ�Ú�ù)â�Ú$ç�Ù0á`î©Ù)Ü�â�Ü�Þ1ìöÕí�Õ�ÞBÕ�ã©Õ0Ö�â`Ö�â�ÚBù0â�Ú�ì²á`àÖ�Ö�à�Ù0Ü�Ö�ç;ÓÏâ�í�æ�Û�â�Ú;Ö�Õ�Ü©íÕzç�Û�à�â�Ü�Þ Ö�â�ÚBù0â�ÚEçtÙ)á î�Ù0ÜÏâ�Ü�Þ Õ0Ú�â¤à�ÜÏçtÛ�æÏí�â1í²èüÜÏÙ�Þ�ÓÏâ�Ú ç�Ù0á`î�Ù0Ü�â�Ü)Þÿà�Ü Þ�ÓÏâ,r^|\}Z ,|\}~�Y�Þ�ÙAÙ)Û0Öëâ�Þ�àÖ�Þ�Ó�â~§Aæ�á`á¤Õ�ÚBà�Ö�â�Ú1ì�ä�ÓÏà�ç;Ó àÖ�ç�Õ0îÏÕ¥þã�Û�âIÙ0×dá¤Õ�ûAà�Ü�Ý Ö�æ�á`á¤Õ�ÚBà�â1Ö~Ù0×wÞBâ��AÞBÖÔãÏÕ)Öëâ1íÙ0ÜÈÕj�Èà¾�Èà�ÚrÖ�æ�î�î�Ù0Ú�Þ�â1ífÕ0ÜÏÕ�Û�ï�ÖëàÖ�Ù0×hä�Ó�â�Ú�âÞ�ÓÏâ,ÚBâ�Õ0Û à�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Ü�ç�Ù0Ü�Þ�â�Ü�ÞBÖfÚ�â1Öëàí�â,à�ÜÞ�ÓÏâªí�Ù�çtæ�á`â�Ü)Þ1è8~Û¯ÞBâ�ÚBÜÏÕ¥ÞBà�ù)â�Û�ïbÖ�æÏç;Ó�Öëæ�á þá¤Õ�ÚBà�â1ÖIç�Õ�Ü�ã©âZá¤Õ0í�âfà�Ü)ÞBâ�ÚBâ�ÖëÞëþ�í�ÚBà�ù)â�Üÿì�ãAï

æ©Öëà�Ü�ÝÈÕ�Üöà�Ü)ÞBâ�ÚBâ�ÖëÞEî�ÚBÙ�ÏÛ�â � à�ÜöÞ�Ó�â`×.Ù)Ú�á Ù�×ÕøÜ©Õ¥Þ�æÏÚBÕ0ÛEÛ�Õ0Ü�Ý0æ©Õ�Ý0â,ÞBâ��AÞ/� Õ�ÜÏí Ý0â�Ü�â�Ú;Õ¥ÞBâÖ�æ�á`á¤Õ�ÚBà�â1Ö"Õ0ç�ç�Ù0Ú;í�à�Ü�ÝkÞBÙ Þ�Ó�â1Öëâ)è}üâ�¡Ïâ1ç6Þ�à�Ü�Ý,Ù0Ü�ÞBÓ�âÈÕ�ã�Ù¥ù0â)ì�à¯Þ¤ç�Õ0Üúã�âZÖBÕ�àíÞBÓÏÕ¥ÞZÞBÓ�Ú�â�âbá¤Õ0à�ÜSÖ�ç�â�ÜÏÕ0Ú�à�Ùµw Öª×.Ù0ÚÈÕ�î�îÏÛ�àç�Õ�þÞBà�Ù)ÜòÙ�×Ù�Èàe�zà�Ú�Õ�ÚBâ á`Ù)ÖëÞöî©â�ÚëÞBà�Ü�â�Ü�Þ ½ Õq�ÇSÚ)Î?�BÌ0Ñ�Î�ÐÒ÷�ÍfÙ0×�à�Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)Ü,×.Ú�Ù)áWÞ�â��Þ;Ö�×.Ù)ÚãÏæ�à�Ûí�à�Ü�Ý`ûAÜ�Ù¥ä�Û�â1í�Ý0ârãÏÕ0Ö�â�Ö�ìAão�Y��Ç�Î?�6ÐÒÇÒ3ÌuAÙ�×à�Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)Üø×.Ú�Ù)á Ù�Þ�ÓÏâ�ÚZÖëÙ)æ�ÚBç�â�Ö � Ö�â�Õ�Ú;ç;ÓÖBçtâ�ÜÏÕ�ÚBà�Ù)Ö/��Õ�ÜÏíbçO�rÖ�ÞBÚBÕ�Þ�â�Ý0à�â�Ö~Þ�ÙZçtÙ0á`îÏÕ)ç6Þ1ìùAàÖëæ©Õ�Û�à�Ö�ârÕ0ÜÏíªí�àÖBÖëâ�á à�ÜÏÕ�Þ�âÔà�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Ü¤ÞBÙî�â�Ù)î�Û�â � í�àÖ�Ö�â�á`à�ÜÏÕ¥ÞBà�Ù)ÜòÕ�ÜÏí ÜÏÕ3ùAà�Ý)Õ�Þ�à�Ù0Üo�tè�»à�Þ�ÓNÞ�Ó�âýÊtÇ�Ë¤Ì¥ÍÏÎ°ÐÒÑÛtwÇ�Éúî�Ó�à�Û�Ù)Ö�Ù0î�ÓAïðÙ�×Õ0Ü â�AîÏÛ�àçtà�Þ�Û�ïzÚBâ�î�ÚBâ�Ö�â�Ü�Þ�â1í Ö�â�á¤Õ�Ü�ÞBà�ç�ÖÔÕ)Ö ÕÝ)à�ù)â�Ü � }Zª\«f©�| z ¤��tì²Þ�ÓÏâoÖ�ç�â�ÜÏÕ0Ú�à�Ù)ÖÔão� Õ0ÜÏíç��ã�â�ç�Ù0á`ârÛ�â�ÖBÖdà�á`î�Ù0Ú�ÞBÕ�Ü�Þd×.Ù0ÚdÞ�Ó�â çtæ�ÚBÚBâ�Ü�Þí�à�ÖBçtæÏÖBÖ�à�Ù)Ü Õ0ÜÏí ä�â|ä�à�Û�Û)Þ�Ó�â�Ú�â�×.Ù0ÚBâ�Ù0Ü�Û�ïrîÏÚ�Ù0þ

Ü4

AIFB AIFB

«hà�Ý0æ�ÚBâ Øg½ ~ÜÙ¦\�¥¤öâ��î�Ù0Ú�Þ�ãÏÕ0Ö�â�í Ù0ÜZÕj§Aâ�á�§�ÞBÚ�æÏç0è

ùAà�í�âfÖ�Ó�Ù0Ú�Þ`â��Õ�á`î�Û�â�ÖIÙ�×rÞ�Ó�â�áfèÝ«�Ù�çtæÏÖ`Ù0×Þ�ÓÏà�ÖdÞ�â�AÞ�ä�à�Û�Û²ã�âÔÙ0ÜoÞBÓ�âÔÝ0â�ÜÏâ�Ú;Õ¥Þ�à�Ù0Ü¤Ù0×�â��Aþî�Û�à�ç�à¯Þbû�ÜÏÙ¥ä�Û�â1í�Ý0â Õ0ÜÏíSà�Ü�×.Ù0ÚBá`Õ�Þ�à�Ù0Ü ×.ÚBÙ0áÕØÖëî�â�ç�à��©ç Þ�â�AÞ�ì ÖëÙ�Þ�ÓÏÕ�Þfà�Þzç�Õ0Ü$ã�â,æ©Öëâ1í×.Ù0Ú ã�æ�à�Ûí�à�ÜÏÝbûAÜ�Ù¥ä�Û�â�í�Ý)âªãÏÕ)Öëâ1ÖIÕ0ÜÏíÏ��æ�â1Ö�þÞ�à�Ù0ÜfÕ�ÜÏÖ�ä�â�ÚBà�Ü�Ý � ç6×�èf}�ªY«Þ��æ�â�Ú�ï`ÛÕ�Ü�Ý)æÏÕ�Ý)âÕ�Ü©íªÞBÙAÙ0ÛÖ ° ±kr~ ��´·´*³��tèX |ù)â�Ü�Þ�æ©Õ�Û�Û�ïoÝ)â�Ü�â�ÚëþÕ¥ÞBâ�íöûAÜ�Ù¥ä�Û�â�í�Ý)â`ãÏÕ0Ö�â�ÖEç�Ù0Ü�ÞBÕ0à�ÜöÚBâ�Ö�æ�Û�ÞBÖEÙ0×Ö�â�á¤Õ�Ü�Þ�àç�Õ0Û|Õ�ÜÏÕ0Û�ï�Ö�à�Ö Ù0× � ädâ�ão��í�Ù�ç�æ�á`â�Ü�ÞBÖÕ�Ü©íEÞBâ�ç;Ó�Ü�à¾��æ�â�Ö�Þ�Ù�ßëá`à�Ü�âOàrÞ�ÓÏâ�æ�ÜÏí�â�Ú�Û�ïAà�Ü�ÝçtÙ)ÜÏçtâ�î�ÞBÖ�Õ0ÜÏífÚ�â�Û�Õ�Þ�à�Ù0ÜÏÖ�è

á,âBã äæåJç^èBé�ê�ëÕì�éXíHî<é�íïîWð,ñ\òBè�ë�èCí

«�Ù)Ú»Þ�ÓÏâÐ��æ�â�ÖëÞ�à�Ù0ÜWÕ�Ü©Öëä�â�ÚBà�ÜÏÝðÖBçtâ�ÜÏÕ�ÚBà�Ù�Ö�ìã�æ�Þ�â�ù0â�Üo×.Ù0Ú�ù�àÖ�æÏÕ�Û�à�ÖBÕ¥ÞBà�Ù)ÜoÙ�×hçtÙ0Ü�ÞBâ�Ü�ÞBÖw×.Ù0Úâ�Õ)Öëï�Ý0Ú;Õ�î�ÓÏà�ç�Õ�Û�à�Ü)ÞBâ�ÚBî�ÚBâtÞBÕ�Þ�à�Ù0Üÿì ÞBÓ�âúç�Ù0Ü�þÞ�â�Ü�ÞkÙ�×�ÞBÓ�â¤Þ�â�AÞBÖ�×.Ù)æ�ÜÏí�Ù)Üö×�è â)èfÞ�Ó�â ädâ�ãÖ�Ó�Ù0æ�ÛíZã�âká¤Õ0í�âIâ��î�Û�à�ç�à¯Þ1è~é"Ó�â�Ú�âkÕ0Ú�âEÖ�â�ù�þâ�Ú;Õ�ÛAä"Õ3ï�Ö Ù0×©î�â�Ú�×.Ù0ÚBá`à�Ü�Ý�ç�Ù0Ü�Þ�â�Ü)Þ|Õ0ÜÏÕ�Û�ï�ÖëàÖ�ìÕ�Û�Û�ÓÏÕ3ùAà�ÜÏÝ¤Þ�Ó�â�à�Ú Ù¥ä�Ü,í�â��ÏÜÏà¯ÞBà�Ù)Ü Ù�×rË`Ç;Ì¥Ío�Ð Íqv0è�~Ü Ù�× ÞBâ�Ü»×.Ù)æ�ÜÏí�Õ�î�î�ÚBÙ)Õ)ç;ÓØÞ�Ùúç�Ù0Ü�þÞ�â�Ü�Þ Õ0ÜÏÕ�Û�ï�ÖëàÖüàÖüÞBÓ�â`Ö�Þ;Õ¥Þ�àÖëÞ�àç�Õ�Û�Õ�îÏî�Ú�Ù�Õ0ç;Óÿè

z ÜmÖ�æÏç;ÓSÕ�î�î�ÚBÙ)Õ)ç;Ó�â�Ö�ìÔädÙ)ÚBí�ÖZÕ0Ú�âöÜ�Ù0ÞzÚ�â�þÝ�Õ�Ú;í�â�íSÕ)ÖzÚBâ�î�ÚBâ�Ö�â�Ü�Þ�à�Ü�Ý»ÚBâ�Õ0Û¯þ�ädÙ)Ú�Ûí Õ�Ú�Þ�à�þ×ÒÕ)ç6Þ;ÖöÙ�×fÖ�î�â�çtà��©çøÖ�Ù0Ú�ÞBÖ�ì ãÏæ�ÞúÕ�ÚBâøá â�Ú�â�Û�ïÖ�â�â�ÜòÕ0Ö îÏÕ¥Þ�Þ�â�ÚBÜÏÖ ä�à¯ÞBÓßÖëÞBÕ¥ÞBà�ÖëÞ�àç�Õ0Ûkî�ÚBÙ0î�þâ�ÚëÞBà�â1Ö � ×.ÚBâO��æ�â�ÜÏçtà�â�Ö~Õ�ÜÏíZçtÙ0þ°Ù�ç�çtæ�Ú;Õ�ÜÏç�âÔ×.Ú�â�þ��æ�â�ÜÏçtà�â�Ö/�6è éwïAî�àç�Õ0Û�Û�ïýÕ�ÜåÕ)í�ù¥Õ�Ü�ÞBÕ0Ý0âØÙ�×Ö�æÏç;ÓªÕ0ÜªÕ�îÏî�Ú�Ù�Õ0ç;Ó¤àÖ|ÞBÓÏÕ¥Þ"à�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0ÜªÚ�â�þÞBÚ�à�â�ù¥Õ0Ûdç�Õ0Ü�ã©â á¤Õ0í�âoÚBâ�ÛÕ¥ÞBà�ù)â�Û�ïbÛÕ�ÜÏÝ0æÏÕ0Ý0âà�ÜÏí�â�î©â�ÜÏí�â�Ü)Þ � î©Õ¥ÞëÞBâ�ÚBÜýá¤Õ¥Þ;ç;Ó�à�Ü�Ý$àÖ æÏÜ�à¯þù)â�Ú;Ö�Õ0Û¾�tì�Õ�ÜÏíZàÖüà�á`î�Û�â�á`â�Ü�Þ�â1íZÚBÕ�Þ�Ó�â�Ú~çtÙ)á þîÏæ�ÞBÕ�Þ�à�Ù0ÜÏÕ0Û�Û�ï�â�oçtà�â�Ü�Þ�è z ÜÏÖ�ÞBâ�Õ)í»Ù�×�æÏÖ�à�ÜÏÝîÏæ�Ú�âöÖëÞBÕ¥ÞBà�ÖëÞ�àç�Õ0Û�á`âtÞ�ÓÏÙAíÏÖ�ìbó�â�çtÞ�Ù)Ú�§�îÏÕ0ç�â�zÙAí�â�ÛÖzî�Ù)ÖBÖëà�ã�Û�ï�çtÙ)ákã�à�Ü�â1íSä�à¯ÞBÓòÜ�â�æÏÚBÕ0ÛÜÏâtÞ�Þ�â1ç;Ó�Ü�Ù0Û�Ù0Ý)ï Ù0Ú�Ý0â�Ü�âtÞBà�ç�Õ�Û�Ý0Ù)Ú�à�Þ�Ó�á¤Ö�ì�Õ�ÚBâÕ0Û�Ö�Ù æÏÖ�â�íÿè á¤Õ3ï0Ù0Úfí�Ú;Õ3ä�ãÏÕ0ç;û�Ù�×kÖ�æÏç;ÓÕ0î�î�ÚBÙ)Õ)ç;Ó�â�Ökà�ÖIÞ�ÓÏâ ×ÒÕ)ç6Þ Þ�ÓÏÕ�Þ¤â�Û�â�á`â�Ü�ÞBÖ à�Üä�Ù0Ú;íAþ�ù0â1ç6Þ�Ù)ÚBÖ�Þ�ï�îÏà�ç�Õ�Û�Û�ïUÓÏÕ3ù)â Þ�Ùðã�â â�AþÕ)ç6Þ"á¤Õ¥ÞBç;ÓÏâ�Ö�ì�ç�Õ�æ©Öëà�Ü�Ý`çtâ�Ú�ÞBÕ0à�Ü ädÙ)ÚBío×.Ù0ÚBá¤ÖÜÏÙ�ÞÈÞ�Ù$ã©â ÚBâ�ç�Ù0Ý)Ü�à�ê�â�í Õ0ÖÈã�â�à�Ü�Ý$Ö�à�á à�ÛÕ�Ú1ìâ�ù0â�Üýà�× ÞBÓ�â�ïßî�Ú�à�ÜÏç�à�îÏÕ0Û�Û�ïòÕ�ÚBâ � çt×�èqîÏÛ�æ�þÚ;Õ�Û�Õ0ÜÏíúÖ�à�ÜÏÝ0æ�ÛÕ�Ú ×.Ù0ÚBá¤Ö Ù�×rä�Ù0Ú;í�Ö�ìwí�à��â�Úëþâ�Ü�ÞkÞ�â�ÜÏÖ�â�ÖEÙ0×üù0â�Ú�ãÏÖ�ì�âtÞ;ç�è%�6è5�»Ó�â�ÚBâ�Õ)Ö�Þ�ÓÏâîÏÚ�Ù)ã�Û�â�á�ä�à�Þ�Ó í�à��â�Ú�â�Ü)Þ�Öëæg�&��â1Ö � Õ0Ö à�ÜEîÏÛ�æ�þ

ô5

AIFB AIFB

«hà�Ý)æ�Ú�â Ü ½ SùAàÖëæ©Õ�Û�à�Ö�â�í�§Aâ�á�§�ÞBÚ�æ©ç Ý0â�ÜÏâ�Ú;Õ¥Þ�â1í¤×.ÚBÙ0áñÕ{rdÙ)Ý·¤�âtÞ � Ö�à�á`î�Û�à��©â�íZù0â�ÚBÖ�à�Ù0Üo�

Ú;Õ�Û¹©¥Ö�à�Ü�Ý)æ�ÛÕ�ÚªÕ0ÜÏí ù)â�ÚBã�þ�ÞBâ�ÜÏÖ�â�Ö/�¤Ó©Õ0ÖfÖëÙ)á âÖ�Ù0Û�æ�Þ�à�Ù0ÜÏÖ�ì©Þ�Ó�â�Ú�âIà�ÖrÛ�â1Ö�ÖÔçtÛÕ�ÚBà�Þ�ï Ù)ÜzÓ�Ù¥äSÞ�Ùí�â1Õ�Û²ä�à�Þ�ÓÈÖëïAÜ�Ù)ÜAï�á¤Ö�ì�Õ�Ü�Þ�Ù)ÜAï�á¤Ö�ìAâtÞBç0è|rÜUÞ�Ó�â Ù0Þ�Ó�â�ÚúÓÏÕ0ÜÏí²ìªÞ�Ó�â�Ú�â$Õ0Ú�â»îÏæ�Ú�â

×.Ù0ÚBá¤Õ�ÛüÝ)ÚBÕ0á`á`Õ0Ú¤Õ�î�î�ÚBÙ)Õ)ç;Ó�â�Ö�ìdÕ0à�á`à�Ü�ÝöÞ�ÙÝ0Ú;Õ0Ö�îÈá`â�Õ0Ü�à�ÜÏÝfÕ0ÜÏí Ö�â�á¤Õ�Ü�ÞBà�ç�Ö~à�ÜbÕªá Ù)Ú�â×.Ù0ÚBá¤Õ�ÛÈÖ�â�ÜÏÖ�â0è é"Ó�âSí�â�ÏÜ�à�Þ�à�Ù0Ü Ù�×bç�Ù0Ü�þÞ�â�Ü�Þ�Õ)ÖöæÏÖ�â�íßà�ÜßÞ�Ó�â Ê�Ç�Ë`Ì�ÍÏÎ°ÐÒÑ�twÇ�É í�âtþ�ÏÜ�â1ÖEá`â�Õ0Ü�à�Ü�Ý Õ0ÖIÕ�Ü�ÇSÚ��:u=ÐÒÑ�Ð Îk��ÇC�:��ÇtÊ�ÇtÍ©Î�Ì�Î°ÐÒ÷¥Í»÷4õ Î õÏÇªÐ Í©Î�Ç�ÍÏÎ°ÐÒ÷¥Í»÷4õZÌ Î�ÇSÚ)ÎÒÊ�Å�ÌÔ�Î õÏ÷-�6èUÜÏÕ�Þ�æ�Ú;Õ�Û|ä"Õ3ïÈÞBÙÈÚ�â�î�Ú�â1Öëâ�Ü�Þ�Ö�æÏç;ÓöÕ�Üöâ��Aþî�Û�à�ç�à¯Þ Ú�â�î�ÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0ÜöçtÙ0æÏÛ�íöã�âªÕ � Ý0Ú;Õ�î�Ó�þàç�Õ�Û�� ÖëÞ�ÚBæÏçtÞ�æ�ÚBâ$çtÙ)Ü�ÞBÕ�à�Ü�à�Ü�ÝýÕ0Û�ÛZçtÙ)ÜÏçtâ�î�ÞBÖÞ�Ó©Õ¥Þ�î�Û�Õ3ïEÕ~Ú�Ù)Û�âdà�Ü`Õ~ç�â�Ú�ÞBÕ0à�Ü í�à�ÖBçtÙ)æ�Ú;Öëâ)ì1à�Ü�þçtÛ�æÏí�à�Ü�ÝÈà�Ü�Þ�â�Ü©í�â�íöçtÙ)ÜÏçtâ�î�ÞBÖ�Õ0ÜÏí,Ú�â�Û�Õ�Þ�à�Ù0ÜÏÖã�âtÞ�ä�â�â�ÜòÞ�Ó�Ù�Öëâ � ç6×�è ÞBÓ�Ú�â�â�ÚBâ�ÛÕ¥ÞBâ�ímç�Ù0Ü�þçtâ�î�ÞBÖ>ßS¤ÿæ�ÞBÓ�â�Ú/à�ìÞßS�zÕ�Ú�Þ�à�Üµà Õ�ÜÏíÍß�± à�Ü�Ýqàù�Ö�è ÕÈÖëà�Ü�Ý0Û�â�w à�Ü�Þ�â�ÜÏí�â1í�w�çtÙ)ÜÏçtâ�î�Þ¥ß�� Õ�Ú�Þ�à�Ü¤ÿæ�Þ�Ó�â�Ú~± à�Ü�Ý·àEã©â1Õ�ÚBà�Ü�ÝIÕ�Û�Û�Ö�â�á¤Õ�Ü�ÞBà�ç�à�Ü�×.Ù)Úëþá¤Õ¥ÞBà�Ù)Üfà�ÜzÕ¤Ö�à�Ü�Ý)Û�âEÕ0ÚëÞBà¯×ÒÕ)ç6Þ1èdé"Ó�âkç�Ù0ÜÏç�â�î�ÞÙ�×��¥¤�±ßç�Ù0æ�ÛífÞ�Ó�â�ÜZã©âEÚBâ�ÛÕ¥ÞBâ�í Þ�Ù`×.Ù)Úüâ��Aþ

Õ0á`î�Û�âwÞ�Ó�â�ç�Ù0ÜÏç�â�î�Þ�Ù�×XßBçtà�ùAà�Û�Ú�à�Ý0Ó�Þ;Ö�Û�â�Õ)í�â�Ú/àÞBÓ�ÚBÙ0æ�Ý)ÓoÝ0Ú;Õ�á`á¤Õ¥ÞBà�ç�Õ�Û©Ö�â�Ü�ÞBâ�ÜÏç�ârÕ�Ü©Õ�Û�ïAÖ�àÖS�tè|rÜzÞ�Ù0îbÙ�×�ÞBÓ�à�ÖÔãÏÕ)ÖëàçkÖëÞ�ÚBæÏçtÞ�æ�ÚBâ0ìÿç�Ù0ÜÏç�â�î�Þ;Öç�Õ�Ü�ã©â�çtÛÕ0ÖBÖ�à��Ïâ1í � ×�è â0èöw à�ÜÏÖëÞBÕ0ÜÏçtâ·w�ì&w çtÙ0Ü�þç�â�î�Þ;Öw�ì�w Ü�æÏákã�â�Ú;ÖwkÕ0ÜÏíxw ÜÏÕ0á`â�ÖOw �tè÷rdæ�Ú�þÚBâ�Ü�ÞBÛ�ï��Èà¾�Èà�Ú àÖrÕ0ã�Û�âEÞ�ÙfÝ0Ú;Õ0Ö�îfÞ�Ó�â`í�à��â�Úëþâ�ÜÏçtâ ã©â�Þ�ädâ�â�ÜØÖëî�â�ç�à��©çªÞ�ïAî©â1Ö Ù�×üÚBâ�ÛÕ¥Þ�à�Ù0Ü©ÖÞBÓÏÕ¥Þ Ó�Ù0ÛíúÕ)ÖEä�â�Û�Û�Õ)ÖIÕbç�Õ�Þ�â�Ý)Ù0ÚBà�ÖBÕ¥ÞBà�Ù)ÜöÙ�×ÞBÓ�âzçtÙ)ÜÏçtâ�î�ÞBÖ`à�Þ í�â1Õ�ÛÖ`ä�à¯ÞBÓÿè$é"Ó�àÖoç�Õ�îÏÕ�þãÏà�Û�à¯Þ�ï¤á¤Õ0û0â�Ö,�Èà¾�Èà�Ú�Ü�Ù�Þ�Ù)Ü�Û�ïoÖ�æ�à�ÞBÕ0ã�Û�âÔ×.Ù)ÚÞ�ïAî�àç�Õ0Û z Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0Ü�}�âtÞBÚ�à�â�ù¥Õ�Û�Þ;Õ0Ö�û�Ö�ì�ã�æ�ÞÕ0Û�Ö�ÙEÖ�æ�î�î�Ù0Ú�ÞBÖ|ûAÜ�Ù¥ä�Û�â�í�Ý0âüã�æÏà�Ûí�à�Ü�ÝE×.Ù)Ú|Þ�ÓÏâÖ�â�á¤Õ0Ü)ÞBà�ç ä�â�ãzÕ0ÜÏífî�ÚBÙ¥ù�àí�â1Ö"Þ�Ó�âkà�Ü�×.Ù)Ú�á¤Õ�þÞBà�Ù)Ü ÜÏâ�â�í�â�í ×.Ù0ÚÂøÔÏÇ6Ê6Î°ÐÒ÷¥ÍöÌ¥Í�Ê�twÇ�6Ð Íqv0è

ù�ú1ûGüoýGú1þOÿOú��ÿ�{û� ��Wþ�ú1þrÖdí�à�ÖBçtæÏÖBÖ�â�íoÕ0ã©Ù¥ù)â0ì�ÞBÓ�âÔãÏÕ0Ö�à�ç~Õ�ÜÏÕ0Û�ï�Ö�à�ÖwÙ�×ÕzÞ�â��Þ¤Õ0ÖIî�â�Ú�×.Ù0ÚBá â1í�ãAïÞ�Èàe�zà�Ú`àÖIã©Õ0Ö�â�íÙ)ÜðÕ�ÞBÙ0û0â�Ü�àÖëâ�Ú�ì¤Õ> �Õ�Ú�ÞëþB|~× þC§�î©â�â�ç;ÓòÞ;Õ�Ý0þÝ)â�Ú1ì1ÖëÞ�â�á`á à�Ü�Ý~Õ0Û�Ý)Ù0ÚBà¯ÞBÓ�á¤Ö�ì » Õ0á`â�í� wÜ)ÞBà¯Þ�ïÚBâ�ç�Ù0Ý)Ü�à¯ÞBà�Ù)Üf×ÒÕ0çtà�Û�à¯ÞBà�â1ÖrÕ)Öüä�â�Û�Û Õ0ÖrÕ¤î�ÚBÙ0î�ÚBà�þ

�6

AIFB AIFB

âtÞ;Õ�ÚBï Õ�Û�Ý0Ù)Ú�à�Þ�ÓÏáW×.Ù0ÚEÝ)â�Ü�â�ÚBÕ�Þ�à�Ü�ÝÈçtÙ)ÜÏçtâ�î�ÞBÖÙ0æ�ÞÔÙ0×dÖëà�Ü�Ý)Û�âIädÙ)ÚBí�Ö � çt×�èk�ÏÝ)æ�Ú�â ¼ �tè�«�ÚBÙ0áÞ�ÓÏâ¤à�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0ÜbÞ�ÓÏÕ�ÞEà�Ö Ý�Õ¥Þ�ÓÏâ�ÚBâ�í,í�æ�ÚBà�Ü�ÝÞ�ÓÏâ�Ö�â`Õ0ÜÏÕ�Û�ï�Öëâ1Ö�ì Õ�rdÙ)Ý·¤�âtÞ Ú�â�î�Ú�â1Öëâ�Ü�ÞBÕ¥ÞBà�Ù)ÜàÖkÝ)â�Ü�â�ÚBÕ�Þ�â�íöä�ÓÏà�ç;Ó�î�æ�ÞBÖ`Õ�Û�Ûwà�Ü�×.Ù0ÚBá¤Õ¥ÞBà�Ù)Üà�ÜßÚBâ�ÛÕ¥ÞBà�Ù)ÜßÕ�Ü©íßí�â�ÏÜ�â1Ö,Þ�Ó�â»çtÙ)Ü�Þ�â��AÞöà�Üä�Ó�àç;Ókà�Ü�×.Ù0ÚBá¤Õ¥ÞBà�Ù)ÜkÖëÓÏÙ0æ�Ûí�ã�âdà�Ü�Þ�â�ÚBî�ÚBâtÞBâ�í²èé"Ó�â�à�Ü�×.Ù0ÚBá`Õ�Þ�à�Ù0Ü Ú�â1Öëàí�à�Ü�Ýøà�ÜmÕøÖ�à�Ü�Ý)Û�ârdÙ0Ýq¤ÿâ�Þªç�Õ�ÜøÜ�Ù¥ä ã�âzæÏÖ�â�íØÞ�Ù�â��î©Ù)ÚëÞªÖ�âtþá¤Õ�Ü�Þ�àç~ÖëÞ�ÚBæÏçtÞ�æ�ÚBâ�Ö � Ö�Ù�þ�ç�Õ�Û�Û�â�í��Ç�Ë��Î?��ÔÏÑtÊB�6è§Aâ�áj§AÞ�ÚBæÏç�Ö ç�Õ�Ü»ã�â ÚBâ�îÏÚ�â1Öëâ�Ü)ÞBâ�íØà�Ü�¦Y�¥¤×.Ù0ÚBá¤Õ¥Þ1ì�ä�Ó�à�ç;ÓØçtÙ)æ�Û�íúã©â ×.â�íúà�Ü�ÞBÙ ùAà�Ö�æÏÕ0Û¯þàÖ�Õ�Þ�à�Ù0ÜòÕ�Û�Ý0Ù)Ú�à�Þ�ÓÏá`Ö�è£rdÙ)Ý·¤ÿâ�ÞBÖzÕ0Û�Ö�Ù$ç�Ù0Ü�þÞBÕ0à�Ü�Þ�ÓÏâfÜÏâ�çtâ1Ö�ÖBÕ�ÚBïöà�Ü�×.Ù0ÚBá`Õ�Þ�à�Ù0ÜúÞ�Ù�Õ�ÜÏÕ0Û¯þï�Öëâ~ädâ�ãoîÏÕ�Ý)â�Ö|Õ0ÜÏíoÕ0æ�Ý0á`â�Ü�Þ|ÞBÓ�â�áUä�à�Þ�ÓªÕ}�â1ÖëÙ)æ�ÚBç�âZª~â1Ö�ç�Ú�à�î�ÞBà�Ù)Ü�«�Ú;Õ�á`â�ä�Ù0ÚBû � }�ªY«��îÏÕ0ÚëÞ�í�â�ÖBçtÚBà�ã�à�ÜÏÝrí�Ù�çtæÏá â�Ü�Þhá`âtÞ;ÕríÏÕ¥ÞBÕ � Õ)ç6þçtÙ)ÚBí�à�Ü�ÝòÞ�ÙÐªræ�ã�Û�à�ÜÇrdÙ0ÚBâ��Õ0ÜÏíNÕmÛ�à�Ý0Ó�Þëþä�â�à�Ý0Ó�ÞoÙ0Ü�Þ�Ù)Û�Ù)Ý0ï�ãÏÕ0Ö�â�íØÙ0ÜøÞBÓ�âbÕ0ÜÏÕ�Û�ï�Öëâ1íÜÏÕ�Þ�æ�Ú;Õ�ÛrÛ�Õ0Ü�Ý0æÏÕ0Ý0âZÞBâ��AÞfç�Ù0Ü�ÞBÕ0à�Ü�â1íøà�Ü Þ�Ó�âí�Ù�çtæÏá â�Ü�Þ�è

� ��gûÕÿOú�� ÿ��-ý��ÿOý��þ�ú1û��Ðùz Ü�×.Ù)Ú�á¤Õ¥ÞBà�Ù)Ü�ç�Ù0Ü�ÞBÕ0à�Ü�â1í�à�ÜØÕÞrdÙ0Ýq¤ÿâtÞ`ç�Õ0ÜÕ�ÛÖ�Ù,ã�âzâ�Aî�Ù0Ú�Þ�â1íøà�Ü�ÞBÙ Õ�Ü�¦Y�¥¤m×.Ù0ÚBá`Õ�Þ�ìÖ�ÙbÞBÓÏÕ¥Þ¤à�Þªç�Õ0Ü�ã�âÈæÏÖ�â�íØÕ0Ö`Öëâ�á¤Õ�Ü�Þ�àçZÕ�Ü�þÜ�Ù0ÞBÕ¥ÞBà�Ù)Ü$Ù)Ü Õ�ädâ�ã Ö�à�Þ�â,Ù0ÚZà�ÜSÕØûAÜ�Ù¥ä�Û¯þâ�í�Ý0â,ãÏÕ0Ö�â0èË¦\�¥¤ðÓÏÕ0Öfã�â�â�Ü ç;Ó�Ù�Öëâ�Ü ã�âtþç�Õ0æÏÖ�âÔà�ÞüàÖ�Ú�â�Ý)Õ0ÚBí�â1íªÕ)Ö�Þ�Ó�â�Ü�â�AÞ~Ö�ÞBâ�îZæ�î�þä"Õ�Ú;ík×.ÚBÙ0áðÖ�Þ;Õ�ÜÏíÏÕ�Ú;í&¢ré^�¥¤zÕ0Ü�Ü�Ù�Þ;Õ¥ÞBà�Ù)ÜÏÖ�è|rÜ�Û�ï ÕfÖëæ�ã©Öëâ�Þ Ù0×wÞBÓ�â`à�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Ü à�ÜbÞ�Ó�ârdÙ0Ýq¤ÿâ�Þ"àÖüæ©Öëâ1íª×.Ù)Ú�Þ�ÓÏâk¦Y�¥¤�Ý)â�Ü�â�ÚBÕ�Þ�à�Ù0Üÿìä�Ó�à�Û�âEç�Ù0Ü�ÞBÕ0à�ÜÏà�Ü�Ý¤â�Ü�Ù0æÏÝ0Ófà�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Üf×.Ù0ÚÞ�ÓÏâ�Ý0â�ÜÏâ�Ú;Õ¥Þ�à�Ù0Ü Ù�×�Ý0Ú;Õ�î�ÓÏà�ç�Ö�è« à�Ý0æÏÚ�â Ø î�ÚBÙ¥ùAà�í�â�ÖÔÕ�Ü â��Õ�á`î�Û�âIÙ0×dÖëæ©ç;Ó

Õ�Ü�¦Y�¥¤ ÚBâ�îÏÚ�â1Öëâ�Ü)Þ;Õ¥ÞBà�Ù)Üÿè }�â�ÛÕ¥ÞBà�Ù)ÜÏÖ,à�ÜÖ�æÏç;Ó ùAà�Ö�æÏÕ0Û�àÖ�Õ�Þ�à�Ù0ÜÏÖ Õ�ÚBâzÜ�Ù0Þ Ù0ÜÏÛ�ïØÞ�ïAî©â1í²ìã�æ�Þ$Õ0Û�Ö�ÙåÕ�Ü�ÜÏÙ�ÞBÕ�Þ�â1íñä�à�Þ�ÓWÕUç�Õ0Û�ç�æ�Û�Õ�Þ�â1íÓ�â�æ�Ú�àÖëÞ�àçòÖ�ÞBÚ�â�Ü�Ý�ÞBÓÿè é"Ó�â¬¦\�¥¤ ÚBâ�î�ÚBâtþÖ�â�Ü�Þ�â1í à�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0Üúà�ÜÞ�©Ý0æ�ÚBâ Ø çtÙ)æ�Û�í ã©âæÏÖ�â�íZÕ0Ö"à�Ü�î�æ�Þ�×.Ù0Ú"Þ�ÓÏâ�Ý0Ú;Õ�î�Ó ù�àÖ�æÏÕ�Û�à�Ö�â�Ú1è

�8ú?þOý� �?ú1þ�ú1ûGü � �� gûÕÿOú�� ÿ��-ý��ÿ�ý!��þ~Ö á`â�Ü�Þ�à�Ù0Ü�â1í ã�ât×.Ù0ÚBâ0ì Ù0Ü�â Ù0× Þ�Ó�âÖëÞ�ÚBâ�Ü�Ý0Þ�ÓÏÖdÙ�×G§�â�á�§�Þ�ÚBæÏçtÞBÖ�à�ÖdÞBÓÏÕ¥Þ�Þ�Ó�â�ïoç�Õ0Üã�âúæÏÖ�â�íS×.Ù)Ú ùAàÖëæÏÕ0Û�àÖBÕ¥Þ�à�Ù0Ümà�Ü�Þ�â�Ú�â1Ö�Þ;Ö Õ�Ü©íçtÙ)Ü�Þ�â�Ü�Þ;Ö�èÈé"Ó�àÖkç�Õ�îÏÕ0ã�à�Û�à�Þ�ïbÕ�Û�Û�Ù¥äüÖÔ×.Ù0Úkæ©Ö�þÕ�Ý)âøà�ÜðùAà�Ö�æÏÕ�Û¤ãÏÚ�Ù¥äüÖ�â�Ú;ÖöÕ�ÜÏíßÜÏÕ3ùAà�Ý�Õ¥ÞBÙ0Ú;ÖãÏÕ)Öëâ1íøÙ0Ü»ÛÕ�ÚBÝ0â�Úoí�ÙAç�æ�á`â�Ü�ÞfÖ�âtÞ;Ö�ìüÕ0ÜÏíØÞ�ÙÙ��â�Úwî©â�Ù0î�Û�âüÕ�ÜoÕ�Þëþ�Õ¥þ�Ý0ÛÕ�ÜÏç�âdÙ¥ù)â�ÚBù�à�â�äúÙ¥ù0â�ÚÞ�ÓÏâ�à�Ü�×.Ù)Ú�á¤Õ�Þ�à�Ù0Ü Þ�Ó�â�ïªÓÏÕ3ù)âÔÕ)ç�çtâ1Ö�Ö�ÞBÙÏè

«hà�Ý)æ�ÚBâ Ü ÖëÓ�Ù¥äüÖ�ÕßÖ�à�á îÏÛ�à��Ïâ�í#"$ÖëÞ�ÚBæÏçtþÞBæ�ÚBâòçtÚBâ�Õ¥ÞBâ�íR×.ÚBÙ0á ÕË§Aâ�á�§�ÞBÚ�æÏçòÝ0â�ÜÏâ�Ú�þÕ�Þ�â1í�ãAïSÕÛrdÙ0Ýq¤ÿâ�Þ Õ�ÜÏí ùAà�Ö�æÏÕ0Û�àÖëâ1íSä�à¯ÞBÓr^r~~ùAà�â�ädâ�Ú%$¥èmé"Ó�âbÖ�ÞBÚ�æÏçtÞ�æ�ÚBâzÖ�Ó�Ù¥äüÖ`Þ�ÓÏâÖ�â�á¤Õ0Ü)ÞBà�ç¿çtÛ�æÏÖ�ÞBâ�Ú;ÖRÕ0Ú�Ù)æ�ÜÏí ÞBÓ�â î�â�Ú;Ö�Ù0Üß � ÓÏÕ�Ü©í�àeàÏè é"Ó�â�Ú�â"Õ�ÚBâdÞ�Ó�ÚBâ�â"á¤Õ�à�Ü ç�Û�æ©Ö�ÞBâ�Ú;ÖÚBâ�ç�Ù0Ý)Ü�à�ÖBÕ�ãÏÛ�â)ì©Ù)Ü�â¤í�â�Õ0Û�à�Ü�Ýfä�à¯ÞBÓ � ÓÏÕ0ÜÏí�à2w ÖÚBÙ0Û�â�Ö �'&)( *,+�-/.10 243#5�6,2/7�8 ì &�7�9�2�3/940:8 Õ0ÜÏí&,; 2/7�9)-/;:8 �6ì�Ù0ÜÏâoí�â�Õ0Û�à�Ü�ÝÈä�à�Þ�ÓöÓ�àÖEî�Ó�à�Û�Ù)ÖëþÙ)î�ÓAï �'& <)24;�( 24.�0 2,=:2/8 ì &)-:*)-?>,@A5)*�7�9,-A6,9�8Õ0ÜÏí &:5B-#</5C.�=/; 8 � Õ�ÜÏíSÙ)Ü�âúí�â1Õ�Û�à�ÜÏÝ ä�à¯ÞBÓÞBÓ�â�ùAà�Ù)Û�â�Ü�Þ"ä�Ù0ÚBÛí � ÓÏÕ0ÜÏí�à�×.Ù0æ�Ý)Ó)ÞüÕ0Ý)Õ�à�ÜÏÖëÞ�'& 6)*/7�*,-D5�2/7:5�<�EA8 ì &,@A5)*/749,-A6,9/8 ì &); *4; 2/70 9�F�9:6C;A5)*,-?8 ì &�G:9:<�; 9,0�-�EA24;/940A5)2/7 5�<�ED8 �6è |àç6ÞBæ�ÚBâ�ÖªÞ�ÓÏÕ�ÞZÕ0Ú�â Þ�ÓAæÏÖZÕ�æ�ÞBÙ0á¤Õ¥ÞBà�ç�Õ�Û�Û�ï

