2nd workshop semantic web mining

13th European Conference on Machine Learning (ECML'02) 6th European Conference on Principles and Practice of Knowledge

Discovery in Databases (PKDD'02)

2nd Workshop Semantic Web Mining

Bettina Berendt, Andreas Hotho, Gerd Stumme

August 20, 2002 Helsinki Finland

Foreword

The Semantic Web and Web Mining are two fast-developing research areas,which have many points in contact. A growing number of Web resources aresemantically annotated, usually in XML or RDF. Data mining algorithms arebecoming standard modules in data analysis packages, and their application isbecoming a basis for Web reporting and controlling measures. However, inte-grated theories and software solutions are still lacking. 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.

Web mining applies data mining techniques on Web content, usage, and struc-ture. Methods of Web content mining can, on the one hand, be used to createsemantic annotations from Web page content; on the other hand, content miningcan profit from content that is already structured in XML, RDF, or ontologicalformat. Methods of Web usage mining can profit from semantically enricheddescriptions of the Web pages visited; this will provide for the identificationof more meaningful patterns within site visits, and better site improvements,recommendations, and personalization options based on these patterns. Usagepatterns can in turn serve to improve the semantic annotations of pages. Thethird form of Web Mining, Web structure mining, utilizes the hyperlink struc-ture. Crawlers that take into account structure as well as semantic content cansignificantly improve search engine results. The rapid development of the fieldmeans that Semantic Web Mining now plays a wide range of different roles.

Our workshop aims to bring together researchers and practitioners from theinvolved fields, and to foster the exchange of ideas, the discussion of currentlyproposed solutions, and the establishment of an agenda for further emerging is-sues. In the contributions to this workshop, four categories can be distinguished:

Two invited papers give an overview on different uses of data mining tech-niques within Semantic Web Mining: Claire Nedellec discusses the use of Ma-chine Learning and Information Extraction for annotating text, with the aimto build the Semantic Web. The approach is illustrated by an example takenfrom genomics. Nada Lavrac discusses Inductive Logic Programming techniquesfor learning from the Semantic Web. She focusses especially on Relational DataMining and Subgroup Discovery.

The invited talks are complemented by three full papers which focus on basicresearch: Andreas Faatz and Ralf Steinmetz explain how an existing ontologycan be enriched by mining the web. The authors gather a text corpus using theGoogle search engine, and compare it with the concepts of the ontology based

on statistical information on word usage. Honghua Dai and Bamshad Mobasherdiscuss a Semantic Web Usage Mining approach. They present a way of usingdomain ontologies to automatically characterize usage profiles. Multi-facetedlearning for building web taxonomies is the topic of the paper by Wray Buntineand Henry Tirri. They present results for multi-faceted clustering of bigramwords.

Two application papers discuss the use of existing Data/Web mining meth-ods to Semantic Web related problems/data: Chung-Hong Lee and Hsin-ChangYang consider multilingual text corpora and cluster them using self-organizingmaps. The documents are re-categorized based on their semantic relatedness inthe corpus. Peter Edwards, Gunnar AAstrand Grimmes, and Alun Preece em-pirically investigate the hypothesis that learning from the Semantic Web willoutperform traditional learning from today’s Web. They compare the perfor-mance and accuracy of ILP against k-Nearest Neighbor and Naıve Bayes on twodatasets.

Lastly, an important aim of this workshop is to promote the exchange ofideas for the benefit of ongoing research efforts. This is reflected in four projectdescriptions.

With this collection of research papers, we aim to further the convergence ofthe Semantic Web and Web mining. We wish to express our appreciation to theauthors and to the members of the program committee, for making the workshopa valuable contribution to Semantic Web Mining.

Last but not least we wish to thank KDNet for the support provided for theworkshop. Our aim is to use this workshop as a first discussion forum for furtheractivities within KDNet. We want to clarify and discuss further interests andneeds concerning Semantic-Web-related activities within KDNet. These couldinclude the organization of further workshops, as well as the participation inproject initiatives, mailing lists and community fora. Of major interest in thisdirection is the initiative of the Web Mining Forum as planned by KDNet.

July 2002 Bettina BerendtAndreas HothoGerd Stumme

Program Committee

Stefan Decker(Stanford University, Palo Alto)Ronen Feldman(Bar-Ilan University, Ramat Gan)Stefan Haustein(Universitat Dortmund)JÄorg-Uwe Kietz(kd labs ag, Zurich)Nada Lavra·c(Jozef StefanInstitute, Ljubljana)Klaus-Peter Huber(SAS, Heidelberg)Alexander MÄadche(FZI Research Center for InformationTechnologies at the University ofKarlsruhe)

Bamshad Mobasher(DePaul University, Chicago)Claire N¶edellec(Universite Paris Sud)John R. Punin(Rensselaer Polytechnic Institute,Troy, N.Y.)Myra Spiliopoulou(Handelshochschule Leipzig)Rudi Studer(Universitat Karlsruhe)Stefan Wrobel(Rheinische Friedrich-Wilhelms-Universitat Bonn)

Table of Contents

Invited Talks

Machine Learning applied to Information Extraction in specific domainsan example, gene interaction extraction from bibliography in genomics . . . . 1

C. Nedellec

Relational Data Mining and Subgroup Discovery . . . . . . . . . . . . . . . . . . . . . . . 8N. Lavrac

Research Papers

Ontology Enrichment with Texts from the WWW . . . . . . . . . . . . . . . . . . . . . . 20A. Faatz, R. Steinmetz

Using Ontologies to Discover Domain-Level Web Usage Profiles . . . . . . . . . . 35H. Dai, B. Mobasher

Multi-faceted Learning for Web Taxonomies . . . . . . . . . . . . . . . . . . . . . . . . . . . 52W. Buntine, H. Tirri

Application Papers

Acquisition of Web Semantics based on a Multilingual Text Mining Technique 61C.H. Lee and H.C. Yang

An Empirical Investigation of Learning from the Semantic Web . . . . . . . . . . 71P. Edwards, G. A. Grimmes, A. Preece

Project Descriptions

Semantic Methods and Tools for Information Portals - The SemIPort Project 90J. Gonzalez-Olalla, G. Stumme

Basic Techniques for the Extraction and Annotation of Machine Under-standable Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

M. Kunze, D. Rosner

Semantic Web and Retrieval of Scientific Data Semantics . . . . . . . . . . . . . . . 92G. Soldar, D. Smith

Data Fusion and Semantic Web Mining: Meta-Models of Distributed Dataand Decision Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

V. Gorodetski, O. Karsaev, V. Samoilov

1

Machine Learning applied to Information Extraction in specific domains — an example, gene interaction extraction from bibliography in genomics

Claire Nédellec

Laboratoire Mathématique, Informatique et Génome (MIG), INRA

[email protected]

1. Introduction As well as the generalization of multimedia communication, the volume of textual information is exponentially increasing. Today mere Information Retrieval technologies are unable to meet specific information needs because they provide information at a document collection level. Developing intelligent tools and methods, which can give access to document content, is therefore more than ever a key issue for knowledge and information management. Text content access is a crucial issue as much in the document engineering system of a small firm as in the document management of a whole scientific domain, whichever the source of information is: an Intranet information system or the "semantic web". As soon as one wants to automate access to the content of texts in electronic form, one needs semantic knowledge to localize and interpret the relevant information. The acquisition of semantic knowledge is a well-known bottleneck for real-world applications, whichever technology is used (Information Extraction, Question/Answering, and more generally document engineering). There are two main reasons. Firstly, little semantic knowledge specific to application domains has been available because, until now, effort has been mainly devoted to the definition of formal languages for the representation of ontology and to the acquisition of generic knowledge bases, either lexical databases such as WordNet or EuroWordNet or general ontologies; CYC, for instance. In contrast, almost no community effort has been devoted to the acquisition of specific semantic knowledge that is required for particular applications and to the design of the acquisition methods that could be applied. We claim that no generic knowledge can be used as such and that the required semantic knowledge, even if it is derived from a generic source, must be specifically tuned to the application, domain and task that it will be used for. Although the process of acquiring this specific semantic knowledge cannot be fully automatic, methods and tools can be designed to efficiently help its acquisition. Secondly, it is also noticeable that there has been little dialogue between the various disciplines involved in knowledge acquisition and text analysis, although the integration of methods and tools from various disciplines is obviously needed. These disciplines include Information Science, Linguistics, Natural Language Processing, Knowledge Acquisition, Knowledge Representation, Machine Learning, Information Retrieval and Information Extraction. The Caderige project (http://www-caderige.imag.fr) is an example of such a collaboration in the domain of functional genomics. It involves four French laboratories, IRISA (ML and NLP), LIPN (KA, KR and NLP), LRI (ML) and MIG (genomics, ML and IE) and more recently a biotechnology company, Hybrygenics. After sequencing, the next challenge in genomics is identifying the role of genes in interaction networks. Genome research projects have resulted in new experimental approaches, such as using DNA chips, at the level of whole organisms. Such chips provide comprehensive data about gene activity, so a research team can quickly produce thousands of measurements. More than ever, these new lab technologies are calling for fast and efficient access to previous results to interpret elementary measurements from the laboratory. Unfortunately, most functional genomics knowledge is not described in databanks; it is only available in scientif ic abstracts and articles written in natural language. For instance, the main generalist bibliographic database, Medline, contains approximately 12 millions entries. Efficiently using previous research results requires automating access to bibliography content. Therefore, exploring bibliographies and extracting knowledge from literature is a major milestone toward developing functional models of gene interactions.

1

2

In our opinion this new challenge offers as many benefits and present the same level of technical difficulty as other more popular bioinformatics challenges such as designing predictive algorithmic models. Moreover, AI research in natural language processing (NLP), information extraction (IE), machine learning (ML), and genomics have now reached the stage where automating IE from genomics literature is a realistic and exciting research goal. The specificity of the genomics bibliography, compared to other domains, justifies the expectation for short-term and high-quality results. We will illustrate this claim in the following by a genomic example about information extraction of gene interaction in Bacillus subtilis.

2. An example of the IE problem in genomics Biologists can search bibliographic databases via the Internet using keyword queries that retrieve a

large superset of relevant papers. Alternatively, they can navigate through hyperlinks between genome databanks and the corresponding papers. To extract the requisite gene interaction knowledge from the retrieved papers, they must identify the relevant fragments (see the bold text in Figure 1). Such manual processing is time consuming and repetitive, because of the bibliography size, the relevant data sparseness, and the database continuous updating. UI - 99175219

AB - Ger E i s a t r anscr i pt i on f act or pr oduced i n t he mot her cel l compar t ment of spor ul at i ng Baci l l us subt i l i s . It is a critical regulator of cot genes encoding

proteins that form the spore coat late in development. Most cot genes, and t he ger E

gene, ar e t r anscr i bed by si gmaK RNA pol ymer ase. Previously, it was shown that the GerE protein inhibits transcription in vitro of the sigK gene encoding sigmaK. Her e, we show

t hat Ger E bi nds near t he si gK t r anscr i pt i onal st ar t si t e, t o act as a r epr essor [ …]

Figure 1. An extract of a Medline abstract on transcription in Bacillus subtilis.