Ý)â�Ü�â�ÚBÕ�Þ�â1í$×.ÚBÙ0á ÜÏÕ�Þ�æ�Ú;Õ�ÛIÛ�Õ0Ü�Ý0æ©Õ�Ý0âöÞBâ��AÞBÖîÏÚ�Ù¥ùAàí�âzÕ�Ü$Õ�Þëþ�Õ¥þ�Ý0ÛÕ�ÜÏç�âÈÙ¥ù0â�Ú�ùAà�â�ä Ù¥ù0â�Ú ÕîÏà�â1çtâ`Ù�×"à�Ü�×.Ù0ÚBá`Õ�Þ�à�Ù0Üÿè�§Aæ©ç;Óbî�àç6ÞBæ�ÚBâ�Ö�ç�Õ0ÜÞBÓ�â�Ü,ã�â æ©Öëâ1í à�ÜbÙ)ÚBí�â�ÚÔÞ�ÙÈÕ�æÏÝ0á`â�Ü�Þ â��â�çtþæ�Þ�à�ù0âöÖëæ�á`á¤Õ�ÚBà�â�ÖfÕ�ÜÏí$Ú�â1Õ0í�â�ÚBÖfÕ�àí�Ö�ì~ã�æ�ÞÞBÓ�â�ïoÕ0Ú�â Õ�ÛÖ�ÙIæÏÖ�â�í Õ0Ö�ùAà�Ö�æÏÕ0Û�à�Ü�Þ�â�Úë×ÒÕ)çtâ�ÖdÞBÙíÏÕ¥ÞBÕ0ãÏÕ0Ö�â�Ö � îÏÚ�â�×.â�Ú;Õ�ã�Û�ï»à�ÜSçtÙ)Ú�î�Ù0Ú;Õ¥ÞBâ Ö�âtÞëþÞBà�ÜÏÝ)Ö/�6èæ~ÖbÖëæÏç;ÓSÞBÓ�â�ïSÕ�æÏÝ0á`â�Ü�ÞzûAÜ�Ù¥ä�Û�þâ1í�Ý0â á`Õ0ÜÏÕ�Ý)â�á`â�Ü�Þ Ö�ï�Ö�ÞBâ�á¤Ö�ì�ä�Ó�â�ÚBâzÞBÓ�â�ïîÏÚ�Ù¥ùAàí�â~ÕIù�àÖ�æÏÕ�Û�â�Ü�Þ�Ú;Õ�Ü©çtâüÞBÙIîÏà�â1çtâ�ÖdÙ�× à�Ü�þ×.Ù)Ú�á¤Õ�Þ�à�Ù0Ü$î�â�Ú�Þ�à�Ü�â�Ü)ÞZÞBÙØÖ�î©â1çtà��©ç,à�Ü�ÞBâ�ÚBâ�ÖëÞÝ)Ú�Ù)æ�îÏÖ�ä�à�Þ�Ó�à�ÜÈÕ�Üfâ�Ü�Þ�â�Ú�î�ÚBàÖëâ)è~Ö�Õ0ÜIâ��Õ�á`î�Û�â�Ù0×©ÞBÓ�â�â��î�ÚBâ�ÖBÖ�à�ù)âwî�Ù¥ädâ�Ú

Ù0×"Þ�Ó�â8§�â�á�§�Þ�ÚBæÏç�Ö�ì�Ù0ÜÏâoá`à�Ý)Ó)ÞkÞBÕ�û)âªÖ�Ù0á`âÞBà�á`â�ÞBÙkÕ�Ü©Õ�Û�ïAÖ�â,�ÏÝ)æ�Ú�â Ü Õ�ÜÏí ÞBÚ�ïIÞ�ÙEà�á¤Õ�Ý0þà�Ü�âIä�ÓÏÕ¥ÞÔÞ�Ó�â Ù0ÚBà�Ý0à�ÜÏÕ�Û Þ�â��Þ àÖÔÕ0ã©Ù)æ�Þ�ì²Õ0ÜÏíä�ÓÏà�ç;Ó�w í�àÖ�ç�Ù0æ�Ú;Öëâ1Öw�Þ�Ó�ârÙ0ÚBà�Ý)à�ÜÏÕ0ÛÏí�Ù�çtæÏá â�Ü�Þç�Ù0Ü�ÞBÕ0à�ÜÏâ�í²è

�8ý�ü#� ��ûÕÿOú?û�üIH��JÓþOú1ÿK��þLHjú¹ÿKM�N�O�P�»à�Þ�ÓÏà�Ü ÞBÓ�â |rÜ©é�Ùq± Ü�Ù¥ä�Û�â1í�Ý0âPî�ÚBÙ � â1ç6Þ1ì}�ªY«øä�à�Þ�Ó,â��AÞ�â�ÜÏÖ�à�Ù)ÜÏÖrÕ0Ú�âIæÏÖ�â�í Õ)Ö~ÚBâ�îÏÚ�â�þÖ�â�Ü�Þ;Õ¥Þ�à�Ù0ÜfÛÕ�Ü�Ý)æÏÕ�Ý)âr×.Ù)Ú"Þ�Ó�âoÊtÇ�Ë¤Ì¥ÍÏÎ°ÐÒÑÂtwÇ�É� ç6×�èf|Ôé^± ½ ° « ùH¢Y±k�²�²-³°ìH| z ¤ ½ ° ¢�« ¶ ¸f²�²*³¹�6èé"ÓÏâ r^|\}Z ,|\},��RQ�S�TVUXW�Y�T�Z\[^]_T çtÙ)á`î©Ù0þÜÏâ�Ü�ÞªàÖ í�à�ÚBâ�ç6ÞBâ�íØÞBÙ�Þ�Ó�âbÝ0â�Ü�â�Ú;Õ¥ÞBà�Ù)ÜØÙ0×EÕ`ba " NO34]�N02.?(*A4" .234+*.234"?3B��0?3B= #%�

cedgf h\i!j�k?h FG� f �/.23Ö'��.2]Z�/)¥.234+*.?34"234�02�S0?#%��*"ml026*3onÂK<#%)%):3B�/"2#%) nÂ��.2�BKÞ02��k)`�/.2��3,'��.f#%�*A4)%(*"2#%��$#%��Y+-�/+34.�!qp�34�*A43,0?6*3^h�3B.VnÂ"26*��.10sr �WF i n�F i 9 d r023ot0^3ot*�/]^+*)%3�!j�W6*3�h�#%"?(-�/)%#%"Q�S0?#%��j�/'fN3B]�N�02.?(*A4"u'vxw)y�zXw�{ozm{ 0?3otO0?"mlÕ�/�*=�A4��(*)%=i026�3B.?3C'e��.?3Yp3�p-�/"234=��b��)`�/.?��34.?;�"?#}|B34=Z023ot0�~e�/p��(�0��G6*�/�*=*#��Q!� �W6*3Y�W� a hO#%3CK�34.~#%"��b+*.2�=*(*AC0,p^n a #%=*]^#%��;#%"?0?.Q�S0?��.Y_<34=*34.2)`�/�*=Ùa:c~!Õ�¾0�(*"?34"b�Õ�� f 3C0Y��3B��3B.1;�S0234=ÙNO34]�N02.?(*A4"~0?�&��34�*34.Q�S0?3b+*#%AC0?(*.234"Zp-�/"234={��"2�/;�AB�/)%)%34=k�/(*��]^34��0234=ÂNO+*.?#%�*�^7:]�p3B=�=*34.�0?3BA26*�*��) ;��\n�~�AC'?! l ��(*.?3Wm��C!:�W� a "?02�/�*=*"o'e��.:�Õ3B��0?.Q�/)·�Õ��;AB34+�0 a .23B�Bl·.23C'�3B.?.2#%�*�,0?�Z026*3�#%��'��.2]Z�S0?#%��bA4.23B�S0234=p�nZ0?6*3XNO34]�N0?.2(*A4"�!

�7

AIFB AIFB

w Û�à�Ý0Ó�Þëþ�ä�â�à�Ý0Ó�Þ~Ù0Ü�ÞBÙ0Û�Ù0Ý0ïÕw©ãÏÕ0Ö�â�íÈÙ)ÜzÛ�à�Ü�Ý)æ�àÖ�þÞ�àçfÕ�ÜÏÕ0Û�ï�Ö�à�Ökã�ïÝr^|\}� ~|\},�� à�Ü�ç�Ù0ákãÏà¯þÜÏÕ�Þ�à�Ù0Ü»ä�à�Þ�Ó»ÞBÓ�âzà�Ü�×.Ù0ÚBá¤Õ¥Þ�à�Ù0ÜøÞBÓÏÕ¥Þ�§Aâ�á þ§�ÞBÚ�æÏç�Ööç�Õ�Üýî�Ú�Ù¥ùAàí�â0è é"Ó�àÖ,á`â�Õ�Ü©ÖzÞBÓÏÕ¥Þ×.Ù0ÚBá¤Õ�Û)ÞBÕ�AÙ)Ü�Ù0á`àç�ÚBâ�ÛÕ¥Þ�à�Ù0Ü©ÖëÓ�à�îÏÖ�Þ�ÓÏÕ�ÞhÓÏÙ0Ûíà�ÜSÞ�Ó�âúí�à�ÖBçtÙ)æ�ÚBÖ�â�Õ¥ÞzÓÏÕ0ÜÏímÕ�ÚBâ�í�àÖ�ç�Û�Ù�Öëâ1íÕ�Ü©í¿á¤Õ0í�âðâ��î�Û�àçtà�ÞýÕ0ÖmÕ Ö�âtÞmÙ�×�}Zª\«Þ�æÏî�Û�â1Ö�è ~í�í�à¯ÞBà�Ù)ÜÏÕ�Û�Û�ï)ì ÞBÚBÕ)í�à�Þ�à�Ù0ÜÏÕ0ÛÈädâ�ãîÏÕ0Ý0â�ÖIÕ�ÚBâoÕ�æ�Ý)á`â�Ü�Þ�â1í�ä�à¯ÞBÓïªræ�ã�Û�à�Ü�rdÙ0ÚBâá`âtÞ;Õ íÏÕ¥ÞBÕÏì�Õ�ÛÖëÙ Ý0â�Ü�â�Ú;Õ¥ÞBâ�í Õ�æ�ÞBÙ0á¤Õ¥ÞBà�ç�Õ�Û�Û�ïãAïZÞ�Ó�â�r^|\}Z ,|\}~�Y� Q!S)TVUXW�Y�T�Z\[^]_T çtÙ0á`î�Ù�þÜ�â�Ü)Þ\��è« à�Ý0æÏÚ�â ô î�ÚBÙ¥ùAà�í�â�ÖÔÕ�Ü â��Õ�á`î�Û�âIÙ0×dÖëæ©ç;Ó

Õ�æ�Þ�Ù0á¤Õ�Þ�àç�Õ�Û�Û�ï»Ý0â�ÜÏâ�Ú;Õ¥Þ�â1íMªkrWÕ0ÜÏí$Ù)Ü�Þ�Ù�þÛ�Ù0Ý0àçbû�ÜÏÙ¥ä�Û�â1í�Ý0â)èRé"Ó�âöÕ¥Þ�Þ�â�Ü�ÞBà�ù)â,ÚBâ�Õ)í�â�Úä�à�Û�Û�Ü�Ù�ÞBà�ç�âWÞ�ÓÏÕ�ÞNÞ�ÓÏâ�ÚBâvÕ�ÚBâ Þ�ädÙ ç�Ù0Ü�þÖëÞ�ÚBæÏç6Þ;Öúí�â1çtÛÕ�ÚBâ�íðÞ�Ó©Õ¥ÞØÕ0Ú�â$Ü�Ù�ÞØæÏÖëâ1íUà�ÜÞ�ÓÏâðâ��Õ�á`î�Û�â0ìúà°è â0è & 5�<�� 9/7424; 9�3 8 Õ�Ü©í&)=:2 <,��*CEA9,��0 *,�:9,0�;�( 8 è é"Ó�â1Öëâ Þ�ädÙòç�Ù0Ü�þÖëÞ�ÚBæÏç6Þ;ÖüÕ�ÚBâEí�â��©Ü�â�íZà�ÜfÞ�Ó�â&| z ¤öÛÕ�ÜÏÝ0æÏÕ0Ý0â0è�»Ó�â�Ú�â1Õ0Ö Õ Þ�ïAî�à�ç�Õ�Û Ù0Ü�Þ�Ù)Û�Ù)Ý0ï�Ù�× Þ�â�ÜòÚBâ�î�þÚBâ�Ö�â�Ü�ÞBÖSÕRÞBÕ�AÙ)Ü�Ù0áIï � Þ�ÓÏâåÙ0Ü�Þ�Ù)Û�Ù)Ý0ï à�ÜÞ�ÓÏâØâ��Õ�á`î�Û�âØàÖ,Ü�ÙSâ��çtâ�î�Þ�à�Ù0ÜßÙ0ÜßÞ�Ó�àÖ/�6ì&:5�<C� 9/7�24;/9�3 8 ÚBât×.â�ÚBÖfÞBÙ»çtÚBÙ)ÖBÖ�þ°ÞBÕ�AÙ)Ü�Ù0á`àçÛ�à�Ü�û�ÖøÞBÓÏÕ¥Þ$ç�Õ�Ü Ó�Ù0ÛíRä�à¯ÞBÓ�à�ÜvÕUí�Ù)á¤Õ�à�ÜÕ�Ü©í²ìüà�×EÚ�â�î�Ú�â1Öëâ�Ü�Þ�â�íÿìüç�Õ�Ü$á`Õ0û0â Õ�í�à��â�Úëþâ�Ü©çtâ"à�ÜÂ�ÏÜÏí�à�Ü�Ý Ü�â�â�í�â1íkà�Ü�×.Ù0ÚBá¤Õ¥ÞBà�Ù)ÜkãÏÕ)Öëâ1íÙ0Ü,çtÙ)Ü)ÞBâ��AÞÔí�â�ÖBçtÚBà�î�ÞBà�Ù)ÜÏÖ�èbrÖrÕ0Ü â��Õ0á`î�Û�âÙ0ÜÏâIç�Õ0Üzà�á`Õ0Ý0à�Ü�âEÞ�ä�ÙÙr^r~ðçtÙ)ÜÏçtâ�î�ÞBÖrÛ�à�û0â&�*?5)7�>,0A5C. 8 Õ�ÜÏí & <C=D5B�:8 èd§Aæ©ç;ÓúçtÙ)ÜÏçtâ�î�ÞBÖÕ�ÚBâ¤Ü�Ù0ÞEÞ�ïAî�à�ç�Õ�Û�Û�ïEw çtÛ�Ù)Ö�â�wÿà�ÜúÕfÞ�Ú;Õ0í�à¯ÞBà�Ù)ÜÏÕ�ÛÙ0Ü�ÞBÙ0Û�Ù0Ý0ï)ìüä�Ó�â�Ú�âbÞ�Ó�â�ï Õ�ÚBâ Ü�Ù0ÞZ×.Ù0æ�ÜÏí�Õ0ÖÖ�æ�ã�þ�çtÛÕ0ÖBÖëâ1ÖkÙ0×~ù0â�Ó�à�ç�Û�â1Ö �_&�*?5)7�>,0A5C.?<48 Õ�ÚBâÜ�Ù0Þ¤Þ�ï�îÏà�ç�Õ�Û�Û�ï�á â1Õ�ÜÏÖoÙ�× Þ�Ú;Õ�ÜÏÖ�î�Ù0Ú�ÞBÕ¥ÞBà�Ù)Üo�6ìÜ�â�à¯ÞBÓ�â�ÚøÕ0ÖØÖëæÏã�þ�ç�Û�Õ)Ö�Ö Ù�× Õýç�Ù0ÜÏç�â�î�ÞØÛ�à�û0â&)��+A5,743#5B-�.:8 ì &4� 7�*�2,;A5B-/.�3�94@A5�6,9�8 ì âtÞ;ç�è» â�ù0â�ÚëÞBÓ�â�Û�â�ÖBÖ�ì�î�â�Ù0îÏÛ�âÈä�Ù0ÚBûAà�Ü�Ý,à�Ü�ÞBÓ�âzÙ)à�Ûà�ÜÏí�æÏÖëÞ�ÚBï Þ�ïAî�àç�Õ0Û�Û�ï$ÚBâ�Ý�Õ�Ú;í»Þ�ÓÏâ�Þ�ädÙ»ç�Ù0Ü�þçtâ�î�ÞBÖkÕ0Ö Ó�à�Ý0Ó�Û�ï ÚBâ�ÛÕ¥ÞBâ�í²ì Ü�Ù�ÞEà�Ü,Þ�Ó�â¤Û�â�Õ)Ö�Þí�æ�âöÞBÙØÞ�Ó�â�à�ÚÈÜÏÕ�Þ�æ�Ú;Õ�Ûjw Ö�ï�áIã�à�Ù)Ö�à�ÖOw"à�ÜSâ�ù�þâ�ÚBï�í�Õ3ïMw Û�à¯×.â,Ù0Ü»ÞBÓ�â ÚBà�Ýµw�è�r^|\}Z ,|\}~�Y�àÖ ÓÏÙ¥ädâ�ù0â�ÚZÕ�ãÏÛ�âbÞ�Ùøç�Õ0î�Þ�æ�ÚBâöÖëæÏç;ÓåÑ��÷3Ê;Ê��Î�ÌÚA÷¥Í�÷¥Ë ÐÒÑ Û�à�Ü�û�Ö�Õ�Ü©íbÚBâ�îÏÚ�â1Öëâ�Ü)ÞÔÞ�ÓÏâ�á¿æ©Ö�þà�Ü�Ý Þ�Ó�â &:5�<��/9/7�24; 943 8 Ö�ÞBÚ�æÏçtÞ�æ�ÚBâ0èé"Ó�â Ù�ÞBÓ�â�Ú ç�Ù0ÜÏÖëÞ�ÚBæÏç6Þ�_&)=:2:<,��*CEA94��0/*,�:940�;�(:8 � à�Ö Þ�ÓÏâ á Ù�Ö�ÞÝ0â�Ü�â�Ú;Õ�Û°ì æ�ÜÏà�ù)â�Ú;Ö�Õ0ÛRÚBâ�ÛÕ¥ÞBà�Ù)Ü Þ�ïAî©â ÚBâtþ¡Ïâ1ç6Þ�à�Ü�Ý �ÏÌ-�6Î2�Ct õ�÷u�Ç Ú�â�Û�Õ�Þ�à�Ù0ÜÏÖ ä�à¯ÞBÓ�à�ÜÕýÞBÕ-��Ù)Ü�Ù0áIï0è z Þ»à�Öøç�æ�Ú�ÚBâ�Ü�ÞBÛ�ïåæÏÖ�â�íñà�Ü� j � ]^3C02�Ñ=-�S02�÷#%�*A4)%(*=*34" #%��'��.2]Z�S0?#%��/p��(�0Ð�/(�0?6*��.mlÛ[�3on×A4��*A43B+O02"Xl>"2(�]^]Z�/.�nx�/'0?6*3ÓAB��0?3B��0��/'��Ö=��OA4(*]^34��0ml�# 02"Ý9 i �'lj3C0?A�!j (*p�)%#%� �Õ��.23 ]^3402�Ç=-�S02� #%"�=*34"2A4.2#%p34= �S0m��^�x�\�:��x�m�x�^��^�\�^��%�x��%�K� !

r^|\}� ~|\},��RQ�S�TVUmW!YKT�Z%[^]_T,ÞBÙðí�â�ÏÜ�âSÜ�Ù0Þ×.æÏÚëÞBÓ�â�Ú Öëî�â�ç�à��Ïâ1íÑ��Ì�6Î?�Ct õÏ÷-u�Ç ÚBâ�ÛÕ¥Þ�à�Ù0Ü�þÖ�Ó�à�îÏÖ ã�âtÞ�ä�â�â�Ü Õ & 6,*,-A6)9,�/;:8 Õ�Ü©í Õ& <C�:9:6�5��#5)940 8 Ù�×,Þ�ÓÏÕ�Þ$çtÙ)ÜÏçtâ�î�Þ�è ¢~Ù¥ä"þâ�ù0â�Ú�ì"à�Ü�Ö�Ù0á`â ç�Õ)Öëâ1Ö¤ÞBÓ�â�ÚBâ àÖoûAÜ�Ù¥ä�Û�â�í�Ý)âÕ3ù¥Õ0à�ÛÕ�ã�Û�âÔ×.Ú�Ù)áñÞBÓ�âÂ± Ü�Ù¥ä�Û�â�í�Ý0â ¶ Õ0Ö�â Þ�ÓÏÕ�ÞÕ0Û�Û�Ù¥äüÖ æÏÖfÞ�ÙØ×.æ�Ú�Þ�Ó�â�ÚÈÚ�â�ÏÜ�â,Þ�ÓÏâbÞ�ïAî�âöÙ�×îÏÚ�Ù)î©â�ÚëÞBà�â1ÖrÕ0Ú�â`Õ0çtÞ�æÏÕ0Û�Û�ïÈî�ÚBâ�Ö�â�Ü�Þ�è z ÜbÖ�æÏç;Óç�Õ0Ö�â�Ö�ì r^|\}Z ~|\},�Y� Q!S�TVUmW!YKT�Z\[^]_T ä�à�Û�Û��æ�â�Ú�ïbÞ�ÓÏâ8± ¶ à�ÜúÙ0Ú;í�â�Ú�ÞBÙ¥�ÏÜ©í Ù0æ�ÞIÓ�Ù¥äà�Þ ç�Õ�Ü â�Ü�ÓÏÕ�Ü©çtâzà�ÞBÖªû�ÜÏÙ¥ä�Û�â1í�Ý0âzÚBâ�îÏÚ�â1Öëâ�Ü�þÞ;Õ¥ÞBà�Ù)Üÿè��Þ çtæÏÚ�ÚBâ�Ü�Þ�Þ�ÓÏà�ÖIî�ÚBÙAç�â�ÖBÖ�â��AÞBâ�ÜÏí�Ö|rÜ�ÞBÙ0Û�Ù0Ý0ï`Ý)â�Ü�â�ÚBÕ�Þ�à�Ù0Üoà�ÜÙ}�ªY« ×.Ú�Ù)áNã©â�à�ÜÏÝÖ�à�Ü�Ý0Û�âEÞBâ��AÞrãÏÕ)Öëâ1ízÛ�à�ÜÏÝ0æ�àÖ�ÞBà�çIÕ�ÜÏÕ0Û�ï�Ö�à�Ö~à�Ü�ÞBÙÕ0Ü Õ�æ�Ý)á`â�Ü�ÞBÕ�Þ�à�Ù0Ü î�ÚBÙAç�â�ÖBÖåä�Ó�â�Ú�âvîÏÚ�â�þùAà�Ù0æÏÖ�Û�ïðÝ)â�Ü�â�ÚBÕ�Þ�â1íNÙ0Ü�ÞBÙ0Û�Ù0Ý0àç»ûAÜ�Ù¥ä�Û�â�í�Ý)â� çtÙ)Ü)Þ;Õ�à�Ü�à�Ü�Ý ß�ãÏÕ0ç;ûAÝ)Ú�Ù)æ�ÜÏígà ûAÜ�Ù¥ä�Û�â�í�Ý)âÕ0ã©Ù)æ�ÞöÞ�Ó�â í�Ù0á¤Õ�à�Üo�,àÖ,ÞBÕ�û)â�Üðà�Ü�ÞBÙòçtÙ0Ü�þÖ�àí�â�Ú;Õ¥ÞBà�Ù)ÜSÕ)ÖÈákæ©ç;ÓòÕ0ÖÈî�Ù)ÖBÖëà�ã�Û�â0è ü× Þ�â�ÚÓ©Õ3ù�à�Ü�ÝöÕ�æ�Ý)á`â�Ü�Þ�à�Ü�Ý,Þ�Ó�âÈÙ)Ü�Þ�Ù0Û�Ù0Ý)ïöÝ0â�ÜÏâ�Ú�þÕ�Þ�â1íòãAïÐr^|\}Z ~|\},��IQ!S�TVUmW!YKT�Z%[�]_T ÞBÓAæÏÖ�ìÞBÓ�âbç�Ù0á`î�Û�âtÞBâ�}Zª\« � §g�ªÚ�â�î�ÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0Üøà�ÖÖ�â�Ü©íÈÞ�ÙfÞBÓ�âj}ZªY«$Ú�â�î©Ù�Öëà�Þ�Ù)Ú�ïZá¤Õ�à�Ü�ÞBÕ0à�Ü�â1í×.Ù)ÚoÙ0Ü�ÞBÙ0Û�Ù0Ý0ï ÖëÞ�Ù)ÚBÕ0Ý0â � |rÜ�ÞBÙ·± Ü�Ù¥ä�Û�â�í�Ý)âªþ§�â�ÖBÕ�á`â�Õ¥Þ"Þ�Ó�â�á`Ù)á â�Ü�Þ/�6è

� � Á � Â�� À � �&�s¡ �&� � ��`ÁhÀ

�»ÓÏà�Û�âbæ©Öëâ1í à�Ü�Õ�ù3Õ0Ú�à�âtÞ�ïøÙ0×IçtÙ)á á`â�ÚBç�à�Õ0ÛîÏÚ�Ù�í�æÏçtÞBÖ�ì)Ú;Õ�ÜÏÝ0à�Ü�Ý ×.Ú�Ù)á£¢6ÍÏÎ�Çu�u=Ð�v)Ç�Í�Ñ;Çs¤ü÷-��Î�Ì-u=Ê;ì)¢tÍÏÎ�Ç�u�u=Ð�v�Ç�ÍÏÎ�¥J��Ìt~u�Ç��o®3Ê6Î�ÇtËIÊ�Þ�Ù¦�:Ô�Ë$�Ë¤Ì-�6ÐÊ6Ð Í�v1§Ï÷1÷-u ÊEÕ0ÜÏí©¨�ÐÊ�ÔÏÌu=ÐÊ6Ð Íqvª¥|÷¥Ë��÷�Í�ÇtÍ©ÎÒÊ6ì�ÞBÓ�â$r^|\}Z ,|\}~�Y� Þ�ÙAÙ0Û�Ö�âtÞrà�ÖüÖ�æ�ã�þ� â1ç6ÞZÞ�Ù»ç�Ù0Ü�Þ�à�ÜAæ�Ù0æÏÖ à�á îÏÚ�Ù¥ù)â�á`â�Ü�Þ�èÇrdæ�Ú�þÚBâ�Ü�ÞBÛ�ïøÞBÓ�â Öëï�ÖëÞ�â�á à�ÖÈÕ0ã�Û�âbÞBÙ í�â�Õ0Û�ä�à¯ÞBÓ wÜ�Ý0Û�àÖëÓÿì � â�ÚBá¤Õ�Ü»Õ�Ü©í » Ù0ÚBädâ�Ý0àÕ�Ü�Þ�â��Þ;Ö�ìä�ÓÏâ�ÚBâ�Õ0Ö á`Ù0ÚBâdÙ0×ÏÞBÓ�â^ |æ�ÚBÙ0î�â�Õ0ÜEÛ�Õ0Ü�Ý0æ©Õ�Ý0â1Ö� «�ÚBâ�ÜÏç;Ó�ì*§Aî©Õ�Ü�àÖëÓ�ì�ªræ�ÞBç;Ó:�ÿÕ�ÚBâ�â��î�â�ç6ÞBâ�íÔÞBÙã�âEÕ0íÏí�â�ífÖëÙAÙ)Üÿèé"Ó�âÂ�zàe�Èà�Ú çtÙ0á`î�Ù0Ü�â�Ü�ÞràÖ~Õ�ÛÖ�ÙªÖëæÏã � â�çtÞ

ÞBÙfç�Ù0Ü�Þ�à�ÜAæ�Ù0æ©Örà�á`î�ÚBÙ¥ù0â�á`â�Ü)Þ1ìÿÖëÙ Þ�Ó©Õ¥Þ�Þ�ÓÏârdÙ)Ý·¤�âtÞ Ý)â�Ü�â�ÚBÕ�Þ�â1í�§Aâ�á�§�ÞBÚ�æÏç�ÖhÝ0âtÞ�Õ�ÜIâ�ù)â�ÜÚBàç;Ó�â�Ú�w á`â�Õ�ÜÏà�Ü�Ýow�Ú�â�î�ÚBâ�Ö�â�Ü�ÞBÕ�Þ�à�Ù0Üðá Ù�í�â�Û�è�Þ�ÞBÓ�â�Ö�Õ0á âdÞ�à�á`â�Þ�Ó�â�×.æ�ÜÏçtÞ�à�Ù0ÜÏÕ0Û�à�Þ�ïEÙ0×ÏÞ�ÓÏâç�Ù0ÚBâÏ�Èà¾�Èà�Ú,çtÙ)á î�Ù0ÜÏâ�Ü�Þ àÖzâ�ÜÏÓÏÕ�ÜÏç�â�íSà�ÜÖ�æÏç;ÓZÕ`ä�Õ3ïoÞ�ÓÏÕ�Þüà¯Þrç�Õ�ÜÈÖ�â�ÚBù0âÔá`Õ0ÜAïªá`Ù0ÚBâÞ;Õ0Ö�û�Ö � Þ�Ó�à�Ü�û$Ù0×Iâ�Ü�ÓÏÕ�Ü©çtâ�í�ÖëæÏá á¤Õ0Ú�àÖëà�Ü�Ý©ìà�ÜÏç�Û�æÏí�à�Ü�ÝöÖ�á ÙAÙ0Þ�Ó�à�Ü�ÝÏì�í�àÖ�ç�Ù0æ�Ú;Öëâ Ú�â1çtÙ)Ý0Ü�à�þÞBà�Ù)Üÿì�Õ0ÖIädâ�Û�Û�Õ0ÖIÕzá`Ù0ÚBâj¡Ïâ�Aà�ã�Û�â » Õ¥Þ�æÏÚBÕ0Û¤ Õ�Ü�Ý)æÏÕ�Ý)â~ãÏÕ)Öëâ1íªÚBâ�Õ)í�â�Ú;Ö"Õ�àí²è%�6èüÜ�à�ÖBÖ�æ�âªÞBÓÏÕ¥Þ¤àÖ`çtæ�ÚBÚ�â�Ü�Þ�Û�ïöí�â�Õ�Û�Þ¤ä�à¯ÞBÓ

àÖüí�à�ÚBâ�ç6ÞBâ�í Þ�Ù¤ÖBçtâ�ÜÏÕ�ÚBà�Ù�Ö"ä�Ó�â�ÚBâ Ù0Ü�â�ä"Õ�Ü�Þ;ÖÞBÙxw Õ�Ü©Öëä�â�Ú��æ�â�ÖëÞ�à�Ù0ÜÏÖOw�è z Ü ÖëæÏç;Ó ç�Õ)Öëâ1Ö

«8

AIFB AIFB

«hà�Ý)æ�ÚBâ ôg½ üÜØâ��çtâ�Ú�î�Þ Ù0×br^|\}Z ~|\},�Y�RQ!S�TVUmW!YKT�Z\[^]_T Ý0â�ÜÏâ�Ú;Õ¥Þ�â1íÏ}Zª\«òÕ�ÜÏÜ�Ù�Þ;Õ¥Þ�à�Ù0Üúà�Ü�þçtÛ�æÏí�à�Ü�Ý�ªræ�ã�Û�à�ÜdrdÙ)Ú�â á`âtÞ;Õ íÏÕ¥ÞBÕÏè

á`Ù0ÚBâØÖ�â�î©Õ�Ú;Õ¥Þ�â1íòà�Ü�×.Ù0ÚBá¤Õ¥ÞBà�Ù)ÜßÖ�Ó�Ù0æ�Ûíßã©âÕ3ù¥Õ�à�Û�Õ0ã�Û�âúÞBÓÏÕ¥Þ�×�è â0è ç�Õ0Üßá¤Õ0û0âØÞ�Ó�â»í�à�× þ×.â�ÚBâ�Ü©çtâ»ã�âtÞ�ä�â�â�ÜNÕNÑ;÷�Í�Ñ;ÇQ��Î � à�è â0è & 6)240EA2,-�+ �/2:6�;4+/0 940:8 ��Õ0ÜÏí`Õ�ÜfÐ Í�Ê6Î�Ì¥Í�Ñ;ÇhÞBÓ�â�ÚBâ�Ù0×� à°è â0è{w }�â�ÜÏÕ�æ�Û�ÞOw%�6èIé"Ó�âj�)æÏâ�ÖëÞ�à�Ù0Ü ä�Ó©Õ¥ÞÔÞ�Ó�âí�à��â�Ú�â�ÜÏçtâ ã�âtÞ�ä�â�â�Ü Ð Í�Ê6Î�Ì¥Í�ÑBÇtÊúÕ�Ü©í Ñ;÷¥Ío�Ñ;ÇQ��ÎÒÊkÕ)ç6ÞBæÏÕ�Û�Û�ï�àÖ à�Ö Ü�Ù0Þ`Õ0Û�ä"Õ3ï�ÖkÖëÞ�Ú;Õ�à�Ý0Ó�Þëþ×.Ù0ÚBä"Õ�Ú;í�Û�ï»Õ0ÜÏÖëä�â�ÚBâ�íÿìrÕ0ÖZç�Õ�Ü ã©âöÛ�â�Õ�ÚBÜ�â1í×.ÚBÙ0á$Ù)Ü�Ý0Ù)à�ÜÏÝ"í�à�ÖBçtæÏÖBÖ�à�Ù)ÜÏÖÿÕ¥Þ²ÞBÓ�âwÕ)ç�Õ0í�â�á`à�çÛ�â�ù0â�ÛIÙ)ÜmÞ�ÓÏà�ÖbÞ�Ù)î�àç�è é"Ó�â�ÚBât×.Ù)Ú�â�×.æÏÚëÞBÓ�â�Úí�â�ù0â�Û�Ù0îÏá â�Ü�Þ�Ù�×:r^|\}Z ,|\},��RQ�S�TVUXW�Y�T�Z\[^]_Tä�à�Þ�Ó�à�Ü,Þ�Ó�â{|rÜÏé Ù·± Ü�Ù¥ä�Û�â�í�Ý0âEî�ÚBÙ � â1ç6Þ ä�à�Û�Ûã�âfí�à�ÚBâ�ç6ÞBâ�íöÞBÙ¥ä�Õ0ÚBíÏÖ�Þ�ÓÏâ � Öëâ�á à�þ4��ýÕ�æ�ÞBÙ�þá¤Õ¥ÞBà�ç�Ý)â�Ü�â�ÚBÕ�Þ�à�Ù0Ü Ù�×µ}�ªY«fÚBâ�î�ÚBâ�Ö�â�Ü�Þ�â1í&w Ö�âtþá¤Õ�Ü�Þ�àçw ûAÜ�Ù¥ä�Û�â1í�Ý0â)ì�ä�Ó�àç;Ó�àÖ�Þ�ÙÈã�âªæ©Öëâ1íãAï$ÚBâ�Õ)ÖëÙ)Ü�à�ÜÏÝØÕ0ÜÏí>�)æÏâ�ÚBï$â�ÜÏÝ0à�Ü�â�ÖÈí�â�ù0â�Û¯þÙ0î�â�íkä�à¯ÞBÓ�à�ÜEÞ�ÓÏâdù)â�ÚBï ÖBÕ�á`âdî�Ú�Ù � â�çtÞ�è<�ÈÙ)Ú�âÖ�î©â1çtà��©ç�Õ0Û�Û�ï0ìwÞ�Ó�âÈÕ�Û�Ý0Ù)Ú�à�Þ�ÓÏá`Ö í�â�ÏÜ�à�Ü�Ý,Þ�Ó�â&:5�<C� 9/7�24;/9�3 8 ÚBâ�ÛÕ¥ÞBà�Ù)ÜÏÖ�Ó�à�î©Ö�ä�à�Û�Ûzã©âSÚBâtþ�ÏÜ�â1í»à�Ü»Ù0Ú;í�â�Ú¤ÞBÙúá`Ù0ÚBâÈî�ÚBâ�ç�à�Ö�â�Û�ï�ÚBâ�¡©â�ç6ÞÞ�ÓÏâ Ö�î�â�çtà��Ïâ�ÚBÖdÙ�×�çtÙ0Ü©çtâ�î�ÞBÖdÓ�Ù0Ûí�à�Ü�ÝIà�ÜZÖëî�âtþçtà��©çÈí�Ù)á¤Õ�à�ÜÏÖ � à°è â0è à�ÜÏÖëÞ�â1Õ0íØÙ�×Eç�æ�Ú�ÚBâ�Ü�ÞBÛ�ïÖëÞBÕ¥ÞBà�ÜÏÝoÞ�Ó©Õ¥Þ�ÕªÖ�î©â1çtà��©çEà�ÜÏÖëÞBÕ�Ü©çtâ & 6,240 ¬#�8ÓÏÕ)ÖÔÕ î�ÚBÙ0î�â�Ú�Þ�ï & 5�<�� 9/7424; 9�3 8 ä�à¯ÞBÓbù¥Õ0Û�æ�â&,0 943 8 ìWà�Þ á à�Ý0Ó�Þ ã�â Õ�ã�Û�â Þ�Ù ÚBâ��ÏÜ�âÞ�ÓÏâ &:5�<C� 9/7�24;/9�3 8 î�Ú�Ù)î©â�ÚëÞ�ï�ä�à¯ÞBÓ$Õ Ö�æ�ã�þÚBâ�ÛÕ¥Þ�à�Ù0Ü &�=:2:<)®/*�7�*,+/0:8 ä�à¯ÞBÓ ÞBÓ�â Ö�Õ0á âù¥Õ�Û�æ�â0èé"Ó�â>r^|\}Z ,|\},�� ÞBÙ�Ù)ÛfÖ�âtÞøÕ0ÖØÖëæ©ç;Ó

Þ�â�ÜÏí�Ö ÞBÙRÝ)Ú�Ù¥ä ä�à�Þ�Ó Þ�Ó�âð×.æ�ÜÏçtÞ�à�Ù0ÜÏÕ0Û�à�Þ�ïÙ�×¤à�ÞBÖbçtÙ0ÚBâ çtÙ)á`î©Ù)Ü�â�Ü�ÞbÕ)ÖÈädâ�Û�Û Õ0ÖÈä�à�Þ�Ó