Type: negative Agent: GerE protein

Interaction

Target: Expression Source: sigK gene Product: sigmaK protein

Figure 2. Information extracted from the second selected fragment

For example, the query “Bacillus subtilis and transcription” retrieves 2,209 abstracts such as the one of Figure 1. We chose this query example because Bacillus subtilis is a model bacterium and transcription is a central phenomenon in functional genomics. Gene functions are realized through gene transcription and protein production. The example of Figure 1 represents the problems posed by applying IE to a bibliography in genomics. Extraction involves understanding and requires expertise in biology. The information to be extracted is sparse in the document set. For instance, in the set of 2,209 abstracts I mentioned, only 3 percent of the sentences contain relevant information on gene interaction—that is, text that mentions the interaction’s agents and type. Hopefully, in biology the bibliography is well structured and the information is local, mainly located in a single sentence or in a part of it as opposed to other domains where it is spread over the document. Many other biological phenomenon, such as translation or gene homology, raise similar IE problems.

3. L imitations of usual IE methods Up to now, DARPA’s MUC (Message Understanding Conference) program has defined automatic IE

as the task of extracting specific, well-defined types of information from natural language texts in restricted domains. The objective is to fill predefined template slots and databases, such as shown in Figure 1b. In functional genomics, even such a restrictive view of IE is useful. Until now, no operational

2

3

IE tool has been made available in genomics, and extraction has not been automated. However, applying IE à la MUC to genomics and more generally to biology is not an easy task because

deep text analysis methods are needed to handle the relevant fragments. IE systems should combine the semantic-conceptual analysis of text understanding methods with IE through pattern matching, [Thomas et al., 2000], [Blaschke et al., 99], [Sekimizu et al., 98], [Ono et al., 2001]. Indeed, IE approaches to genomics, based either on predefined sets of f ixed patterns, or on shallow representations of the text, yield limited results with either a bad recall or a low precision.

Hand-coded sets of patterns based on significant interaction verbs, gene names, or even syntactic tags and dependencies, [Blaschke et al., 99], [Thomas et al., 2000], [Ono et al., 2001], retrieve little high-quality information. Our experiments with such patterns (described in the IE literature in genomics)—for example, [(Protein1/Gene1) *1 (interact/associate/bind) * (Protein2/Gene2) * ]—yield a precision around 98 percent with a recall between 0 and 20 percent. The reason is that, even in technical and scientif ic domains, there are many ways to express given biological knowledge in natural language. Manually encoding all patterns encountered in a corpus is thus unfeasible due to cost and unreliability. Therefore, automatically learning such IE patterns or rules from corpus seems to be an appropriate solution. Additionally, building IE systems is time consuming if they rely on manually encoded dictionaries and extraction rules or patterns that are specific to the domains and tasks at hand and they are not easily portable.

At the opposite end, some methods are based on statistic measures of keywords and gene name co-occurrences, [Craven, 99] (for example, shallow information-retrieval-based techniques), [Blaschke et al., 99]. They yield high recall and low precision because they assume that any pair of genes encountered in the retrieved sentences interact, which is not always true. Many false positives are thus retrieved because potentially discriminant keywords and gene names occur in sentences where the genes mentioned are not semantically related. The following example (Figure 3.a and 3.b) illustrates some of the problems encountered by both hand-coded patterns and statistic –based approaches. Figure 3.a gives an example of a sentence that cannot be handled by these approaches and Figure 3.b represents the correct gene interaction network that should be extracted from this sentence.

"GerE st i mul at es cotD t r anscr i pt i on and i nhi bi t s cotA t r anscr i pt i on i n vi t r o by sigma K

RNA pol ymer ase, as expect ed f r om i n vi vo s t udi es, and, unexpec t edl y, pr of oundl y i nhi bi t s i n vi t r o t r anscr i pt i on of t he gene ( sigK) t hat encode sigma K. ". The sentence describes five interactions, s i gma K with cotA and cotD, Ger E with cotD, with cotA and with s i gK.

Figure 3.a An example of sentence that cannot be handled by hand-coded patterns and pure statistic–based approaches

An intuitive pattern, such as the one mentioned above, (i. e. [(Protein1/Gene1) * (interact/associate/bind) * (Protein2/Gene2) * ]), that would match any pair of gene or protein names and interaction verbs or nouns (framed in the figure 3.a), [Craven & Kumlien, 1999], [Blaschke et al., 99], would retrieve many erroneous interactions from this sentence, such as cotD [...] inhibits [...] cotA. Additional criterion such as a maximum number of words between gene names would yield a better precision but would miss some interactions such as the inhibition of sigK gene transcription by GerE (28 words apart). Statistics and keyword-based approaches would select the relevant sentences but would not be able to determine the right interactions between the five different gene and protein names cited in Figure 3.a (in bold-faced text).

1 * matches any string of any length (including zero).

3

4

POSITIVE

INTERACTIONGer E

transcription Agent cot DAgent cot Atranscription

Sigma K

NEGATIVE

INTERACTIONRegulator

POSITIVE

INTERACTION

Target

Factor

Agent sig Kexpression

Target

ProductTarget

POSITIVE

INTERACTION

Target

Target

Factor

Regulator

denotes a process

Figure 3.b The gene interaction network to be extracted.

Extracting relevant knowledge in the selected documents thus requires deeper syntactic and semantic analysis based on lexical and semantic resources specific to the domain. For instance, in Figure 2.a, identifying that GerE is the subject of the verb “ inhibit” and that sigK is its direct object, and given that these relations are compatible with the conceptual agent and target roles, would improve the extraction’s quality. To summarize, extraction patterns should be learned, because their manual development is unfeasible, and the learning should be based on syntactic–semantic regular expressions.

4. ML and IE today Since the beginning of the nineties, automatically learning extraction rules from examples of pairs of

filled patterns and annotated documents seemed like an attractive approach.6 However, by the end of the decade, people were questioning the relative merits of the trainable and the knowledge engineering approaches—Doug E Appelt and David J. Israel, for example, discussed this issue at an IJCAI-99 (Int’ l Join Conference on Artificial Intelligence) tutorial on IE (http://www.ai.sri.com/~appelt/ie-tutorial/). According to them, trainable (that is, statistics and ML-based) approaches should be preferred when the training data is cheap and plentiful, the extraction specifications are stable, and obtaining the highest possible performance is not a critical issue. They consider that the best recall the ML-based systems obtained is quite low compared to hand-coded IE systems. Appelt and Israel’s analysis is based on the current state of the art in IE, in which existing ML-based systems exploit little, if any, background knowledge for guiding learning. The systems are often applied to a rather shallow representation of the training texts, and most of them are based on general-purpose ML algorithms —mainly K nearest-neighbor, grammatical inference, naïve Bayes methods, and top-down or bottom-up relational learning based on an exhaustive search or a local information gain measure.