ÞBÓ�âªà�á¤Õ�Ý)à�ÜÏÕ�Þ�à�Ù0Ü�Ù0×~Õ�ÜÏíö×ÒÕ0á`à�Û�à�Õ0Ú�à�Þ�ïöä�à¯ÞBÓ± ÜÏÙ¥ä�Û�â1í�Ý0âÉ�zÕ�ÜÏÕ0Ý0â�á â�Ü�ÞUÖBçtâ�ÜÏÕ�ÚBà�Ù�ÖýãAïû)â�ï�w ± Ü�Ù¥ä�Û�â�í�Ý)âÙ�zÕ0ÜÏÕ�Ý)â�Ú;Öwhà�Ü�ÞBÓ�âÈÛÕ�ÚBÝ0ââ�Ü�Þ�â�ÚBî�ÚBà�Ö�â�ÖØä�âmçtÙAÙ0î�â�Ú;Õ¥ÞBâ ä�à�Þ�Ó�è z Þ à�ÖÙ)æ�Ú`â��î�â�ÚBà�â�ÜÏçtâZÞBÓÏÕ¥Þoà�Ü�á¤Õ0Ü�ïØÖëà�Þ�æ©Õ¥Þ�à�Ù0Ü©ÖÞBÓ�â�ÚBâÔà�Ö"ÕEÛÕ�ÚBÝ0â�ÚdîÏÚ�Ù)ã�Û�â�áUà�Üfá¤Õ�ûAà�ÜÏÝkî�â�Ù�þîÏÛ�â�æ�Ü©í�â�Ú;Ö�Þ;Õ�ÜÏí�ÞBÓ�â î�Ù�Þ�â�Ü�Þ�àÕ�ÛEÙ)ÜSÕøÓAæ�þá¤Õ0ÜúÕ�ÜÏí Ù0ÚBÝ)Õ�ÜÏà�ÖBÕ¥ÞBà�Ù)ÜÏÕ�Û|Û�â�ù)â�Û�Ù�×~Ö�â�á¤Õ�Ü�þÞBà�çbÞBÙ�Ù)Û�Ö�ì Õ0ÖfÞ�Ó©Õ¥ÞZÞ�Ó�â�Ú�âöàÖÈÖëÓÏÙ¥ä�à�Ü�Ý�Þ�ÓÏâÞBâ�ç;Ó�ÜÏà�ç�Õ�Û|î�ÚBà�ÜÏçtà�î�Û�â�ÖEã�â�ÓÏà�ÜÏí à¯Þ1è5|rÜ�âoç�Õ0Üí�à�ÖBçtæÏÖBÖ ä�Ó�ïØÞBÓ�à�ÖfàÖ Þ�Ó�âöç�Õ)Öëâ)ì�Þ�Ó�â,á¤Õ�à�ÜÚBâ�Õ)ÖëÙ)Üöî©Ù�Ö�Ö�à�ãÏÛ�â ã©â�à�Ü�Ý Þ�ÓÏÕ�Þ ÛÕ�ÚBÝ0â�ÚEâ�Ü�ÞBâ�Ú�þîÏÚ�àÖëâ1Ö�Þ�â�ÜÏí`ÞBÙEÓÏÕ3ù0ârç�Õ�î©Õ�ã�Û�â�î©â�Ù0î�Û�â~ä�Ù0ÚBû�þà�Ü�ÝSä�à�Þ�Óåä�ÓÏÕ�Þ ädâ»ä�Ù0æ�ÛíUç�Õ0Û�Û�w ± Ü�Ù¥ä�Û�þâ1í�Ý0â8�zÕ�ÜÏÕ0Ý0â�á â�Ü�ÞOw�ì�Õ0Û¯ÞBÓ�Ù0æÏÝ0ÓöÞ�ÓÏâªâ�Ü�ÞBâ�Ú�þîÏÚ�àÖëâzÕ)Ö¤Õ�ä�Ó�Ù)Û�âzí�ÙAâ1Ö¤Ü�Ù�ÞfÕ�Û�ä�Õ3ï�Ö`Öëâ�â�áÞBÙøÚ�â1Õ�Û�à�Ö�âbâ�Ü�Ù0æÏÝ0Ó ÞBÓ�â�ã�â�Ü�â��ÞBÖZÙ0×`Õ)ç6Þ�æ�þÕ0Û�Û�ïØà�Ü�Þ�â�Ý0Ú;Õ¥ÞBà�Ü�ÝH©3à�á`î�Û�â�á`â�Ü�Þ�à�Ü�ÝúÖ�Ù0Û�æ�Þ�à�Ù0Ü©Öí�â�ù0â�Û�Ù)î©â1íúãAïúÖ�æÏç;Ó�±k�qí�â�îÏÕ0ÚëÞBá`â�Ü�ÞBÖ¤Õ�ÞÕÈç�Ù0á`îÏÕ0Ü�ïbä�à�í�Þ�Ó ÖBç�Õ�Û�â0è�~Ü Õ0ç�ç�â�î�Þ;Õ�ÜÏç�âÙ0×�ÞBÓ�â à�ÜÏí�æÏÖ�ÞBÚ�ïªÙ�×�ÞBÓ�â�î�Ù)ÖBÖëà�ã�à�Û�à�Þ�à�â�Ö"Ù0×�Þ�ÓÏâÖ�â�á¤Õ0Ü)ÞBà�ç ä�â�ãöÖëÓÏÙ0æ�Ûí ã�â¤ã�Ù�Ù�Ö�ÞBâ�íbãAïzÞ�ÓÏâÕ3ù¥Õ0à�ÛÕ�ã�à�Û�à¯Þ�ïØÙ�×�ÞBÙ�Ù)Û�ÖoÞ�ÙøÖëæÏî�î©Ù)ÚëÞfà¯Þ1è ~Û¯þÞBÓ�Ù0æÏÝ0Ó ç�æ�Ú�ÚBâ�Ü�ÞBÛ�ïZÙ0ÜÏÛ�ïfÞ�â1Ö�ÞBâ�í à�Ü,Öëá¤Õ�Û�Û�â�Ú1ìç�Ù0Ü�Þ�ÚBÙ0Û�Û�â�í â�ÜAùAà�Ú�Ù)Ü�á`â�Ü�ÞBÖ�ì3Þ�Ó�âdÞ�ÙAÙ)ÛAÖ�âtÞ�í�àÖ�þç�æÏÖBÖëâ1íSà�ÜSÞBÓ�à�ÖbîÏÕ0î©â�ÚbÖ�â�â�á¤ÖzÞ�Ù$Õ)í�í�ÚBâ�ÖBÖá¤Õ0Ü�ïªÙ�×�ÞBÓ�â�àÖ�Ö�æ�â�Ö"Ú;Õ�àÖ�â�í à�Ü Þ�Ó�àÖ�îÏÕ0î©â�Ú�è

²9

AIFB AIFB

Å � Á �~Â��ZÄ^�üÀ¯ DB°dJ1J�± ²�F�³x´�µ=9x¶)?�9 ·`D�F�D�?�9%¸

C�9;F�¹ @d8b³�>�º D�¹ °�»�7D¼¾½¿ F)¹b³KµÀ³1?�G Á�F)?1:=F�9;9�8�:¯F)? ÁÿF4¸Â : 8b³�F�C�9;F�¹ ¿ F)¹b³BÁ�¶0:}¹x½ Ã�9\ÄXÅ4¸F�:ÀÄ6DKµ 8�9\º/³�8o¹65 °dF)?19%´�D�F�¶,¹�9»°F4ÆÇ³�8�CED^¹�:�È >�F�¶ 74³�8�CEDKµ=9¼h9\ÉoÄXÅ)8�9;:=<�>)F�?KÉ Â 9B8oÆ D�Å)8�9;FÏ5ËÊ�F):À¸Â 9B8bÉ�:}¹°G ³�ÆÌ²üD�8bµÀÉ�8�>4Å�915 ÍÔ5Å)¹o¹bº:Î ÏBÏ�³�F�¹b³BÉ�9B8 Â 9�8x½ D�:}Æ<?½ >)F�:}¸È1Dt8bµ�É�8�>4Å�9B½ ¶0915�H�J1J�J,½

¯ ¼�@"J1J^± Íq½¦¼ 8�:ÀÄXÈ)µ¯9�GýD�F�¶ÑÐÒ½ Ó�½k@�>4ÅAD)½Ðd9\Éo³�>)8bÄ;9ÔÍ�9\ÉoÄB8�:Àº4¹�:À³1Fv708ëD�C�9%¸´e³�8bÈÖÕ×Ð¾Í�7ÙØÛÚ)ÄmÅ)9;CEDÜÚ,º�9\Ä%¸:ÀÝ�Ä6D�¹�:À³�F KK½ J,½ Ã©9\ÄmÅ)F�:ÀÄ6D�µ8�9\º/³�8o¹65 Þ ßCà à�³�F�Éo³�8o¹�:=>)C`5Å)¹o¹bº:Î ÏBÏ^´á´á´ ½ ´¾ß)½ ³�8�?�Ï�ÃeÐâÏ68X¶,ÆÇ¸ÉoÄmÅ)9;CEDCÏ¥5AH�J1J�J,½

¯ ¼eãCäKä�± ¼�½á¼ 8�9BCå¶�DKµdD�F�¶ö7�½áãK³BÅAD�F�É�9;F:½à ¿ Ð�æ ¿ ÐáÊ ·�çBÃ�9\ÄXÅ�F4³BµÀ³1?�G�D�F�¶°�º�º4µ=:�Ä;D�¹�:À³1F4É\½�Ã�9%ÄmÅ)F�:ÀÄ6DKµ�8�9\º/³�8o¹65à�³1?1F4»'ÃñD,½ É;5gè�DKµ�¶)9BF©5�é�³�8o´�D;G¥5Å)¹o¹bº:Î ÏBÏ^´á´á´ ½ Ä\³�?1F):À¹x½ F4³�Ï�5²K\äBäKä,½

¯ ¼eã�Ú4ê!äKä�± ¼�½ °Ò½â¼ 8�9BCå¶�DKµ 5Ô7�½âãK³BÅAD�F�É�9;FÏ5à�Å?½!Ú)ºAD�?�?1:¯D�8�: 5áÐÒ½ÙÁÿF�?�9\µÀÉ;5�D�F�¶Ðs½�ãB³�F�9\É\½Ûà 8�96D^¹�:¯F)? Dìë©9;D�8�F4¸:¯F)? ¿ 8�?1D�F):�É�D^¹�:À³1Fí¹bÅ08b³1>)?BÅíà�³1F,¸¹�9;F�¹Ô¼�D�É�9x¶îÍ�³�Ä;>�C�9;F�¹Ñ·`D�F,¸D�?19BC�9;F�¹x½ Ã�9\ÄXÅ�F�:ÀÄ6D�µ,8�9\º/³�8o¹65à�³1?1F4»'ÃñD,½ É;5gè�DKµ�¶)9BF©5�é�³�8o´�D;G¥5Å)¹o¹bº:Î ÏBÏ^´á´á´ ½ Ä\³�?1F):À¹x½ F4³�Ï�5²K\äBäKä,½

¯ Á�¼hJ1J�± Ðs½ è�½ æA½ïÁÿF�?�9\µÀÉ D�F�¶ ¼�½ °Ò½¼ 8�9;Cå¶�D�µ×½ »°F,Æð³�8�CED�¹�:À³1FñÁDò,¸¹�8ëDKÄ%¹�:À³1F:½ Ã©9\ÄmÅ)F�:ÀÄ6D�µ�8�9\º/³�8o¹65¿ F�Ã#³K²�F�³x´�µ=9x¶)?�9óà�³�F�Éo³�8o¹�:=>)C`5Å)¹o¹bº:Î ÏBÏ^´á´á´ ½ ³�F)¹b³KÈ�F�³x´�µ=9x¶)?�9B½ ³�8�?ÏK¶)9%µ�½ ÉoÅ�¹�C¦µ 5�H�J1J�J,½

¯ 7Aè�è ê J1J�±ôÍq½Ï7�9;F4É�9\µ 5A»m½:è�³�8�8b³�ÄXÈ,É;5�7�½:ÓhD�Fè"D�8�C�9%µ¯9BF©5!Ú/½DÍ"9\ÄXÈ39B865�·�½ÙÁ²8X¶,¸CED�F�FÏ5ZD�F�¶£·�½L² µ¯9B:¯F:½ ¿ »_ë:¯F D F�>,¹bÉoÅ�9\µÀµ×½ »°FõÐÒ½ìÍ�:À¸9;F�? 9%¹ DKµ×½=5N9x¶0:À¹b³�865R�ü�A%�#�ö}÷��\ø¥��ù¾úbûB��'�� .%��4ü�� %6�¥�%ö ��ø �� \ø¥�mýE�� %xú��ø�'$%��(��%�*)��ð�ü��%�#�ö¯��\ø¥�þù�ú\÷ûB�0��'ë�� .%�� ÿ�%��6�6��/úë��+�� âù � ÷-0/3/3/�4x½KÚ)º)8�:=F�?�9B865K¼h9B8bµ=:=F©5Kè�9;:�¶)9%µÀ¸<�9B8�?)54é�9%´��³�8bÈ�5�H�J�J1J,½

¯ 7 Â è�²Ò°�äKä�±qÍq½ 7)9;F�É�9\µ 5o7�½ Â D�FÑè"Dt8�C�9\µ=9;F©5·�½�² µ=9;:=FÏ5�D�F�¶Iès½!°�È)È39B8�CED�F�É\½¿ F4¸�Ã#³K¸V²�F4³^´h9\µ�¶)?�9BÎ ¿ F�¹b³BµÀ³1?�G,¸<ADKÉ�9\¶ Ã#³�³KµÀÉ ÆÇ³�8 ²�F4³^´�µ=9x¶0?19·`D�FAD�?19;C�9BF)¹x½C»°FÔ��%xú��ø�'w%��(��E��B��'��'�'o��o��úë%�ýqýE��'úë�ÿ�%��6�6��/úë��5 ·`DB¶�8�:�¶Ï5îÚ)ºAD�:=FÏ5KxäBäKä,½

¯ èd9;FAJ�J�± ã,½ èd9;F�¶,µ=9B8x½ Í�°¾Ð�æ °°w?19;F�¹ ·`D�8bÈ�>�º ë©D�F)?1>AD�?19B½Ã©9\ÄmÅ)F�:ÀÄ6D�µ 8�9\º/³�8o¹65 Í"9%¸Æ�9;F�É�9 °�¶ Â D�F4Ä;9x¶ Ðd9\É�9;D�8bÄXÅæ²8b³��9\Ä%¹bÉ °d?�9;F�ÄBG Õ×Í�°�Ð�æ:°�Ø�5Å�¹o¹bº:Î ÏKÏ^´á´á´ ½ ¶�D�C¦µ×½ ³�8�?�Ï¥5©H�J1J�J,½

¯ è�7D¼áêDäBä^± »X½ è�³�8�8b³�ÄXÈ)É;5 Íq½ 7�9;F�É�9%µ.5ã,½Ö¼ 8b³¥9\È)É'¹�8ëD05 Ú ½ Í"9\ÄXÈ39B865·�½4Á²8X¶)CED�F�F©54à�½�@�³�<�µ=915�7�½ Â D�Fè�D�8�C�9\µ=9;FÏ5#·�½?² µ=9;:=FÏ5AÚ/½#Ú)¹ëD1D�<©5D�F�¶ ÐÒ½ Ú�¹�>�¶09B8x½ ÃeÅ)9 ¿ F4¸¹b³KµÀ³1?�Gî»�F4Æ9�8�9;F�Ä;9 ë�D;G¥9B8 ¿ »_ë!½Ã©9\ÄmÅ)F�:ÀÄ6D�µ�8�9\º/³�8o¹65 ¿ F�Ã#³B²�F4³^´�µ}¸9\¶)?19 Á!Ê¾¸V»oÚ4Ã�¸�K;J0Kxß1H æ²8b³��9\Ä%¹Í�9\µ=: Â 9B8ëD�<4µ¯9 é�³4½ ¿ Ãe²â¸VÍÔK15Å�¹o¹bº:Î ÏKÏ^´á´á´ ½ ³1F�¹b³KÈ0F4³x´�µ¯9\¶)?19K½ ³�8�?)5K\äBäKä,½

¯ ²âàeæ °�J1J^± @�½ ²üDt8 Â ³1>)FAD�8ëD�È�:�É;5Ós½ à�Å)8�:ÀÉ'¹b³Kº�Å�:�¶09\É;5ÛÍ�½1æ!µ=9%ò)¸³�>�É�D�É�:�È�:ÀÉ;5qD�F�¶ Ú ½ °�µ=9%ò)D�È0:×½� >)9B8�G0:=F�? à�³1C�CÔ>)F�:}¹°G ÞZ9;<æA³�8o¹ëDKµÀÉ\½DÃ©9\ÄmÅ)F�:ÀÄ6D�µ�8�9\º/³�8o¹65,»'àáÚ)¸7 ¿ Ð!Ãeèr5�èd9B8ëDKÈ)µ=:À³1FÏ5R@d8�9;9\Ä;915Å�¹o¹bº:Î ÏKÏ^´á´á´ ½ :ÀÄ\É\½ Æð³�8o¹bÅ?½ ?�8mÏ^º)8b³��bÏ:ÀÉoÉ'¹mÏ^Ð�Í�7ÙÏ^Ð � ëAÏ68)µ×½�Õ×Å�¹�C¦µ 54º ¶)Æ°5º4É;5�¶ Â :ÇØ�5�H�J1J1J)½

¯ ³C· °w@"J1J�± ÊdF�: Â 9B8bÉ�:}¹�G ³KÆ ·`D�F�ÄXÅ�9\É'¹�9B865708�9;9 Ê�F�: Â 9�8bÉ�:À¹�G °wC¦É'¹�9B8X¶�D�C`5D�F�¶ »°F�¹�9B8bº08�:ÀÄ;9I@�CÔ<�ès½ ¿ :Àµ Á!¶:½Ã©9\ÄmÅ)F�:ÀÄ6D�µ 8�9\º/³�8o¹65 Ê�F): Â 9B8o¸É�:}¹�G ³�Æ ·`D�F�ÄXÅ�9%É'¹�9B865 Ê�² 5Å�¹o¹bº:Î ÏKÏ�:=C�?4½ Ä\É\½ CED�F?½ DKÄB½ >4È�Ï^³1:Àµ Ï�5H�J1J�J,½

� ´10

AIFB AIFB

�� !��"!#��"�$%��'& �( ��) ��*,+-�/.0�1��2��"

3547684:9<;>=@?BA8CEDF4%GH68?JILKB4:MBNPO�4%Q7DF;RO�4:9S4STU4:V 4%MBNPWX4%9Y4%Z[;\A]4�^_4:DF4:`:?Ba]ZB;\VbFc<dEe]f7gihkj2lEmon%lEm\hqp�dsrutvdw<x2ykw7m\lEruz�dshu{|lEw}n%e]{~t]wUe]tE�<yqp!dhut��rutHm�tHeHhuf7r�dgJ��w7{~�StHru��{ihkj

�s�E��:�E�@� dE�%{~��d]p�d��fU�El � dE�%{~��d]p�d@yqp!dhut�� q� ��L��S� � ��%f7rut��lEm\h�� {ip�dshutH��UfB� deL� � �

¡�¢/t]�:h��U£XlLz¤�Uf:hutHr�n:eH{|tHwUe]tL�<n:¥7{~�]fUlE�Ld��wU{~�St8ru��{~hqj�s�:�]�<�� lE¥UlE�:f'¦/dEz�dz�dhu��f'n:¥7{|�HfUlL�Edv§ �S� �k¨E� �E� �L� � ��

© e]��ªE�E«L¨%�<j%dEz�d�Lf:hu{�¬��e]�]� {~w7mk� ��¥7{|�HfUlL�Ed:� deL� � �

�®�¯]°]±s²<³E°L´ ¦�lspµhul0tH¶:�Ug~lL{ih¤d}z�de8¥7{|w7tH�·rut�dLx7dE¸7g|t�x7{~eHhu{~lEw<dr�jº¹�»'¼�¢�½�dEwUxhutH¶%h2e]lru�UfU�2{~w¾��fU�7�¿lEr�hu{~wU�Phu¥7t0e]lLw7� h�rufUeHhu{~lLw¾lm�x7lEz�dE{~wºlLwYhulEg|lE�L{~t]�'hu¥Udh��¿t]eH{~mÀj h�ds¶7lEwUlEz¤{|e!dEw<x�wUlEw7�·h�d¶:lLw7lLz¤{~eXrut]g|dshu{|lEwU��¥7{|�7�Xdz¤lLw7�/�L{~�St]w x7lEz�dE{~we]lEwUe]tH�7hu�/dsrut x:{~��e]fU��t�x}¥7tHrutL�[dL½_ykw}¸7fU{~g|x7{~wU�@h�ds¶:lLwUlEz¤{~e�rutHgÀdshu{~lLwU��¥7{~�U�@¹Á¥7{~�tHr�dsrue8¥Yj�� h�rufUeHhuf:ruts½�lEmRx7lEz�dE{~wFe]lEwUe]tH�7hu�]�<��lEz¤t�¥U{~tHr�drue�¥Yj'� h�ruf7eHhuf7rut�e]dEw}¸¿ttH¶%h�r�deHhut�x�m\rulEz�»'¼_¢Âp�{ihu¥5z�dru�Lt�xP��fU¸:�·h�rut]t]�vhu¥Udh z�d�j�¸¿t2z¤l%x7{iÃ<t�x�¸Yjd�x7lLz�d{~wPtH¶:�¿tHr�h��[f7��{|w7�'¸¿lhu¥Pz�dhue�¥U{~w7�2rut]��f7gih�dEw<dgij7��{~��dEwUx0h�ru{~z¤z¤t�xFrutH��f7g~h>dEwUdEgij:��{|�]�¢/lEz�dE{~w7��¿t]eH{~ÃUeÄ¥U{~tHr�drue�¥U{~e�dg%� h�rufUeHhuf:rutRe]dEwvdEg~��l�¸¿tRtH¶%h�r�dEe8hut�xmÀrulLzÅhutH¶%h�eHlEru�Uf7�]�fU��{~wU�_�<d{~ru��lm<e]lEwUe]tH�7hu�>hu¥<dsh>huf7ruw�hul�¸¿tXg~lYe�dhut]xvwUt]drRdEwUx¥Uds�Ltv��{~z¤{~g|dr/eHlLwYhutH¶%h�¸Yj�Æ�lEr�xUn:�UdEe]tL� � ¥:f7��hqp!l2x7{iÇ[tHrut]wYh��%{|wUx7��lEmR¥U{~tHr�dsr��e�¥U{~e�dg7� h�ruf7eHhuf7rutoe8¥UdEwU�EtX{~wYhul�f7wU{iÃ<t�x�lLw7top�{ihu¥@dExUx:{~hu{~lEw<dEg%z¤l%x:{~ÃUe�dhu{~lEw ¸Yj�dx:lLz�dE{~w@t8¶7�¿t8r�h��L¸<½�yqw ¸7fU{~g|x7{~wU��w7lLw7�·h�ds¶7lEwUlEz¤{|eorut]g|dhu{~lLw7��¥U{~�U��¹Á��¿t]e]{iÃ<e�dshu{~lLwhut]z¤�7g|dhut]��½�lEmox7lEz�dE{~wFe]lEwUe]tH�7hu�]�<pXtve]lEwU� h�ruf7eHh�eHlLwUeHt]�7h��¿t]e]{iÃ<e]dhu{~lLw}hut]z@��7gÀdshut]�ohu¥UdhXe]lLz¤t�m\rulEzµ�UdE{iru�XlmBe]lLw7e]t]�7hu�Rhu¥<dshohuf:ruw@hul�¸¿t�g~lYe�dshut�x¤wUt�dsr�dEwUx¥Uds�Lto��{~z¤{~g|drÄe]lEw%hut8¶:hJ¸Yj/Æ�lr�x<n:�UdEe]tE� � x:lLz�d{|w�tH¶:�¿tHr�h>x7lYt]�Èhu¥Utoh�d��¸<d��t�xlEw2hu¥7t]zÉg|dhut8r�� ¥7t�e�d��t/� huf<x%j@p�{ihu¥��lEz¤t�g|d]pÊe�dEg~g~t�x'£Xy�nUË��¥Ul�p��fU�Xhu¥Udhhu¥7t�h�r�dEx7tH��lEÇ�¸¿tHhqp!t]t]w0�:rut]e]{~��{~lLw�dEwUx}rutHe�dEg~g�{~��l�{~z¤�¿lEr�h�dw%h�{~w0�7r�dEe8hu{|e]dEg~gij¸7fU{~g|x7{~wU�¤x:lLz�d{|w�lLwYhulLg~lE�L{~t]�]�

Ì ÍUÎ_Ï¿Ð[Ñ�ÒvÓ�ÔJÏ[ÕuÑ_Î

Ö�× G8Z[Q7?[`:Z�Q7M7GHQ × Q:`7;ÀCL6�ZÈ4SØ:C@ÙÈCECM�Ø7CAHT0ÚÈQ7Ú[? × 4%A�;\M�D}4%M<T�4:Ú[Ú × ;\aE4YG8;\Q:M�4:A8CL4:6EK¿Û�C�6�G8; ×À×Ü 4:asC�G8ZBCPÚBA8Q7Ù × CEDÝQ Ü ZB;À`7ZÅaQ76�GF476868Q¿as;�4YG8CLN¾Û�;ÀG8ZÅÙ[?[; × N¿;\M[`Þ?[Ú-G8Z[CEDßD}4%M<?B4 ×\× T:à!á�MÚB4%A8G8;�as? × 4%ALK<6�;\MBasC¤N¿Q7D}4%;\M�Q:MUGHQ × Q:`:;\CE6�ZB4SØ7C�G8ZBC DFCL4%M[;\M[`F6�ÚJCEa;|âÈavGHQF4:Ú[Ú × ;\aE4YGH;ÀQ7M�N¿Q:ãDF4:;ÀMÈ6K[Z<?[D}4:MPCä¿ÚÈCEA�G]6�ZB4SØ:C GHQ}DF4:9:C¤Z<?[`7C¤Cså>Q:A8GH6 Ü Q:AvasQ:MÈ6uGHA8?BasG8;\M[`FG8ZBCDæCMUG8;\A8C × TÙUT�ZB4%MBNÄàá�MçQ:A]N¿CA�G8Q-A8CLN¿?BasCèG8Z[CºasQU6uG]6K/4%?[G8Q:D}4%G8;�a5Q:AP6�CEDF;|ã�4%?¿GHQ:D}4YGH;\aèDFCsG8ZBQ<NB6�ZB4SØ7C

ÙÈCECMçÚ[AHQ:ÚJQ768CENé?È6�;\M[`�9<M[QYÛ × CEN[`:C�CMB`:;\M[CCEA8;\M[`¾GHCEa]Z[M[;�êU?[CE6P4%MÈNëMB4YGH?[A]4 ×�× 4%MB`:?B4:`:C

� ¥7{~�X�<d�¿tHrR{~�X{|x7t]wYhu{~e�dg7hul/hu¥Ut�lEwUt��Uf7¸Ug~{~��¥Ut�x {~w¤y � £ � y�ì � � Æ�lEru�%��¥7lL�@lEw2í�w%hulEg~lL�Ej@îÈt�dsruw7�{~w7�

11

�

Ú[A8Q¿aCE6H6�;\M[`0Q7M[CE6�ï�a Ü àRðvMUG8QU684:?[AH?B6�ñ ò<Û�4:A�GHQ:?¿G�CGEàB4 × à�óLô:ô:õLö·÷�à�W/Z[C'4%?¿GHZ[Q:A]6�ZÈ4SØ:C24 × 68QN¿CØ7C × Q:ÚJCENø4çN¿Q7DF4:;ÀMùQ7MUG8Q × Q:`7TçAHCsâBM[CEDFCMUGè6�?BÚ[ÚÈQ7A�GúCM<Ø<;ÀAHQ:M[DFCEM7GèaE4 ×\× CLNùûXð ü ýñ =@?[AHCD}4%GH68?P4:MBN0^_4:D}4%`:?Èa]Z[;�óEô7ôUþö[4%MBN24vN¿Q7D}4%;\M�Q:MUG8Q × Q7`:T AH4:Ú[;�N�N¿CØ7C × Q:Ú[DFCEM7GXCM¿ãØU;\AHQ:M[DFCMUG@a4 ×\× CENèü¤ð ü üvûoýÿñ ò<CE9U;\?Ba]Z[;�CGEàB4 × à�óLô:ô��LökK�?B68;\M[`�D}47a]Z[;ÀMBC2AHCE4:NB4%Ù × C�N¿;�a�ãG8;\Q:MB4:A8;\CE6Eà��QYÛ�CEØ:CEAEKBG8Z[CL6�C'CM<Ø<;ÀAHQ:MBD'CEMUGH6 Ü 4:as; × ;|G]4YG8C'GHZ[C'aQ:MB6�G8AH?BasG8;\Q:M5Q Ü Q:M × T�4�Z[;|ãCA]4%A]a]Z[;\aE4 ×\× TP6uGHA8?Èa�G8?BA8CLN56�CGvQ Ü N¿Q:D}4%;\MèaQ:MBaCÚ¿G]6KJ;ÀMèQ%G8ZBCA Û�Q7AHN[6EK[G]4Yä¿Q:MBQ:DF;\a2asQ:M[ãasCÚ[G8?B4 × A8C × 4%G8;\Q:MB68Z[;\ÚB6EàÖ 6�N¿Q:D}4%;\M�Q7MUG8Q × Q:`7;ÀCL6�ZB4SØ7C�ÙJCCMP4%Ú[Ú × ;\CEN0G8Q'Û�;\N¿CL6�ÚBA8CL4:N}4%AHCE476K<68?Ba]Z�4:6�9<M[QYÛ × ã

CEN¿`7C�6�ZB4:A8;\M[`ÈK¿9UMBQYÛ × CLN¿`:C@AHC?B68C:KB68Q Ü GuÛ/4%AHC¤4%`7CMUGH6�4:MBNP;ÀM Ü Q7A8D}4%G8;\Q:M�;ÀMUGHC`:A]4YGH;ÀQ7MÄK¿Û�CM[CCLNÊ6�Q Ü GuÛ�4:A8C�CM<Ø<;ÀAHQ:M[DFCEM7G]6@G8ZÈ4YG}6�?[ÚBÚÈQ7A�GF4�Z<?[D}4%MÊCä<ÚJCA8G�;\M�aQ:MB6�G8AH?Ba�GH;ÀMB`5G8ZBCN¿Q:D}4%;\MÅQ:MUG8Q × Q7`:;\CE6�Û�;|GHZÅMBQ%G}Q:M × T¾G]4Yä¿Q:M[Q7DF;\aPasQ7MBasCEÚ¿G8?B4 × A8C × 4%G8;\Q:MB68Z[;\ÚB6'Ù[?¿G�4 × 68QM[Q:M¿ãkGH4%ä¿Q:M[Q7D';�a'Q:M[CL6àoá�M¾Q:A]N¿CA�G8Q5N¿CEØ:C × Q:ÚºGHZ[C0CEMUØ<;\A8Q7M[DFCMUGH6EKÄ;ÀG�68CCD}62ÙÈCG�G8CEA�GHQÚ[?¿G�GHQ:`:CG8Z[CEA!GuÛ�Q@Q7A�D'Q7A8C�GHCEa]Z[M[;�êU?[CE6�68?Ba]Z}4:6!9<M[QYÛ × CEN[`:C/CM[`7;ÀMBCCAH;\M[`BKSMB4%G8?[A]4 ×[× 4%M[ã`:?B4:`:C�Ú[AHQ¿asCE6H68;ÀM[`ÈK�D}4:a]Z[;\M[C × CE4%AHM[;\M[`�4%MÈN¤N[4YG]4/CM[`7;ÀM[CECAH;ÀMB`BK476Ä6�CECM¤;\M@G8Z[C_Û�Q7A89¿68Z[Q:ÚQ:MPQ:MUG8Q × Q7`:T × CE4%AHM[;\M[`F;ÀM�ý�� Ö á��Pï·C:à `BàÄñ 354:CEN[a]Z[C¤4:MBN�òUGH474%Ù��LöÁ÷sà

��CAHC};\MúGHZ[;�6�ÚB4:ÚÈCEAEKÄÛ�C}Csä<GHCMBN¾ü¤ð üvü ûRýø;\MUG8Qèü¤ð ü üvûoýøá�á G8ZB4%G�asQ7MB6�G8AH?Ba�G]6ÙÈQ:G8Z5G]4Yä¿Q:MBQ:DF;\a�4%MBN5M[Q:M¿ãkGH4%ä¿Q:M[Q7D';�a�aQ:MBaCÚ¿GH?B4 × A8C × 4%G8;\Q:MB68Z[;\ÚB6EKÈCä<Ú × Q7;|GH;ÀM[` �úQ7AHN¿ãO�CsG�ñ �[C ×\× ÙB4%?[D óEô7ô�Lö�4%MBNÊN[Q:D}4%;\M¿ã�6�ÚJCEa;|âÈa}G8Csä<G]6�Û�;ÀG8ZÊG8Z[C�4%?¿GHQ:D}4YGH;\a�4%MB4 × T¿68;\62Q Ü× Cä<;�a4 × aQ%ã Q¿aas?BA8AHCMBaC�6�GH4YGH;\6�G8;�a6EK!ÙB4:68CENÊQ:M��ÞQ:A]NBò<ÚÈ4:asCÞñ 3è4%A8G8;Ä4%MÈN�ò[a]ZU?[G��Cö/G8ZB4%GZB4:6�G8ZBC ;�N¿CL4�GHZB4YG�4�ÚB4%;\A/Q Ü GHCAHD}6�Û�;ÀG8ZPZ[;\`:Z Ü A8CLê7?BCMBaT}Q Ü aQ%ã Q¿aas?BA8AHCMBaC�6�GH4YGH;\6�G8;�a6a4%M�ZB4SØ:C�M[Q7M¿ãqG]4Yä¿Q:MBQ:DF;\a�asQ7MBasCEÚ¿G8?B4 × A8C × 4%G8;\Q:MB68Z[;\ÚB6à��B?[A�GHZ[CAHDFQ:AHC:KLÛ�C�CØY4 × ?B4%G8C/Z[QYÛü@ð ü üvûoý á8á�Û�Q:AH9<6�;ÀMèG8Z[C�âBC × N5Q Ü × 4SÛ¤KÈG8ZBC��Q7M7GHAH47a�G]6 Ü Q:AvG8ZBC�á�MUG8CEA8MÈ4YG8;\Q:MÈ4 × ò[4 × CQ Ü�� Q<Q<NB6�ï��á�ò � ÷sàÈW/Z[C2CDFÚ[;\A8;�a4 × A8CL6�? × G]6�6�ZBQYÛµ?B6�G8ZB4%Gvü@ð ü ü ûRýëá8á/aE4%M�68?[Ú[ÚJQ:A8G4 × 4SÛ Cä¿ÚÈCEA�G�;\M�asQ:MÈ6uGHA8?BasG8;\M[`FN¿Q:D}4%;\MPQ:MUG8Q × Q7`:;\CE6Eà

� �� Í¿Í! #"$�ëÑ&%('�Õ8Î)��Î�Ï¿Ñ�*uÑ�+-,/.0'21vÕ�Ò3��465�47*�Ñ819%:4RÎ�Ï��Î&5�ÕuÐ[Ñ_Î9%:4RÎ�Ï

�o;\`:?BA8C}ó�68Z[QYÛ�6�4%M�QYØ7CAHØU;\CÛµQ Ü ü@ð ü ü ûRýëá8á�K�;84FüvQ:D}4%;\MúðvMUG8Q × Q7`:T}A]4%Ú[;�üÂüvCØ:C × ãQ:Ú[DFCMUG�ý_MUØ<;\A8Q7M[DFCMUG=<�G8ZB4%G�ZB4:6�GHZ[C Ü Q ×\× QYÛ�;ÀM[`'GuÛ�Q}asQ:DFÚJQ:M[CEMUGH6?>@ Wo4%ä¿Q:M[Q7D';�a A8C × 4%G8;\Q:MB68Z[;\Ú�47aêU?[;�6�;ÀG8;\Q:MPDFQ<N[? × C¤?È6�;\M[`A�úQ7AHNBO�CsG@ O�Q7M¿ãqG]4Yä¿Q:MBQ:DF;\avAHC × 4YG8;\Q:MÈ6�Z[;\Ú × CL4%AHM[;ÀMB`'DFQ¿N¿? × C@?B68;ÀM[`0N[Q:D}4%;\M¿ã�6�ÚJCEa;|âÈavGHCsä<GH6Ö N¿Q:D}4%;\MPCsä¿ÚÈCEA�G�`7;ÀØ7CE6/4F6�CG�Q Ü N¿Q:D}4%;\MPG8CEA8D}6�G8Q'GHZ[C�6�T¿6�G8CEDPàÖ ÷�W/ZBC}GH4%ä<Q7M[Q:DF;�a}AHC × 4YG8;\Q:MÈ6�Z[;\Ú¾4:aEê7?B;\68;|GH;ÀQ7M¾DFQ<N[? × Cúï�WCB Ö D'Q¿N¿? × CS÷�N¿Q<CE6D;868ÚÈC ×À×DF4%GHa]ZE<�ÙÈCGuÛ�CECM}GHZ[C�;\M[Ú[?¿G�N¿Q7DF4:;ÀMFGHCAHD}6�4:MBNA�úQ7AHNBO�CsGLàUW/ZBCF;868ÚJC ×\× D}4YG]a]ZE< × ;ÀMB9<6G8Z[CL6�CèG8CAHD}6�G8Q��ÞQ:A]N[O�CsGEà/W/Z<?B6�G8Z[CÞ;ÀMB;|GH;\4 × DFQ¿N¿C × Ü AHQ:D G8ZBCG;H6�ÚJC ×\× D}4%GHa]ZE<�A8Cã6�? × GH6�;\6¤4�Z[;\CA]4%A]a]Z[;�a4 ×À× T�6uGHA8?BasG8?[AHCENÞ68CsG�Q Ü 4 ×\× G8ZBC}M[Q¿N¿CE6@Q7MúGHZ[C}ÚB4%G8Z Ü A8Q7D�G8ZBCE68CG8CAHD}6�G8Q�G8Z[C@AHQUQ:G�Q Ü �ÞQ:A]N[O�CGEàH��QYÛ�CØ7CA�GHZ[C@;ÀMB;|GH;\4 × D'Q¿N¿C × ZB4:6�?[M[MBCEasCL686H4%AHTF;ÀMUGHCA8ãMB4 × G8CEA8D}6'ï·M[Q¿N¿CL6H÷sàÄW/Z[CT5N[Q�M[Q:G@aQ:MUG8AH;ÀÙB?¿G8C�G8QP9:CECÚ[;\M[`0GHQ:ÚJQ × Q:`7;\aE4 × AHC × 4YGH;ÀQ7MB68Z[;ÀÚÈ64%DFQ:M[`ºD}4YG]a]Z[CEN-M[Q¿N¿CL6K_68?Ba]Zé4:6FÚB4:A8CEMUG�ã�a]Z[; × N-A8C × 4%G8;\Q:MB68Z[;\Ú-4%MÈNÅ68;\Ù × ;\M[`ºAHC × 4YG8;\Q:M[ã6�Z[;\ÚÄàÈò¿Q'Û�C¤a4:M�GHA8;\D GHZ[C¤?[M[M[CLasCL686H4%AHT};ÀMUGHCAHMB4 × M[Q¿N¿CL6 Ü A8Q7D G8ZBC¤;ÀM[;ÀG8;�4 × D'Q¿N¿C × ;ÀMUGHQ4FG8AH;ÀDFDFCEN5D'Q¿N¿C × K>4:6 68Z[QYÛ�M�;ÀMI�o;\`:?BA8CJ�[àBá�M5Q7AHN[CA�GHQ�AHCsâBM[C2G8Z[C2G8AH;ÀDFDFCEN5D'Q¿N¿C × KÛ�C¤ZB4SØ7C G8Z[C Ü Q ×À× QYÛ�;\M[`'G8Z[AHCC26�G8A]4YGHC`7;ÀCL6�G8ZB4%G�Û�C¤Û�; ×À× N¿CE6HasAH;\ÙÈC × 4YGHCA�;\M�G8Z[C�aQ:MUG8Cä<GQ Ü ;ÀMUG8CEAH47a�GH;ÀQ7M�Û�;ÀG8Z�4%M�?B68CAK>@ 354YG]a]Z[CEN�AHCE68? × G�4:MB4 × T¿6�;�6

12

AIFB AIFB

�

a Set of Domain Terms

WordNet matched result analysis

trimmed result analysis

modification usingthe syntactic strategies

modification usinga candidate for domain-specific

hierarchy structure

Taxonomic Relationship

extraction of 4-grams

constrution of WordSpace

extraction ofsimilar concept pairs

extraction of a candidate for domain-specific

hierarchy structure

Non-Taxonomic Relationship

Domain SpecificTexts

(Text Corpus)

Concept Specification Template

A Domain Ontology

Extension & Modification

A Candidate for Domain-Specific

Hierarchy Structure(Pairs of Concepts)

User

A)Taxnomic Relationship Aquisition Module

B)Non-Taxnomic Relationship Learning Module

MRD

L�MON ´�P<´ ¢�í�¢/¢/î�Q5y�yXls�StHru�%{~tHp

@ WRAH;\D'DFCLN�AHCE68? × G�4%MB4 × T¿68;\6@�R 68;ÀMB`�4Êa4:MBN¿;�N[4YGHC Ü Q7A�N¿Q:D}4:;ÀM¿ã�68ÚÈCLas;ÀâÈa�Z[;\CA]4%A]a]Z<T-6�G8AH?Ba�GH?[A8CèCsä<G8A]4:asG8CEN Ü AHQ:DG8CäUG�aQ:AHÚ[?B6EàS/÷2W/Z[C0M[Q7M¿ãkGH4Yä¿Q7M[Q:DF;�aFA8C × 4%G8;\Q:MB68Z[;\Ú × CE4:A8MB;ÀM[`�DFQ¿N¿? × CúïqO�WCB�û DFQ<N[? × CS÷¤CäUGHAH47a�G]6G8Z[C�ÚB4:;ÀA]6'Q Ü GHCAHDF6'GHZB4YG068Z[Q:? × N�ÙÈC�AHC × 4YG8CLNÊÙ<T�68Q:DFCPA8C × 4%G8;\Q:MB68Z[;\Ú Ü A8Q7D N¿Q7DF4:;ÀM[ã6�ÚJCEa;|âÈa5G8CäUG]6K�4%MB4 × TT�E;ÀM[` × Csä¿;�a4 × asQ:ãkQ¿aEas?[AHA8CEMBasC�6�GH4%G8;�6uGH;\aE6K�ÙB4:68CENéQ7MU�úQ7AHNÈò<ÚB4:aCG8ZB4%G�;\6�4ºD2? × G8;Àã N[;ÀDFCMÈ6�;\Q:MB4 × K_A8CL4 × ãkØY4 × ?[CLN�Ø7CEa�GHQ:A068ÚB4:aC�Û�Z[CAHC�G8ZBCèasQU6�;\M[C�Q Ü G8ZBC4%M[` × C�ÙJCsGuÛ�CCEM�GuÛ�Q¤Ø:CLa�GHQ:A]6!;\6�4¤asQ:MUGH;ÀM<?[Q7?B6!DFCE4:68?[AHC�Q Ü G8Z[CE;ÀA�6�CED}4%MUG8;�a/A8C × 4%G8CLN¿M[CE6H6EàW/ZU?È6_GHZ[C ÚB4%;\AH6�Q Ü G8CAHD}6_Cä<G8A]4:a�GHCEN Ü A8Q7D�N[Q:D}4%;\M¿ã�6�ÚJCEa;|âÈa�G8CäUG]6/4%AHC�G8ZBC@aE4%MBN[;\N[4%G8CL6Ü Q:A'M[Q:M¿ãkGH4%ä¿Q:M[Q7D';�aFAHC × 4YG8;\Q:MÈ6�Z[;\ÚB6Eà7�ÞC�aE4%MÊÙ[?B; × GFaQ:MBaCÚ¿G'6�ÚJCEa;|âÈaE4YGH;ÀQ7M¾G8CED'Ú × 4%G8CE6ÙUTºÚ[?[G�G8;\M[`èG8Q:`7CsGHZ[CA2GH4Yä¿Q7M[Q:DF;�a}4%MBN¾MBQ:M¿ãkGH4%ä<Q7M[Q:DF;�aFA8C × 4%G8;\Q:MB68Z[;\ÚB6 Ü Q7A�GHZ[C�;\M[Ú[?¿GN¿Q:D}4%;\MúG8CAHD}6àÄW/ZBC'AHC × 4YGH;ÀQ7MB68Z[;ÀÚÈ6 68Z[Q7? × NúÙJC';�N¿CEM7GH;|âÈCENÞ;ÀMúG8ZBC};ÀMUG8CEAH47a�GH;ÀQ7MèÛ�;ÀG8Zº4ZU?BDF4:M�Cä¿ÚÈCEA�GLà