Two related facts explain the limited range of these approaches, despite the rich spectrum of the modern state of the art in ML. First, according to the limited experiments performed, [Freitag, 98], on the common and quite simple IE tasks (MUC tasks, IE on the job, and seminar announcements), approaches based on linguistic analysis, lexical semantics, and informative representation of the training data do not perform much better than more shallow approaches. This does not encourage the design and application of novel symbolic and relational ML methods, which would be suitable for richer text analysis. Second, until recently, the main stream in text processing was mainly linguistic and statistic but not ML-based, besides

4

5

some notable exceptions such as S. Soderland’s work and T. Mitchell’s group research, [Soderland, 99] [Freitag, 98]. A large part of the effort in learning for IE, including genomic applications, has also been devoted to lower-level tasks such as named entity recognition, [Fukuda, 98]. This situation is evolving with the growing interest of the ML community in text processing and in IE in particular. Moreover, the growing demand for applications brings many new IE tasks, such as IE in functional genomics, that require a deeper understanding and consequently call for more sophisticated linguistics- and ML-based approaches [Craven & Kumlien, 99]. Additionally, in real-world applications, training complements rather than opposes knowledge engineering, as ontology-based and interactive approaches illustrate.

5. L inguistics- and ML-based approach of IE in future genomics In Caderige, we view the genomics IE of the future as a three-step method. In the first step, we select

the relevant textual fragments from all sentences in the papers, based on shallow criteria (for example, discriminant keywords or gene and protein names) to deal with the relevant data’s sparseness. In the second step, we build a representation of the content of the fragments using successive interpretation operations based on syntactic–semantic lexicon, [Sekimizu et al., 99], [Rindflesh et al., 2000], following a classical approach in text understanding. Figure 4 shows an example of this phase’s output. This step should involve terminology, ontologies, and predicate argument structures to label the relevant terms and syntactic dependencies with the appropriate concepts. In doing so, we rely on the fact that in the language of a given specific domain there exists strong syntactic regularities, which make it possible to build a semantic structure.

[cspAp ] [ directs ] [the expression of the cspA gene ]

DobjSubjectNprepN

Protein Production GenePositive_ interaction

Fragmen t

Semantic labeling

Noun Verb Det Noun Prep Det NounNoun

Syntactic parsing NP NP

Syntactic relation

Figure 4. Example of syntactic-semantic interpretation (NP denotes noun phrases and Dobj denotes the Direct Object).

Finally, we apply extraction rules (see Figure 5) to the resulting text interpretation to identify the relevant information and store it in a database in the suitable format, or to fill forms as in MUC case. In this example, the IE is realized by transducers designed by Intex software, that insert XML labels in the text fragments when the syntactic and semantic conditions are verif ied. For example, the transducer in Figure 5 says states conditions, among others, there must be a noun phrase, subject of the verb and representing a protein (denoted by variable $1), and a noun phrase, direct object of the interaction verb (denoted by variable $2), representing a gene expression (denoted by variable $3). The gray boxes represent subtransducers. If all conditions are true, then XML protein, interaction and gene expression tags should be inserted (for example, see <protein>, <interaction> and <gene_expression> tags in the figure).

<Protein>

</gene_ expression><protein> </protein>

<interaction> </interaction>

<gene_ expression>

Semantic Class :Positive interaction

[NP( )

1 1[Ver b

(2

)2

[ NP(

3)3]

POSITIVE INTERACTIONSubject( $2 ,$1 )

Dobj($2 ,$3 )<Gene

expression>] ]

Figure 5. Example of extraction rule in the form of transducers for extracting gene interactions in functional genomics.

ML methods can help develop the knowledge bases needed for each step. For sentence filtering,

5

6

discriminant keywords are learnable by classification methods such as naive Bayes or Support Vector Machines, [Marcotte et al., 2001], [Nedellec et al., 2001]. For building terminologies and ontologies from parsed corpora or assisting their design, unsupervised methods such as conceptual clustering are appropriate [Nedellec & Faure, 98]. The many methods designed for semantic class learning, query expansion, word sense disambiguation, or for building restrictions of selection are easily applicable to ontology and subcategorization frame learning. Then, predicate argument structures are learnable from subcategorization frame clustering or from semantically labeled corpora. Finally, extraction rules or automata (see Figure 5) are learnable from annotated corpora (see Figure 6) at the suitable level of linguistic interpretation (see Figure 4), [Sasaki & Matsuo, 2000]. The feasibility of such learning tasks from parsed corpora has been shown many times in the framework of specific domains such as scientif ic ones. <SENTENCE name = "2" > <INTERACTION

id = "1" type = "Y" Previous studies showed that <Agent1 type=" Protein" func=" Factor" > spoII ID </Agent1> <Interaction> is needed to produce </Interaction> <Target1 type="SigmaFactor">sigma K</Target1> [...]

</INTERACTION>

Figure 5. Example of annotated sentence for IE rule learning. The highlights indicate the graphic attributes of the XML tags. For example, the regions tagged as "Interaction" are underlined, and the regions tagged as Agent are in bold. Superficial approaches will not sufficiently resolve the problem of building IE systems for genomics. However, given the specificity of the language used in genomics texts, we can solve the task by combining ML from a corpus of annotated and un-annotated texts with syntactic–semantic analysis. Genomics provides demanding problems that will stimulate the development of more sophisticated approaches in IE. Although these aspects of extraction are not yet in the mainstream of IE research, this seems a promising direction not only for genomics but more generally for biology and for other perhaps less technical domains. Preliminary work in this area has produced encouraging results that we should now extend and deepen.

Acknowledgments The author thanks her collaborators in the Caderige project, in particular Philippe Bessières, Gilles Bisson, Adeline Nazarenko and Mohamed Ould Abdel Vetah for their contribution to the scientific program presented in the paper. Caderige project is partially funded by the French inter-EPST program in bioinformatics. This paper is a completed and revised version of the paper published by the IEEE Intelligent System Journal (May-June 2002).

References Blaschke C., Andrade M. A., Ouzounis C. and Valencia A., “Automatic Extraction of Biological Information from Scientific Text: Protein-Protein Interactions,” Proceedings of the International Symposium on. Molecular Biology (ISMB’99), AAAI Press, USA pp. 60-67, 1999.

Craven M. and Kumlien J., “Constructing Biological Knowledge Bases by Extracting Information from Text Sources,” Proceedings of the 7th International Conference on Intelligent Systems for Molecular Biology (ISMB-99), pp. 77-86, AAAI Press, USA, Heidelberg, Germany, 1999.

Faure D. and Nédellec C.,"Knowledge Acquisition of Predicate-Argument Structures from technichal Texts using Machine Learning" in Proceedings of Current Developments in Knowledge Acquisition: EKAW-99, p. 329-334, Fensel D. & Studer R. (Ed.), Springer Verlag, Karlsruhe, Germany, April 1999.

Freitag D., “Multistrategy Learning for Information Extraction,” Proceedings of the 15th International Machine

6

7

Learning Conference (ML'98), A. Danyluk (ed.) Morgan Kaufmann, pp. 100-107, Madison, Wisconsin, 1998.

Fukuda K., Tsunoda T., Tamura A. and Takagi T., “Toward Information Extraction: Identifying Protein Names from Biological Papers,” Proceedings of the 3d Pacific Symposium on Biocomputing (PSB’1998), 3:705-716 (http://www-smi.stanford.edu/projects/helix/psb98/), 1998.

Marcotte E. M., Xenarios I., and Eisenberg D., “Mining Literature for Protein-Protein Interactions,” Bioinformatics Journal, vol. 17, no. 4, pp. 359–363, Oxford University Press Applications, 2001.

Nédellec C., Ould Abdel Vetah M., and Bessières P., “Sentence Filtering for Information Extraction in Genomics: A Classification Problem,” Proceedings of the International Conference on Practical Knowledge Discovery in Databases (PKDD’2001), pp. 326–338, Springer Verlag, LNAI 2167, Freiburg, September, 2001.

Ono T., Hishigaki H., Tanigami A. and Takagi T., “Automated Extraction of Information on Protein-Protein Interactions from the Biological Literature,” Bioinformatics Journal, vol. 17, no. 2, pp. 155–161, Oxford University Press Applications, 2001.

Riloff E., “Automatically Constructing a Dictionary for Information Extraction Tasks,” Proceedings of the 11th National Conference on Artificial Intelligence (AAAI-93), pp. 811–816, AAAI Press/The MIT Press, Cambridge, Mass., 1993.

Rindflesh T., Tanabe L., Weinstein J.N. and Hunter L., “EDGAR: Extraction of Drugs, Genes and Relations from the Biomedical Literature,” Proceedings of the 5th Pacific Symposium on. Biocomputing (PSB’2000), (http://www-smi.stanford.edu/projects/helix/psb00/), pp. 514–525, 2000.

Sasaki Y. and Matsuo Y., “ Learning Semantic-Level Information Extraction Rules by Type-Oriented ILP,” Proceedings of the 18th International Conference on Computational Linguistics (COLING-2000), Morgan Kaufmann, Saarbrüken, Germany, 2000

Sekimizu T., Park H. S. and Tsujii J., “ Identifying the Interaction Between Genes and Gene Products Based on Frequently Seen Verbs in MedLine Abstracts,” Genome Informatics, pp. 62–71, Universal Academy Press, Tokyo, Japan, 1998.

Soderland S., “Learning Information Extraction Rules for Semi-Structured and Free Text,” Machine Learning Journal, vol. 34, pp. 233-272, 1999.

Thomas, J., Milward, D., Ouzounis C., Pulman S. and Caroll M., “Automatic Extraction of Protein Interactions from Scientific Abstracts,” Proceedings of the 5th Pacific Symposium on Biocomputing (PSB’2000), vol. 5, pp. 502–513, http://wwwsmi.stanford.edu/projects/helix/psb00/, 2000.

7

!"#$%&(')*+,

-/.102.435.7698:.<;=>@?2ACBEDGFIHJ4KLJ@MNBEOPBEQ1BEDSR@>7HUTWVXUHYUZ1R\[:]S]U]W^1_NQ@`9a _bHJcHdR2A1aPVX7D:J@OeH

f1gUhdg\iNj7gUk7lCgCmUndoSp7q\ibq7o

rsut:v:wxcyUvSz KJKLJ9|@Q9~BEOeXSDW^VUSOPuEVSHUTWTWOPJ9IKL^\5~R\BE@DEDG:D:JCBMb9OFBVUFHBbBED:JCBEOeVUJ FEVST@EVSHUTMN@JCBE@D:MbOPM5BEV1J@VaeD:|@SD|@OPMb:VX7D~bED:MbQ9aBED|OPJ HS|1XUHUJ@:D| EDGaHBEOPVSJcHa|@HBHTWOPJ9OPJ@BED:9J@OedQ@D:MBEcHBHED)1HUGBEOPHaea4HU9@aPOe:HU`9aPDFVU|@OPMb:VX7D~EOeJ@dJ@VaPD|1SDOPJED:aeHBEOeVUJcHa@|@HBH`cHUMbDGM:?S@OPMcHDGUOeXSD:MuH`@EOPDGFOPJCBbEVd|@Q@GBEOPVSJ4BEVWKN^R91ED:MbD:JCBEMBE@D)MNBHBEDGVFBE9DGHbBKL^\BEDG~@J9OeCQ9D:M!FVEDGaeHBEOPVUJcHUa@dJ9VaPD|@UD|@OPMb:VX7DGbHUJ9|/VSQ1BEaPOeJ@D:MuED::D:JCBH9@EVSHU9D:MuBEVEDGaHBEOPVSJcHaMbQ@`9UEVUQ9|1OPMb:VX7DGbC?