V WX'ZY�Ñ_Î�Ñ�%�Õ8ÔU.047*['XÏ[ÕuÑ_Î9\T]�Õ[1^"ëÔ6_/Ó�Õ�\UÕuÏ[ÕuÑ_Î

Ö Ü G8CEA�`7CsG8G8;\M[`5GHZ[C0GHA8;\DFD'CLN¾DFQ¿N¿C × KXWCB Ö DFQ¿N¿? × C�;�6�AHCsâBMBCEN¾Ù<Tº;ÀMUGHCA]4:a�GH;ÀQ7MºÛ�;|GHZ4PN[Q:D}4%;\MúCä¿ÚÈCEA�GLK�?B6�;\M[`PG8ZBC Ü Q ×À× QYÛ�;\M[`�GHZ[AHCC06�G8A]4YG8CE`:;\CE6?>ÈD}4%GHa]Z[CLNèAHCE68? × G¤4:MB4 × T¿6�;�6EKG8AH;ÀDFDFCEN�A8CL6�? × G�4:MB4 × T¿6�;�6�4:MBN}?B68;ÀMB`�N¿Q7DF4:;ÀM[ã 68ÚÈCLas;ÀâÈa�ZB;ÀCEAH4:AHa]Z<T�6�G8AH?BasG8?[AHC�CäUGHAH47a�GHCENÜ A8Q7D G8CäUG�aQ:AHÚ[?B6EàûÄQ<Q:9<;\M[`�4YG�GHZ[C0GHA8;\DFD'CLNºDFQ<N[C × Ko;ÀG�G8?BA8MB62Q:?¿G2G8ZB4%G�;ÀG2;\6�N¿;\ØU;�N¿CLN¾;ÀMUGHQè4�` Ö S

ï�4A` Ö G8ZP;ÀMÈa × ?BN¿;\M[`}Q:M × TFS�CE6�G�6�ÚJC ×\× ãkD}4YG]a]Z[CEN�M[Q¿N¿CE6]÷/4%MÈN�4Fò¿W�3 ï�4}ò<?[Ù[ã WoA8CEC G8ZB4%G

13

AIFB AIFB

§

a b c d e f d g h b i j k l mn o p q r o s t u v w qx y y z { z | | } ~ � � � � z

� � � � � � � �

L8M�N ´?�B´ � ru{~z¤z¤{~wU�v��rulYe]t]��

� � � �

� � � � � � � � � � � � � � �� ¡ ¢ £ ¤ �¥ ¦ §¨ © ª« ¬ ®¯ ° ± ² « ¬ ®¯ ° ± ²

« ¬ ®¯ ° ± ²

³ ´ µ ¶· ¸ ¹ º

L8M�N ´¼»B´ »Fdshue8¥7t�x�¼_t]��fUgih � wUdEgij:��{~�

;ÀMBa × ?ÈN¿CE6/ÙJCE6�G�6�ÚJC ×\× ã D}4YG]a]Z[CEN�M[Q¿N¿CE6�4:MBN0Q:G8Z[CEA�MBQ<N[CE6�4%MÈN�68Q'aE4%MPÙÈC@DFQYØ:CLNB÷�ÙÈ4:68CENQ:M�G8ZBC2N¿;�6uGHA8;\Ù[?¿GH;ÀQ7M�Q Ü ÙÈCL6uG8ãkD}4%GHa]Z[CLN�M[Q¿N¿CE6Eà>ðvM�Q7M[C�ZB4:MBN�KÈ4�` Ö S ;�6�40ÚB4YGHZ�G8ZB4%G;ÀMBa × ?ÈN¿CE6¤Q7M × TèÙJCE6�G�ã DF4%GHa]Z[CLNÞM[Q¿N¿CE6@GHZB4YG�ZÈ4SØ:C'G8Z[C�6�CEMB6�CL6@`:Q<Q¿N Ü Q:A¤`:;\Ø:CEMÞN[Q:D}4%;\M6�ÚJCEa;|âÈa;|GuT7àS�CEa4:?B68C/4 ×À× MBQ<N[CE6!ZB4SØ7C�4 × AHCE47N¿T@ÙÈCECM}4:N¾½u?È6uGHCEN2G8Q@G8ZBC/N[Q:D}4%;\M�;\MA` Ö S/6K` Ö S/6/a4%M�6uG]4ST';\M�G8Z[CvG8AH;ÀDFDFCEN�DFQ¿N¿C × àBðvM0G8ZBCvQ:G8Z[CEA�ZÈ4%MBN�K¿4Fò¿W�3�;�6/6�?Èa]Z�4268?[Ù¿ãG8AHCCvG8ZB4%G�4:M�;\MUG8CAHMB4 × MBQ<N[Cv;�6/42AHQ<Q%G�4%MBN0G8Z[C¤68?[ÙÈQ7AHN[;ÀMB4%G8CL6�4:A8CvQ:M × T}ÙJCE6�G�ã DF4%GHa]Z[CLNM[Q¿N¿CE6Eà¿S�CEa4:?B68Cv;\MUG8CAHMB4 × MBQ<N[CE6/ZB4SØ7CvM[Q:G�ÙÈCECM�asQ:M[âBA8DFCLN}G8QFZB4SØ7C�GHZ[C¤6�CEMB68CE6/`:Q<Q¿NÜ Q:A�4 `:;\Ø:CEM�N¿Q7DF4:;ÀMRK%4@ò[W�3 aE4%M'ÙÈC�DFQYØ:CLN�;\M'G8Z[C/GHA8;\DFD'CLN2DFQ¿N¿C × à:W/Z<?B6�ü¤ð ü üvûoýá�á0;�N¿CMUGH;|âBCL6 ` Ö S/6�4%MÈN ò¿W�356P;\MëG8Z[CúG8AH;\D'DFCLNëDFQ<N[C × 4%?¿GHQ:D}4YGH;\aE4 ×\× Té4%MBNëG8Z[CEM6�?[ÚBÚÈQ7A�G]6�4¤?È6�CEA�;\M�asQ7MB6�G8AH?Ba�GH;ÀM[`�4�asQ7MBasCEÚ¿G8?B4 × ZB;ÀCEAH4:AHa]Z<T2Ù<TFDFQYØ<;ÀM[`'ò¿W�356EàÀ�X;\`:?[AHCÁ ; ×\× ?B6�G8A]4YGHCE6�GHZ[C�4%ÙJQYØ:CãkDFCMUGH;ÀQ7M[CEN�DF4%GHa]Z[CLN�AHCE68? × G�4%MB4 × T¿68;\6Eàá�MèQ:A]N¿CEA�GHQ�AHCsâBMBC�G8ZBC�GHA8;\DFDFCEN5DFQ¿N¿C × K�ü¤ð üvü ûRýÉá�áva4%Mú?B68C2GHA8;\DFD'CLN5AHCE68? × G

4%MB4 × T¿68;\6�476�Û�C ×À× 4:6�D}4YG]a]Z[CENPAHCE68? × G�4%MB4 × T¿68;\6EàBWX4%9<;ÀMB`F68Q:DFC�68;ÀÙ × ;\M[`}M[Q¿N¿CL6�Û�;ÀG8Z�G8ZBC684:D'C�ÚB4%AHCMUG_M[Q¿N¿C7K:GHZ[CAHC�D}4ST'ÙÈCvD}4%M<T}N¿;|å>CAHCMÈasCE6�4%ÙJQ:?¿G�G8Z[CvM<?[D2ÙJCA�Q Ü GHA8;\DFD'CLNM[Q¿N¿CE6�ÙJCsGuÛ�CCM�G8Z[CED 4%MÈN}G8ZBCvÚB4:A8CEMUG�M[Q¿N¿C7àT�ëZ[CM�6�?Èa]Z�42ÙB;À`FN¿;ÀåJCEA8CEMBasC@asQ7D'CL6�?[ÚQ:MF4¤6�?[Ù[ãqGHA8CEC�;\M�G8Z[C/GHA8;\DFD'CLN'D'Q¿N¿C × K:;|G�;�6�ÙJC�ÙÈCG�G8CEA�G8Q¤a]ZB4:M[`:C�GHZ[C�6uGHA8?Èa�G8?BA8C/Q Ü G8ZBC6�?[Ù[ãqGHA8CEC:àÄü¤ð üvü ûRýÉá�á@4:689¿6vG8Z[C}?B68CA@; Ü GHZ[C068?[Ù¿ãkG8AHCCF6�Z[Q7? × NèÙJC}A8CLasQ7MB6uGHA8?Èa�G8CLN5Q7AM[Q%GLà¿S�476�CLN}Q:M�G8ZBCvCED'ÚB;ÀAH;\aE4 × 4%MÈ4 × T<68;�6KUG8ZBC@68?[Ù¿ãkG8AHCCL6_Û�;ÀG8Z�GuÛ�Q�Q7A�DFQ7A8C N¿;|å>CAHCMÈasCE6

14

AIFB AIFB

�

ÂÃ Ä

Å

Â

Ã Ä

Å

0 0 4

Â

Ã ÄÅÆÇÉÈËÊ2Ê2ÈËÌKÍ Î¿Ï�Ð�ÑÒÌKÓ�ÔÕÇÉÖKÐ�ÔÕÈ×Ì¾Í

L8M�N ´ÙØJ´ � ru{~z¤z¤t�x�¼�t]��f7gih � w<dgij7��{~�

DF4ST�ÙÈC2A8CLasQ:MÈ6uGHA8?BasG8CLN�àE�X;À`7?[AHC#Ú0; ×\× ?B6uGHAH4%G8CL6/G8Z[C�4%ÙJQYØ:Csã DFCMUG8;\Q:MBCEN�G8AH;ÀDFDFCEN�AHCE68? × G4%MB4 × T¿68;\6EàWCB Ö DFQ<N[? × C�a4%M@M[Q:GÄD}4%9:C�68?[`7`:CE6�G8;\Q:MÈ6�4%ÙJQ:?¿GRN¿Q7D}4%;\M¿ã 68ÚJCEas;ÀâÈa!Z[;\CA]4%A]a]Z<T�6uGHA8?Èa�ã

G8?[AHC¤?B68;ÀM[`�4%ÙJQYØ:Csã DFCMUG8;\Q:MBCEN�6�G8A]4YGHC`7;ÀCL6à¿S�CEaE4%?B68C@G8ZBCE68C�6uGHAH4%G8C`7;ÀCL6�N¿Q:M7Û G�9<M[QYÛøN¿Q:ãDF4:;ÀMú6�ÚJCEa;|âÈa2;ÀM Ü Q:AHDF4%G8;\Q:MÄà>ò<QBKJ;\Mú4:NBN¿;|GH;ÀQ7M�G8QPN¿Q0GHZB4YGLKJWCB Ö DFQ¿N¿? × C2D}4%97CE6�68?[`:ã`:CE6�G8;\Q:MÈ6R?B68;\M[`v4va4%MÈN¿;\NB4YG8C Ü Q:A!N¿Q7DF4:;ÀM[ã 68ÚÈCLas;ÀâÈa�ZB;ÀCEAH4:AHa]Z<Tv6�G8AH?BasG8?[AHC_CäUGHAH47a�GHCEN Ü AHQ:DG8Csä<G�asQ7A8Ú[?È6ÄÙ<T¤O�WCB�û�D'Q¿N¿? × C7àSüvQ:D}4:;ÀM¿ã�68ÚÈCLas;ÀâÈa�Z[;\CA]4%A]a]Z<Tv6�G8AH?Ba�GH?[AHC�;\6!4�68CsG!Q Ü ÚÈ4%;\AQ Ü asQ7MBasCEÚ¿GH6EàÀ�úCvÛ�; ×\× N[CE6HasAH;ÀÙJC�Z[QYÛÅGHQ2Csä<GHAH47a�G�GHZ[C 6�G8AH?BasG8?[AHC × 4YG8CEAEàÀ�ëZ[CEM}G8Z[CEA8C 4%AHCGuÛ�QèN[;|å>CAHCMUG�9<;\MBN[6�Q Ü Z[;\CA]4%A]a]Z<Tú6�G8AH?Ba�GH?[AHC�ÙJCsGuÛ�CCMÊGuÛ�QèaQ:MBaCÚ¿G]6KXü@ð ü ü ûRý á�á4:689<6¤G8Z[C0?È6�CEA2; Ü GHZ[C�ZB;ÀCEAH4:AHa]Z<T56�G8AH?Ba�GH?[A8C�6�Z[Q7? × NºÙÈC�a]ZÈ4%M[`7CENÞ;ÀMUGHQ5?[M[;ÀâBCENºQ7M[C0Q7AM[Q%GLà

�o;\MB4 ×À× T�ü¤ð üvü ûRýÅá�á�asQ7D'Ú × CG8CE6�G]4Yä¿Q:MBQ:DF;\avAHC × 4YG8;\Q:MÈ6�Z[;\ÚB6�Q Ü G8Z[C¤;\M[Ú[?[G�N[Q:D}4%;\MG8CAHD}6/D}4%M<?B4 ×\× T Ü A8Q7D G8ZBC¤?B6�CEAEà

Ü ÝµÑ_ÎßÞ�WX'àY�Ñ_Î�Ñ�%�Õ�Ô�.á4à*['XÏ[ÕuÑ_Î9\À]vÕ[1â��4ã'!ÐBÎvÕ�Îä+

O�Q:M[ãqG]4Yä¿Q:M[Q7DF;\aIB�C × 4YGH;ÀQ7MB6�ZB;ÀÚÉûÄCE4:A8MB;ÀM[`Å4 × DFQ76�G�aQ:DFCE6 Ü A8Q7Då�úQ7AHNÈò<ÚB4:aC:K/Û�Z[;�a]ZN¿CAH;ÀØ7CE6 × Csä¿;�a4 × aQ%ã Q¿aas?BA8AHCMBaC_;\M Ü Q:AHD}4YGH;ÀQ7M Ü AHQ:D 4 × 4%AH`:C�G8Cä<G_asQ:AHÚ[?B6�4:MBN�;\6�4 D�? × GH;|ãN¿;ÀDFCEMB6�;\Q:M�Ø7CEa�GHQ:A�68ÚB4:aC}ï�4}68CsG�Q Ü Ø:CLa�GHQ:A]6H÷sà[W/Z[C¤;ÀMBM[CA�ÚBA8Q¿N¿?BasG�ÙÈCGuÛ�CECMPGuÛ�Q'Û�Q:A]NØ:CEasG8Q7AH6¤Û�Q:AH9¿6�4:6�G8Z[C�DFCE4:68?[AHC0Q Ü GHZ[C;\A'68CD}4:M7GH;\a0A8C × 4%G8CEN[M[CE6H6à7�ëZ[CM¾GuÛ�Q5Û�Q:A]N[6?Û;ÀM[MBCA�Ú[AHQ¿N¿?Ba�G/;�6/ÙÈCET:Q:MÈN�68Q:DFC ?[Ú[ÚJCA�ÙJQ:?[MBNÄK<G8Z[CEA8C¤4%AHC ÚÈQU6868;ÀÙB; × ;|GH;ÀCL6�G8QFZB4SØ:C@68Q:DFCM[Q:M¿ãkGH4%ä¿Q:M[Q7D';�avA8C × 4%G8;\Q:MB68Z[;\Ú�ÙJCsGuÛ�CCEM�GHZ[CD�à

æ7çéè êJë�ìàí?î¼ï�ðZñî?òéë!ì0ë�óCôõë!ï�ö2÷ãøZùEñú

�úQ7AHNBò¿ÚB4:aC ;�6�asQ7MB6�G8AH?Ba�GHCENP4:6�68Z[QYÛ�M�;\M��X;À`7?[A8C�û[àüýCþ=ÿ�� þ��Eþ�� !��"�#� $&% ò<;\MBasC × CsG8G8CA8ã ÙUTUã × CG�GHCA¤asQ:ãkQ¿aa?[AHA8CEMBasC�;ÀM¿ãÜ Q:AHDF4%G8;\Q:M�ÙJCEasQ7DFCE6�GHQ<Q0D2?Ba]Z54:MBN�68Q0Q Ü GHCM�;\A8AHC × CØY4%MUGLK[Û�C�GH4%97C¤G8CEA8D'ã Ù<T7ãkG8CEA8DæasQ:ãQ<aEas?[AHAHCMBaC¤;ÀM Ü Q7A8D}4%G8;\Q:M�;\M Ü Q:?[AvÛ�Q7AHN[62ï Ú%ã `:A]4%D0÷/476/G8Z[C�Ú[AH;ÀDF;ÀG8;\Ø:C¤GHQ0D}4%9:C¤?BÚèasQ:ãQ<aEas?[AHAHCMBaC�D}4YGHA8;Àä�?B68C Ü ? × GHQ�AHCÚBA8CL6�CEM7G�asQ7MUG8Csä<G/Q Ü 4�GHCsä<GEà¿�ÞC GH4:9:C Z[;À`7Z Ü AHCEêU?[CEMBasTÚ%ã `:A]4%D}6_;\M�Q:A]N¿CA�GHQ}DF4:9:C@?[Ú �ÞQ:A]NBò<ÚB47asC:à

15

AIFB AIFB

«

'( ( ) * + , - . / 0 1 21 3 1 4 5 6 5 7 5 8 9 :; < = > = ? = > = < @ @A

A@ @ = B = C = D = ? = B= C = > = D = C = E = F= < = > = ? = > = < @ @A

GG H H II J J K LK L M N O PM N O P Q Q J JR ST S

T U

V UW XY Z [\

]^^

_ _`

`

a b c d e f g h i j k l k j j k m n o p q r s t u v w x y z y x x y { | o p q r s t u v } ~ � � � � � � � ~ � � � � � �� ¡ ¢ ¡ � £ £¤ ¥ ¦ § ¨ © ª « ¬ ® ¬ ª «

¯ °± ²³ ´µ ¶

·

¸ ¹ º º ¹ ¸ » ¼ ½ ¹¸ ¹ º º ¹ ¸ » ¼ ½ ¹ ¾ ¿¾ ¿ ¸ ¹¸ ¹ ÀÀ Á Á Â Ã Ä Å Æ ÇÂ Ã Ä Å Æ Ç Å Å È ÂÈ Â Ã Ã É É ÆÆ

Ê Ë Ì Ì Ë Í Î Ï Ð Ë Ñ Ò Î Ï Ó Ð Ô

L8M�N ´ÖÕB´ £XlLw7� h�rufUeHhu{~lLw0c>g~l�p-lEm>ÆPlEr�x<n%�<de]t

×ÀýØ�� !%��Ù�!��Ú�"� Û�Ü�� ÝÞÝßà�� á$â�"�Ú�Ù ÿ Ö �� ÝÞÝßà�� ã$â�"�Ú�Ù ÿ ;�6!asQ:MÈ6uGHA8?BasG8CLN�;\M2Q:A]N¿CEAG8Q@asQ:DFÚB4:A8C�G8ZBC/aQ:MUG8Cä<GXQ Ü GuÛ�Q Ú%ã `:A]4%D}6EàEý × CED'CEMUGåä�æ�ç è�;\M2GHZ[;\6!D}4%G8AH;|ä�;�6oG8Z[C/M<?[D2ÙJCAQ Ü Ú:ãk`7AH4:D Ü æoÛ�Z[;�a]Z�aQ:DFCE6�?[Ú ½u?B6uG�ÙJC Ü Q7A8C Ú%ã `:A]4%D Ü è'ï�aE4 ×\× CLN �� ÝÞÝéÖ�� ê� ��þ�� ÷�àBW/ZBCasQ ×\× Q<aE4YGH;ÀQ7MúD}4YGHA8;ÀäÞaQ:?[MUG]6@Z[QYÛ D}4:MUTúQ%GHZ[CA#Ú:ãk`7AH4:DF6@aQ:DFC}?[ÚºÙÈC Ü Q:AHCFG8Z[C0GH4%AH`:CGÚ%ã `:A]4%D�à¿ý_4:a]ZPasQ × ?[DFM�Q Ü G8Z[;�6�D}4YGHA8;Àä�;�6/G8ZBC �ë�� $íì¾þ��î"� Q Ü G8ZBC#Ú%ã `:A]4%D Ü àï�ýð�� ë%Ù��Ù�!�� ñ��ò�#"�!� þ�ÿ��óì�þ��ô ��% Ö �� !� þ=ÿ��õì¾þ��î"� A8CEÚ[A8CL6�CEMUGH6�asQ:MUGHCsä<G�Q Ü 42Û�Q:A]NQ:A_ÚBZ[AH476�C�;ÀM�4�G8CäUGLà Ö 6�?BDÿQ Ü Ú%ã `:A]4%D Ø:CEasG8Q7AH6�4:A8Q7?[MBN}4:Ú[ÚÈCL4%A]4%MBaC�Ú × 4:asCvQ Ü 4¤Û�Q:A]NQ:A2ÚBZ[AH476�C5ïqa4 ×À× CEN ��"�!� þ�ÿ��ò� ��þ�� ÷�;�624úasQ:MUGHCsä<G2Ø:CLa�GHQ:A�Q Ü 4�Û�Q7AHNÞQ7A2Ú[Z[A]4:68C0;\M¾G8ZBCÚ × 47asC7à�!ýå�� !%��Ù�!��Ú�"�ñÛ�÷öõ �#øùì�þ#��ô ��% Ö Û�Q7AHN0Ø:CEasG8Q7A�;�6/4F6�?[D Q Ü aQ:MUG8CäUG�Ø7CEa�GHQ:A]6�4%G/4 ×À×4%Ú[ÚJCE4:AH4:MBasC¤Ú × 4:aCE6�Q Ü 4}Û�Q7AHN�Q:AvÚ[Z[A]4:68C¤Û�;|GHZ[;\M�G8Csä<G]6K>4%MBN�aE4%M�ÙJC2Csä¿Ú[AHCE6H68CEN�Û�;|GHZG8Z[C Ü Q ×\× QYÛ�;ÀM[` Ü Q:AHD2? × 4BàT��CAHC:KûúÄï�üv÷�;\6�4'Ø7CEa�GHQ:A�A8CEÚ[AHCE68CMUGH4%G8;\Q:M0Q Ü 4�Û�Q7AHN�Q:A/Ú[Z[A]4:68Cü�KðýFï�üv÷2;\6'4%Ú[ÚJCE4:AH4:MBasC}Ú × 47asCE62Q Ü 4�Û�Q:A]NºQ:A�ÚBZ[AH476�C ü ;ÀM-4�G8CäUGLKX4:MBNÿþ/ï��÷�;\6�4Ú%ã `:A]4%D Ø:CLa�GHQ:A/Q Ü 4AÚ%ã `:A]4%D��Xà Ö 6�CG�Q Ü Ø:CLa�G8Q7A úÄï üv÷�;\6ß�ÞQ:A]NBò<ÚB47asC7à

úÄï üv÷�� æ��

ï �� a × QU6�C G8Q æ

þ/ï��÷�÷

�ý!��"�ë%Ù�Ú�Ù�!�� ùÛ�óì�þ��ô � ��þ�� þ�%¼þ�� î�"�Ú�"�ë% Û� � ÝÞÝû�� Òþ�� Ú% W/ZBC�ÙÈCL6uGRD}4%GHa]Z[CLN ;H6�T<MB68CsGÒ<Q Ü CE47a]Z0;\M[Ú[?¿G�G8CEA8D}6�;ÀM �úQ7AHN[O�CsG�;\6�4 × A8CL4:N¿TF68ÚÈCLas;ÀâBCEN�K¿4:MBN04268?[D Q Ü G8Z[C Û�Q7AHNFØ7CEasãG8Q:A'asQ7M7G]4%;\M[CENÞ;\M¾G8ZBCE68C�6�T<MB68CsG]6�;�6268CsG�G8Q�GHZ[C0Ø:CLa�GHQ:A�AHCÚ[AHCE68CMUG]4YG8;\Q:MºQ Ü 4èasQ:MÈasCÚ[GasQ:AHAHCE68ÚÈQ7MBN¿;\M[`2GHQ}4�;\M[Ú[?[G�G8CEA8D�à[W/Z[C�aQ:MBaCÚ¿G × 4%ÙJC × ;�6/G8ZBC¤;ÀM[ÚB?¿G�G8CEA8D�à

æ7ç�� êJë�ìàí?î¼ï�ðZñî?òéì�� ù¿ìZö�� ëãöZò ó��ãòéì�� êJë!ìZñúÀøZîF÷6ø�úTñ�ò��2ñùHîKò ë�ì��ú! ø#" ùHîKúÀí$�CEa�GHQ:A�AHCÚ[AHCE68CMUG]4YG8;\Q:MÈ6_Q Ü 4 ×\× asQ7MBasCEÚ¿GH6�4:A8C Q:Ù[GH4%;\M[CLN�Ù<T0asQ7MB6�G8AH?Ba�GH;ÀM[`A�ÞQ:A]NBò<ÚB47asC7àò<;ÀDF; × 4:A8;ÀGuTèÙJCsGuÛ�CCEMÊasQ7MBasCEÚ¿GH6¤;�6¤Q7Ù¿GH4:;ÀM[CLN Ü A8Q7D1;\M[M[CEA�Ú[AHQ<N[?Ba�G]6�;\Mº4 ×À× G8Z[C�asQ:D�Ù[;ÀãMB4YGH;ÀQ7MúQ Ü G8ZBCE68C'Ø7CEasG8Q:A]6EàJW/Z[CEMºÛ�CFN¿CsâÈM[C0asCEA�G]4%;\M5G8ZBA8CL6�Z[Q × N Ü Q:A GHZ[;\6¤68;\D'; × 4%AH;|GuT7à Ö

16

AIFB AIFB

ª

asQ:MÈasCÚ[G�ÚB4%;\A�Û�;|GHZ56�;\DF; × 4%AH;ÀGuT0ÙÈCET:Q7MBN0GHZ[C¤GHZ[A8CL6�ZBQ × N�;�6�Csä<GHAH47a�G8CLN�4:6�4F6�;\DF; × 4%AvasQ:M[ãasCÚ[G!ÚÈ4%;\AEà Ö 6�CGXQ Ü 68;ÀDF; × 4:AXaQ:MBaCÚ¿G!ÚB4%;\A]6oÙJCEaQ:DFCE6!aQ:MBaCÚ¿G�6�ÚJCEa;|âJa4YGH;ÀQ7M�GHCDFÚ × 4YG8CL6àS�Q:G8Z�Q Ü GHZ[C2aQ:MBaCÚ¿G�ÚÈ4%;\AH6EK[Û�Z[QU6�C�DFCL4%M[;\M[`};�6�68;ÀDF; × 4:A�ï·Û�;ÀG8Z5GH4Yä¿Q7M[Q:DF;�a@A8C × 4%G8;\Q:MÈ÷sK4%MBN-ZB4:6}68Q:DFCG8Z[;\M[`ºA8C × CEØY4%MUGFG8QºCE47a]Z-Q:G8Z[CEA5ï·Û�;ÀG8ZéM[Q:M[ãqG]4Yä¿Q:M[Q7DF;\a�A8C × 4%G8;\Q:MÈ÷sK�4%AHCCsä<G8A]4:asG8CENº4762asQ7MBasCEÚ¿G2ÚB4:;ÀA]6¤Û�;ÀG8Z�aQ:MUG8Cä<G26�;\DF; × 4%AH;|GuTÞ;\MÊ4�D}4:6H6à7��QYÛ�CEØ:CEAEKÄÙ<Tú?B6uã;ÀM[`PG]4Yä¿Q:M[Q7DF;\a�;ÀM Ü Q:AHDF4%G8;\Q:M Ü A8Q7D1WCB Ö DFQ¿N¿? × C'Û�;ÀG8Z¾aQ%ã Q<aEas?[AHAHCMBaC2;\M Ü Q:AHD}4YG8;\Q:MRKü@ð ü üvûoý�á�áFN¿;�6uGH;ÀM[`7?[;�6�Z[CL6'G8Z[C5aQ:MBaCÚ¿G}ÚB4:;ÀA]6'Û�Z[;�a]Zé4:A8CPZ[;\CA]4%A]a]Z[;�a4 ×\× Tºa × QU6�CPGHQCE4:a]ZPQ%GHZ[CA Ü AHQ:D G8Z[C¤Q:G8Z[CEA�ÚB4%;\AH6�476�W Ö&% ð O ð 3�^'àÖ ?È6�CEA@aQ:MB6�G8AH?BasGH6@4�N¿Q:D}4%;\MèQ7MUG8Q × Q:`7T�ÙUTúasQ:MÈ6�;�N¿CAH;\M[`0GHZ[CFA8C × 4%G8;\Q:MèÛ�;ÀG8ZºCL4:a]Z

asQ:MÈasCÚ[G_ÚB4%;\A�;\M}G8ZBC�asQ7MBasCEÚ¿G�68ÚÈCLas;ÀâÈa4%G8;\Q:MFG8CEDFÚ × 4%G8CL6KU4%MBN}N¿C × CG8;\M[`24:M}?[M[M[CLasCE6H6H4%AHTasQ:MÈasCÚ[G/ÚÈ4%;\AEà

æ7ç(' )+*6î¼ï�ù¿ñî?ò ì,�.- ë/ ùEò ì10 ÷ãø�úTñò��2ñ32�òéúÀï¾ù¿ï¾ñ546��÷ãî¼ï�ðàñî?ðàï�úá�M�Q:A]N¿CA/GHQ0D}4%9:C¤68?[`:`7CE6�G8;\Q:MB6�4:ÙÈQ7?¿GvN¿Q7D}4%;\M¿ã 68ÚJCEas;ÀâÈa@Z[;\CA]4%A]a]Z<T�6�G8AH?Ba�GH?[A8C7KBO�WCB�ûD'Q¿N¿? × C�GHA8;\CE6'GHQÞCä<G8A]4:a�G}ÚB4:;ÀA]6'Q Ü asQ7MBasCEÚ¿GH6FÛ�Z[;�a]Z Ü Q:AHDßÚB4%A8G}Q Ü 4ÞaE4%MBN[;\N[4%G8C Ü Q7AN¿Q:D}4%;\M¿ã�6�ÚJCEa;|âJa�Z[;\CA]4%A]a]ZUT�6uGHA8?BasG8?[AHC:à7á�M0Q:A]N¿CEA�G8Q2N[Q¤G8ZB4%GEKUÛ�C�ÚB4ST'4%G�GHCMUG8;\Q:MFGHQ�G8ZBCN¿;\6�GH4:MBasC�ÙJCsGuÛ�CCEM@GuÛ�Q�aQ:MBaCÚ¿G]6Ä;\M@4�N[Q<a?[DFCMUGEàEá�M@G8Z[;�6ÄÚÈ4%ÚJCALK�G8Z[C�N¿;\6�GH4:MBasC�ÙJCsGuÛ�CCEMGuÛ�Q�aQ:MBaCÚ¿G]6�DFCE4:MB6�GHZ[C�M<?[D2ÙJCA�Q Ü Û�Q:A]N[6�ÙJCsGuÛ�CCMFGHZ[CD�à7á Ü G8Z[C�N[;\6�GH4:MBasC�ÙJCsGuÛ�CCEMGuÛ�QFasQ7MBasCEÚ¿GH6�;�6�6�D}4 ×\× 4%MBN�G8ZBC@68;\D'; × 4%AH;|GuT}ÙJCsGuÛ�CCEM0G8ZBCD ;�6�a × Q768C:K<Û�C 68?[Ú[ÚJQ768C G8ZB4%GQ:M[CèaQ:MBaCÚ¿GPCsä¿Ú × 4%;\MB6}GHZ[C5Q:G8Z[CEAEà�á Ü G8Z[CÞN¿;\6�GH4:MBasCú;\6 × 4:A8`7C54%MBNëG8Z[Cè68;\D'; × 4%AH;|GuTÅ;\6a × Q768C:K!Û�C�68?[Ú[ÚJQ768C�GHZB4YGFG8ZBCT Ü Q7A8DÝÚB4:A�GFQ Ü N¿Q:D}4:;ÀM¿ã�68ÚÈCLas;ÀâÈa�ZB;ÀCEAH4:AHa]Z<Tº6uGHA8?BasG8?[AHC:àÖ aaQ:A]N¿;\M[`�G8QÞ4%ÙJQYØ:Csã DFCMUG8;\Q:MBCENú;\N¿CL4[KoÛ�C�aE4 × as? × 4YGHC}G8ZBC�Ú[AHQSä¿;ÀD}4 ×\× T5AH4%G8C0ÙJCsGuÛ�CCEMGuÛ�QèaQ:MBaCÚ¿G]62Û�;ÀG8ZB;ÀM�4úasCEA�G]4%;\MÊ68aQ:ÚJC:àXá G�;�6�GHZ[C�M<?[D�ÙÈCEA�Q Ü G8;\D'CL62ÙJQ%GHZ�asQ7MBasCEÚ¿GH6Q<aEas?[A�Û�;ÀG8ZB;ÀM�G8Z[C�6HasQ7ÚÈC¤N¿;\Ø<;\N[CEN�ÙUT0G8Z[C¤M<?[D�ÙÈCEA�Q Ü GH;ÀDFCL6/Q:M × T0Q:MBC@aQ:MBaCÚ¿G�Q¿aa?[A]6Û�;|GHZ[;ÀMP;ÀGEàT�úC¤N¿CâBM[C@asCEA�G]4%;\M}G8Z[AHCE68Z[Q × N Ü Q:A�G8Z[;�6/Ú[AHQSä<;\D}4 ×\× T'A]4YG8C7àÀ`�4:;ÀA]6�Q Ü asQ7MBasCEÚ¿GH6Û�Z[Q768C@Ú[A8QSä¿;\D}4 ×\× T0A]4YGHC@;\6�Û�;ÀG8ZB;ÀM�GHZ[;\6�G8Z[AHCE68Z[Q × N�4%MBN�GHZ[C�6�;\DF; × 4%AH;ÀGuT0ÙÈCGuÛ�CECMPG8ZBCD;\6¤ÙJCT7Q:MBNÞGHZ[CFG8ZBA8CL6�Z[Q × N Ü Q7A26�;\DF; × 4%AH;ÀGuTú4%AHCFCsä<G8A]4:asG8CLNÞ476¤ÚB4:A�G�Q Ü 4�aE4%MBN¿;�N[4%G8C Ü Q7AN¿Q:D}4%;\M¿ã�6�ÚJCEa;|âJavZB;ÀCEAH4:AHa]Z<T}6uGHA8?Èa�G8?BA8C7à

7 8I'-\À4:9�Ï[Ó�ÒvÕ[47\<;]Ñ�ÐáWX'ZY�Ñ_Î�Ñ&%ÿÕ8ÔU.047*['XÏ[ÕuÑ�Î \T]�Õ�1â"ëÔ6_/Ó�Õ�\<Õ�Ï[ÕuÑ_Î

á�M0Q:A]N¿CA_G8Q'CØY4 × ?B4%G8C�Z[QYÛçü¤ð ü üvûoý�;�6�N[Q:;\M[`�;\M�Ú[A]4:asG8;�a4 × âBC × N[6EKUaE4:68Cv6�G8?BN[;ÀCL6_ZB4SØ7CÙÈCECM�N¿Q:M[C ;\M�4�ÚB4%A8G8;�as? × 4%A�âBC × N�Q Ü × 4SÛ a4 ×À× CEN ��Q:MUG8A]4:asGH6 Ü Q:A�G8Z[C á�MUG8CEA8MB4%G8;\Q:MB4 × ò[4 × CQ Üß� Q<Q¿N[6Fï��á�ò � ÷�àRW/Û�Q × 4SÛ�T:CEAH6�½uQ:;\M[CLN�G8ZBC0a4:68C�6�G8?ÈN¿;ÀCL6à�á�MúG8ZBC�âÈAH6�G�a4:68CF6uGH?BN¿T:K;ÀM[ÚB?¿G�G8CEA8D}6�4:A8CßÚ7õ × C`U4 × G8CEA8D}6 Ü AHQ:D ��á�ò � `�4%A8G�ã á�á�àSá�M'G8Z[C�68CEaQ:MBNFa476�C�6�G8?BN[T:KYGHZ[CT4%AHC}ó?� Á GHCAHDF6�;ÀMBa × ?ÈN¿;ÀMB`0`:CEM[CA]4 × GHCAHD}6�;ÀM54:M�Csä[4%DFÚ × C2a476�C�4:MBN × CE`74 × G8CEA8D}6 Ü AHQ:D��á�ò � 4:A�GH;\a × CL6_AHC × 4YGHCEN0Û�;ÀG8Z�G8ZBC@aE4:68C:àBðvM[C × 4SÛ�T7CA�N¿;\N�G8ZBCvâÈAH6�G/aE4:68C 6�G8?ÈN¿T04%MBN�G8ZBCQ%G8ZBCA × 4SÛ�T:CA�N¿;�N�G8Z[C�68CEaQ:MBNÄàWo4:Ù × C¤ó�68Z[QYÛ�6XGHZ[C�A8CL6�? × G�Q Ü GHZ[C�a476�C�6�G8?ÈN¿;ÀCL6�à��X;\`:?[AHC�õ¤6�ZBQYÛ�6�Z[QYÛ�D�?Ba]ZFQ Ü G8ZBC

;ÀMUG8CEA8DFCLN¿;\4%G8C�Ú[A8Q¿N¿?Èa�GH6�;�62;\MBa × ?BN¿CLN¾;\MÊâBMB4 × N[Q:D}4%;\M¾Q:MUG8Q × Q7`:TÞ4YG�CL4:a]Z�ü¤ð ü üvûoý4:a�GH;ÀØ<;ÀGuT:à

� CM[CEAH4 ×À× T56�ÚJCE4:9<;ÀM[`ÈK>;ÀMºasQ7MB6uGHA8?Èa�G8;\M[` × CE`74 × Q7MUG8Q × Q:`7;ÀCL6KRþ�>=æQ7A@D'Q7A8CF68?[Ú[ÚJQ:A8GasQ:DFCL6 Ü AHQ:D�ü@ð ü ü ûRý�à Ö ÙJQ:?[GvZB4 × Ü ÚJQ:A8G8;\Q:M�Q Ü G8Z[C2âBMB4 ×X× C`74 × Q7M7GHQ × Q:`7T�AHCE68? × G]6�;\MG8Z[C@;\M Ü Q:AHD}4YG8;\Q:M�Csä<G8A]4:asG8CLN Ü Q:AHD �úQ7AHNBO�CsGLà¿S�CEa4:?B68C�G8ZBCvGuÛ�QF6uGHAH4%G8C`7;ÀCL6ã½u?È6uG�;\D'Ú × T

17

AIFB AIFB

¨

? @A@BA@CA@D @EA@F @G @H @IJ@? @@ K�L M NONQPARSUT R PWVYX�Z [ PW\�RJP]RK�L M NON^P]RSUT RJP V`_ \ [ P]L \�a]Vb�ced _ \ [ P]L \�a]Vb�c�f bgT N^aAM \c \ [ T V T]h]ij^k�l m�m [ L n�o [ n�L P p q]r]s d t�uv N^a]V V b K�p wAs q tgu v [ L a [ P h]i f p x]s d tguv [ L a [ P h]i�d p wAs q tgu y T

L L P]o [ M T \{z i n m PWL]p d | s fWtgu} RJRJM [ M T \{z i n m PWL]p ~]s r tgu

�^�Q�Q�Q�� J��]�

� �� 5�g��A�O�O�1 ?A@A@BA@CA@D @EA@F @G @H @IJ@? @@ K�L M NON^PARN T RJPWVYX�Z [ P]\�RJPARK�L M NQNQPARSUT RJPWV¡_ \ [ P]L \(aAVb�ced _ \ [ PWL \(a]Vb�c¢f bgT NQa M \c \ [ T V TAh]i

jQkgl m�m [ L n�o [ n�L PAp w | s £ tguv NQa]V V b K�p d]d s w tgu v [ L a [ P h]iJf p wAs ~ tguv [ L a [ P hAi(d p rWs ¤ tgu y T

L L PAo [ M T \Oz i n m PWL]p f wAs |Wtgu} R(RJM [ M T \Oz i n m P]LWp wAs ~ tgu

�^�O�Q�Q�g�e� ��A��J��A�

� �� 5�g��A�O�O�¦¥

L8M�N ´¨§[´ � ¥Ut�£XlLz¤�¿lEwUt]wYh�¼�dhut/lm�hu¥Ut�c>{~wUdEgJ¢�lLz�dE{~wFí�w%hulEg~lL�Ej