¡9¢£¤2¥)¦/§¨\£2©E¥¢

ªb«<0¬ =~® 61¯°±@² ®=³ 8~±@²18:.d´4´ ® «2²¶µªE35·¸/¹ º@º2»º@¼2»~½1¼2»¾7½7¿ ®À ."8G¯ À ¯U.d8 =:Á .18~¯S. ~Á . +Á . À®GÀ8~±c± :À+® « ® «<0¬ =~® 61¯"´4. =:Á2® «2¯ °¯U.18~« ® «2²Â.1«<0°±1² ®=W³ 8~±@²18:.d´4´ ® «2²Ã<ªE35·Ä8~¯ À ¯S.d8 =:Á . ® ´ À. .ÆÅ±@8~´.d°(Å8:.d´4¯Ç±18GÈ. À Ç¯U°°!. À³ 8:. =~®= .1°.1°²@±18 ®eGÁ ´ À Å±@8 ® «<0¬ =~® 61¯4°¯U.18~« ® «2²±dÅ8~¯U°. ~® ±1«<.1°c02¯ ÀG= 8 ®³G® ±@« À(~Á . !EÉc³2®= .d°° ÉÁ .761¯ ~Á ¯ÊÅ±18G´Ë±dÅ°±@² ®=³ 8~±@²18:.d´ À ÃCÌ<8~±@´Ë°±@² ®=³ 8~±@²18:.d´4´ ® «<²<»<ªE35· Á . À® « Á ¯U8 ® ¯S0 ®GÀÀ ±1¬2«<0 ~Á ¯U±18G¯ G®= .d°(Í<. À~®À ».d«<0ÎÅ8~±@´Ï´. =:Á2® «<¯°¯S.d8G« ® «<²<».1«Ð¯Ñ ³ ¯U8 ® ´4¯« .1°. ³<³ 8~±9. =:Á .d«<0Ò±@8 ® ¯« . ~® ±1« ±CÇÊ.d8:0 À³ 8:. =~®= .1°. ³2³ ° ®= .dÓ~® ±1« À Ã1ªE35·Ô8~¯ À ¯S.d8 =:Á"Á . À Í\¯¯U« ÀE 8~±@«2²1° É ® «Õ<¬<¯« = ¯U0Æ.1° À ±WÍ ÉÖ ±@´ ³ ¬ . ~® ±1«<.1°2°¯S.d8G« ® «2²~Á ¯U±18 É »Ê.d«<0Ä8G¯ = ¯« ° É ».d° À ±×Í ÉÄØ «2±CÇ)°¯U02²1¯Ù ®ÀG= ±C6@¯8 É×® «Ù. .dÍ<. À ¯ À µ Ø ÙÙ¸¹¾SÚC¿Ç Á2®=:Á °¯U0 ± GÁ ¯0¯6@¯°± ³ ´4¯U« ±1Å5«2¯UÇ ¯ =:Á « ®Û ¬2¯ À Å±@8)8~¯U°. ~® ±1«<.1°Ü02. .´ ® « ® «2²<Ãªb«Ý²@¯«2¯U8G.1°»u.d«ÐªE3·°¯U.d8G«2¯U8 ®À ² ® 61¯U«Ý.1« ® « ®~® .d° GÁ ¯±@8 ÉÎÞ µÍ<. = Èc²18G±1¬<«<0Èc«2±CÇ)°Ó

¯U0²@¯S¸".d«<0 À ±@´4¯¯U6 ® 0¯U« = ¯ßàµ¯Ñ2.d´ ³ °¯ À ¸»!.1«<0 ®GÀ . ® ´ ®À" ± ® «0¬ = ¯. ~Á ¯U±18 Éâáµ ÁcÉc³ ± ~Á ¯ À®À ¸ GÁ . ) ±1²@¯ ~Á ¯8)Ç ®~ÁÞäã~åæçèCéIê<ëÊÀ ±1´4¯ ³ 8G± ³ ¯8 ~® ¯ À ±dÅ!ßÃ2ªb«´4± À+= . À ¯ À~Á ¯ ÁcÉc³ ± ~Á ¯ À®ÀWáìÁ . À ± À . ~®À Å ÉÐ= ¯8 . ® «Ð8G¯ À 8 ®=~® ±1« À »Ç Á2®=:Á Ç¯ ÀÁ .1°°!8G¯Å¯U8 ±Î. À.Ýí éè7ë ÃÜî ® . À+® « = °¬<0¯ À/³ 8 ® ±18/¯Ñ ³ ¯ = . ~® ±1« À .1«<0. À~À ¬2´ ³~® ±1« À ».d«<0 = .d« ~Á ¯U8~¯Å±18G¯ Í\¯= ±1« À® 0¯8G¯U0. À!GÁ ¯°±1² ®= .d°° É ¬2«CïE¬ À~®ð ¯U0 ³ .d8 ±1Å GÁ ¯Í. = Èc²18G±1¬2«0"Èc«2±CÇ)°¯U02²1¯1Ã@î ® . À®À«2¯¯S0¯U0 ±8~¯S0¬ = ¯ GÁ ¯«c¬2´"Í¯U8±dÅ = .d«<0 ® 02. ¯ ÁcÉ9³ ± GÁ ¯ À ¯ À ÃSª != ±1« À~®ÀE:À ±1Å ~Á ¯°.d«<²1¬<.1²1¯Í ® . Àñ »d02¯ ¯U8~´ ® « ® «2² ~Á ¯ Á9Éc³ ± GÁ ¯ À~®À!À~³ . = ¯1»d.1«<0 ~Á ¯ À ¯S.d8 =:Á Í ® . À Ç Á<®=:Á 8G¯ À 8 ®=GÀ(~Á ¯À ¯S.d8 =:Á ±dÅ ~Á ¯ À³ . = ¯±dÅ ³ ± À~À~® Í2°¯ ÁcÉc³ ± ~Á ¯ À ¯ À Ãò Á2®ÀÊ³ . ³ ¯8 ð 8 ÀE ² ® 6@¯ À .Í28 ® ¯Å ® « 8G±c02¬ =G® ±@« ±ªE3·.d«<0 ³ 8G¯ À ¯« GÀ . À ¯°¯ =~® ±1«¶±dÅ

8~¯ = ¯« ° É 0¯6@¯°± ³ ¯U0ÐªE3· ¯ =:Á « ®Û ¬2¯ À Å±188G¯°. G® ±@«<.d°0<. .´ ® « ® «2²×¹¾S½7¿N»uÅ±@°°±CÇ¯S0Í É.d«Î±@¬ ° ® «<¯"±1Å8~¯ = ¯« . ³2³ 8G±@. =:Á ¯ À) ±¶8~¯U°. ~® ±1«<.1° À ¬2Í2²18G±1¬ ³ 0 ®ÀG= ±C6@¯8 É Ã\ò Á ¯ ±C6@¯8G6 ® ¯Ç®À 8~¯ ÀE 8 ®= ¯U0 ± ¯ =:Á « ®Û ¬2¯ ÀÀ . ~®À Å É9® «2² GÁ ¯ ÀE 8~±@«2² = 8 ® ¯8 ® ±1«4Å±18G´ ¬<°. ¯U0Å±18´. =:Á2® «<¯°¯S.d8G« ® «<²Í É¶ó®=:Á2® ¯¹ º<¾¿ GÁ . 8G¯ Û ¬ ® 8G¯ À ¯Ñ ³ ° ®=®/ÀÉ ´"Í±@° ®= Å±18G´ô±1Å ® «<02¬ = ¯S00¯ ÀG= 8 ®³ Ó~® ±1« À Ã

8

£(£7¥U£ ¤£©~¢Ï¡

ò Á2®À À ¯ =~® ±1«ÒÍ28 ® ¯Õ ÉÝ® « 8G±0¬ = ¯ À Ç±Í<. À~®=GÁ ¯±@8~¯ ~®= .1° À ¯ ~® «2² À »Ü² ® 6@¯ ÀW³ ± ® « ¯U8 À ±À ¬ =U= ¯ À~À Å¬2°ÊªE3· . ³2³ ° ®= . G® ±@« À .d«<0 ³ 8~¯ À ¯U« GÀ 8~¯ = ¯« " ¯ =:Á «2±1°±1² ®= .d°Ê0¯6@¯°± ³ ´4¯« :À ® «~Á ¯.18~¯S.2» = . ¯²@±18 ® ¯U0 ® « ± GÁ ¯ Ç±4´. ® « ~Á ¯U±18G¯ G®= .d° À ¯ G® «2² À Ã

"!$#&%('*)+-,./)021*3"045.1#768 « ® «0¬ =G® 6@¯W°±@² ®=³ 8~±@²18:.d´4´ ® «<² . À È = .d«¶Í\¯Å±18G´4.1°° É 0¯ ð «2¯S0¶. À Å±1°°±CÇ À.9: 1<;7)6=> . À ¯ ±1Å5¯Ñ2.1´ ³ °¯ À ß> .Í<. = Èc²18G±1¬2«<0 ~Á ¯U±18 ÉÆÞ> .°.d«2²@¬<.d²@¯/Í ® . À)ñ×GÁ . 02¯ ð «2¯ ÀÊ~Á ¯ = °.d¬ À ¯ À .d°°±CÇ¯S0 ® « ÁcÉc³ ± ~Á ¯ À ¯ À> .«2± ~® ±1«¶±dÅ ã~åæçèCêè$?Lé*@Cê