© ²7®�ªJ« P<´ � ¥7t�£!dE��tvn%hufUx7{~t]��¼_t]��fUgihu�¬®U¯1°²±�³�´ ¬/U¯�³�¯�µ�¶�·¨¸¬®¨¯¹·QºU»,¼5¯�±®¶�½�¾ µ�¿�³�¯�³�´�º¨¸^À µ�¿�³�¯�³�´�º¨¸^ÀÁ(·UÂUºU´®´�¯�±�»�³ Ã{Ä IgÅ{ÆÇ »�¿{È ÈÊÉ�¬ Ë ¶�»�Â5¶{·U¯�·O´®´�¯�±�»�³ � V Ä � Ä V{Ì �Í6¶Q¸U¯�³®»�¿�´�µ�U¯�¸ÏÎ/Ð ´��Ñ+¶{±�¸¨ÍÒ¯¢´ AÓ ·U»�¿�´�µ�¨¯�¸ ��Ô Ã V Å � Õ I Ã �Ç ¿{È Ð ¯�·O´ÒÁ(·O´�¯�±�·U¿{ÈÊÍ6¶Q¸U¯�³ ¬±�Ð »�»�¯�¸Ö·¨¶Q¸U¯�³ � IeÆ Ì{× � V Õ^ ×{Æ �Ç »�¿{È ÈÊÉ�¬ØÐ ·O´�¯�Ù�±�¿�´�¯�¸ÏÐ ·O´�¶�¿�´�±�Ð »�»�¯�¸Ö»�¶Q¸¨¯�È AÓ ·UÐ ·O´�¯�Ù{±�¿e´�¯�¸ � V Å � Ì I �ÚÛ¶Q¸UÐ °5µ�¿e´�Ð ¶{·Ö¼OÀÏ´�U¯¹º¨³�¯�± AÜ ¸¨¸UÐ ´�Ð ¶�· � I ÕU Ì � Ã�Ã Õ{�Ý!Þe¿{È º¨¿e´�Ð ¶{·�¶e½!³�´�±�¿�´�¯�Ù�À I ÔQÔ ÃOß I Ä V{ÌOà ÅQá � â ß V âU Æ�I�à ÅQá �Ý!Þe¿{È º¨¿e´�Ð ¶{·�¶e½!³�´�±�¿�´�¯�Ù�À V ÔQÔ Æ ß I�Å Æ�ÅQà ÅQá � Ã{ß IuV Æ{ÆOà ÆOá �Ô&ã Í6¶Q¸¨¯�³®»�¿�´�µ�¨¯�¸ÖÎÒÐ ´��Ñ+¶{±�¸UÍ6¯¢´�ä#Ð ³Ò´�U¯¹·Qº¨»�¼5¯�±/¶e½!Ð ·UÂUºU´®´�¯�±�»�³®ÎÒUÐ µ��U¿gÞ{¯�¼²¯�³�¯�È ¯�µ¢´�¯�¸ÖÂ¨±�¶{Â²¯�±³�¯�·U³�¯�³Ð ·�Ñ+¶{±�¸UÍ6¯¢´Ò¿�·¨¸ ã�Ó ·U»�¿�´�µ�¨¯�¸Uä#Ð ³/·¨¶e´Ò´�¨¯1µ�¿{³�¯ àÔQÔ ¬/U¯�·QºU»,¼²¯�±/¶e½!³�º¨Ù�Ù{¯�³�´�Ð ¶{·U³Ò¿�µ�µ�¯�ÂU´�¯�¸Ö¼OÀ�¿�ºU³�¯�±�ß�¬®¨¯1·Qº¨»�¼5¯�±å¶�½�³�ºUÙ{Ù�¯�³�´�Ð ¶{·U³6Ù{¯�·U¯�±�¿�´�¯�¸Ï¼OÀÉ¹æ�É�É�çÝ

G8Z[C¤ÚB4:A�G/Û�Z[CEA8C¤asQ7MBasCEÚ¿G�N¿AH; Ü G�D}4ST�asQ7DFCv?[ÚRK¿G8Z[C@ÚB4:A�G�`7CM[CEAH4%G8CEN�Ù<TFG8Z[CED ZB476 × QYÛasQ:DFÚJQ:M[CEMUG!A]4YGHCE6!4%MÈN�4:ÙÈQ7?¿G Á �è=øZ[;ÀG�AH4%G8CL6�à:ò¿Q Q7M[C/Q:?¿G�Q Ü G8ZBA8CEC�;\MBN¿;�a4%G8;\Q:MB6!ÙÈ4:68CENQ:MúG8Z[C'GuÛ�QP6�G8A]4YGHC`:;\CE6 Û�Q:AH9�Û�C ×À× ;\MºQ:A]N¿CAvG8Q�D}4%MB4:`:CFasQ7MBasCEÚ¿G�N¿A8; Ü GEààS�CLa4:?B6�C�G8ZBCGuÛ�QF6uGHAH4%G8CE`:;\CE6�?B68C 68?Ba]Z�6�T<MUGH47a�G8;�a4 × Ü CE4YGH?[AHC 476/DF4%GHa]Z[CLN�4:MBN0GHA8;\DFDFCEN�A8CL6�? × GH6EK<G8ZBCZ[;|G/A]4YG8CL6�4:A8CvM[Q%G/6�Q�ÙB4:NÄàUá�M�Q:A]N¿CA_G8Q'D}4%MB4:`:C�aQ:MBaCÚ¿G�N¿AH; Ü G�6�D}4:A�G × T7K7Û�C�D}4STFM[CCLNG8Q'?B6�C@DFQ:AHCv68CD}4:M7GH;\av;\M Ü Q:AHD}4YG8;\Q:M�G8ZB4%G�;\6�M[Q:G/CL4:68T'G8Q}asQ7D'C ?[Ú�;ÀM�4:N¿ØY4:MBasC ;ÀM�G8ZBC6uGHAH4%GH4%`7;ÀCL6à

é " 8I'-\À4Y9�Ï¿ÓvÒC,ê;]Ñ�ÐõÝµÑ_ÎCÞ¾WX'ZY/Ñ�Î�Ñ�%�Õ�Ô:.047*�'oÏ[Õ�Ñ_Î \T]�Õ�1� 4ã'!ÐBÎ�Õ8Îä+

ü@ð ü üvûoýøá�á�KRN[Q:D}4%;\MúQ7MUG8Q × Q:`7T�AH4:Ú[;\NÞN¿CEØ:C × Q:ÚBD'CEMUG@CEM<ØU;\AHQ:M[DFCMUGLKJÛ�ZB;\a]ZÞA8C Ü CA@GHQ3 B�üÿ4%MBN�N¿Q7D}4%;\M¿ã 68ÚJCEas;ÀâÈa GHCsä<GH6EKÈ;�6�ÙJC;\M[`0;\DFÚ × CDFCMUG8CLN�Q:M `!CEA ×�ë W/9�M[QYÛ¤à!�X;\`:?[AHCþ�68Z[QYÛ�6�G8Z[C}Q7MUG8Q × Q:`7T�CEN¿;ÀG8Q7AFï × C Ü G�Û�;ÀMÈN¿QYÛ�÷ 4%MBNèG8Z[C}aQ:MBaCÚ¿G¤`:A]4%ÚBZèCEN[;|GHQ:A}ï·AH;À`7ZUGÛ�;ÀMBN[QYÛ�÷�à

18

AIFB AIFB

�

L�MON ´²ì[´ � ¥Ut�í�wYhulEg|lE�EjXQox:{ihulEr

© ²7®�ªJ«F�B´ ��{~�Lw7{iÃ<e�dw%h_§L«@e]lLw7e]t]�7hu��{~w}£Xy nUËé�<dsr�h�y�y¿�µ�µ�¯�ÂU´�¿{·Uµ�¯ ¸U¯�È Ð Þ{¯�±�À ¶eíÊ¯�± ±�¯�ÂUÈ À¿�µ¢´ ¸UÐ ³�µ�±�¯�Â¨¿�·¨µ¢À ¶eíÊ¯�±�¯�¯ ±�¯�³�Ð ¸U¯�·Uµ�¯¿�¸¨¸UÐ ´�Ð ¶�· ¸UÐ ³�ÂU¿�´�µ� ¶eíÊ¯�±�¯�± ±�¯¢Þ{¶Qµ�¿�´�Ð ¶�·¿�¸¨¸U±�¯�³�³ ¯¢í5¯�µ¢´ ÂU¿{±�´(À ³�Ð È ¯�·¨µ�¯¿�³�³�¯�·O´ ¯�·OÞ{¯�È ¶{Â²¯ ÂU¿�À^»�¯�·O´ ³�Â²¯�¯�µ��¿�µ¢´µ�Ð ±�µ�ºU»�³�´�¿{·Uµ�¯ Ù�¶Q¶Q¸¨³ Â²¯�±�³�¶�· ´�¯�È ¯�Â¨U¶{·U¯µ�¶�»�»,ºU·¨Ð µ�¿�´�Ð ¶{·Ö³�À^³�´�¯�» U¶{È Ð ¸¨¿�À ÂUÈ ¿�µ�¯1¶e½!¼Uº¨³�Ð ·U¯�³�³ ´�¯�È ¯¢îµ�¶�·¨¸UºUµ¢´ Ð ·¨¸UÐ µ�¿�´�Ð ¶�· ÂU±�Ð µ�¯ ´�Ð »�¯µ�¶�·O´�±�¿�µ¢´ Ð ·O´�¯�·O´�Ð ¶�· ÂU±�¶{Â5¶�³�¿{È ´�±�¿�·¨³�»�Ð ³�³�Ð ¶{·µ�¶�º¨·O´�¯�±�¶�í5¯�± Ð ·OÞ^Ð ´�¿�´�Ð ¶{· ïQºU¿{È Ð ´(À ÎÒÐ ´�U¸¨±�¿�Îå¿{È¸U¿�À È ¯¢´�´�¯�± ïQºU¿{·O´�Ð ´JÀ¸U¯�È ¿�À »�¶Q¸UÐ °²µ�¿�´�Ð ¶{· ±�¯�ð�¯�µ¢´�Ð ¶{·

ò<?[ÙB68CEêU?[CEM7G × T7KU476�4�a476�C�6�G8?ÈN¿T Ü Q:A�M[Q:M¿ãkGH4%ä¿Q:M[Q7D';�a�AHC × 4YGH;ÀQ7MB6�ZB;ÀÚ�47aêU?[;�6�;ÀG8;\Q:MRK:Û�CasQ:MÈ6uGHA8?BasG8CLN�GHZ[C2aQ:MBaCÚ¿G�N¿CâBM[;ÀG8;\Q:M Ü Q7A�6�;\`:M[;ÀâÈaE4%MUGäÚ7õFasQ7MBasCEÚ¿GH6�Q Ü ZB4SØ<;ÀMB`'?È6�CLN�Q7MG8Z[CPâBA]6uG}a476�CP6uGH?BN¿T ïqWo4:Ù × C �7÷�Û�;ÀG8ZÅCEN¿;ÀG8;\M[`ÞG8ZBC�aQ:MBaCÚ¿G}68ÚÈCLas;ÀâÈa4%G8;\Q:MÊGHCDFÚ × 4YGHC?B6�;\M[`¾ü¤ð üvü ûRýÿá�á�K�4:MBN-Ø:CAH;ÀâBCEN�?B6�C Ü ? × M[CE6H6à�W/Z[C�aQ:MBaCÚ¿G}Z[;\CA]4%A]a]Z<T:KXÛ�Z[;\a]Z-G8ZBC× 4SÛ�T7CA}4:asG8?B4 ×À× T�asQ7MB6�G8AH?Ba�GHCEN-?B6�;\M[`Êü¤ð üvü ûRý�;\MÅG8Z[C�âÈAH6�G0a4:68C�6uGH?BN¿T-Û�476F?B6�CLNZ[CAHC}ï �X;À`7?[AHC �U÷�à

ñãçéè êJë�ìàí?î¼ï�ðZñî?òéë!ì0ë�óCôõë!ï�ö2÷ãøZùEñú��;\`:Z¿ã Ü A8CLêU?[CMBaT Ú%ã `:A]4%D}6�Û�CAHC�CäUGHAH47a�GHCEN Ü AHQ:D ��á�ò � ï�4%ÙJQ:?[GPóK�[K ��Û�Q:A]N[6]÷�4:MBNû�%õ�9<;\MBN[6�Q Ü Ú:ãk`7AH4:DF6�Û�CEA8C}Q7Ù¿GH4:;ÀM[CLN�àXá�MÊQ:A]N¿CA�G8Qè97CCEÚÊN¿CMB68;ÀGuTÞQ Ü 45asQ ×À× Q¿a4%G8;\Q:MDF4%G8AH;|ä¾Z[;\`:ZRKoGHZ[CPCsä<G8A]4:asG8;\Q:M Ü AHCEêU?[CEMBasTºQ Ü Ú%ã `:A]4%D}6�D�?B6uGFÙJCP47N¾½u?B6�G8CEN-4:aEasQ:A]N¿;\M[`G8Q�G8Z[CF68aE4 × C2Q Ü G8CäUG¤asQ7A8ÚB?B6à Ö 6#��á�ò � ;\6�G8Z[CFasQ7DFÚB4%A]4YGH;ÀØ7C × TP68D}4 ×\× ã�68aE4 × C¤G8Cä<GEK>G8ZBCCsä<G8A]4:asG8;\Q:M Ü AHCEêU?[CMÈasTPÛ�476�68CsG 476 �FGH;ÀDFCL6�GHZ[;\6 a476�C7àÈW/Z[CEMÄKÈG8Z[C'asQ ×\× Q<aE4YGH;ÀQ7M�DF4%G8AH;|äÛ�476�aQ:MB6�G8AH?Ba�GHCEN�ÙUT�asQ:?BM7GH;ÀMB`�G8ZBC M<?[D�ÙÈCEA/Q Ü CL4:a]Z û��%õ29U;\MBN[6ßÚ:ãk`7AH4:D�½u?È6uG�ÙJC Ü Q:AHC@4Ú%ã `:A]4%D Ü Q7A_CL4:a]Z}9<;ÀMÈN�à<ò¿;ÀMBaC9û�:õ¤9<;ÀMÈN[6�Q Ü Ú%ã `:A]4%D}6!Û�CEA8C�Csä<G8A]4:asG8CENÄK%G8ZBCvasQ ×À× Q¿a4%G8;\Q:M

19

AIFB AIFB

� �

statement reply

speech actassentinvitationproposal

quality pricediscrepancy

party offereeofferor

offer counteroffer

communication systemtelex

instrumentality

inanimate object

goods

person

entity

departure dispatch

deed transmission

content

indication of intention

letter

changeadditionmodification

revocation

attributesilence

time delay

address residence

act

intentionconduct

acceptance

abstraction

CONCEPT

day holiday

relation regal relation contract

withdrawal

rejection

location place of business

state circumstanceeffect

paymentdelivery

envelopetelephone

communication

quantity

L8M�N ´¨òB´ x7lEz�dE{~w'e]lEwUe]tH�7h_¥U{~tHr�dsrue8¥Yj2lEmo£Xy nUËÅ�<dsr�h_y�y

DF4%G8AH;|ä�4 × 68Q�ÙJCEa4:DFC�û�:õ�N¿;\DFCMB68;ÀQ7MB6EàEá�M�Q:A]N¿CARG8QvaQ:MB6�G8AH?BasGX4�aQ:MUG8Cä<GoØ:CLa�GHQ:ALKE4�68?[DQ Ü Ú%ã `:A]4%D Ø:CLa�G8Q7AH6�4%AHQ:?[MÈN}4%Ú[ÚJCE4:AH4:MBasC�Ú × 47asCva;ÀA]as?BD Ü CEA8CEMBasC�Q Ü CL4:a]Z}Q Ü ÚUõ�asQ7MBasCEÚ¿GH6Û�476�aE4 × as? × 4YGHCEN�àBðvM[C 4%A8G8;�a × CvQ Ü ��á�ò � aQ:MB68;\6�GH6�Q Ü 4%ÙJQ:?¿G@ó¼ÚÀ�#Ú%ã `:A]4%D}6Eà7W/Z[C M<?[D2ÙJCAQ Ü Ú%ã `:A]4%D Ø7CEa�GHQ:A]6�;ÀM5aQ:MUG8CäUG�4:A8CL4'Û�476�6�CG�4:6�õ�� Ü AHQ:D Csä¿ÚJCAH;ÀCEMBasC7àE�[Q7A�CE47a]ZPQ Ü ÚUõasQ:MÈasCÚ[GH6EK7GHZ[C@6�?[D Q Ü aQ:MUG8CäUG�Ø:CLa�G8Q7AH6�;\M�4 ×\× G8Z[C¤4:Ú[ÚÈCL4%A]4%MBaC�Ú × 4:aCE6�Q Ü GHZ[C@asQ:MÈasCÚ[G;ÀM��á�ò � Û�476}a4 × a? × 4%G8CLN�K_4%MÈNÊGHZ[C5Ø:CLa�GHQ:AFA8CEÚ[AHCE68CMUGH4%G8;\Q:MB6'Q Ü G8ZBC5asQ:MÈasCÚ[GH6}Û�CAHCQ:Ù¿G]4%;\M[CEN�àUW/Z[C�68CsG_Q Ü G8Z[CL6�C�Ø7CEasG8Q:A]6!;�6�?B68CEN}4768�ÞQ:A]NBò<ÚB47asC/G8Q2Csä<G8A]4:asG_asQ:MÈasCÚ[G_ÚB4%;\A]6Û�;|GHZ�aQ:MUG8Cä<G�68;\D'; × 4%AH;|GuT7à

ñãç�� êJë�ìàí?î¼ï�ðZñî?òéì�� ù¿ìZö�� ëãöZò ó��ãòéì�� êJë!ìZñúÀøZîF÷6ø�úTñ�ò��2ñùHîKò ë�ì��ú! ø#" ùHîKúÀí��4SØ<;ÀMB`úa4 × a? × 4YG8CLNºG8ZBCP68;ÀDF; × 4:A8;ÀGuT Ü A8Q7D G8Z[C�;\M[M[CA'Ú[AHQ¿N¿?Ba�G Ü Q:A�G8ZBCúóK� Á û5asQ:MÈasCÚ[GÚB4%;\AH6/Û�ZB;\a]Z�;�6�4 ×\× G8Z[C�aQ:D�Ù[;ÀMÈ4YG8;\Q:MPQ Ü Ú7õFaQ:MBaCÚ¿G]6K[4:MBNPZB4SØ<;ÀM[`F?B68CEN�G8ZBA8CL6�Z[Q × N�476�[à ô:ô:ô Á Ksô��asQ7MBasCEÚ¿GRÚB4:;ÀA]6ÄÛ�CAHC�Csä<G8A]4:asG8CLN�KE4:MBN¤asQ7MBasCEÚ¿Go68ÚÈCLas;ÀâÈa4%G8;\Q:M@G8CEDFÚ × 4%G8CL6RÛ�CAHCasQ:MÈ6uGHA8?BasG8CLN�àJWX4%Ù × C Á ;\6�G8ZBC × ;\6�GvQ Ü G8Z[C'Csä<GHAH47a�G8CLN�6�;\DF; × 4%A asQ7MBasCEÚ¿GH6 asQ7A8AHCE68ÚÈQ7MBN¿;\M[`G8Q'CE4:a]Z�asQ:MÈasCÚ[GEà Ö aQ:MBaCÚ¿G/;\M�ÙJQ × N × CsG8G8CA]6�;\6�C;ÀG8ZBCA�4%MP4%MBaCE6�G8Q:ALK<N¿CE6HasCEMBN[4%MUG�Q:A/46�;\Ù × ;ÀM[`�G8Q'GHZ[C × C Ü G�aQ:MBaCÚ¿G�;\M0GHZ[C¤asQ:MÈasCÚ[G/ZB;ÀCEAH4:AHa]Z<TFasQ:MÈ6uGHA8?BasG8CLN0?B68;ÀMB`Fü¤ð ü üvûoý;ÀMëGHZ[CúâBAH6�GPaE4:68Cè6�G8?BN¿T7à�á�MÉasQ7MBasCEÚ¿GP68ÚÈCLas;ÀâÈa4%G8;\Q:MéGHCDFÚ × 4YG8CL6K�68?Ba]Zç4�aQ:MBaCÚ¿GP;\6N¿;\6�G8;\M[`7?[;\68Z[CLNÅ476}W Ö&% ð O ð 3�^�AHC × 4YGH;ÀQ7MÄà Ö 6}GH4Yä¿Q7M[Q:DF;�a�AHC × 4YGH;ÀQ7MB6�ZB;ÀÚB6}4:MBN-M[Q:M¿ãGH4Yä¿Q7M[Q:DF;�a!A8C × 4%G8;\Q:MB68Z[;\ÚB6>D}4ST�ÙJC�DF;|ä¿CEN@;\M@G8ZBC × ;�6�GRÙB4:68CEN Q7M Q7M × T aQ:MUG8CäUGR68;\D'; × 4%AH;|GuT7KG8Z[C�aQ:MBaCÚ¿G�ÚB4%;\AH6�Û�Z[;\a]Z}D}4ST�ÙÈC�aQ:MBaCAHM[CEN'Û�;ÀG8Z}M[Q7M¿ãkGH4Yä¿Q7M[Q:DF;�a�AHC × 4YGH;ÀQ7MB68Z[;ÀÚÈ6�4%AHCQ:Ù¿G]4%;\M[CEN�ÙUT¾AHCDFQYØ<;\M[`5GHZ[C�asQ:MÈasCÚ[G'ÚB4:;ÀA]6�Û�;|GHZ�G]4Yä¿Q:M[Q7DF;\a�A8C × 4%G8;\Q:MB68Z[;\ÚB6à-�X;\`:?[AHCôP68Z[QYÛ�6@asQ7MBasCEÚ¿G¤6�ÚJCEa;|âÈaE4YGH;ÀQ7MèGHCDFÚ × 4YG8CL6@Csä<G8A]4:asG8CLNú4:ÙÈQ7?¿G@G8Z[C}asQ7MBasCEÚ¿G�<:476868CMUG=<BàW/Z[C�asQ7MBasCEÚ¿GH6/?[MÈN¿CA × ;\M[CLNP4:A8C ;\M�G]4Yä¿Q:MBQ:DF;\a AHC × 4YGH;ÀQ7M�Û�;ÀG8Z�CE47a]Z�Q:G8Z[CEAEà

20

AIFB AIFB

�E�

© ²7®�ªJ«�»[´ hu¥Ut!e]lEwUe]tH�7hÄ�<d{iru�Rt8¶:h�r�deHhut�xvdEe]e]lr�x7{~wU��hul�e]lEw%hut8¶:hÄ��{~z¤{~g|dru{ihkj�¹�hu¥7rutH��¥UlLg|xv�:� �L�E�L� ½Ë æ¹Í Ë Ý¦ó¦¬ Ë æ�Í Ë Ý¦ó¦¬ØçôÁ Ç ¬õÁ(Í Ç Á�ÚÛÁ(ç Ü6ö÷Ë æ�Í6¬®Ý¦¾6¬¿{µ�µ�¯�Â^´�¿{·Uµ�¯ ø{ù5ú+úüûþýÿWø��gÿWù5ý��Uù��Êÿ]ý �þÿWø��gÿWù5ý��¨´�¯�È ¯¢î¿{µ¢´ ¶�íÊ¯�±�¶{±��Qý��/¯¢íÊ¯�µ¢´��®ÂU¿�À^»�¯�·O´��®Â5¯�±�³�¶{·��®ïQº¨¿�·O´�Ð ´(À��®´�Ð »�¯��®Ù�¶Q¶Q¸¨³�� ]ÿ��ÿ�� ¢ø��Â¨±�Ð µ�¯��¨µ�¶�·O´�±�¿�µ¢´��¨¸U¯�È ¿�À��UÎ/Ð ´�¨¸U±�¿gÎå¿�È��²¶�íÊ¯�±�¯�¯��¨ÂUÈ ¿{µ�¯��¨ïQºU¿{È Ð ´(À¿{³�³�¯�·O´ ¶�íÊ¯�±�¶{±��Uø��5¯¢íÊ¯�µ¢´��¨¶�íÊ¯�±��²Â5¯�±�³�¶{·��¨¶�íÊ¯�±�¯�¯��UÎ/Ð ´�¨¸U±�¿�Îå¿{È��²´�Ð »�¯�� gù��!ù��µ�¶{»�»�º¨·UÐ µ�¿e´�Ð ¶{· �¨ø�ø��!��¨ýø��ôù��" �5µ�¶{·U¸¨ºUµ¢´��¨ÿ]ý �þÿWø��gÿWù5ýµ�¶{·U¸Uº¨µ¢´ Â¨¿�±�´JÀ��²´�¯�È ¯¢î��^µ�¶{»�»,ºU·UÐ µ�¿�´�Ð ¶�·µ�¶{·O´�±�¿{µ¢´ ¯¢íÊ¯�µ¢´��¨¿{µ¢´��²Â5¯�±�³�¶�·#�¨¸U¯�È Ð Þ{¯�±�À��UÂ¨¿�À^»�¯�·O´��¨ïQºU¿{·O´�Ð ´(À¸U¯�È ¿�À ¸¨¯�È Ð ÞO¯�±�À��U¶�í5¯�±��²¿�µ¢´��²ÂU¿�À^»�¯�·O´¸U¯�È Ð ÞO¯�±�À � ��ú$�Qý��¨ïQºU¿{·O´�Ð ´(À��¨Ù�¶Q¶Q¸¨³��²Â¨È ¿{µ�¯��Uø��²¸U¯�È ¿�À��U´�Ð »�¯��¨µ�¶{·O´�±�¿{µ¢´��¨Â5¯�±�³�¶{·��5¯¢í5¯�µ¢´��¨ïQºU¿{È Ð ´(À¸UÐ ³�Â¨¿e´�µ� Ù{¶Q¶Q¸U³��²ÂU±�Ð µ�¯��¨ø%��²Â5¯�±�³�¶{·��¨ïQº¨¿�·O´�Ð ´(À��²¶�íÊ¯�±�¶{±¯¢íÊ¯�µ¢´ Â5¯�±�³�¶�·#�O¿{³�³�¯�·O´��U¿{µ¢´��Q¶�íÊ¯�±�¶{±��Qµ�¶{·O´�±�¿{µ¢´��^ÂU±�¶{Â5¶�³�¿�È��^ÂU¿�À^»�¯�·O´��O´�Ð »�¯��{ÎÒÐ ´�U¸U±�¿�Îå¿�È��UÂU¿{±�´(À��O¸¨¯�È Ð Þ�&¯�±�ÀÙ{¶Q¶Q¸U³ ¸¨Ð ³�ÂU¿�´�µ��¨ïQºU¿{·O´�Ð ´(À��¨¸U¯�È Ð Þ{¯�±�À��^Â¨¿�À^»�¯�·O´��¨¿{µ¢´��²Â5¯�±�³�¶{·��¨Â¨±�Ð µ�¯��¨ïQºU¿{È Ð ´(ÀÐ ·U¸¨Ð µ�¿�´�Ð ¶{· Ð ·O´�¯�·O´�Ð ¶{·��¨ø�ø��!��¨ýø��Êø�ù5ú+ú&ûýÿ]ø��ÿ]ù²ýÐ ·O´�¯�·O´�Ð ¶{· Ð ·¨¸UÐ µ�¿e´�Ð ¶{·¶�í5¯�± �¨ø�ø��!��¨ýø��ô¿�³�³�¯�·O´��²ø{ù²ú úüûþýÿWø��gÿWù5ý��U¸U¯�È ¿�À¶�í5¯�±�¯�¯ ÎÒÐ ´�U¸¨±�¿�Îå¿{È��5ù��gù��5¿{³�³�¯�·O´��5¿�µ¢´��5ÂU±�Ð µ�¯¶�í5¯�±�¶�± ¿{µ¢´��²¿{³�³�¯�·O´��¨Î/Ð ´�¨¸U±�¿�Îå¿{È��²ù��'��(��!��(��ù5ý��¨¯¢íÊ¯�µ¢´��U´�Ð »�¯��¨Â¨±�Ð µ�¯��¨¸UÐ ³�Â¨¿e´�µ�ÂU¿{±�´(À µ�¶{·U¸¨ºUµ¢´��U¯¢íÊ¯�µ¢´��¨Â¨È ¿{µ�¯��'��(��ù²ýÂU¿gÀ^»�¯�·O´ ïQº¨¿�·O´�Ð ´(À�� Aÿ��5ÂUÈ ¿�µ�¯��¨¿{µ¢´��²Ù{¶Q¶Q¸U³��²ïQºU¿{È Ð ´(À��¨¸¨¯�È ¿gÀ��¨¯¢í5¯�µ¢´��¨Â²¯�±�³�¶�·#�¨µ�¶�·O´�±�¿�µ¢´��U´�Ð »�¯Â5¯�±�³�¶{· ¯¢íÊ¯�µ¢´��ù��eù��þ¿{µ¢´��ÂU±�¶�Â5¶{³�¿�È��þÙ{¶Q¶Q¸U³��¿{³�³�¯�·O´��ÎÒÐ ´�U¸¨±�¿�Îå¿{È��þµ�¶{·O´�±�¿{µ¢´��¸UÐ ³�Â¨¿e´�µ�#�ôÂ¨¿�À^»�¯�·O´��¸¨¯�È Ð ÞO¯�±�À�� )��5ÂUÈ ¿�µ�¯��UÂ¨±�Ð µ�¯ÂUÈ ¿{µ�¯ Â¨¿�À^»�¯�·O´��¨¸¨¯�È Ð ÞO¯�±�À��^´�Ð »�¯��¨ïQºU¿{·O´�Ð ´(À��¨ÂU¿{±�´(À��²¿�µ¢´��¨Â²¯�±�³�¶�·ÂU±�Ð µ�¯ ¸¨Ð ³�ÂU¿�´�µ��¨¿{µ¢´��²¶�í5¯�±�¶�±��5Ù�¶Q¶Q¸¨³��¨Î/Ð ´�¨¸U±�¿gÎå¿�È��²¶�íÊ¯�±�¯�¯��¨Â5¯�±�³�¶{·ÂU±�¶�Â5¶{³�¿�È Â5¯�±�³�¶�·#�¨¯¢í5¯�µ¢´��UÎ/Ð ´�U¸¨±�¿�Îå¿{È��Qý��ïQºU¿{È Ð ´JÀ ïQº¨¿�·O´�Ð ´(À��¨Â¨¿�À^»�¯�·O´��¨Ù{¶Q¶Q¸U³��²¿{µ¢´��²¸¨¯�È Ð ÞO¯�±�ÀïQºU¿{·O´�Ð ´(À Â¨¿�À^»�¯�·O´��¨¸¨¯�È Ð ÞO¯�±�À��UÙ�¶Q¶Q¸¨³��²¿{µ¢´��²ïQº¨¿�È Ð ´(À��¨¸UÐ ³�ÂU¿�´�µ��¨Â¨È ¿{µ�¯��¨µ�¶{·O´�±�¿{µ¢´��gÿ]ú$�´�¯�È ¯¢î µ�¶{·U¸¨ºUµ¢´��Uø{ù²ú úüûþýÿWø��gÿWù5ý��¨¿{µ�µ�¯�Â^´�¿{·Uµ�¯´�Ð »�¯ ¿{µ¢´��²¶�íÊ¯�±�¶{±��²¸U¯�È Ð ÞO¯�±�À��UÂUÈ ¿�µ�¯��¨¯¢íÊ¯�µ¢´��¨Â¨¿�À^»�¯�·O´��*ôû �¨ý��gÿ �)��¨¿{³�³�¯�·O´Î/Ð ´�¨¸U±�¿gÎå¿�È ¶�íÊ¯�±�¯�¯��¨¶�í5¯�±�¶�±��5Â²¯�±�³�¶�·#�¨ÂU±�Ð µ�¯��¨¿{µ¢´��²¿{³�³�¯�·O´��5¯¢í5¯�µ¢´��¨ÂU±�¶{Â5¶�³�¿{È

W/Z[CèâÈMB4 × asQ7MBasCEÚ¿G�N¿CsâBMB;|GH;ÀQ7M ;�6PaQ:MB6�G8AH?BasG8CEN Ü A8Q7D asQ7MB6�;�N¿CEAH4%G8;\Q:MëQ Ü asQ:MÈasCÚ[GÚB4%;\AH6@;\MºG8ZBC}G8CED'Ú × 4%G8CE6Eà7�X;\`:?[AHC�ó?��68Z[QYÛ�6 G8Z[C�N¿CsâBMB;|GH;ÀQ7M¾Q Ü G8Z[C�aQ:MBaCÚ¿G <74:6H6�CEMUG=<asQ:MÈ6uGHA8?BasG8CLN Ü AHQ:DùGHZ[C/G8CEDFÚ × 4%G8CL6à Ö�× GHZ[Q:?B`:ZFA8C × 4%G8;\Q:M Ö � ý_O�W-Csä¿;�6uG]6�4 × 6�Q ;\MF476868CMUG8ãQ%å>CAHCCè4:MBNç4:6H68CMUG�ã Q%å>CAHQ:ALK�;ÀG�;�6�AHCÚ[AHCE68CMUGHCENéÙ<TëN¿CsâBMB;|GH;ÀQ7Mç;ÀMBZ[CAH;|G]4%MBaCú4:MBNëM[Q:GN¿CE6HasAH;ÀÙJCENÄà

ñãç(' )+*6î¼ï�ù¿ñî?ò ì,�.- ë/ ùEò ì10 ÷ãø�úTñò��2ñ32�òéúÀï¾ù¿ï¾ñ546��÷ãî¼ï�ðàñî?ðàï�ú�úC@ZB4SØ7CvN¿CâBM[CEN0G8Z[C GHZ[A8CL6�ZBQ × N Ü Q:A�GHZ[CvÚBA8QSä¿;\DF4 ×À× TFAH4%G8C@4:6��[à~þ��BK:GHZ[C¤asCA8GH4:;ÀM�68aQ:ÚJC4:6_G8Z[C¤6H4%DFCv68CMUG8CEMBasC@4%MÈNFG8AH;ÀCLN}GHQ�CäUGHAH47a�G/N¿Q:D}4%;\M¿ã�6�ÚJCEa;|âJa�Z[;\CA]4%A]a]Z<T'6�G8AH?Ba�GH?[AHCv;\MG8Z[C�âBAH6�G'aE4:68C�6�G8?BN¿T7à Ö 624èA8CL6�? × GEKXü@ð ü ü ûRýÂá8á�Csä<GHAH47a�G8CLNéóK��ÚB4%;\AH62Q Ü asQ7MBasCEÚ¿GH6A8CE`74%A]N¿CLN�476�ÚB4%A8G_Q Ü N[Q:D}4%;\M¿ã�6�ÚJCEa;|âÈa/ZB;ÀCEAH4:AHa]Z<T26�G8AH?Ba�GH?[AHC Ü A8Q7D GHCsä<G�asQ7A8ÚB?B6à��¤ÚB4%;\A]6Q:?¿G/Q Ü GHZ[CD ZB4SØ7C�Q¿aa?[AHA8CLNF;ÀM�G8Z[C@asQ7MBasCEÚ¿G�Z[;\CA]4%A]a]Z<T'aQ:MB6�G8AH?BasG8CEN0ÙUT'GHZ[Cv?È6�CEA�4:MBNZB4SØ:C�M[Q:G_Q¿aa?[A8AHCEN';\MFG8Z[C�G8AH;ÀDFDFCENFDFQ¿N¿C × àUW/ZB4YG_;�6EK%G8ZBCTF4%MBNFDFQ¿N¿;ÀâÈa4%G8;\Q:MB6_ÙUT�G8ZBC?B6�CEA@Û�CEA8C'684:D'C7àÈá G�68Z[QYÛ�6�G8ZB4%G@ü¤ð üvü ûRýÉá8á�a4:MèD}4%97C2?B68C Ü ? × 68?[`7`:CE6�G8;\Q:MÈ6�4%ÙJQ:?[GN¿Q:D}4%;\M¿ã�6�ÚJCEa;|âJa�Z[;\CA]4%A]a]Z<Tv6�G8AH?Ba�GH?[A8C�?B68;ÀM[` a4:MBN¿;�N[4YGHC Ü Q7ARGHZ[CDùCäUGHAH47a�GHCEN Ü AHQ:DøG8Cä<GasQ:AHÚ[?B6EàãS�?¿G�GHZ[C}AH4%G8C}Q Ü 6H4%DFC}6�?B`:`:CL6uGH;ÀQ7MB6¤4:6@DFQ¿N¿;ÀâÈa4%G8;\Q:MºÙ<T5GHZ[C0?B68CA¤;�6�4%ÙJQ:?[Gõ�=�ï � ë óK��7÷�4%MBN�;�6/M[Q%G�`7QUQ¿N�àBò¿QBK¿Û�C¤ZB4SØ:CvG8Q};\DFÚ[A8QYØ7CvCä<G8A]4:a�GH;ÀQ7M�Q Ü aE4%MBN¿;�N[4%G8C:à

21

AIFB AIFB

��

²U¯�¯O« +B°-,�./,10/243�5�6�7�6�8�9�:<;�=�>�«%±�=<±¬ Ü ¾Òæ�Í�æ¹Ú@? ;U²U³L°,�./,10/243�5�6�7�6�8�9�:<;�«#>�«7³E°,�./,10/243�5�6�7�6�8�9�:<;�=�>�«%±,�./,10/243�5�6�7�6�8�9�:<;�A¹«%±¯�=+,�./,10/243�5�6�7�6�8�9�:<;�=�>�«%±^««,�./,10/243�5�6�7�6�8�9�:<;�B M °�CEDJ±s²!Bv²�ª,�./,10/243�5�6�7�6�8�9�:<;U° MGF «¬ Ü ¾Òæ�Í�æ¹Ú@? ;�A>±�='A<=¿¯]²�ª

L�MON ´�HB´ � ¥Ut�eHlLwUeHt]�7h_��¿t]e]{iÃUe�dhu{~lLwFhutHz¤�Ug|dhut]�!m\lErJI d��t]wYh(K

²U¯�¯O« +[° ÜML Ý¦Í6¬ ;�A¹«%±¯�=+çôÝ L¹Ü ç!& Ç ÝON Ó Ý¦Í Ë Ý ;�= >,«%±Ü Í6¬/æ�ÍM?¹Ú ;�B M °�CEDJ±s²!B�²!ª

L�MON ´¾P!PÈ´ � ¥7t�e]lLw7e]t]�7h_x:tHÃ<w7{ihu{|lEw'm�lrQKSdE��tHw%h(K@p�{ihu¥Ft]x7{ihu{~wU� hu¥7t�hut]z¤�Ug|dshut]�

ñãç�æ R úÀíKð," î?íXù¿ìZö�)$S¿ù®" ðZùHî?òéë!ìW/Z[Cv?È6�CEA�Û�;ÀG8Z × CE`74 × 9<M[QYÛ × CEN¿`7C�N¿;�N0CØY4 × ?B4%G8;\Q:M�4:ÙÈQ7?¿G/Csä<G8A]4:asG8;\Q:M0Q Ü asQ:MÈasCÚ[G�ÚB4:;ÀA]6EàO�Q%GHC�G8ZÈ4YG_GHZ[CvaQ:MBaCÚ¿G�N¿CsâBMB;|GH;ÀQ7M�asQ7MB6uGHA8?Èa�G8CLNF;ÀMFGHZ[;�6�aE4:68C�6uGH?BN¿T';\6_Q:M × T Ü Q:A�GHZ[C ÚUõasQ:MÈasCÚ[GH6@4:6 ;\M[Ú[?¿G¤GHCAHDF6EK�4%MBNú;�6@M[Q%G G8Z[C}Û�ZBQ × C'aQ:MBaCÚ¿G¤N¿CsâÈM[;|GH;ÀQ7MºÛ�Z[;\a]ZÞ6�ZBQ:? × NÙÈC�aQ:MB6�G8AH?BasG8CEN Ü AHQ:D ��á�ò � à<W/Z[C�N¿CGH4%; × Q Ü GHZ[C�Cä<G8A]4:a�GHCENFasQ7MBasCEÚ¿G_ÚB4:;ÀA]6�;\MFG8Z[;�6_aE4:68C6uGH?BN¿T�4:A8C@68Z[QYÛ�M�;\M�Wo4:Ù × C ÚÈàWo4%ä<Q7M[Q:DF;�a�Q:A!M[Q7M¿ãkGH4Yä¿Q7M[Q:DF;�a�A8C × 4%G8;\Q:MB68Z[;\ÚB6oCsä¿;\6�G8CLN�;\MXû:ô�= Ü A8Q7DøG8Z[C�GHQ:Ú2Q Ü G8ZBC

× ;�6uG Q Ü asQ7MBasCEÚ¿G�ÚB4:;ÀA]6�Û�;ÀG8Z5ZB;À`7Z5asQ:MUGHCsä<Gv68;ÀDF; × 4:A8;ÀGuT�ÙJCsGuÛ�CCEM�G8Z[C'asQ:MÈasCÚ[GH6EàJò<;\MBasC�4asQ:MÈasCÚ[G�ÚB4:;ÀA�Û�;ÀG8ZPZ[;À`7Z�aQ:MUG8CäUG/6�;\DF; × 4%AH;ÀGuT'ZÈ4:6/42ZB;À`7Z�ÚJQ76H6�;\Ù[; × ;ÀGuT'G8ZB4%G/;ÀG/ZB4:6/68Q:DFC9U;\MBNPQ Ü A8C × 4%G8;\Q:MÄK[aQ:MBaCÚ¿G�N[CsâBM[;ÀG8;\Q:MÈ6�a4%MPÙJC × CLN�Ù<TPasQ:MÈ6�;�N¿CAH;\M[`2GHZ[CE68C@ÚB4%;\A]6àW/Z[C@Ú[A8Q7Ù × CD}6/Q:Ù¿G]4%;\M[CEN Ü AHQ:D GHZ[;�6�a476�C@6uGH?BN¿TP4%AHCvGHZ[C Ü Q ×\× QYÛ�6à

T þ��Éþ��Ù$ Þ� � �� VU �û��þ�%��! Ýßø W/Z[AHCE68Z[Q × NÊQ Ü G8ZBC�aQ:MUG8Cä<GF68;ÀDF; × 4:A8;ÀGuTÊa]ZB4:M[`:CL62;\MCså>CEa�GH;ÀØ7C@ØS4 × ?BCvÛ�;ÀG8ZPCE47a]Z�N¿Q7DF4:;ÀMRàUá G�;�6�ZB4%A]N}GHQ'68CsG�?[Ú�G8Z[C@DFQ76�G/Cså>CEa�GH;ÀØ7CvØY4 × ?[C@;\M4:N¿ØY4%MÈasC:àÀ�X;À`7?[A8C¤ó7ó�;�6�G8ZBC�A8C × 4%G8;\Q:MFÙJCsGuÛ�CCM0G8Z[C�M<?[D�ÙÈCEAH6_Q Ü G8ZBC�Csä<GHAH47a�G8CLNFasQ:MÈasCÚ[GÚB4%;\AH6�4:MBN�A8CLa4 ×\× 4%MÈN�Ú[AHCEa;\68;\Q:MP;ÀMPG8ZB;\6�aE4:68C 6�G8?ÈN¿T:àW �Eþ#�� X ��"�Ú�"� Û�á�ZYó ��=þ�� \[ þ�Ýß�"�Ú�"� ��Q:MBaCÚ¿G�68ÚÈCLas;ÀâÈa4%G8;\Q:MÞG8CEDFÚ × 4%G8CL6¤ZB4SØ:C'Q:M × TasQ:MÈasCÚ[G�ÚB4%;\AH6!ÙÈ4:68CEN'Q:M'G8ZBC�asQ:MUGHCsä<G_6�;\DF; × 4%AH;ÀGuT:KY;ÀG_A8CLêU?[;ÀAHCE6�6�G8; ×\× Z[;\`:Z}asQU6uG�GHQ�6�ÚJCEa; Ü TA8C × 4%G8;\Q:MB68Z[;\ÚB6�ÙJCsGuÛ�CCEMçGHZ[CD�à�á G�;\6PM[CECEN¿CLNçGHQÅ68?[Ú[ÚJQ:A8G�6�ÚJCEa;|âÈaE4YGH;ÀQ7M Q Ü asQ:MÈasCÚ[GA8C × 4%G8;\Q:MB68Z[;\ÚB6�Q:MPG8Z[;�6�68T¿6uGHCD ;\MPG8Z[C Ü ?¿GH?[AHC¤Û�Q7A89>à] $_^��"�ûÞ��òÛ�Q` �ûÝ��Ú ��ÝOþaU!þ��Ù$ Þ� "Ýß#�"� �BQ:AXCsä[4%DFÚ × C:KG8Z[C�G8CEA8D <%G8A]4%MB68DF;\6H6�;\Q:M!<�;�6o?B6�CLN;ÀMúGuÛ�Q�DFCE4%MB;ÀM[`U6K6<%G8A]4%MB68DF;\6H6�;\Q:M-ï·Q Ü `:Q<Q¿N[6H÷�<�4:MBND<%G8A]4%MB68DF;\6H68;ÀQ7MÊï·Q Ü asQ7D'D�?[M[;�a4%ãG8;\Q:MÈ÷�<[KR;\MÞG8Z[C�4%A8G8;�a × C:KÄÙ[?[G�ü@ð ü ü ûRýùá8á@aQ:MB68;�N¿CA]6 G8Z[CL6�C0G8CAHD}6�4:6¤G8Z[C�6H4%DFC04:MBNasAHCE4YGHCE6��ÞQ:A]NBò<ÚB47asC@4:6�;|G�;�6à[W/ZBCAHC Ü Q:AHC@asQ:MÈ6uGHA8?BasG8CLN0Ø:CLa�GHQ:A/Csä¿Ú[AHCE6H6�;\Q:M�DF4ST�M[Q%G�ÙJCCsä[4:asGEà:á�M�Q:A]N¿CA�GHQ�Csä<GHAH47a�G_DFQ:AHC�?B68C Ü ? × asQ7MBasCEÚ¿G_ÚB4:;ÀA]6EK768CD}4%MUG8;�a�68ÚÈCLas;�4 × ;O�L4YGH;ÀQ7M'Q Ü 4D2? × G8;�6�CEMB68CvÛ�Q:A]NF;\6�M[CLasCL686H4%AHT:KU4%MBN�;|G�6�ZBQ:? × N0ÙJC aQ:MB68;�N¿CAHCENFGHZB4YG�G8ZBC9Ú:ãk`7AH4:DF6�Û�;|GHZ684:D'C¤4:Ú[ÚÈCL4%A]4%MBaC 4:MBN�N¿;ÀåJCEA8CEM7G�DFCE4:M[;\M[`}4%AHC@N[;|å>CAHCMUGäÚ%ã `:A]4%D}6Eà

22

AIFB AIFB

�]�© ²7®�ªJ«�ØJ´ � ¥7t�x7tHh�d{|gBlEm�hu¥7t/tH¶%h�r�dEe8hut�x'e]lEwUe]t]�:h��<d{~ru�

� ¥:rut]��¥UlEgÀx Q�¶:h�r�deHhut�x2e]lLw7e]t]�7h_�UdE{ir � x7�%{~�udE¸Ug~t ��w7�:w7l�p�w yqz¤�7rulE�¿tHr�:� �L�E�L� � � �E� � § �E�

L8M�N ´�PTP<´ rut]e�dg~g>dEwUx'�7rutHe]{~��{|lEw

b .047*�'oÏH4RÒdc Ñ�ÐOe

á�MÊG8ZBC�AHCE68CE4:AHa]Zº?B68;ÀMB`èØ:CEA8Ù[ãkQ7A8;\CMUG8CLNºDFCsG8ZBQ<NÄKoGHZ[CPA8C × 4%G8;\Q:MÊQ Ü 45Ø:CEA8Ù-4%MBN¾MBQ:?[MB6D'Q¿N¿;ÀâBCLNéÛ�;|GHZë;|G�;�60N¿CL68aA8;\ÙJCEN�K_4:MBN-G8Z[CúasQ:MÈasCÚ[G�N¿CsâBMB;|GH;ÀQ7Më;\60aQ:MB6�G8AH?BasG8CEN Ü AHQ:DG8Z[CL6�CP;ÀM Ü Q:AHDF4%G8;\Q:M ï�C:à `Bà�ñ ��4%ZBMÞóLô:ô��LöÁ÷sàXá�Møñ �B4%?BA8C¤4%MÈN�OgfCEN¿C ×À× CEa'óLô:ô:ôLö KXGH4%ä<Q7M[Q:DF;�aA8C × 4%G8;\Q:MB68Z[;\ÚB6�4:MBNøò<?[Ù>a4YGHC`7Q:AH;O�L4YG8;\Q:MG�[A]4%DFCúQ Ü Ø:CEA8ÙÈ6ÞïkòH��÷P4%AHCúCäUGHAH47a�GHCEN Ü AHQ:DG8CEa]ZBM[;\aE4 × GHCsä<GH60?B6�;\M[`Ê4ÞD}4:a]Z[;\M[C × CE4:A8M[;\M[`ºDFCsGHZ[Q¿N�à�W/Z[C�MBQ:?[MB60;ÀMÅGuÛ�Q¾Q:A}DFQ:AHC9U;\MBN[6�Q Ü N[;|å>CAHCMUGvòH��Û�;|GHZ�4F6H4%DFC Ü AH4:DFCsã MB4%DFC@4%MBN�6 × Q:G�ã MB4%DFC@;\6/`U4YGHZ[CAHCEN�4:6/Q:MBCasQ:MÈasCÚ[GEK[ÙB476�C¤a × 47686Eà Ö MBNPQ7M7GHQ × Q:`7T}Û�;|GHZ�Q:M × T}GH4%ä¿Q:M[Q7D';�a A8C × 4%G8;\Q:MB68Z[;\ÚB6/;�6�Ù[?[; × G�Ù<Ta4%AHAHTU;\M[`}Q7?¿G a × ?B6�G8CEA8;\M[`0Q Ü G8Z[C�ÙB4:68C�a × 47686 Ü ?BA�GHZ[CALàJ35Q:AHCQYØ:CEAEK¿;\M5ÚB4:AH4 ×À× C × K!B�CE6�G8AH;\asãG8;\Q:M0Q Ü ò<C × CEasG8;\Q:MèïéB�ò[÷_Û�Z[;\a]Z};�6�6 × Q%G8ãkØY4 × ?[C�;\M�òH�º;�6�4 × 6�Q¤AHCÚ × 4:aCENFÛ�;ÀG8Z�G8Z[C asQ:MÈasCÚ[GÛ�;|GHZPÛ�ZB;\a]ZP;|G�;�6�6H4YG8;�6�âBCEN�;ÀMB6�GH4:MUG8;�4YG8CLNPò¿��à!��QYÛ�CØ:CEAEKUÚ[AHQ:ÚJCA/CØY4 × ?B4%G8;\Q:MP;\6/MBQ%G�T:CGN¿Q:M[C7à!ò¿;ÀMBaCPòH�øAHCÚBA8CL6�CEM7G]6¤GHZ[C�68T<M7G]4:asG8;�a0AHC × 4YG8;\Q:MÈ6�Z[;\ÚB62ÙÈCGuÛ�CECMÊØ7CAHÙÊ4%MBN¾MBQ:?[MÄKG8Z[C�6�G8CEÚ Ü Q7A/G8Z[C�aQ:M<Ø:CEAH68;\Q:M}GHQ}M[Q:M¿ãkGH4%ä¿Q:M[Q7D';�avA8C × 4%G8;\Q:MB68Z[;\ÚB6�;\6/M[CLasCL686H4%AHT:à

ðvM�G8Z[C�Q%G8ZBCA�ZB4:MBN�KJ;ÀMèQ:MUG8Q × Q7`:T × CE4:A8M[;\M[`0?È6�;\M[`�N[4%GH4%ãkDF;\M[;ÀMB`}D'CG8Z[Q¿N�K>N¿;�68aQYØUãCAH;ÀM[`@M[Q7M¿ãkGH4Yä¿Q7M[Q:DF;�a�AHC × 4YGH;ÀQ7MB68Z[;ÀÚÈ6!?B6�;\M[`�4¤476868Q¿as;�4YG8;\Q:M�A8? × C�4 × `7Q:AH;|GHZ[DÂ;�6�Ú[AHQ:ÚJQ768CENÙUTÞñ 3è4%CENBa]Z[C¤4%MBN�òUG]4:4%Ù��Lö àÈW/Z[CET�CäUGHAH47a�G�aQ:MBaCÚ¿G�ÚÈ4%;\AH6/ÙB476�CLN�Q:MPG8ZBC�DFQ<N[;|â[ãa4YGH;ÀQ7M�;ÀM Ü Q:AHDF4%G8;\Q:M�ÙJCsGuÛ�CCM�GHCAHD}6�68C × CEa�GHCEN�Û�;ÀG8Z5ÚÈ4%A]6�;\M[`BKÈ4%MBN�D}4:N[C¤G8Z[C�asQ:MÈasCÚ[GÚB4%;\AH6/42G8A]4%MB6H4:asG8;\Q:MÄàHS�T}?B68;ÀMB`'ZBC?[AH;\6�G8;�a6�Û�;ÀG8Z�6�ZÈ4 ×\× QYÛëGHCsä<G�Ú[AHQ<aCE6H6�;\M[`ÈK:GHZ[C@`:CMBCA8ã4YG8;\Q:MFQ Ü 4 GHAH4:MB6847a�GH;ÀQ7M�DFQ:AHC/A8C�hBCEasGH6!GHZ[C�68TUMUG]4Yä2Q Ü GHCsä<GH6Eà:3�Q7A8CEQYØ:CEAEK¾B�û Ö K:Û�Z[;�a]ZF;\6G8Z[CE;ÀA Q:AH;À`7;ÀMÈ4 ×Ä× CE4:A8M[;\M[`�47aa?[AH47asT�Q Ü M[Q:M[ãqG]4Yä¿Q:M[Q7DF;\a@AHC × 4YGH;ÀQ7MB6�ZB;ÀÚB6�?B68;ÀM[`0G8Z[C�Csä¿;\6�G�ã;ÀM[`�GH4Yä¿Q7M[Q:DF;�a/A8C × 4%G8;\Q:MB6EKY;\6_Ú[AHQ:ÚJQ768CEN�à:W/ZBC�asQ7MBasCEÚ¿G_ÚB4:;ÀA�Cä<G8A]4:a�GH;ÀQ7MFD'CG8Z[Q¿N};\M}Q:?BA

23

AIFB AIFB

� §

ÚB4%ÚJCA�N[QUCL6�M[Q:G�M[CCLN�ÚB4%A]68;ÀM[`ÈKB4%MÈN�;ÀG aE4%M�4 × 68Q}A8?[M�Q:åÞasQ7MUG8Csä<G�68;ÀDF; × 4:A8;ÀGuT�ÙJCsGuÛ�CCEMG8Z[CvG8CEA8D}6�4%Ú[ÚJCE4:A8CLN}4%ÚB4:A�G_CL4:a]Z}Q:G8Z[CEA�;\M0GHCsä<GH6�Q:A�M[Q%G�D'CLN¿;�4YG8CLN0Ù<T'G8Z[C 684:DFC�Ø:CEA8ÙRà

i 8úÑ�Î�Ô6*�Ó9\<ÕuÑ_Îá�M5G8ZB;\6 ÚB4:ÚÈCEAEK>Û�C�N¿;�68a?B6H6�CLN5Z[QYÛøGHQPaQ:MB6�G8AH?Ba�G@4�N¿Q7D}4%;\MèQ:MUGHQ × Q:`:TP?B68;ÀMB`�Csä¿;�6uGH;ÀMB`3 B�ü 4:MBN}N¿Q7DF4:;ÀM[ã 68ÚÈCLas;ÀâÈa�G8Cä<GH6Eà%á�M}Q:A]N¿CA!GHQ24:aEêU?[;ÀAHC�GH4%ä<Q7M[Q:DF;�a/A8C × 4%G8;\Q:MB68Z[;\ÚÄKYGuÛ�Q6uGHAH4%G8C`7;ÀCL6!ZB4SØ:C�ÙJCCEM}Ú[AHQ:ÚJQ768CEN6>YD}4YG]a]Z[CENFAHCE68? × G_4%MB4 × T¿68;\6�4:MBN'G8AH;\D'DFCLNFA8CL6�? × G_4%MÈ4 × ãT<68;�6à��[?[A8G8Z[CEA8DFQ7A8C7KB;\M�Q:A]N¿CA�GHQ × CE4:A8M�M[Q7M¿ãqG]4Yä¿Q:MBQ:DF;\a@AHC × 4YGH;ÀQ7MB68Z[;ÀÚÈ6KJasQ:MÈasCÚ[G�ÚB4%;\A]6DF4ST�ÙÈCvA8C × 4%G8CLN2GHQ2asQ7MBasCEÚ¿G_N¿CâBM[;ÀG8;\Q:MÄKUCsä<G8A]4:asG8CLN�Q7MFG8Z[C�ÙB4:68;\6�Q Ü G8Z[CvasQ%ã Q¿aa?[A8AHCMÈasC;ÀM Ü Q7A8D}4%G8;\Q:MF;ÀM�N¿Q:D}4%;\M¿ã�6�ÚJCEa;|âJa�G8Csä<G]6KU4%MBN}4�N[Q:D}4%;\MFQ:MUG8Q × Q7`:T�;�6_N¿CEØ:C × Q7ÚÈCLNFÙUT�G8ZBCD'Q¿N¿;ÀâÈaE4YG8;\Q:Mç4:MBNë68ÚÈCLas;ÀâÈa4%G8;\Q:MëQ Ü aQ:MBaCÚ¿G�A8C × 4%G8;\Q:MB6}Û�;ÀG8ZµasQ:MÈasCÚ[G�6�ÚJCEa;|âJa4YGH;ÀQ7MG8CDFÚ × 4YGHCE6EàÈá G@6�CEA8Ø7CE6v4:6�GHZ[C2`7?[;�N¿C × ;ÀMBC Ü Q:A MB4:A8AHQYÛ�;ÀMB`FN[QYÛ�M5Z<?[`:C268ÚB47asC�Q Ü asQ:MÈasCÚ[GÚB4%;\AH6�GHQ}asQ:MÈ6uGHA8?BasG�4'N[Q:D}4%;\M�Q:MUG8Q × Q7`:T:àá G�;�6�4 × DFQ76�G�aAH4 Ü G�ã Û�Q7A892G8Q�aQ:MB6�G8AH?BasG�42N¿Q7D}4%;\M0Q:MUG8Q × Q7`:T7K:4%MBN0;|G�;\6�6uGH; ×\× N¿;�j0a? × G

G8Q¤Q:Ù[GH4%;\M�G8Z[C�Z[;\`:Z}68?[Ú[ÚJQ:A8G�AH4%G8C/Q:MF68T<6�G8CED�à:W/Z[C�ü@ð ü ü ûRý¾á8áoD}4%;\M × T268?[Ú[ÚJQ:A8GH6 Ü Q7AasQ:MÈ6uGHA8?BasG8;\Q:MFQ Ü 4�aQ:MBaCÚ¿G�Z[;ÀCEAH4:AHa]Z<T�Û�;ÀG8Z0GH4Yä¿Q7M[Q:DF;�a/A8C × 4%G8;\Q:MB68Z[;\ÚB6�4%MÈNFCsä<G8A]4:asG8;\Q:MQ Ü aQ:MBaCÚ¿G@ÚB4%;\A]6�Û�;ÀG8ZúM[Q:M¿ãkGH4%ä¿Q:M[Q7D';�a¤AHC × 4YGH;ÀQ7MB6�ZB;ÀÚB6vM[QYÛ¤à ��QYÛ�CØ7CAv4P68?[Ú[ÚJQ:A8G Ü Q7A6�ÚJCEa;|âÈaE4YGH;ÀQ7MPaQ:MBaCÚ¿G�AHC × 4YG8;\Q:MÈ6�Z[;\ÚP;\6/;\MBN¿;�6�ÚJCMÈ684:Ù × C7à[W/Z[C Ü ?¿GH?[A8C¤Û�Q:AH9 Ü Q ×\× QYÛ�6¼><;\D�ãÚ[A8QYØ7CDFCMUG�;\M5GHZ[C}68aE4 × 4%Ù[; × ;|GuTPQ Ü GHZ[CFN¿CsâBMB;|GH;ÀQ7MÞ68?[Ú[ÚJQ:A8GvÙ<T × CE4%AHM[;\M[`�Q Ü Z[C?BA8;�6uGH;\aE6K;ÀMUG8AHQ¿N¿?BasG8;\Q:M5Q Ü G8Z[C'?B68C Ü ? × N[4%GH4Yã DF;ÀMB;ÀM[`�D'CG8Z[Q¿Nè;ÀMB6�G8CL4:N�Q Ü �úQ7AHNÈò<ÚB4:aC:K>4%MBNè6�T¿6�ãG8CDø;\MUG8C`7AH4%G8;\Q:M@Q Ü G]4Yä¿Q:M[Q7DF;\a!AHC × 4YGH;ÀQ7MB6�ZB;ÀÚ¤4:aEêU?[;\68;ÀG8;\Q:M¤DFQ¿N¿? × C�4%MÈN@M[Q:M¿ãkGH4%ä¿Q:M[Q7D';�aA8C × 4%G8;\Q:MB68Z[;\Ú × CE4:A8M[;\M[`'DFQ<N[? × C0ï�M[QYÛ ;\DFÚ × CDFCMUG8;\M[`<÷�à

"ëÔOe�Î�Ñlk *[4oÒä+�%:4RÎ�Ï¿\

�úC�Û�Q:? × N × ;\9:C�G8Q@Csä¿Ú[AHCE6H6XQ:?[AXGHZB4%M[9¿6XGHQ@35AEà%WX4%9Y4%D}47684vá�Û�47N¿C@ï�4 `:A]4:N¿?È4YG8C/6�G8?BN¿CEMUGQ Ü 6�Z[;��?BQ:9Y42?BM[;ÀØ7CA]6�;ÀGuT[÷�4%MBN�G8ZBC¤D'CED2ÙJCA]6/;ÀMPGHZ[C@^_4%D}4%`7?Ba]Z[;Àã ûo4%ÙÄà

.á41;=4RÐ¿4RÎ�Ô 4à\m nYp�dr�hulEf7h_tHh��UdEgÁ� �]�L� «�npoX{~g~g7n%p!dr�hulLf:h��s¼�dz¤t]��¥��dhu{~gÁ��q�t]�%{~wrq�wU{~�L¥Yh�dw<x � lEz-¼_fU�� ; I � lE�p�dsr�x�¢/{~� h�ru{~¸Uf:hut�xF��tvlEmXîJdsru�LtH�kn:e]dEg~t�í�wYhulLg~lL�E{~t]��K7�È��rulYeL�BlEmRhu¥Ut � �hu¥sq�w7lsp�g~t�x:�Lt � eH�t fU{~��{ihu{|lEw'Æ�lru�%��¥UlL��¹uq � Æéì � «L½8�J¹ �]�L� «S½

m q�f:rut]z�dhu��f'dEw<xpv!dz�dE�EfUe�¥U{ ��E� ª�n�»FdE�udE�%{\q�f:rut]z�dhu��f¾dw<x � d�Ld¥U{ir�dgv!dz�dE�EfUe�¥U{ ; I �îÈt]�LdEg[í�wYhulEg|lE�Ej ¼�tHÃUwUt]z¤tHw%hÄn:fU�7�¿lEr�hàQÄw:�%{irulLw7z¤t]wYhÄ��{~wU��d�»Fde8¥7{~wUtH��¼_t�dLx7dE¸Ug~t_¢/{~eHhu{~lE�w<dsr�j K:�wI�xQy�zu{ |w}�{4~��%��z1��G� {��#��}%��~��'��~��K:� �E�� ]� ª%�>¹ ��L� ªL½

m n%t]�%{|f7e8¥7{[tHh��<dgÁ� ��E� ¨�n¤¼�{~t]�Ll n%t]�%{~fUe�¥U{Á�%£X¥7{|�Hf7ruf � lE�:{Á�:»Fd�udE�%{q�f7rutHz�dhu��f¤dw<x � d�Ld¥U{ir�dv!dz�dE�EfUe�¥U{ ; Ik¢�í�¢/¢�î�Q ; � ¢/lLz�d{~wÞí�wYhulEg|lE�Ejè¼�dE�U{|x�¢/t]�Lt]g~lL�Uz¤tHw%h QRw%�%{irulLw7z¤t]wYh(K7��R¼_y�£ � y � ¨:�È¹ ��L� ¨S½

m »Fdt�x7e�¥Ut�dw<x}nYh�dLdE¸ � �E�L��n � g~tH¶7dEw<x:tHr »FdEt�x:e8¥7tL�RnYhutHÇ[t]wèn%h�dEdE¸ ; Ik¢/{~��e]ls�St8ru{|w7��£XlLwUeHt]�7�huf<dgÈ¼_t]g|dhu{~lLw7��m\rulLz � t8¶:h(K:�TQo£ � y � �E�L�:�U�U�B� �L�%� � �L�L� ¹ � �E�L�L½

m cUtHg|g~¸UdEfUz �]�L� ¨�n�£�� cUt]g~g~¸<dfUzßt�x ; IqÆPlEr�x:wUtHh(K:� � ¥7tÞ»'y � �Ärut]��]� ��L� ¨:��t]tèdEg~��l¾��¼�î ;¥Yh�hu� ; �� p�p�p�� eHlL�L��eH{·� �:ru{|w7e]tHhulEwÈ� t�x7f �� p�w �

24

AIFB AIFB

��

m »Fdsr�hu{>dEwUxFn%e8¥%f7hu�Ht�n¤»Fdsr�hu{ � �s¦�t�dsru� h��¦�{~w7ru{~e�¥@n%e8¥%f7hu�Ht ; I�£Xf7� hulLz¤{~�]{~wU��d�îÈt8¶7{~e]lEw�hul@oXtHh��hutHr�n%fU{ih�dv£XlLz¤�7f7h�dshu{|lEw<dg � dE��!K:�7{~w�£XlEru�7fU��rulYe]t]��{~w7�vm�lr�îÈt8¶7{~e�dg � e t fU{~��{ihu{~lLw�t�x:{~hut]x¸Yj�oRr�dEwU{~z¤{irQoXlE�Lf7r�dt]�_� � dEz¤t]�_�ÄfU� hutk�kl�%��YjY�7�U�È� ªEªs� � «

m ¦�dE¥Uw ��E� ¨�n��x7l�¦�dE¥UwB��q�g~t]z¤t]w7�>n%e8¥7w<dsh�hu{|w7�LtHr%��I��w~�y��O��z'�r�'��#�<��#��{G�O�(��y%{G�!��K:�� y � ¨:�:y �� yq� � ¨��7rulYe]tHt�x7{~wU�E�]�<�7�È� �L� §E� �E�%� ¹ ��L� ¨S½

m c<df7rut�dw<x ��t�x7tHg|g~t]e �]�L�E� n�¢�d��%{|xºc<dEf:rutL�!£XgÀd{irut ��t]x7t]g~g~t]eL��I�q�wUl�p�g~t�x:�Lt � e t fU{~��{ihu{~lLwºlm�Ärut]x7{~e�dhut � ru�EfUz¤t]wYh�nYh�rufUeHhuf:rut]��m\rulLz � t]e�¥Uw7{~e�dEg � tH¶%hu��{~wU�2»FdEe�¥U{~w7t îBt�druw7{|w7� ; � ¥7tn%j:� hut]z � n:yq��»�K7��Q<q � Æéì �L�

m n%lLwUl@dEw<xpv!dz�dhut �]�L�E� n�q�d�]f<d�:{7n%lLwUl:�E»FdE�ud��{ vXdEz�dhut ; ��w7{~hut]x � dhu{~lLw7�>e]lLw%�StHw%hu{~lEwvlEw£XlLwYh�r�dEe8hu��m\lEr_hu¥Ut/yqwYhutHruw<dshu{|lEw<dg>n:dEg~t�lEmoË�lYl%x:�]�[n:tH{~ru{~w:� n%¥UlE{|wÈ¹ ��E�L� ½

25

AIFB AIFB

Ontology Discovery for the Semantic Web Using Hierarchical Clustering

Patrick Clerkin, Pádraig Cunningham, Conor Hayes

Department of Computer Science Trinity College Dublin

[email protected]@cs.tcd.ie

[email protected]

Abstract. According to a proposal by Tim Berners-Lee, the World Wide Web should be extended to make a Semantic Web where human understandable con-tent is structured in such a way as to make it machine processable. Central to this conception is the establishment of shared ontologies, which specify the fundamental objects and relations important to particular online communities. Normally, such ontologies are hand crafted by domain experts. In this paper we propose that certain techniques employed in data mining tasks can be adopted to automatically discover and generate ontologies. In particular, we focus on the conceptual clustering algorithm, COBWEB, and show that it can be used to generate class hierarchies expressible in RDF Schema. We consider applica-tions of this approach to online communities where recommendation of assets on the basis of user behaviour is the goal, illustrating our arguments with refer-ence to the Smart Radio online song recommendation application.

1 Introduction

Tim Berners-Lee has proposed and extension to the existing World Wide Web known as the Semantic Web (Berners-Lee, 1998, Berners-Lee et al, 2001). Most of the Web’s existing content is designed to be read and understood by humans, and cannot readily be parsed and processed by software agents. The central idea behind the Semantic Web is to develop and use machine-understandable languages for the expression of the semantic content of Web pages. This promises to enhance the ability of software agents to navigate the Web’s information space and carry out tasks for humans, with-out the need for sophisticated artificial intelligence.

Central to the Semantic Web project is the concept of an ontology. Web pages are conceived of as being composed of statements relating objects. The denotations of the terms making up the statements need to be fixed relative to a particular universe of discourse, which is represented in an ontology. The ontology codifies a shared and common understanding of some domain. An ontology is usually constructed by domain experts. In this paper, we examine the possibility of generating ontologies automatically using hierarchical conceptual clustering, and consider certain online

27

communities where such methods are highly appropriate, since there is no existing conceptualisation of the site resources. It is important to emphasis at this point that we are concerned with generating ontologies from behavioural and usage data relating to resources of interest, rather than from the free text data that might be found on web pages.

Since we aim to demonstrate how this technique may be practically implemented, we first present an overview of some of the technologies being used to build the Semantic Web, focusing in particular on how basic ontologies can be represented using RDF Schema (Brickley and Guha, 2000). In subsequent sections we discuss the concept formation system, COBWEB (Fisher, 1987), and demonstrate how the concept hierarchies discovered by this algorithm can be represented as ontologies with RDF Schema. We conclude with a discussion of the application of this approach to online communities - dealing in particular with the Smart Radio system (Hayes and Cunningham, 2000), developed as a test bed for our ideas - and point to some further research directions.

2 Implementing the Semantic Web

The Uniform Resource Identifier (URI)1 provides the foundation for the Web, since it allows us to give any object or concept a uniquely identifying name. URIs are decentralized, in the sense that no one person or organisation controls their defintion and use. Since anyone can create a URI, we inevitably end up with multiple URIs representing the same thing, so it is important for the Semantic Web to provide a means for resolving names correctly.

This is provided for in the use of an ontology, which usually takes the form of a taxonomy defining classes and relations among them. The meaning of terms can now be resolved if they point to particular ontologies, and if equivalence relationships are defined between ontologies.

The Resource Description Framework (RDF)2 provides a means for software agents to exchange information on the Web. It defines a simple model for describing relationships between Web resources in terms of properties and their values. However, RDF itself provides no means of declaring such properties. This task is left to RDF Schema, which can be used to represent simple ontologies.

RDF can be written using XML tags, but it is important to note that XML is not sufficient for building the Semantic Web. XML facilitates the arbitrary creation of tags that can be used to annotate Web pages. If a programmer knows in advance what these tags signify, then it is possible for her to write software to process these Web pages automatically. However, in the absence of such knowledge, it is not possible to write such programs, since XML builds no semantics into its structures. RDF, on the other hand, encodes machine-processable structures into its statements. An RDF

1 See http://www.w3.org/Addressing/ for an overview of naming and addressing schemes used

on the World Wide Web. 2 See http://www.w3.org/RDF/.

28

statement consists of a triplet, which asserts that a particular thing has a certain prop-erty with a certain value. For example, the sentence

Ora Lassila is the creator of the resourcehttp://www.w3.org/Home/Lassila.

can be represented in RDF by

<?xml version="1.0"?><rdf:RDF

xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"xmlns:s="http://description.org/schema/"><rdf:Description

about="http://www.w3.org/Home/Lassila"><s:Creator>Ora Lassila</s:Creator>

</rdf:Description></rdf:RDF>

The RDF Schema defines a collection of RDF resources that can be used to

describe properties of other RDF resources that define application-specific RDF vocabularies. The RDF Schema type system is similar to the type systems of object-oriented programming languages such as Java. For the purposes of this paper, it is sufficient to present this system in the context of an example. Consider the following class hierarchy. We first define a class MotorVehicle. We then define three subclasses of MotorVehicle, namely PassengerVehicle, Truck and Van. We then define a class Minivan which is a subclass of both Van and PassengerVehicle. In representing this hierarchy we must make use of some core classes and properties defined by RDF Schema. In particular: all things being described by RDF expressions are called resources, and are considered to be instances of the class rdfs:Resource; when a resource has an rdf:type property whose value is some specific class, we say that the resource is an instance of the specified class; when a schema defines a new class, the resource representing that class must have an rdf:type property whose value is the resource rdfs:Class; the rdfs:subClassOf property specifies a subset/superset relation between classes. Our example class hierarchy is therefore represented by the following diagram:

29

Fig. 1 Class Hierarchy for MotorVehicle class and its subsets (Brickley and Guha, 2000)

This model can be rendered in XML as follows: <rdf:RDF xml:lang="en"

xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"

xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#">

<rdf:Description ID="MotorVehicle"><rdf:type resource="http://www.w3.org/2000/01/rdf-

schema#Class"/><rdfs:subClassOfrdf:resource="http://www.w3.org/2000/01/rdf-

schema#Resource"/></rdf:Description>

<rdf:Description ID="PassengerVehicle"><rdf:type resource="http://www.w3.org/2000/01/rdf-

schema#Class"/><rdfs:subClassOf rdf:resource="#MotorVehicle"/>

</rdf:Description>

<rdf:Description ID="Truck"><rdf:type resource="http://www.w3.org/2000/01/rdf-


</rdf:Description>

30

<rdf:Description ID="Van"><rdf:type resource="http://www.w3.org/2000/01/rdf-


</rdf:Description>

<rdf:Description ID="MiniVan"><rdf:type resource="http://www.w3.org/2000/01/rdf-

schema#Class"/><rdfs:subClassOf rdf:resource="#Van"/><rdfs:subClassOf rdf:resource="#PassengerVehicle"/>

</rdf:Description>

</rdf:RDF>

3. Hierarchical Conceptual Clustering

Data mining is the extraction of implicit, previously unknown, and potentially useful information from data. The objective is to build computer programs that automatically detect regularities or patterns in databases. Useful patterns, if found, should generalise to make accurate predictions on future data. Thus, the final objective of data mining activity is knowledge discovery.

Machine learning provides the technical basis of data mining. It is used to extract information from the raw data in databases. The process is one of abstraction in order to find patterns. Usually, we also require that the system should provide us with an explicit structural description, so as to provide the observer with an explanation of what has been learned and an explanation of the basis for new predictions.

Clustering is a data-mining task that has at its goal the unsupervised classification of a set of objects. Classification is unsupervised in the sense that there are no a priori target classes used during training. Clustering techniques rely on the existence of some suitable similarity metric for objects. Clustering algorithms may be classified according to a number of criteria. Some are distance-based and describe clusters purely by enumerating their members, while others represent the clusters by means of a description. This description may take the form of a set of necessary and sufficient conditions for membership of a given cluster, or it may be a probabilistic description where no such set of conditions is tenable. Furthermore, a set of clusters may be “flat” in the sense that no cluster is “contained” in any other cluster, or it may be hierarchical, providing a taxonomy of clusters with definite relationships between them.

COBWEB is an incremental conceptual clustering algorithm with represents concepts probabilistically. It was initially inspired by research on basic level effects. For example, humans can typically verify that an item is a bird more quickly than they can verify the same item is an animal, vertebrate, or robin. Thus, the concept of birds is said to reside at the basic level. COBWEB’s design assumes that principles which dictate basic concepts in humans are good heuristics for machine concept formation as well.

31

COBWEB is designed to produce a hierarchical classification scheme. It carries out a hill-climbing search - which consists of taking the current state of the search, expanding it, evaluating the children, selecting the best child for further expansion, etc, and halting when no child is better than its parent – through a space of schemes, and this search is guided by an heuristic measure called category utility. The category utility metric was originally developed by Gluck and Corter (1985) to predict the basic level in human classification categories. In adopting it as a criterion for evaluating concept quality in AI systems, Fisher notes that it can be viewed as a function that rewards traditional virtues held in clustering generally – similarity of objects within the same class, and dissimilarity of objects in different classes.

n

VAPCVAPCPCCCCU

k i jjikijik

n

∑ ∑∑

=−=

=

22

21

)()|()((}),...,,({

(1)

COBWEB performs its hill-climbing search of the space of possible taxonomies and uses category utility to evaluate and select possible categorisations. It initialises the taxonomy to a single category whose features are those of the first instance. For each subsequent instance, the algorithm begins with the root category and moves through the tree. At each level it uses CU to evaluate the taxonomies resulting from:

1. Classifying the object with respect to an existing class. 2. Creating a new class. 3. Merging: combining two classes into a single class. 4. Splitting: dividing a class into several classes.

4. Ontology generation using COBWEB

We propose that COBWEB may be used to automatically generate ontologies. Let us consider the following artificial and simple example. We have a domain consisting of only four resources, namely, the following four cells:

Fig. 2 The four cells to be clustered by COBWEB. (Gennari et al, 1989, Luger and Stubblefield, 1998)

COBWEB generates the following hierarchy of concepts:

32

Fig. 3 The concept taxonomy produced by COBWEB. (Gennari et al, 1989, Luger and Stubblefield, 1998)

Just as in our previous example, we can represent this hierarchy in RDF Schema

with a diagram:

33

Fig. 4 The Class Hierarchy corresponding to the COBWEB concept taxonomy

The actual XML looks like this:

<rdf:RDF xml:lang="en"xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-

ns#"xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#">

<rdf:Description ID="C1"><rdf:type resource="http://www.w3.org/2000/01/rdf-

schema#Class"/><rdfs:subClassOfrdf:resource="http://www.w3.org/2000/01/rdf-

schema#Resource"/></rdf:Description><rdf:Description ID="C2"><rdf:type resource="http://www.w3.org/2000/01/rdf-

schema#Class"/><rdfs:subClassOf rdf:resource="#C1"/>

</rdf:Description><rdf:Description ID="C3"><rdf:type resource="http://www.w3.org/2000/01/rdf-


</rdf:Description><rdf:Description ID="C4">

34

<rdf:type resource="http://www.w3.org/2000/01/rdf-schema#Class"/><rdfs:subClassOf rdf:resource="#C1"/>





</rdf:Description></rdf:RDF>

5. Applications

Now we turn to the question of an application for this approach to ontology generation. We will discuss this in the context of Smart Radio, which is a web-based song recommendation system that relies on users' ratings of songs. The user builds playlists consisting of ten songs that they choose from the database. They may then listen to the songs and rate them on a scale of one to five, where one indicates a strong dislike for song, and five indicates a strong liking.

Fig. 5 An example Smart Radio playlist showing the songs rated by the user.

This results in something like the following matrix of values:

35

Table 1. Smart Radio data showing how users have rated songs.

Song 1 Song 2 Song 3 Song 4 Song 5 Song 6 User 1 5 2 4 2 User 2 2 5 1 2 5 User 3 1 1 5 3 User 4 4 2 User 5 3 1 5 5 User 6 4 1 1 5 3 User 7 1 1 User 8 3 1 4 1 5

The Smart Radio system currently relies solely on Automated Collaborative Filter-

ing (ACF) to make recommendations. We are experimenting with using knowledge discovery techniques to enhance the quality of these recommendations. In particular, we have employed the COBWEB algorithm to generate a hierarchy of concepts. To do this, we define a song to be good for a user if and only if she gives that song a rating of 4/5 or more; below this, the song is bad for the user. Then, we characterise each song in the database as an object with the same number of attributes as there are users in the database. Each attribute can take on the value good or bad according to how the user rated the song. The following is an example song object, where question marks symbolise missing values: good bad ? ? good ? good ? ? ? ? ? ? ? good ? ? ? ? ? ?? ? good ? ? ? ? ? good bad ? ? ? ? good good ? ? ? bad? ? bad ? bad good ? ? ? bad ? bad ?

When COBWEB is run on the complete set of these objects, we acquire a hierarchy

of clusters and associated probabilistic descriptions. An example cluster, arbitrarily named ‘C112’, is presented below:

['Break On Through (To The Other Side)', 'Mr Tambourineman', 'Say hello wave goodbye', 'Hallelujah', 'Unfin-ished Sympathy', 'Where The Wild Roses Grow', 'PleaseForgive Me', 'The Girl From Ipanema', 'Gin Soaked Boy','Street Spirit (Fade Out)', "La Femme D'Argent", 'RightHere, Right Now', 'Redemption Song']

We can then go on to translate the output of COBWEB into a class hierarchy that

can be rendered in RDF Schema as outlined above. The advantages of this approach to this sort of domain is best discussed in light of

the fundamental goal of the Semantic Web project, which is to create ‘an environment where software agents roaming from page to page can readily carry out sophisticated tasks for users’ (Berners-Lee et al, 2001). In the context of an online music commu-nity, one of these tasks is going to be the recommendation of songs to users. A user might, for example, request that an agent finds her new songs similar to those she has liked in the past. Or she might request that new songs be provided which are some-

36

what dissimilar to those that she has previously encountered, but which she may nevertheless like. An ontology facilitates the fulfilment of these requirements, because similar songs will fall under the same concept, and degrees of similarity/dissimilarity will hopefully be captured in the relationships between concepts.

The usual way of creating an ontology is for domain experts to establish the fun-damental concepts, objects, relations, etc, which exist for a given community. This presupposes that these ontological elements can be uncovered a priori. However, in domains such as that of Smart Radio, it is not at all clear that any a priori analysis by a team of experts could yield the sort of concepts important to recommendation tasks. While songs may be categorised according to artist, and while to a much lesser extent, genres and sub-genres may be employed, this approach is inadequate, since it does not account for the fact that many people like songs that are widely divergent according to the artist and genre criteria. By using such algorithms as COBWEB to cluster songs based on user ratings, we hope to discover structures more truly reflective of the simi-larities and dissimilarities between songs. We need only evaluate the resulting concep-tual structures in terms of their impact on recommendations, and we need not worry that users may be unable to articulate the hypothesised perceived similarities and dis-similarities between songs. Furthermore, we do not expect that the discovered concep-tual hierarchy will map onto any existing and already familiar network of human con-cepts. Rather, we expect to discover structures that it was never feasible for human experts to detect.

A further advantage conferred by the automatic generation of ontologies using COBWEB and related systems is that such concept formation algorithms are incre-mental, in the sense that observations are not processed en masse. There is a stream of objects, which is processed over time. In the case of Smart Radio, this means that the conceptual hierarchy can automatically evolve over time as new songs are added to the database, and as new users join the system. This is something that would be very costly if human experts were involved, and yet such a capacity to evolve over time is essential to a constantly expanding online community resource.

Finally, we may wonder at how such automatically generated ontologies, which do not map onto any existing human understandable ontologies, can fulfil the requirement for interoperability across web sites. So far, we have considered how one online com-munity can be held together and enhanced by such ontologies, but we now turn to a consideration of how other online communities with similar assets (in this case, songs) could exchange information, via agents, with our site. In traditional ontology engineer-ing, collaboration is required between the people who run Web sites and online com-munities. It is no different in the case where we are employing automatic ontology generation techniques. The only difference is that it is not human beings who collabo-rate, but, rather, machines. If the Smart Radio database is accessible to a COBWEB-based agent, and another, different, database of songs and user ratings from a hypo-thetical Smart Radio II, is also accessible to the same agent, then there would be no problem in that agent constructing, maintaining, and evolving a shared ontology for both sites. The only limitation is that the agent must be able to understand the struc-ture of all such databases. Collaboration requires a set of standards and conventions for the construction or description of such databases, and such collaboration is entirely within the spirit of the Semantic Web project.

37

Summary and conclusion

We have discussed how hierarchical clustering algorithms may be employed to auto-matically construct basic ontologies, and illustrated this in the context of COBWEB and RDF Schema. We have argued that such an approach is highly appropriate to domains where no expert knowledge exists, or where it proves inadequate, and have gone on to propose how we might employ software agents to collaborate, in place of human beings, on the construction of shared ontologies. This benefits recommenda-tion tasks, in particular, by allowing for the evolution of concept hierarchies which do not match any articulated human conceptual structures, but which are, hopefully, closely reflective of the criteria that people employ in rating online assets. If the task of Semantic Web project is to render human understandable resources processable by machines, we might say that the task envisaged here is to extend the resources proc-essable by machines beyond the domain of human understanding – but always with a view to helping humans carry out their online tasks.

References

1. Berners-Lee, T. (1998). Semantic Web Road map. Available online at http://www.w3.org/DesignIssues/Semantic.html.

2. Berners-Lee, T., Hendler, J., Lassila, O. (2001). The Semantic Web. Scientific American feature article, May 2001. Available online at http://www.scientificamerican.com/2001/0501issue/0501berners-lee.html.

3. Brickley, D., Guha, R.V., (Eds.) (2000). Resource Description Framework (RDF) Schema Specification 1.0. W3C Candidate Recommendation 27 March 2000. Available online at http://www.w3.org/TR/rdf-schema/.

4. Fisher, D. H. (1987). Knowledge acquisition via incremental conceptual clustering. Machine Learning, 2, 139-172.

5. Gennari, J.H., Langley, P., Fisher, D. (1998). Models of incremental concept for-mation. Artificial Intelligence, 40(1-3):11-62.

6. Gluck, M.A., Corter, J.E. (1985). Information, uncertainty, and the utility of cate-gories. Proceedings of the Seventh Annual Conference on Artificial Intelligence, pp. 831-836, Detroit, MI: Morgan Kaufmann.

7. Hayes, C., Cunningham, P. (2000) Smart Radio: Building Music Radio on the Fly. Expert Systems 2000, Cambridge, UK, December 2000.

8. Luger, G.F., Stubblefield, W.A., (1998). Artificial Intelligence: Structures and Strategies for Complex Problem Solving. Third Edition. Addison-Wesley, 1998.

38

Web Directories as Training Data forAutomated Metadata Extraction

Martin Kavalec, Vojtech Svatek and Petr Strossa

Department of Information and Knowledge EngineeringUniversity of Economics, Prague, 13067 Praha 3, Czech Republic

e-mail: [email protected], [email protected], [email protected]

1 Introduction

Although man-made annotations are considered as the main ‘knowledge fuel’ forthe Semantic Web, the majority of existing commercial pages are still poorlyequipped with any kind of metadata, never mind the forthcoming standardssuch as the RDF syntax or the Dublin Core semantics. Information Extraction,relying on characteristic patterns in text, can be applied even on such ‘legacy’pages, in order to obtain metadata containing, for example, the names, types,and domains of activity of the WWW subjects (companies).

The two decades of development of Information Extraction techniques haveshown that extraction patterns applicable on real-world, unstructured text datacannot be satisfactorily prepared by hand; instead, Machine Learning (ML) be-came the enabling technology. The common assumption in ML-based Informa-tion Extraction is that the training cases are chunks of text pre-labelled by ahuman indexer. This is acceptable for domain-specific resources with limitedvocabulary and more-or-less conventional structure, such as computer sciencedepartment pages [2] or housing advertisement pages [7]. However, if we proceedto broad categories such as ‘pages of companies offering products or services’, thenumber of training cases needed will explode, the acquisition of text fragmentsbecomes difficult, and their manual labelling simply infeasible. Yet, there is apromising resource of web data that has already undergone a process of humanindexing, of a sort: web directories such as Open Directory or Yahoo!

In this paper, we analyse the possibility of reusing the knowledge embeddedin the structure of the directories in order to obtain labelled training data forWeb Information Extraction with limited human effort. In section 2 we showthe results of preliminary experiments consisting in mining the fragments of webpages, obtained with the help of web directory information, for indicator termsusable for subsequent extraction of semantic information from other pages. Insection 3 we outline an ontology of web directories, and suggest the way it canbe used to refine the above process. In section 4, our approach is compared tosome other projects. Finally, in section 5, we summarise our plans for the future.

39

2 Mining Indicator Terms through Directory Headings

Our assumption is that the directory headings (such as ... /Manufacturing/Materials/Metals/Steel/...) coincide with the generic names of productsand services—let us nickname them informative terms in this paper—offered bythe owners of the pages referenced by the respective directory page. By matchingthe headings with the page fulltexts, we obtain sentences that contain the infor-mative terms. The terms situated near the informative terms in the structure ofthe sentence are candidates for indicator terms, provided they occur frequentlyon pages from various domains. The resulting collection of indicator terms can,conversely, play the role of ‘extraction patterns’ for discovering informative termsin previously unseen pages.

The knowledge asset embedded in web directories is the judgement of humanindexers who have assigned the pages under the particular heading(s). Naturally,informative terms on the page need not always correspond to the existing direc-tory headings, e.g. due to synonymy. As consequence, our method will extract(without the help of a thesaurus) only a fraction of the sentences with informa-tive terms. This however does not disqualify the method, since, in this trainingphase, we aim at discovering indicator terms rather than at identifying the in-formative terms themselves. The small degree of completeness of the method isactually compensated by the hugeness of the material available1 in the direc-tories. Namely, the ‘Business’ subhierarchy of Open Directory (www.dmoz.org),which we have exploited in our experiments, points to approx. 150,000 pagesoverall, each of these containing the ‘heading’ terms (from the referencing nodeor one of its ancestors) in two sentences, on the average.

We have tested the training phase of our method on a sample of 14,500sentences2 containing the ‘heading’ terms. The syntactical analysis has beencarried out using the Link Grammar Parser [6]. The verbs which occurred theclosest (in the parse tree) to informative terms have been counted, and arrangedinto a frequency table. In Table 1, the essence of the table is shown, mostlyfeaturing verbs that are likely to be associated with the informative terms (e.g.‘our assortment includes...’, ‘we manufacture...’, ‘in our shop you can buy...’).The table contains only the verbs that occurred in at least 50 sentences3. Wehope to build a more comprehensive collection using a larger sample of pages.Furthermore, the plain verbs (in particular ‘to be’, which has no significanceof its own) can be extended to more complex phrases, again via selecting theneighbouring terms with frequent occurrence.

1 As we dispense with manual labelling, processing a larger sample of data is merelythe matter of computer time/storage.

2 I.e. about 5% of the total of such sentences.3 In this display, they are however not arranged according to the relative frequencyof occurrence in the neighbourhood of the informative term (Pn), but according tothe ratio of this frequency to the relative frequency of occurrence in the whole of theextracted, possibly compound, sentence (Ps). This visibly pushes down the universalverbs such as ‘to be’.

40

Pn/Ps Verb Pn Ps2.23 includes 0.0048 0.00212.20 manufacture 0.0038 0.00172.17 buy 0.0038 0.00172.09 including 0.0057 0.00272.08 supply 0.0036 0.01751.96 offers 0.0119 0.00601.92 provides 0.0135 0.00701.92 offer 0.0200 0.01041.89 include 0.0062 0.00321.85 specializing 0.0051 0.00271.78 providing 0.0091 0.00511.70 specializes 0.0045 0.00261.66 specialize 0.0053 0.00321.56 using 0.0037 0.0024

Pn/Ps Verb Pn Ps1.55 provide 0.0191 0.01221.40 use 0.0046 0.00331.39 sell 0.0039 0.00281.26 see 0.0046 0.00361.25 are 0.0740 0.05891.24 were 0.0042 0.00341.23 made 0.0040 0.00321.22 make 0.0066 0.00541.15 need 0.0053 0.00461.12 is 0.0988 0.08801.05 get 0.0043 0.00411.03 find 0.0060 0.00581.03 meet 0.0043 0.00411.00 related 0.0035 0.0035

Table 1. Frequent verbs in sentences containing the headings

3 Ontology of Web Directory Headings

As we have shown in the previous section, interesting pieces of information canbe extracted from web directories even without specific assumptions about thenature of the heading terms. Nevertheless, we believe that only deeper ontologicalanalysis of the headings can bring the automated discovery of indicators to itsfull potential, in particular for complex terms spanning across multiple levels ofheadings. We will thus now outline an ontology of web directory headings.

The semantic information associated with the particular page, in the contextof a web directory, is defined by the sequence (or, several sequences, in the caseof a non-tree hierarchy) of headings preceding the node pointing to that page.The headings essentially belong to one of the following classes:

1. ‘Entity’ terms (most often nouns), which correspond to real-world entities.Note that their meaning may depend on the preceding terms: for example,pages referenced by the node preceded by the subpath Cranes/Accessoriesare likely to offer accessories for cranes but not clothing accessories. Never-theless, the word ‘accessories’ can possibly be found on the page even withoutthe attribute ‘for cranes’, since the latter can be assumed by context.

2. ‘Property’ terms (most often adjectives), which correspond to properties ofentities. They are restrictive rather than descriptive, since they (usually)restrict the scope of the immediately preceding ‘entity’ term to denote anarrower class of entities. For example, the pages referenced by the node pre-ceded by the (sub)path Telecommunications/Wireless can be viewed as‘indexed’ by the compound term ‘wireless telecommunications’. The ‘prop-erty’ term is completely dependent on the given ‘entity’ term, seeking itindependently on a page would be spurious.

41

The ‘entity’ terms can be further refined to:

1. Subjects (active entities) such as Manufacturers, Publishers, orAssociations.

2. Objects (passive entities) such as Materials, Aircraft or Textiles.3. Domains of activity such as Telecommunications or Publishing.

In addition, we can identify a common subclass of both object and domain,which can be denoted as activity. An activity is a domain, since it fits intothe generality hierarchy of domains, but it is also an object, since it can beviewed as a ‘commodity’ offered by a certain subject, e.g. Manufacturing orConstruction. Furthermore, a distinct feature of an activity is the aptitude ofbeing applied on an (other) object.The diagram at Fig. 1 depicts the essence of the ontology. Boxes correspond

to classes, full edges to named relations, and dashed edges to the class-subclassrelationship. Reflexive binary relations are listed inside the respective boxes.

S applies A on O

Property

Subject

is-a

Object

is-ais-part-of

Domain

is-subdomain-of

Activity

is-subactivity-of

Entity

P restricts E1 to E2

S is active in D

D is destination for O

S acts on O

Fig. 1. The ontology of web directory headings

Given the ontology, the method described in section 2 can be enhanced inthe following way:

1. Select the most promising paths and nodes in the directory structure.2. Assign class labels to the headings from the selected paths, and arrange theminto a semantic network of relations.

3. Generate full-text queries based on the headings (and their classes), andapply them on the pages to extract sentences.

42

4. The mining of indicator terms could then be done separately for each classof ‘entity’ terms, thus obtaining collections of ‘extraction patterns’ specificfor different type of information to be extracted.

The structure over the headings should enable to generate4 finer queries,and thus to obtain a more comprehensive sample of training data for indicatorlearning. As an example, given the path segment Cranes/Accessories, in whichboth Cranes and Accessories were pre-classified as ‘objects’, and Cranes iden-tified as ‘destination’ for Accessories, the training cases containing e.g. theexpression ’accessories for cranes’ might me the most desirable ones.

The importance of step 1 (selection) follows from the fact that step 2 (la-belling) has to be done manually, but, in distinction to ‘classical’ labelling, asingle labelling action can lead to class (or, relation) assignment to several train-ing cases. The efficiency of manual labelling is thus closely related to the numberof pages referenced by the node being labelled, as well as to the number of sen-tences (from these pages) containing the headings. In order to obtain a high‘assignments-to-actions’ ratio, we have to trade off the high number of pages(for general headings close to the root, pertaining to huge subhierarchies) withthe higher number of sentences per page (for the headings close to the leaves,which are better tuned to the page content, and even subsume several ‘ancestor’terms). The parameters of the respective utility function could be determined inthe future.

4 Related Work

The common approach to overcome the lack of classified training examples intext categorisation is to apply statistical techniques consisting in iterative auto-mated labelling of unclassified examples based on a few classified ones (boot-strapping, see [1], [4], [5]). So far, we have not considered such techniques, andinstead rely on the prior work of a human indexer of the web directory. Whiledirectories have already been used for learning to classify whole documents [3],their use for information extraction seems to be rather innovative.

Our work is actually rather similar to Brin [1], which targets on automateddiscovery of extraction patterns using search engines. The patterns can be usedto find relations, such as books, i.e. pairs (author, title). The patterns are basedsimply on characters surrounding the occurence of investigated relation. In com-parison, we aim at finding less structured information, for which such simple pat-terns wouldn’t be sufficient; we therefore search for linguistic indicators, whichare based on syntax analysis. (The indicators themselves can be thought of as‘syntactic patterns’.)

4 We are currently working on a rewriting grammar that will automatically convertthe set of relational expressions on headings into a layered set of query terms.

43

5 Conclusions and Future Work

We have suggested a novel method for learning indicative terms, which can be,in turn, used to extract important terms (in fact, meta-data) from web pages.The source of learning cases is a web directory: thanks to the prior work ofhuman indexers of the directory, the burden of manual case labelling is eithercompletely removed, or significantly reduced. Preliminary results in a ratherrestricted setting suggest that the method may be viable.

As we have mentioned in the end of section 2, we will soon extend the non-interactive method of mining the indicators by searching forth in the parse trees,beyond the neighbouring verb (in particular for the ‘unclear’ verbs). The accu-racy on unseen pages also has to be thoroughly tested. Furthermore, the prospec-tive use of the web directory ontology has been described in section 3.Finally, we anticipate that best results could be obtained by combining our

reuse of human effort (with rather precise but incomplete results) with bootstrap-ping techniques mentioned in section 4 (more complete but possibly imprecise),in a more distant future.

The research has been partially supported by the grant no. 201/00/D045(Knowledge model construction in connection with text documents) of the GrantAgency of the Czech Republic.

References

1. Sergey Brin. Extracting Patterns and Relations from the World Wide Web. In:WebDB Workshop at EDBT’98.

2. Dayne Freitag. Information Extraction From HTML: Application of a GeneralLearning Approach. In Proc. 15th National Conference on Artificial Intelligence(AAAI-98).

3. Dunja Mladenic. Turning Yahoo into an Automatic Web-Page Classifier. In Pro-ceedings of the 13th European Conference on Aritficial Intelligence, ECAI’98, pp.473-474.

4. Andrew McCallum and Kamal Nigam. Text Classification by Bootstrapping withKeywords, EM and Shrinkage. In ACL’99 Workshop for Unsupervised Learning inNLP, 1999.

5. Ellen Riloff and Rosie Jones. Learning Dictionaries of Information Extraction byMulti-Level Bootstrapping. In Proc. 16th Nat. Conf. Artificial Intelligence (AAAI-99).

6. Daniel Sleator and Davy Temperley: Parsing English with a Link Grammar. InThird International Workshop on Parsing Technologies, August 1993.

7. Stephen Soderland. Learning Information Extraction Rules for Semi-Structured andFree Text. Machine Learning Vol. 34, 1999, pp.233-272.

44

Multiagent Cooperative Learning of UserPreferences

Adorjan Kiss and Joel Quinqueton

LIRMM, Multiagent Team, 161 rue Ada, F-34392 Montpellier Cedex, France{kiss,jq}@lirmm.fr,

WWW home page: http://www.lirmm.fr/~kiss

Abstract. We present in this paper a Machine Learning method de-signed to predict preference knowledge in a multi-agent context. In thefirst part we show some theoretical properties of our learning scheme andthen present an application of it to a corporate knowledge managementsystem.

1 Introduction

In this paper we will present some attempts to design a Machine Learningmethod to predict preference knowledge in a multi-agent context. Here we definepreference knowledge as knowledge about a preference between elements of a set.

For instance, the documents found by a search engine on the web are orderedaccording to a preference function computed from the user request. Thus, theycan be considered as ordered according to a preference relation.

The framework that gave birth to this work is a joint research project,CoMMA1, dedicated to corporate memory management in an intranet. Themain objective of the project is to implement and test a Corporate Memorymanagement framework integrating several emerging technologies in order tooptimize its maintenance and ease the search inside it and the use of its contentby members of the organization.

The main challenge is to create a coherent system that relies upon severalpromising new technologies which are in the middle of their struggle to becomestandards:

– Multi-agent architecture: it is well suited to the heterogeneity of the Cor-porate Memory; its flexibility eases the system maintenance and keeps therhythm with the dynamics and evolution of the Corporate Memory; coop-erating and adaptive agents assure a better working together with the userin his pursuit to more effectively achieve his goals. The FIPA standard, sup-ported by the CoMMA project, offers the specifications for interoperableintelligent multi-agent systems.

1 This work was supported by the CoMMA (Corporate Memory Management throughAgents) project [Con00] funded by the European Commission under Grant IST-1999-12217, which started beginning of February 2000.

45

2

– XML: is a standard recommended by the World Wide Web Consortiumintended to offer a human and machine understandable description language:a good choice if it is important to ensure an easy maintenance, and seamlessflow through various information processing systems that are expected toevolve in time.

– RDF/RDFS another W3C recommendation, that creates a semantic levelon the top of XML formal description. RDF annotations allow having anintegrated, global view of the Corporate Memory keeping untouched (interms of storage, maintenance) the heterogeneous and distributed nature ofthe actual info sources. RDF also allows us to create a common ontology torepresent the enterprise model. The ontological commitment, a fundamentalchoice in our approach to design the Corporate Memory, is motivated by ourbelief that the community of corporate stakeholders is sharing some commonglobal views of the world that needs to be unified and formalized (RDFS)to form the basis of the entire information system.

– Machine Learning Techniques make the system adaptive to the user, andcome even more naturally due to the previous choices, as presented in thefollowing section.

2 The use of Preference knowledge

Here we define preference knowledge as knowledge about a preference betweenelements of a set. Such knowledge can be stated in various forms: a numericalvalue assigned for each item, a total ordering relation, a partial ordering relationor even a preordering of the set.

Logical models have been proposed to deal with such knowledge, some dealingdirectly with the comparison abilities [Sch96] and others inspired from discretelinear-time temporal logics [SR99].

It is generally admitted that a preference stands for an order that maybepartial, even a preorder, but that it is often convenient to represent it by alinear extension (which is a total order) or a numeric value compatible with theknown orderings.

Then, in terms of Machine Learning, different strategies may be used, de-pending on the form of the preference knowledge.

2.1 Numeric labelling

A numeric labeling, i.e. a mapping of our set of examples into a set of realnumbers, is a convenient way to summarize a preference relation. Some MachineLearning methods are available to learn numerical variables [Gas89,Bre96b,Bre96a].

Generally, the methods for learning to predict a numerical variable v measurethe quality of a predictive rule R by the standard deviation σ2(v) of the valueof the variable among the set of objects verifying the concept R to be tested.

Q(R, v) =1

|R(x)true|

R(x)true∑x

(v(x) − v)2

46

AIFB AIFB

3

The lower Q(R, v) is, the better R is to predict v, because the mean value ofv can be used with less error. With such criteria, any learning method will leadto grouping several neighbour values around their mean. Then, the learnt ruleswill not be very different from rules learnt from examples roughly rated with afinite set of values.

2.2 The order relation

By definition, a binary relation, which we note /, is an order if it has the followingproperties:

– reflexive x / x,– transitive: if x / y and y / z, then x / z,– antisymmetric: if x / y and y / x, then x = y.

Then, we can imagine to learn the binary relation by learning each of itselements, that is, learn on each couple of objects (a, b) such that a / b. Then, letus summarize the suitable properties of such a learning set for this approach towork correctly.

First, if (a, b) with a / b is an example, then (b, a) is a counter-example.Then, what happens to (a, a)? We can see that they would be both examplesand counter-examples, then it is better to consider the strict order relation, andeliminate diagonal elements.

With these hypotheses, the description of an example is made of 2 parts: theattributes which are modified between a and b, and those which keep the samevalue. We can notice here that these attributes are the same as those involvedin the sorting tree of our examples.

Then, our method appears to be “half lazy”, in comparison with lazy learningmethods, like kNN or LWR [Aha92]. Our learned knowledge is partly explicit,but in the classification step, we need to compare a new instance with severalelements of the learning set (maybe in a dichotomic way) to put it in the rightplace.

2.3 Statistical evaluation criteria

Usually in Machine Learning, particularly for building of decision trees, thelearned classification rules are evaluated by their similarity to the desired clas-sification.

We can use the same principle here, and we have two possible families ofcriteria. If we can compute a rank for each element, the similarity is computedby measuring the rank correlation to the expected ranking. Otherwise, each pairmust be given: then we use a pairwise comparison between the expected orderand the learnt order.

Several measures of similatity between 2 different orderings of the same datahave been proposed. In each case, one has to deal with tied elements.

47

AIFB AIFB

4

The Spearman rank order correlation The Spearman rank order correlationrs is a correlational measure that is used when both variables are ordinal. Thetraditional formula for calculating the Spearman rank-order correlation is

Corr(r, r′) = 1 −6

∑i=ni=1 (ri − r′

i)2

n(n2 − 1)

where r and r′ are the ranks to compare of paired ranks. When there are tiedcases they should be assigned the mean of their ranks. The mean of the ranksfrom p+1 to p+n is 1

n ( (n+p)(n+p+1)2 − p(p+1)

2 ), which become after simplificationn+2p+1

2 .

The Kendall pairwise τ criterion When we have to compare each pair ofdata, they can be classified as either tied (T), concordant (P), or discordant (Q).

The best measure for this case is Kendall’s τb which takes into account acorrection for tied pairs. Its formula is

τb =P − Q√

(P + Q + Tx)(P + Q + Ty)

where Tx is the number of pairs tied on X but not Y, and Ty is the numberof pairs tied on Y but not X.

2.4 Verifying the consistency of the method

In the case we perform a pairwise learning of the order relation, we can noticethat a fundamental property, the transitivity, can be guaranteed by the learningprocess itself, as we show below for a version space method [Mit97].

We can check that, if, for 3 examples the transitivity holds, then it is notnecessary to add the 3rd pair as example to learn the relation :

– let (a, b) = (a1 . . . an, b1 . . . bn) and (b, c) = (b1 . . . bn, c1 . . . cn). Then, S is ofthe form L∧R, with the left part L as a generalisation of both a and b, andthe right part R of both b and c. Then, as L is a generalisation of a and Rof c, S is a generalisation of (a, c).

– with the same conventions, G has a disjunctive form whose elements rejectall the examples, then, if we represent any of its elements as L ∧ R. If (b,a)and (c,b) are rejected, it means that L rejects b or R rejects a, and L rejectsc or R rejects b. But G must also be a generalisation of S.

Of course, this is only a scheme of the proof, and is, strictly speaking, onlyavailable for version-space-like learning. In a more general case, like decision treelearning, we can only make the hypothesis that it is true. We concluded thatwe could learn directly a sorting rule (in a greedy way, like decision trees) andevaluate the obtained rule with the τ criteria defined in section2.3.

Let us now describe more widely our application.

48

AIFB AIFB

5

3 Using preference in Knowledge Management

This section aims to present and discuss how our work on preference learningfits into CoMMA project’s Knowledge Management System [Con00].

3.1 Getting the user profile

One of the advantages of an enterprise that should be exploited by such a corpo-rate information management system is that the users (i.e. the employees) canbe known (their domains of interest/competence, their current activities/tasks).This can be especially useful in some cases where users are likely to be over-whelmed by the quantity of information to process and navigate themselvesthrough (new employees during accommodation, technology monitoring scien-tists) who would appreciate personalized automated help in their process ofinformation retrieval. Nevertheless, using Machine Learning to reach this goalcan present some challenges. Some generic solutions are presented in [WPB01].

3.2 Using semantic annotations

Secondly, we have “human and machine understandable” semantic informationupon the corporate knowledge offered by the RDF formalization, based upon an“enterprise ontology” (RDF schema).

The combination of these two sources of information can provide a richground to infer knowledge about the users’ probable/possible preferences. Thiscombination is made possible due to the fact that we use the same RDF stan-dard for formalizing the user profile; the same base ontology for the enterpriseand user models.

It can be imagined that the info combined from these sources form sets ofattributes that will be used as input for an ML mechanism.

In order to set up such a ML mechanism, there are two main tasks to com-plete:

1. Getting and formalizing the information to be decomposed as attributes tofeed the ML mechanism.

2. Defining the ML methodology to process this info

3.3 Collecting the information to create a set of most meaningfulattributes

We will need to answer the following question: Why does a user prefer a docu-ment?

In our attempt to give an example of some possible answers, we are graduallygoing deeper and deeper into details in case of complex answers: The documentis interesting.

– Because it has been stated so:

49

AIFB AIFB

6

• By the user himself (the user has already seen the document, and “told”the system, that he is interested in)

• By someone else (someone, maybe “close” to the user, wanted to sharea favorable opinion about a document)

– Because it concerns a topic close to the user’s interest fields:• by the relation with the user:

∗ Personal interest fields∗ Professional interest fields (known by his role in the enterprise)

• by the way they are obtained:∗ Declared interest fields (the user has stated his interest documents

concerning a topic)∗ Implied interest fields (the user is included in a community of interest

which is close to a topic)

The second question, that introduces the notion of temporality into the pref-erence: Why does a user prefer a document at a given moment?

In other words, to make the difference from the first question: Why does auser prefer a document at a given moment, and does not prefer it at anothermoment?

– The document is interesting only if seen the first time (or the first few times)– It is interesting during a certain period (when the user performs a certain

activity, etc.)

These answers are just some samples, one can think of many other possiblereasons. Though, we realize that it is a very important to find the right questionsand answers, that include the majority of possible situations. Indeed, getting theright questions and answers and translating them into quantifiable attributes,and making sure that the highest number of possible situations are observed isa key to the success of such a learning mechanism, that may even outclass inimportance the chosen learning technique.

Nevertheless, we will present our approach in the Comma project to choosesome typical answers and attributes, but we keep more focused on the secondissue: the preference learning methodology.

3.4 Learning in the CoMMA system

After a short presentation of the design of the CoMMA system, this sectionpresents the multiagent interaction in which Machine Learning is performed inCoMMA.

The chosen MAS consists of a society of coarse-grained agents, that fulfill ingeneral multiple roles, and are organized in a small number of functional sub-societies. The MAS architecture was designed in order to optimize task-division,flexibility and robustness of the system, and network layout (extensibility, scal-ability, traffic optimization).

For the implementation of the prototype system, the Jade agent platformwas chosen, which is an Open Source Project developed by project partners,

50

AIFB AIFB

7

University of Parma and CSELT. Jade is a FIPA compliant agent platform,implemented in Java, and has also the advantages of a wide opening towardsInternet and the Web, interoperability with other MAS-s, and future systems.

In the current status of implementation, the CoMMA system will help theuser in three main tasks:

– insertion and RDF annotation of documents,– search of existing documents, and– autonomous document delivery in a push fashion to provide her/him with

information about new interesting documents.

We have already experimented such an architecture in Network Supervi-sion [EDQ96], with Machine Learning abilities [QEN97]. Here, for the MachineLearning part, we choose to use the Weka open source Java library [WF99],which enables us to experiment various learning methods and frameworks.

3.5 The learning agent

The first context to assess preference learning was chosen to be the documentretrieval scenario, via semantic annotations. The search engine used for docu-ment retrieval in the CoMMA system is an inference engine called CORESE[CDH00] developed by INRIA, one of the partners of the project. CORESE usesConceptual Graphs and combines the advantages of using the RDF language forexpressing and exchanging metadata, and the query and inference mechanismsavailable in CG formalism. In order to produce inferences, CORESE exploits thecommon aspects between CG and RDF: it defined a mapping from annotationstatements (RDF triples) to Conceptual Graphs and vice-versa.

One of the shortcomings of such a query retrieval engine is that there is nostandard method to sort the information returned, such as keyword frequencyin keyword-based search engines. The returned data set must be post-processed,filtered and sorted to present the user with the relevant information. Here comesthe aid offered by our ML mechanism.

In the CoMMA system, information that feeds the ML comes from severalsources: The document sub-society (the annotations accompanying a query re-sponse), the user sub-society (user monitoring and explicit user feedback), andontology sub-society (to help getting the ”meaning” of the results). And of coursethe user profile. Therefore, the learning behavior was ”awarded” to the User Pro-file Manager agent, which belongs to the connection dedicated sub-society, andperforms notably a role of middleman between agents. This decision was justi-fied also by network traffic optimization reasons, especially because in reactionto a user action (query), several interactions can be triggered between agents ofdifferent roles.

For example, during a query retrieval, the main interactions are as describedin the following diagram.

In this scenario, the role of the ML component starts when the query an-swers are collected and transmitted to the profile magager agent. First, the

51

AIFB AIFB

8

Archivist Archivist...

User agents

Document agents

UPM

Query Processed answer

Identify userPreprocess query:

enrich with information from the profile

Post-process answer:Filter, sort using ML prediction

Analyze the response:eventually create extra queries to collect more details for the ML

Distribute the query between archivists, then collect the answers

Enriched query Answer

Fig. 1. The main interactions in CoMMA during a query

semantic annotations are extracted and analysed. They are combined with therelevant fields from the user profile in the attempt to compute the values ofthe input attributes (such as the ones listed in the section about attributes). Incase the system finds that there could be more semantic annotations related tothe documents retrieved that could help in computing the input attributes, asupplementary query can be formulated to retrieve them before going on withprocessing and forwarding the answer.

The second step is to use the learnt preference measures to rank the docu-ments and sort the answers list to be returned by these predictions.

3.6 The learning cycle

The goal of the ML component is to produce a set of rules that will be used toproduce predictions about user preferences. It can be supposed that the systemstarts with a set of predefined rules, that will be gradually improved during theprocess of adaptation to users (method known as theory refinement). Otherwise,the system may start with an empty knowledge base, and will be undergo atraining period, to accumulate sufficient knowledge to allow its deployment.

52

AIFB AIFB

9

In our approach, user adaptability comes from both implicit observation ofthe user (user monitoring subsystem), and explicit user feedback.

We use a ”predict-or-learn” type protocol, that is, when the input is comingfrom query answers, the system tries to use its knowledge to predict, otherwise,when the input comes from the user in the form of negative feedback, the systemtries to update its rules set.

3.7 A sample set of attributes

A sample set of attributes we used to create instances from the answers wegave as examples to the question of preference is listed in the followings. Theseattributes may not be the most relevant ones, or adapted to any case, we onlytried to make it diverse and for most of them restricted the scope as much aspossible for the sake of simplicity of our prototype.

Is the document related to the user’s role We suppose that the users,depending on their roles, will be assigned certain interest fields, that will berecorded in their profiles. Then, the documents can have a generic Concernsproperty, that associates them with such topics. In this scenario, the notions likeTopic, Concerns, Interestedby, etc are defined as concepts in the ontology thatconstitutes the basis of the enterprise model.

As an observation, the notion topic or interestfield is used in a general senseto categorise documents.

Then we can extract this information from the user profile and documentannotation and combine them to create an instance attribute. We can definethis attribute as taking a binary value.

In a complete implementation it is likely that this relationship can be seen asmore complex and eventually be split into several attributes, and/or may takea wider range of values.

Is the document related to the user’s COINs (communities of inter-ests) This attribute reflects a further differenciation we have made in interesttopics when answering the preference question: topics can be assigned or chosenby the user; inherited or inferred by the system, etc.

In the same conditions as for the attribute above, the attribute will take abinary value, and in case we define relationships, we can use the same procedureas above.

User experience at the company Since in our project addressing the NewEmployee scenario is particularly focused upon, we considered important mak-ing the system behave differently towards new versus experienced employees.We have segmented the range of values so that this may take values such as:lessthan1week, lessthan1month, lessthan2months, ... all these regarding thenovice user, and morethan3months (or simply emphexpert, or whatever the op-posite for NE is). A tip for implementation, if such a segmentation is desired, is

53

AIFB AIFB

10

to define these intervals in the enterprise ontology, so it can adapt to its specificneeds.

The following subset of attributes are related to user monitoring.In our example we have imagined a simple user monitoring scenario, that

supposes tracking each consultation of a document by the user, and building anavigation history (or consultation trace).

Document last seen Usually it is important if a document was seen beforeor not, and if it was then how long before. After extraction from the user’snavigation history it should also be discretised into several intervals: ¡1hour,¡1day, ¡1week, ... , never.

Average return frequency In certain cases the user may return more oftento a document, in other cases an information can only present interest when firstseen. This attribute may also tell something about the situation for the currentcase. The value will also result from processing the navigation history, and itwould probably be enough to use some rough intervals (like once, small, large,etc.)

Document category touch frequency It might present an interest to extendthe above attribute for categories the document belongs to. In this case the valuewill be processed the same way as for the previous attribute. In case a specificstrategy for creating implicit communities of interest is envisaged, it should bechecked if there are possible conflicts with the use of this attribute.

The next attributes contain specific information about the particular docu-ment being analysed:

The user’s rating for the document For some specific documents, the usermay explicitly wish to manifest his interest or non-interest. In this case theuser profile should allow storing this information. It is a general vote for thedocument, and does not have the temporal aspect implied by the output of thisclassifier.

Public ratings Sometimes a user would like to share his opinion about a doc-ument with others. In this case it must be foreseen in the enterprise model sothat it may be stored in the form of an annotation, that can be used also byour classifier. A method should be formalised to allow storing information aboutpeople possibly interested about this opinion.

This was a list we used in our first trials. Once again we make the remarkthat the list of attributes should be open, and checking the completeness andexhaustiveness of it has to be an important and ongoing task. In other words,

54

AIFB AIFB

11

watching that the factors that contribute to a document being seen as more orless relevant in a situation have been well captured, and that there are no otherdecisive factors that were omitted, or can not be represented. Because in eithercase, the system might make major mistakes in certain situations, whateverlearning algorithm was used.

3.8 Document ranking and sorting

In our case, the first goal is to sort documents in a query response by the orderof predicted user preference. The two principal strategies that can be used toachieve that, are: grouping documents using numeric labeling or learning theorder relation (or the sorting rule).

For the first trial of our system, we choose to learn a rough labelling in afinite set of classes (namely 5), to classify the documents retrieved by the systemafter a user query. Then, we clearly fall in the first kind of strategy, whose goalis not to obtain a fine grained ranking, but only a coarse grained rating.

This is equivalent to the use of numeric labeling, associating numeric valuesfrom a finite set to categories of documents representing their importance. Butthe drawback is that the system will not be aware of the semantic of the orderrelation. And there is also a risk, that in case of large number of items to classify,many of them will fall into the same class, and there will be no further meansto differentiate them.

On the other hand, if an order relation is used (either by pairwise learningor by learning the sorting rules), we will have no distance measure to separatethe values. That is, to give an idea about for instance “how much a documentis more important then another”.

Then, a perspective for document retrieval systems of this kind can be to useboth methods in a complementary way:

1. in serial-coupling (one method used to pre-process, the second to post-process)

2. in parallel (eventually in distinct agents) then putting together the resultsand solving conflicts.

In order to evaluate the contribution of learning to the efficiency of the re-trieval system, we have designed an experimental protocol. It consists on eval-uating the learning step by comparing the use trace with and without learning,in the various scenarii taken in account in the CoMMA system.

4 Conclusion

In the CoMMA project, the Machine Learning and user adaptability componentis one of the main features. In this paper we presented the advances that we havemade in this domain, especially focussing on the learning of preference data fordocument retrieval.

55

AIFB AIFB

12

We proposed a specific method to learn preference data, and a framework toexperiment and evaluate it. The main choice we focus on does not only presentthe usefulness of Machine Learning, but also tries to overcome some of the lim-itations of semantic information retrieval systems.

In our opinion, the choice we made here will give interesting results duringthe first trial we planned in the project, then allows an experimental evaluationthrough feedback from the user.

The implementation is currently well advanced, and we begin to have someexperimental results.

References

[Aha92] D. Aha. Tolerating noisy, irrelevant, and novel attributes in instance-based learning algorithms. International Journal of Man Machine Studies,36(2):267–216, 1992.

[Bre96a] L. Breiman. Bagging predictors. Machine Learning, 24:123, 1996.[Bre96b] L. Breiman. Stacked regression. Machine Learning, 24:49, 1996.[CDH00] Olivier Corby, Rose Dieng, and C. Hebert. A Conceptual Graph Model for

W3C Resource Description Framework. In the 8th International Conferenceon Conceptual Structures (ICCS’00), number LNCS 1867 in Lecture Notes inArtificial Intelligence, Darmstadt, Germany, 2000. Springer Verlag, SpringerVerlag.

[Con00] CoMMA Consortium. Corporate Memory Management through Agents. InE-Work and E-Business conference, Madrid, October 2000.

[EDQ96] Babak Esfandiari, Gilles Deflandres, and Joel Quinqueton. An interface agentfor network supervision. In Intelligent Agents for Telecommunication Appli-cations, Budapest, Hungary, 1996. ECAI’96 Workshop IATA, IOS Press.

[Gas89] O. Gascuel. A conceptual regression method. In Katharina Morik, editor,EWSL-89, 4th European Working Session on Learning, pages 81–90, Mont-pellier, France, Decembre 1989. Jean Sallantin and Joel Quinqueton, CRIM,Pitman, Morgan Kaufman.

[Mit97] Tom M. Mitchell. Machine Learning. Mac Graw Hill, 1997.[QEN97] Joel Quinqueton, Babak Esfandiari, and Richard Nock. Chronicle learning

and agent oriented techniques for network management and supervision. InDominique Gaiti, editor, Intelligent Networks and Intelligence in Networks.Chapman & Hall, September 1997.

[Sch96] P-Y. Schobbens. A comparative logic for preferences. In Pierre-YvesSchobbens, editor, Working Notes of 3rd ModelAge Workshop: Formal Modelsof Agents, Sesimbra, Portugal, January 1996.

[SR99] P.-Y. Schobbens and J.-F. Raskin. The logic of “initially” and “next”:Complete axiomatization and complexity. Information Processing Letters,69(5):221–225, March 1999.

[WF99] Ian H. Witten and Eibe Frank. Data Mining: practical machine learning toolsand techniques with Java implementations. Morgan Kaufmann, 1999.

[WPB01] G. I. Webb, M. J. Pazzani, and D. Billsus. Machine learning for user mod-eling. User Modeling and User-Adapted Interaction (UMUAI), 11, 2001.

56

AIFB AIFB

ARCH: An Adaptive Agent for Retrieval Based

on Concept Hierarchies

B. Mobasher

We present a client-side agent, named ARCH, for assisting users in one ofthe most diÆcult information retrieval tasks, i.e., that of formulating an e�ec-tive search query. The agent utilizes a hierarchically-organized semantic knowl-edge base in aggregate form, as well as an automatically learned user pro�le,to enhance user queries. In contrast to traditional methods based on relevancefeedback, ARCH assists users in query modi�cation prior to the search task. Theinitial user query is (semi-)automatically modi�ed based on the user's interac-tion with an embedded, but modular, concept hierarchy. The modular design ofthe agent allows users to switch among the representations of di�erent domain-speci�c hierarchies depending on the goals of the search. ARCH passively learnsa user pro�le by observing the user's past browsing behavior. The pro�les areused to provide additional context to the user's information need represented bythe initial query. The full system also incorporates mechanisms for categorizingand �ltering the search results, and using these categories for performing re-�ned searches in the background. Preliminary experiments have shown that theagent can substantially improve the e�ectiveness of information retrieval bothin the general context of the Web, as well as for search against domain-speci�cdocument indexes.

57

Utilising an Ontology Based Repositoryto Connect Web Miners and Application Agents

Stefan Haustein

University of Dortmund, Computer Science VIII,Baroper Str. 301, D-44221 Dortmund, Germany,

[email protected] ,http://www-ai.cs.uni-dortmund.de

Abstract. Ontologies are important for providing a shared understanding of adomain for web mining agents and other agents accessing the gathered informa-tion. When the information access is decoupled from the mining process – forexample when building a semantic web server – an additional storage compliantto the application ontology is needed. The COMRIS information layer was builtto serve that purpose for a system supporting conference participants. It is ableto provide permanent access to gathered or aggregated information suitable forboth, humans and software agents by providing FIPA ACL and HTML interfaces.

1 Introduction

The goal of the COMRIS project was to design an agent based conference support sys-tem. Conference participants were equipped with a wearable electronic device that wasable to recognise other participants wearing a similar device. The purpose of the devicewas to introduce participants toeach other, to filter requests and to provide backgroundinformation depending on the current context [1]. In order to perform its task, the agentcontrolling the device needed background information during the short period of time acertain context was valid: When a person has passed by, it is too late to introduce thatperson.

The background information should be provided by a web mining process. Sincestarting mining on demand seemed too slow for the given application, web mining wasalready performed beforehand. The approach is similar to using web spiders for searchengines: If they were launched just when somebody enters a keyword, web searchingwould not be really practical.

2 The Mining Task

The gathering agents enrich the conference information by gathering information fromdifferent sources in the WWW. In our case, the agents just collect all information avail-able about the registered conference participants, and new persons discovered in thegathering process were not investigated further. The information was used to enrichthe knowledge about a person and its relations to other persons (e.g. co-author, projectpartner).

59

2

In the conference scenario, we were using three different types of gathering agents:The CORDIS collector is able to query the CORDIS project database of the Euro-pean Union, the KA (Knowledge Acquisition) and ILP (Inductive Logic Programming)agents are able to query two different bibliography databases for the correspondingcommunity. Each agent takes into account the special structure of its source, but theyall stem from one generic agent. Together, the gathering agents are able to find Europeanprojects the conference participants were involved in and most of their publications.

Before the actual start of the conference, a learning step was applied to informa-tion gathered for a set of training persons. The Rule Discovery Tool (RDT [2]) of theMOBAL machine learning system [3, 4] was used to learn indicators for a “may-want-to-meet” relation between participants. While the learning step itself was performedoff-line, the rules learned were applied to the information gathered in the runtime sys-tem, in order to create default instructions for the personal representation agents of theparticipants.

3 Complex Mining Tasks require Ontologies

When operating on a highly structured information space, it is no longer sufficient tojust store words in a huge database. This is where the application ontology comes intoplay. Both, gathering and application agents need a common language. Also, in order tobe able to perform the gathering beforehand, some kind of repository for the gatheredinformation is needed. For the COMRIS conference scenario, the amount of concepts tobe modelled, like participants, speakers, talks, sessions, rooms, agents, booths, sched-ules etc., became quite large. Moreover, all concepts had a lot of complex relations toother concepts.

Using relational tables for this purpose seemed inadequate because of the compli-cated mapping that is required to transform the ontology to a high number of tables.Also, the table solution seemed inflexible because ontological changes would cause alot of changes in database tables and additional “agentification” wrappers.

Description Logic [5] systems like KL-ONE [6] provide additional features like au-tomated classification that are computational expensive but not required in the system.All reasoning was intended to be performed by the specialised agents. Like for the rela-tional tables, additional wrappers for an Agent Communication Language (ACL) wouldbe required. Also, using Description Logics would require globally unique slot names,leading to additional negotiation efforts between the project partners designing “their”part of the application ontology.

OntoBroker, a system extracting ontologies from the web, is able to automaticallyunfold the stored knowledge and provides persistence for the ontology itself [7]. Whileits centralised structure would be a good starting point for learning mechanisms, thesystem interface is not agent but human oriented.

4 Information Layer System Architecture

For the given reasons, we decided to build a new kind of information system that

60

AIFB AIFB

3

– provides ACL access in the first place,– is agent based itself,– and is built on an ontology that is not hard-wired to the system.

The main purpose of the system was to act as blackboard [8, 9] for decoupled com-munication between the conference organisers entering the initial participant informa-tion, the gathering agents annotating this information with web content, and the appli-cation agents utilising the gathered information for their tasks helping the conferenceparticipants.

Fig. 1. Information layer architecture overview

The kernel of this system, the COMRIS information layer, provides only a memoryrepresentation of information structured corresponding to a given ontology. All otherfeatures were delegated to additional modules or agents, performing specialised taskslike:

– Handling communication with other agents– Applying the learned rules to transform gathered data to default agent instructions– Synchronisation with the underlying persistent data storage– Building a generic HTML presentation from the ontology and the actual informa-

tion layer content

The HTML presentation was not an initial part of the system, but once the systemwas built, it seemed a waste of resources to set up a separate conference web site builton traditional techniques. Instead, a wrapper agent transformed HTTP requests to ACLmessages and forwarded them to the corresponding agents. Figure 1 shows an overviewof the system architecture.

61

AIFB AIFB

4

Fig. 2.Sample UML ontology diagram

5 Ontology and Data Model

The information layer uses an object-oriented model for data representation. Objectsconsist of atomic attributes and relations to other objects. The consistency of relationsin both directions is ensured automatically, avoiding inconsistencies inside the system.The concepts and relations are defined application-dependent in an external ontologydefinition file. All files used by the information layer are stored as XML documents.

The ontology used in the COMRIS information layer is defined using an UnifiedModelling Language (UML) model [10,11] encoded in a simple XML format. Com-pared to other languages suitable for ontology modelling, UML currently still lacksclearly defined semantics. However, there are significant efforts to solve this problems[12,13].

Figure 2 shows the UML diagram of the shared parts of the COMRIS ontology.Gathered information about publications and projects was transformed to templates forthe Personal Representative Agents (PraTemplate ) by applying the learned rules.The raw data gathered was also stored in the information system, but was not sharedamong all agents.

62

AIFB AIFB

5

6 Communication and Content Languages

The communication and content languages for software agents and system componentsare based on XML, too. An XMLified version of FIPA ACL [14] is used as communica-tion language, whereas the actual content language format is derived from the ontologyautomatically. Figure 6 shows the content language encoding of Tanja Katschenko andCarlos Gomez working at IBM corresponding to the ontology example in the previoussection.

<Organization id="555777"> <name>IBM</name></Organization>

<Person id="888543"> <name>Katschenko</name> <firstName>Tanja</firstName> <memberOf idref="555777"/></Person>

<Person id="878653"> <name>Gomez</name> <firstName>Carlos</firstName> <memberOf idref="555777"/></Person>

<Organization id="555777"> <name>IBM</name> <members> <Person id="888543"> <name>Katschenko</name> <firstName>Tanja</firstName> </Person> <Person id="878653"> <name>Gomez</name> <firstName>Carlos</firstName> </Person> </members></Organization>

Linked Structure Nested Structure

Fig. 3.Content language examples

Relations between instances can be described using theidref attribute, or by em-bedding related instances in the relation element. The encoding used for sending in-stances to software agents or other entities can be controlled by the corresponding entityto fit its particular needs best.

Readers familiar with the Resource Description Format (RDF) will have noticed astrong similarity of the formats. While it would be possible to migrate to RDF, therewould be no improvement concerning human readability, which turned out crucialfor system integration and maintenance. Moreover, RDF uses a property-centric datamodel, causing compatibility issues with traditional object oriented systems. The highnumber of RDF syntax variants leads to integration problems with other XML buildingblocks like XML Schema and XSLT [15] [16]. For those reasons, we will replace thecurrent XML representation by the serialisation format of the Simple Object AccessProtocol (SOAP) [17], improving the compactness and readability of the format as wellas compatibility to SOAP based third party systems. However, migration to SOAP doesnot exclude building an additional RDF based interface if required.

7 Query Interface

The information layer supports a subset of OQL [18] as query language for agents. Ad-ditional languages may be plugged in by adding corresponding agents. By subscribingto the information layer, it is possible to keep an agent up to date without polling [19].

63

AIFB AIFB

6

8 HTML Generation

The information layer contains a module that provides built-in web-server functionality.Since XML is not fully supported by web browsers yet, the server is able to generateHTML dynamically: For any object, the attributes are simply displayed, and the rela-tions are converted to sets of hyperlinks to the related objects (figure 4). The HTMLinterface can also be used to edit the content of the system using forms generated dy-namically based on the ontology.

Fig. 4. Access to the Information Layer using a Web Browser.

In the COMRIS project, the HTML interface was used for interaction with the enduser as well for as debugging and inspection purposes.

In addition to generic HTML generation, templates can be used in order to generateHTML pages conforming to a given look and feel. In the COMRIS project, we havealso used the template mechanism to generate the input structure required by the textgeneration system TG/2 ([20]) which was used to generate natural language output forthe wearable device.

The template mechanism was also used to generate questionnaires for evaluatingthe mining and learning results of the gatherers [21, 22].

9 Conclusion and Outlook

The main purpose of the implemented system was to provide an ontology based per-sistent blackboard communication mechanisms for connecting mining and applicationagents.

64

AIFB AIFB

7

Using ontologies and agent technologies enabled a simple extension of the systembeyond the original purpose. The system can now also be used to publish structured andmassively linked data to the traditional “human readable” web using template based(X)HTML generation. The system proved useful not only for modelling some aspectsof a conference but also for other applications with many sets of small and massivelylinked objects.

Currently, the COMRIS information layer is used for two internal projects and asthe training server of MLnet1. In the future, it is planned to use the information layer inthe MiningMart project for storing and editing data mining meta information.

The most important future developments are to make the information layer compli-ant to SOAP serialisation [17] and XMI in order to use a standardised XML formatsfor the message content language and for the ontology definition. It is also plannedto include structure translation mechanisms for connecting systems using different butrelated application ontologies.

Acknowledgements

The research reported in this paper was supported by the ESPRIT LTR 25500 COMRISproject.

References

1. Plaza, E., Arcos, J.L., Noriega, P., Sierra, C.: Competing agents in agent-mediated institu-tions. Personal Technologies Journal2 (1998) 1–9

2. Kietz, J.U., Wrobel, S.: Controlling the complexity of learning in logic through syntacticand task–oriented models. In Muggleton, S., ed.: Inductive Logic Programming. Number 38in The A.P.I.C. Series. Academic Press, London [u.a.] (1992) 335–360

3. Sommer, E., Emde, W., Kietz, J.U., Wrobel, S.: Mobal 4.1b9 User Guide. GMD – GermanNational Research Center for Information Technology, AI Research Division (I3.KI), St.Augustin, Germany (1996)

4. Morik, K., Wrobel, S., Kietz, J.U., Emde, W.: Knowledge Acquisition and Machine Learning- Theory, Methods, and Applications. Academic Press, London (1993)

5. Nebel, B.: Reasoning and Revision in Hybrid Representation Systems. Number 422 inLecture Notes in Artificial Intelligence. Springer-Verlag (1990)

6. Brachman, R.J., Schmolze, J.G.: An overview of the KL-ONE knowledge representationsystem. Cognitive Science9 (1985) 171–216

7. Decker, S., Erdmann, M., Fensel, D., Studer, R.: Ontobroker: Ontology based access to dis-tributed and semi-structured information. In Meersman, R., other, eds.: Semantic Issues inMultimedia Systems, Kluwer Academic Publisher, Boston,1999. Kluwer Academic Pub-lisher, Boston (1999)

8. Hayes-Roth, B.: A blackboard architecture for control. Artificial Intelligence26 (1985)9. Kollingbaum, M., Heikkilae, T., McFarlane, D.: Persistent agents for manufacturing systems.

In: AOIS 1999 Workshop at the Third International Conference on Autonomous Agents.(1999)

1 http://www.mlnet.org/training

65

AIFB AIFB

8

10. Object Management Group: OMG Unified Modeling Language Specification, version 1.3.http://www.omg.org/technology/documents/formal/unifiedmodelinglanguage.htm (2000)

11. Cranefield, S., Purvis, M.: Uml as an ontology modelling language. In: Proceedings ofthe Workshop on Intelligent Information Integration, 16th International Joint Conference onArtificial Intelligence (IJCAI-99. (1999)

12. Precise UML Group: The Precise UML Group home page. http://www.puml.org (2001)13. Clark, T., Evans, A., Kent, S., Brodsky, S., Cook, S.: A feasibility study in rearchitecting

UML as a family of languages using a precise OO meta-modeling approach. Report, PreciseUML Group (2000) http://www.cs.york.ac.uk/puml/mml/mmf.pdf.

14. Foundation for Intelligent Physical Agents: FIPA web site (2001)http://www.fipa.org/specs/fipa00023/XC00023F.pdf.

15. World Wide Web Consortium: XSL Transformations (XSLT) version 1.0.http://www.w3.org/TR/xslt (1999)

16. Haustein, S.: Semantic Web languages: RDF vs. SOAP serialization. In: Proceed-ings of the Second International Workshop on the Semantic Web at WWW10. (2001)http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS/Vol-40/haustein.pdf.

17. Box, D., Ehnebuske, D., Kakivaya, G., Layman, A., Mendelsohn, N., Nielsen, H.F., Thatte,S., Winer, D.: Simple Object Access Protocol (SOAP) 1.1. Note, World Wide Web Consor-tium (2000)http://www.w3.org/TR/2000/NOTE-SOAP-20000508.

18. Cattell, R.G.G., ed.: The Object Database Standard: ODMG 2.0. Morgan Kaufmann (1997)19. Haustein, S., L¨udecke, S.: Towards Information Agent Interoperability. In Klusch, M.,

Kerschberg, L., eds.: Cooperative Information Agents IV – The Future of Information Agentsin Cyberspace. Volume 1860 of LNCS., Boston, USA, Springer (2000) 208 – 219

20. Busemann, S.: A shallow formalism for defining personalized text. In: Workshop Profes-sionelle Erstellung von Papier- und Online-Dokumenten at the 22nd Annual German Con-ference on Artificial Intelligence (KI-98), Bremen (1998)

21. Haustein, S.: Information environments for software agents. In Burgard, W., Christaller,T., Cremers, A.B., eds.: KI-99: Advances in Artificial Intelligence. Volume1701 of LNAI.,Bonn, Germany, Springer Verlag (1999) 295 – 298

22. Haustein, S., L¨udecke, S., Schwering, C.: The Knowledge Agency. In Sierra, C., Gini, M.,Rosenschein, J.S., eds.: Proceedings of the Forth International Conference on AutonomousAgents, Barcelona, Spain, ACM SIGART, ACM Press, New York (2000) 205 – 206

66

AIFB AIFB

XML Topic Maps and Semantic Web Mining

Benedicte Le Grand, Michel Soto

Laboratoire d'Informatique de Paris 6 8, rue du Capitaine Scott 75015 Paris, France

[email protected], [email protected]

Abstract. Navigation and information retrieval on the Web are not easy tasks; the challenge is to extract information from the large amount of data available. Most of this data is unstructured, which makes the application of existing data mining techniques to the Web very difficult. However, new semantic structures which improve the results of Web Mining are currently being developed in the Web. This paper presents how one of these semantic structures - XML topic maps – can be exploited to help users find relevant information in the Web. This paper is organised as follows: first, we introduce XML topic maps in the context of Tim Berners-Lee's Semantic Web vision. Then, we show how topic maps allow to characterise and "clean" Web data through the definition of a profile; this is achieved by the analysis of a lattice generated by a classification algorithm - called Galois algorithm. This profile may be used to evaluate the relevance of a web site with regard to a specific request on a traditional search engine. We finally explain how data on the Web can be clustered, organised and visualised in different ways so as to enhance users' navigation and understanding of these documents.

1 Introduction

Navigation and information retrieval on the Web are not easy tasks; the challenge is to extract information from the large amount of data available. Most of this data is unstructured, which makes the application of existing data mining techniques to the Web very difficult. However, new semantic structures which improve the results of Web Mining are currently being developed in the Web. This paper presents how one of these semantic structures - XML topic maps – can be exploited to help users find relevant information in the Web. This paper is organised as follows: first, we introduce XML topic maps in the context of Tim Berners-Lee's Semantic Web vision [2]. Then we show how topic maps allow to characterise Web sites through the definition of a profile. This profile may be used to evaluate the relevance of a web site with regard to a specific request on a traditional search engine. We finally explain how data on the Web can be clustered, organised and visualised in different ways so as to enhance users' navigation and understanding of these documents.

67

mailto:[email protected]

mailto:[email protected]

2 XML Topic Maps and the Semantic Web

Finding information on the Web is very difficult. Search engines may return hundreds or more links to users' queries – provided that the right keywords are used. Choosing the most relevant Web sites to explore is not trivial, because no semantics help evaluate the relevance of each hit. The next step is not easier: once a link is chosen, navigation is not always intuitive. Users can get lost easily: they may not find the information they are looking for even though it does exist. Sometimes they do not manage to go back to a page they have already visited. This is due to the lack of structure of the Web. Therefore it is necessary to add structure and semantics as well as to provide a mechanism which allows a more precise description of data on the Web. According to Tim Berners-Lee from W3C [2]:

"The Semantic Web is a vision: the idea of having data on the web defined and linked in a way that it can be used by machines – not just for display purposes, but for using it in various applications."

This Semantic Web can be achieved by adding semantic structures to the current Web. Many candidate techniques were proposed, such as semantic networks [16], conceptual graphs [5], the W3C Resource Description Framework (RDF) [14] and XML Topic Maps [11]. Semantic networks are basically directed graphs (networks) consisting of vertices linked by edges. Edges express semantic relationships between the vertices.

The conceptual graphs theory developed by Sowa [10] is a language for knowledge representation based on linguistics, psychology and philosophy.

RDF data consists of nodes and attached attribute/value pairs. Nodes can be any web resource (pages, servers, basically anything for which you can give a URI), or other instances of metadata. Attributes are named properties of the nodes, and their values are either atomic (character strings, numbers, etc.), metadata instances or other resource. This mechanism allows us to build labelled directed graphs.

Topic maps, as defined in ISO/IEC 13250 [8], are used to organise information in a way that can be optimised for navigation. Topic maps were designed to solve the problem of large quantities of unorganised information. Information is not useful if it cannot be found or linked. In the paper publishing world, there are several mechanisms to organise and index the information contained within a book or document. Indexes allow readers to go directly to the portion of the document that is relevant to their information needs. Topic maps can be thought of as the online equivalent of printed indexes. Topic maps are also a powerful way to manage link information, much as glossaries, cross-references, thesauri and catalogs do in the paper world. Topic Maps allow users to create a large quantity of metadata and tightly interconnected data. They constitute a kind of semantic network above the data themselves.

A new specification which aims at applying the topic map paradigm to the Web is currently being written; this initiative is called XTM (XML Topic Maps) [11]. XML Topic Maps allow to structure data on the Web and therefore make Web mining more efficient.

It was recently proven that the RDF and Topic Map models could inter-operate at a fundamental level [9]. Both standards are concerned with defining relationships between entities with identity. Each language can be used to model the other.

68

All the techniques described previously have the same goals and many of them are compatible. We decided to further investigate XML Topic Maps and study how they could enhance Semantic Web Mining.

We aim at helping users find relevant information and we contribute at three

levels: 1. by evaluating Web sites relevance to users needs based on semantic criteria, 2. by filtering the topic map; the topic map profile constitutes a reference that can be

used to select the most semantically significant objects (called regular objects). This allows to identify the major subjects which the topic map deals with and to discard less relevant topics.

3. by enhancing navigation on the Web through the aggregation of conceptually related topics and through the visualisation with different scales – or levels of details. The different steps of topic maps – or Web sites1 - analysis are represented in

figure 1:

T o p i c m a p f i l t e r i n g

T o p i c m a p

C o n c e p t u a l c l a s s i f i c a t i o n a l g o r i t h m

G a l o i s l a t t i c e

T o p i c m a p c h a r a c t e r i s a t i o n

T o p i c m a p c l u s t e r i n g

Fig. 1. Web sites analysis algorithm

We propose to achieve the first goal by defining profiles of topic maps – and consequently Web sites. These profiles characterise topic maps – or web sites - and help evaluate their relevance to users' information needs. The computation of this sort of topic map "DNA" and its interpretation are described in section 4.

The topic map may contain topics which are not semantically significant or not much related to others. We call them singular topics. They may be eliminated from the topic map so as to clean it, as explained in section 4.2.

1 In the following, we will use the term "topic map" which is more general than "Web site".

Topic maps may apply to any kind of data.

69

Our third contribution consists in enhancing navigation and information retrieval in a Web site. Information retrieval varies according to the needs of the user. If he looks for an answer to a specific question, query languages (like "tolog" [6]) are adapted. Their strength is to exploit the relationships between objects, which allows to answer questions better. For example, one may seek the Beatles’ songs which were not written by John Lennon. This kind of information would be difficult to find with a traditional search engine.

If the subject of interest is clearly identified, it is easy to explore the corresponding topic in the topic map. This topic can be reached through a list of topics, for example an alphabetical list. Tools to navigate in topic maps have been designed so that any topic can be reached in 7 mouse clicks at most.

If the user has no precise question nor any clear subject of interest, none of the search modes described above can apply. This is the case of a beginner user who wishes to have a global understanding of the topic map so as to decide where to start his navigation. Therefore he first needs a simplified view of the topic map, with no detail, then he can decide to see more precise information as his subject of interest gets clearer. Let us compare this to geographical maps: there is no point in displaying very specific data on a map of the world. However, more and more details may be added as the user focuses on some part of the map. We propose to use clustering algorithms to group semantically related topic together at different abstraction levels. Clusters computation and visualisation are described in section 5.

The figure 1 shows that topic maps – or Web sites – characterization, filtering and

clustering are deduced from the results of a conceptual classification algorithm based on Formal Concept Analysis and Galois connections. This algorithm is presented in section 3.

3 Conceptual classification algorithm

The starting point of our Web analysis is a conceptual classification algorithm based on Formal Concept Analysis and Galois connections. FCA is a mathematical approach to data analysis which provides information with structure. FCA may be used for conceptual clustering as shown in [4] and [12]. Let us first define a few terms: − an object is a topic or an association of the topic map, − the objects have characteristics called properties. We describe how these

properties are determined in 3.1. A profile allows to characterise a topic map in a structural way. With this

footprint, one can tell if the topic map is specific or general. We can also tell if the objects of a topic map are similar or very different. In order to characterise objects, we use a Galois algorithm to classify the objects conceptually. This algorithm groups objects in concepts according to the properties they have in common. It is very powerful because it performs a semantic classification without having to express semantics explicitly. We will first describe how the objects and their

70

properties are generated from a topic map. Then we will describe Galois lattices and detail the statistical computations made on the objects. We will finally explain how the profile is determined.

3.1 Objects and properties generation

The generation of objects and properties is a 2-steps process.

• First step: Every time there is an element with an identifier (that is an id attribute), a new

object is created. The name of the object is the value of the identifier. As stated in the DTD (Document Type Definition), all topics and associations of the topic map have an identifier, so there will be the same number of objects as the number of topics and associations.

An object's properties correspond to the values of this object's attributes (including the value of the id attribute), as well as the values of his children' attributes. These properties are weighted (for instance, the weight of the values of instanceOf attributes may be greater than the weight of the values of href attributes).

Generation of object and properties (first step):

element <topic> attributes id= "topic1" attr2="val2" attr3="val3"

object topic1 Properties topic 1val2val3

Example: consider the following extract of a topic map about music, written by Kal Ahmed2:

<topic id="t-the-clash"> <instanceOf> <topicRef xlink:href="tt-band"/> </instanceOf> <baseName> <baseNameString>The Clash</baseNameString> <variant> <parameters>

<topicRef xlink:href="http://www.topicmaps.org/xtm/1.0/psi-sort"/> </parameters> <variantName> <resourceData>clash the</resourceData> </variantName>

2 Kal Ahmed works for Ontopia, http://www.ontopia.net

71

</variant> <variant> <parameters>

<topicRef xlink:href="http://www.topicmaps.org/xtm/1.0/psi-sort"/> </parameters> <variantName> <resourceData>Clash, The</resourceData> </variantName> </variant> </baseName> </topic> An XML document is made of elements limited by tags and is hierarchically

structured. In the example we studied, topic, instanceOf and baseName are elements. An element may have characteristics called attributes. The attributes of an element are declared inside the opening tag of the element. The element topic has an attribute id with a value tt-clash. The element instanceOf has no attribute.

When parsing the topic map, we find a topic which has an identifier with the value t-the-clash. An object t-the-clash is thus created.

In order to determine the properties of these objects, we look for all the attributes of this element. In this case, the only one is the identifier.

Then, we have a look at the children of this element (that is all the XML elements included in the element) to find their attributes. We repeat this for all the children.

In this example, the analysis of this abstract of the topic map creates an object t-the-clash with the properties t-the-clash (weight e.g. 0.5), tt-band (weight e.g. 2) and http://www.topicmaps.org/xtm/1.0/psi-sort (weight e.g. 0.2). The weights shown here correspond to one possible scenario - in which the type of a topic (weight 2) is more important than its name (weight 0.5), its occurrences (weight 0.2) or the associations it is involved in (weight 1).

In the same way, the analysis of the following abstract: <topic id="tt-band"> <instanceOf> <topicRef xlink:href="tt-music"/> </instanceOf> <baseName> <baseNameString>Band</baseNameString> </baseName> </topic> leads to the creation of an object tt-band with the properties tt-band (weight e.g.

0.5) and tt-music (weight e.g. 2). The last example concerns an association: <association id="assoc6"> <instanceOf> <topicRef xlink:href="at-recorded"/>

72

http://www.topicmaps.org/xtm/1.0/psi-sort

</instanceOf> <member> <instanceOf> <topicRef xlink:href="tt-band"/> </instanceOf> <topicRef xlink:href="t-the-clash"/> </member> <member> <instanceOf> <topicRef xlink:href="tt-track"/> </instanceOf> <topicRef xlink:href="t-i-fought-the-law"/> </member> </association> The object assoc6 is created and has the properties assoc6 (weight e.g. 0.5), at-

recorded (weight e.g. 2), tt-band, t-the-clash, tt-track and t-i-fought-the-law. So far, the properties of an object are only intrinsic properties. Indeed, the object

t-the-clash takes a part in the association assoc6, but this does not appear in its properties yet, since the association is not declared inside the topic which has the identifier t-the-clash. The second step takes these characteristics into account.

• Second step:

Generation principle of the objects and properties (second step):

element <topic> attributes id= "topic1" attr2="val2" attr3="val3"

object topic1 Properties topic 1val2val3

object val3

object val2 Properties topic1

Properties topic1

The second steps adds non intrinsic properties to the objects by “crossing” the data. In fact, for an object O with a set of properties P, each property P becomes an object with O (amongst others) as a property. The properties of an object are its intrinsic properties and all the properties that were added recursively.

73

In the previous examples, the object assoc6 has the properties assoc6, tt-band and

t-the-clash. The property assoc6 is added to the objects tt-band and t-the-clash. So all the objects know the associations they appear in.

Moreover, the object t-the-clash has the property tt-band. The data is crossed by adding t-the clash to the object tt-band. This example illustrates a new type of information, which was not present in the first step: the object tt-band knows it has an instance of t-the-clash. In the preceeding scenario, t-the-clash was the only one to know its superclass.

In the end, tt-band has the properties tt-band (weight e.g. 0.5), tt-music(weight e.g. 2) , t-the-clash (weight e.g. 1), assoc1 (weight e.g. 1), assoc2 (weight e.g. 1) and assoc6 (weight e.g. 1). The object t-the-clash has the characteristics http://www.topicmaps.org/xtm/1.0/psi-sort (weight e.g. 0.2), tt-band, t-the-clash (weight e.g. 0.5), assoc1 (weight e.g. 1), assoc2 (weight e.g. 1) and assoc6 (weight e.g. 1).

Note that the properties assoc1 and assoc2 correspond to other associations in which tt-band and t-the-clash appear. These associations are present in the topic map but not in the extracts we presented.

3.2 Introduction to Galois lattices

The notion of Galois lattice for a relationship between two sets is the basis of a set of conceptual classification methods. This notion was introduced by Birkhoff in [3] and by Barbut and Monjardet in [1]. Galois lattices consist in grouping objects into classes that materialise concepts of the domain under study. Individual objects are discriminated according to the properties they have in common. This algorithm is very powerful as it performs semantic classification. Topic maps are semantic structures themselves, but they may be very large and complex, so this algorithm is interesting to extract more semantics from them. The algorithm we implemented is based on the one that was proposed in [7].

Let us first introduce Galois lattices basic concepts. Let two finite sets E and E’ (E consists of a set of objects and E’ is the set of

these objects’ properties), and a binary relation R ⊆ E x E’ between these two sets. Figure 2 shows an example of binary relation between two sets. According to Wille’s terminology [13], the triple (E, E’, R) is a formal context which corresponds to a unique Galois lattice. It represents natural regroupings of E and E’ elements.

Let P(E) be the powerset of E and P(E’) the powerset of E’. Each element of the lattice is a couple, also called concept, noted (X, X’). A concept is composed of two sets X ∈ P(E) and X’ ∈ P(E’) which satisfy the two following properties :

X’ = f(X) where f(X) = { x’ ∈ E’ | ∀x ∈ X, xRx’ }

X = f’(X’) where f’(X’) = { x ∈ E | ∀x’ ∈ X’, xRx’ } (1)

A partial order on concepts is defined as follows : Let C1=(X1, X’1) and C2=(X2, X’2),

122'1'21 XXXXCC ⊆⇔⊆⇔< (2)

74

http://www.topicmaps.org/xtm/1.0/psi-sort

This partial order is used to draw a graph called a Hasse diagram, as shown on figure 2. There is an edge between two concepts C1 and C2 if C1<C2 and there is no other element C3 in the lattice such as C1<C3<C2. In a Hasse diagram, the edge direction is upwards. This graph can be interpreted as a representation of the generalisation / specialisation relationship between couples, where C1<C2 means that C1 is more general than C2 (and C1 is above C2 in the diagram).

Fig. 2. Binary relationship and associated Galois lattice representation (Hasse diagram)

Fig. 3. Concept lattice of a topic map about music

75

The concept lattice shows the commonalities between the concepts of the context. The first part of a concept is the set of objects. It is called "extension". The second set – the intention - reveals the common properties of the extension's objects. The figure 3 shows the concept lattice generated from our example topic map about music.

4 Topic maps characterisation: conceptual profile

4.1 Calculating the statistics for every object.

We calculate statistics for every object of the topic map. We compute a weighted mean of these statistics. Each object has a weight which is assigned according to its importance in the topic map (the number of occurrences of the object in the XML source file).

Consider an object O. It is characterised by a vector with 6 components: • The first component (A1) is the percentage of concepts of the sub-lattice where

the object is present in the list of extensions. This value tells if O is present in many concepts of the lattice. A low value for A1 may indicate that O has few common characteristics with other objects. However, the other components allow to increase our knowledge.

• The second component is the maximum number of objects with which O is grouped, divided by the total number of objects. We have to select the concept containing O and with the largest number of objects. We add a constraint on this concept: it must contain at least one property. Indeed, we wish to group objects with common properties. The component A2 shows if O is grouped with many other objects. However, this value is a maximal value. The validity of A2 must be checked using A3.

• A3 is the mean number of objects with which O is grouped divided by the number of objects. This time we can tell if O is linked to a large number of objects and determine the significance of A2. If A3 is high, then there is a concept with O and many other concepts. On the other hand, if A3 is low, O is grouped with very few objects. The selected concept is thus an exception and we should not base our analysis on it.

• Let S be the set of objects which are grouped with O in one –or more- concepts of the lattice; these objects have at least one of O's properties. A4 is the maximum number of properties O shares with the objects contained in S, divided by the total number of objects. This component is deduced from the concept containing the object O and which has the greatest number of properties. Again, we add a constraint on this concept: it must contain at least two objects, that is at least one object different from O. We want to evaluate the number of shared properties, thus we need at least one object with which O shares them. A4 tells if the objects which are close to O share many common properties with O or not. Objects are more similar when they share an increasing number of properties. This similarity

76

can either be structural or conceptual. However, this value is a maximum number which must be validated with A5.

• A5 is the mean number of properties O shares with other objects, divided by the total number of properties. This tells the degree of significance of A4.

• Finally, A6 is about the topic map itself, and not about the lattice. It is the number of occurrences of the object in the topic map divided by the number of occurrences of objects of the same type (topic or association). A6 is used to compute the topic map’s profile. This profile represents the characteristics of a mean object. Each component of this vector is the mean of the components of each object in the topic map, with a weight A6 given to each of these objects. Thus, objects with a high number of occurrences in the topic map will influence the profile much more than objects with few occurrences. Note that the five first components are deduced from an analysis of the lattice

whereas the last component only depends on the XML document.

4.2 Topic map – Web site - profile and selection of objects

When the statistics have been computed for every topic and association, the profile can be deduced. It is a vector for which each component is a mean of the components of all the objects with the weight A6 of each object. For N objects O1, O2, … ON, each component Ai of the profile vector P is computed as follows:

∑=

=N

jjiji AOAOAP

16.*..

(3)

where Oj.Ai is the component Ai of the j-th object. We wish to keep the most relevant objects, that is the ones which share "many"

common properties with "many" other objects. These objects are called regular objects, they are semantically more significant than others. The significance of the words "many" (properties) and "many" (objects) is given by the topic map profile. A regular object is associated to at least as many objects and shares as many properties as the profile.

Among the statistics presented in section 4.1, the values A3 and A5 are more

relevant than A2 and A4: maximum values may not give a reliable information because they may correspond to an exception. The comparison between the objects and the profile is thus done using the components A3 and A5.

A regular object O must verify the following conditions:

11 .. AprofileAO ≥ 22 .. AprofileAO ≥

(4)

77

This should be refined using the standard deviation. The standard deviation for A3 is the mean distance between an object’s value of A3 and the profile’s value of A3.

N

APAOAdevstd

N

jj∑

=−

= 133

3

....

(5)

For A5, the standard deviation is computed in the same way:

N

APAOAdevstd

N

jj∑

=−

= 155

5

....

(6)

Thus, a regular object is defined as follows: 111 .... APAdevstdAO ≥+

222 .... APAdevstdAO ≥+ (7)

The regularity conditions can be changed (to be more or less restrictive) with a coefficient (C). Thus, a regular object meets the two following requirements:

111 .... APAdevstdCAO ≥×+

222 .... APAdevstdCAO ≥×+ (8)

A non regular object is called a singular object –it conveys little semantics. When the objects of the topic map are submitted to these conditions, singular objects are eliminated. When C increases, more objects are suppressed since the conditions are harsher.

After this selection, we have a new list of objects which are used as an input for

the Galois classification algorithm. A new lattice is generated and the statistics computed on this new panel of objects provide a new profile. We can thus select once again the regular objects for this new footprint of the topic map. The new regular objects are used again as an input for the Galois algorithm, etc. until all the objects become regular. This happens when no object is eliminated. The algorithm stalls and we get a stable list of regular objects which we must group together.

4.3 Results

Several topic maps – of different sizes and subjects - were analysed. The figure 4 displays the distribution of objets in three topic maps. The coordinates of the center of a disk correspond to the values of A3 and A5 attributes. The diameter of a circle is proportional to the number of objects which have these values for A3 and A5. All the objets of the simple topic map are very close. This topic map is qualified of "homogeneous", which means that all topics have the same semantic significance. Music and icc are "heterogeneous" structures. The objects in the lower left corner have low values for A3 and A5: they are "singular" - not much related to other

78

objects in the topic map. These topic maps can be filtered easily by eliminating these singular objects.

-10

0

10

20

30

40

50

60

-10 0 10 20 30 40 50

A3

A5

music récursificc récursifsimple récursif

Fig. 4. Topics conceptual distribution

0

20

40

60

80

100

120

1 2 3 4 5 6 7 8 9 10 11 12 13

number of iterations

% o

f rem

aini

ng o

bjet

s

iccmusicsimpleSH-CCdiscoveryxmle99mp

Fig. 5. Topic map filtering

79

The figure 5 illustrates the filtering of six topic maps. Some topic maps can be simplified a lot; this is the case of discovery. On the other end, after the last iteration, xmle99mp still contains almost 70% of its topics. This means that it is more difficult to filter this topic map: all topics have the same semantic value.

5 Topic maps clustering and visualisation

5.1 Clustering algorithm

The Galois lattice which is generated from a topic map contains some concepts which are made up of a set of topics which share common properties. The lattice gives an exhaustive description of the input data and the number of concepts generated may be very high. The concept lattice shown in the figure 3 is quite complex although it was generated from a small topic map (which contains 46 objects). We wish to group topics together into clusters in order to provide different level of detail (or scales) of the topic map. We propose to extract a tree from the Galois lattice. The concepts contained in this tree are the clusters. Thus, we have a hierarchy of clusters. The root of the tree contains all the topics; it is a gross cluster which provides no additional information. The next level groups some topics together, the next level executes a finer grouping of topics, etc. The number of levels of detail is given by the depth of the tree.

Many clustering algorithms exist; we chose to implement a clustering algorithm based on Galois conceptual classification. The clusters we generate are thus conceptually and semantically relevant. This algorithm also allows us to use the generalisation/specialisation relationship inherent to the Galois lattice.

To build the tree of clusters, we start from the representation which provides the greatest level of detail. Every cluster corresponds to an object: the objects are not grouped together. We begin to construct the leaves of the tree: these clusters correspond to the fathers of the upper bound of the lattice (which is represented at the bottom of the lattice on the Hasse diagram). This is the most specific level.

For each leaf, we select one unique father which is a generalisation of the concept. This selection is done according to a hierarchy of criteria which will be developed in the following. One father is selected for each selected node, and so on until the lower bound of the lattice is reached. At the end of this process, a tree is created. Each level of the tree contains clusters which correspond to a level of detail.

We defined a hierarchy of selection criteria when a concept has several fathers in the lattice. − first, we consider the distance between each father and the lower bound of the

lattice (this distance corresponds to the minimum number of edges between them).

80

− if one of the fathers' distance to the lower bound is smaller than the others, this node is selected. Being at lower distance from the lower bound means that this concept is semantically richer.

− if several nodes are at a minimum distance from the lower bound, we compare the sum of the weights of the properties contained in their intention. The node with the highest value is selected. − if several fathers meet this requirement, the algorithms chooses the one

which minimises the total number of branches in the tree. If this condition is not unique, different scenarios are considered, one for each possible father.

5.2 Clusters analysis

Once the tree of clusters is generated, different measures may be computed, e.g. the proportion of concepts of the initial lattice which were not selected to be clusters.

The depth of the tree is interesting because it indicates the number of navigation levels that may be provided to the user. We also study the distribution of clusters at each abstraction level. If a cluster has no father, it means that it cannot be generalised. On the other hand, a cluster with no children corresponds to the most specific level.

We may also compute the distances between clusters. The distance between two clusters may be the average – or minimum, or maximum – distance between two objects (one in each cluster). Let O1 and O2 be two objects. Let P1 be the set of properties of O1 and P2 the set of properties of O2. Let INTER be the intersection of P1 and P2, and UNION the union of P1 and P2. The similarity between O1 and O2 is defined as:

∑

∑

=

== )(

1

)(

1

')2,1( UNIONcard

j

INTERcard

i

jw

wiOOS

(9)

The distance between O1 and O2 is given by (2):

)2,1(1100)2,1( OOSOOD −= (10)

5.3 Clusters representation

The levels of detail are symbolised by different colours. At each abstraction level, clusters are represented by portions of a disk, as shown in figure 6. Each cluster's size is proportional to the number of children this concept has. When the pointer of the mouse is over a cluster, its extension – the set of topics contained in this cluster – or intention – the set of these topics' properties – is displayed. When the user clicks on a part of the disk, this cluster becomes the current context – i.e. the whole disk -

81

and its content is displayed in greater detail. The disk in the upper left corner represents a global view of the topic map before focusing on a specific cluster.

The figure 6 shows the results of this clustering algorithm on our example topic map about music. These representations are SVG (Scalable Vector Graphics) graphics [15]. SVG is a language for describing two-dimensional graphics in XML. SVG drawings can be interactive and dynamic. SVG leverages and integrates with other W3C specifications and standards efforts. By leveraging and conforming to other standards, SVG becomes more powerful and makes it easier for users to learn how to incorporate SVG into their Web sites.

Fig. 6. Clusters visualisation

6 Conclusion and further work

This article presented how XML topic maps could be exploited to help users find relevant information on the Web. This contribution is at several levels: first, we characterise Web sites by defining their profile. This may be used to evaluate Web sites relevance with regard to a specific query. Second, our analysis identifies topics that have no interest – semantically speaking – which allows to "clean" the topic map. Finally we showed how we could enhance navigation by clustering Web pages and displaying them with different levels of details.

82

These results were deduced from the analysis of Galois lattices generated from Web sites with a conceptual classification algorithm. This algorithm is very powerful as it groups topics semantically.

In the future, we will study Web sites clusters in more details. For example, we noticed that some of the clusters are less relevant than others; it may thus be possible to further filter the Web site if it is really too large.

We will also investigate how ontologies may be used to characterise our clusters. Galois algorithm generates clusters which have a semantic value without expressing this semantics explicitly. Ontologies may help us make this information explicit.

7 Références

1. Barbut, M., Monjardet, B., Ordre et classification, Algebre et combinatoire, Tome 2, Hachette, 1970.

2. Berners-Lee, T., A roadmap to the Semantic Web, http://www.w3.org/DesignIssues/Semantic.html, Sept 1998.

3. Birkhoff, G., Lattice Theory, First Edition, Amer. Math. Soc. Pub. 25, Providence, R. I., 1940.

4. Carpineto, C., Romano, G., Galois: An order-theoretic approach to conceptual clustering, Proc. Of the 10th Conference on Machine Learning, Amherst, MA, Kaufmann, pp. 33-40, 1993.

5. Chein, M., Mugnier M.-L., Conceptual Graphs : Fundamental Notions, Revue d'intelligence artificielle, Volume 6 - n°4/1992, pp365-406, 1992.

6. Garshol, L. M., "tolog" – A Topic Map Query Language, XML Europe 2001, Berlin, Germany, 21-25 May 2001.

7. Godin, R, Chau, T.-T., Incremental concept formation algorithms based on Galois Lattices, Computational intelligence, 11, n° 2, p246 –267, 1998.

8. International Organization for Standardization, ISO/IEC 13250, Information Technology-SGML Applications-Topic Maps, Geneva: ISO, 1998.

9. Moore, G., RDF and Topic Maps, An Exercise in Convergence, XML Europe 2001, Berlin, Germany, 21-25 May 2001.

10. Sowa, J. F., Conceptual Information Processing in Mind and Machine, Reading, Massachusetts, Addison-Wesley, 1984.

11. TopicMaps.Org XTM Authoring Group, XTM: XML Topic Maps (XTM) 1.0: TopicMaps.Org Specification, 3 March 2001.

12. Wille, R., Line diagrams of hierarchical concept systems, Int. Classif. 11, pp. 77-86, 1984.

13. Wille, R., Concept lattices and conceptual knowledge systems, Computers & Mathematics Applications, 23, n° 6-9, pp. 493-515, 1992.

14. World Wide Web Consortium, Resource Description Framework (RDF) Model and Syntax Specification, W3C Recommendation, 22 February 1999.

15. World Wide Web Consortium, Scalable Vector Graphics (SVG) 1.0 Specification, W3C Candidate Recommendation, 2 November 2000.

16. Woods, W.A., What's in a link: foundations for semantic networks, In D.G. Bobrow and A.M. Collins, (Eds.), Representation and Understanding: Studies in Cognitive Science,. New York: Academic Press. p. 35-82, 1975.

83

http://www.w3.org/DesignIssues/Semantic.html

workshop semantic web mining

Documents