A 1*6"B/= . ÁcÉ9³ ± GÁ ¯ À~®À)áDCñ Ç Á2®=:Á ¯Ñ ³ °. ® « ÀGÁ ¯W¯Ñ2.1´ ³ °¯ À ß Ç ®~Á 8G¯ À~³ ¯ =Ê ±~Á ¯ GÁ ¯±@8 ÉÆÞ Ã

ò Á2®À 02¯ ð « ®~® ±1««2¯U¯U0 À) ±4Í¯ ® « À .1« G® . ¯S0ÂÅ±@8+0 ®E ¯U8~¯U« )EÉc³ ¯ À ±1Å(ªE35· . À È À ¹ º@¼C¿NÃò Á ¯ ® « À .d« G® . ~® ±1«Ç ® °° = ±@« = ¯U8~« GÁ ¯8~¯ ³ 8~¯ À ¯U« . G® ±@«±dÅ 8:. ® « ® «2²¯Ñ2.d´ ³ °¯ À » ~Á ¯ =:Á ± ®= ¯±dÅ). Á9Éc³ ± GÁ ¯ À~®À °.1«2²1¬<.1²1¯.d«0Ð.1«Ô. ³2³ 8G± ³ 8 ® . ¯«<± ~® ±1«Ò±1Å¯Ñ ³ °.1«<. G® ±@«ÜÃî É ¯Ñ ³ °.dÓ«<. G® ±@«×Ç¯ Á ¯U8~¯¶8G¯Å¯8 ±Ò.d« . =U= ¯ ³2 .d« = ¯ = 8 ®e ¯8 ® ±@«â±dÅ ÁcÉc³ ± ~Á ¯ À ¯ À.95GÁ ¯ Á9Éc³ ± GÁ ¯ À~®À¯Ñ ³ °. ® « À~Á ¯Â02. . ® Å ®"À . G®Àð ¯ À . = ¯8 . ® «Ô¬ À ¯U8Ób0¯ ð «2¯U0 = 8 ® ¯8 ® ±1«ÐÇWÃ 8SÃ Ã GÁ ¯Â02. .2ÃFÝ¯4Ç ® °°0 ®ÀG= ¬ ÀGÀÀ ±@´4¯ Å±@8~´.d°. == ¯ ³ .d« = ¯ = 8 ® ¯U8 ® .Â¬ À ¯S0 ® «Ò0 ®GE ¯8G¯« ªE3· À ¯ G® «2² À »Í2¬ Ç¯W.d° À ±4«2¯U¯U0 ±Í\¯U.18 ® «´ ® «<0 ~Á . ªE3·. ® ´ À . )GÁ ¯ ® «<0¬ =~® ±1«±dÅ ÁcÉc³ ± ~Á ¯ À ¯ À~Á . .d8G¯¯Ñ ³ 8~¯ À~À ¯U0 ® «4.d«4¯Ñ ³ ° ®=®ÀÉ ´"Í±@° ®= Å±18G´¶» GÁ . = .d«Í¯)¯S. À~® ° ÉW® « ¯8 ³ 8G¯ ¯S0 Í É~Á ¯¬ À ¯8H7¯Ñ ³ ¯8 .1«<0Æ´. É= ±@« 8 ® Í2¬ ¯ ± GÁ ¯/Í\¯ ~ ¯U8¬<«<0¯8 À .d«0 ® «<² ±dÅ ~Á ¯ ³ 8G±1Í2°¯´.10208G¯ ÀGÀ ¯U0u» ® 0¯U.1°° É Å±@8~´ ® «2². ³<® ¯ = ¯±dÅ«2¯Ç Èc«2±CÇ)°¯U0²@¯0 ®ÀG= ±C61¯U8~¯S0Å8G±1´ GÁ ¯02. .2Ã

I J,K)L5M5.16ONP,ò Á ¯ ÀE . ¯ÓL±1ÅIÓ GÁ ¯Ób.d8 ªE3· À ¯ G® «2² À .d8G¯±C61¯8G6 ® ¯UÇ¯S0Í\¯°±CÇWÃ2Ì2±18 GÁ ¯W¬2«<0¯U8~° Éc® «<² GÁ ¯Ó±18 ÉÂÀ ¯U¯¹ º@¼2»ºQS¿NÃ2Ì2±18+. ³ 8:. =~®= .1° ® « 8G±0¬ =G® ±@« ±4ªE3· À ¯¯¹ ½1¼C¿NÃ

R!)B"1*025.1<;P)SM/UTWV ãYXdé*Z?LéI[Cã]\^ T ®ÀGÁ ¯´4± À= ±1´4´4±1«ÆªE35· À ¯ G® «2²»@±1Å ¯U88G¯Å¯8G8G¯U0 ±â. ÀÎêP@ V`_ èCça\^ T » ãGåæ\çèdê\è$?b@ V`c éIêPX$d&ZM?Lé*@Cê » XCéëMZ V é _ éIêè$?b@ V`c éIêX$d&ZM?é*@dê »±18 ë`? V @CêLe\^ T Ã·8G¯U0 ®=G® 6@¯4ªE3· ®À . ® ´4¯S0Ð. °¯S.d8G« ® «<²¶±1Å = °. À~À~®eð\= . G® ±@«Ð.1«<0 ³ 8~¯S0 ®=~® ±1«Ò8G¬2°¯ À Ãò Á2®À ªE35· À ¯ G® «<² EÉc³2®= .1°° É 8~¯ ÀE 8 ®=:À ß ±²@8~±@¬2«<0Å. =:À ».d«<0 á .d«0 Þô ± À ¯ GÀ ±dÅ0¯ ð « ®e ¯ = °.d¬ À ¯ À Ã2ò Á ¯ À 8 ®= «2± ~® ±1«±dÅ¯Ñ ³ °.d«<. ~® ±1« ® « GÁ2®À+À ¯ G® «2²Â¬ À ¬.d°° É 02¯«2± ¯ À= ±C61¯U8G.1²1¯.d«<08~¯ Û ¬ ® 8G¯ À ²@°±@Í<.d° = ±1´ ³ °¯ ¯U«2¯ ÀGÀ .d«0 = ±1« À~®À ¯U« =É Ãf °±@Í<.d° = ±1´ ³ °¯ ¯«2¯ À~À .d«0 = ±@« À®À ¯« =É¶® ´ ³ ° ®=®e ° É . À~À ¬2´4¯ GÁ ¯«2± G® ±@«±dÅ éIê7?Nãê&g

ë:é*@dê\èdç&ZY@h[7ã V è.e@ã 0¯ ð «2¯U0 . À Å±@°°±CÇ À Ã f ® 6@¯«Í. = Èc²18G±1¬2«0 GÁ ¯±@8 ÉÞ » ÁcÉ9³ ± GÁ ¯ À~®À5á .1«<0¯Ñ2.d´ ³ °¯ À ¯ ßÆ»2.d«¶¯Ñ2.d´ ³ °¯jikß ®À µ ® « ¯U« À® ±1«.d°° É ¸ = ±C6@¯8G¯U0Í ÉÂá® Å Þmláonp i@Ãq Éc³ ± ~Á ¯ À®Àá ®À µ²1°±1Í<.1°° É ¸ = ±@´ ³ °¯ ¯ ® ÅsritkÐßvu 9<Þwláxnp i9Ãq Éc³ ± ~Á ¯ À®Àá ®Àµ²1°±1Í.d°° É ¸ = ±@« À®À ¯« )® Åyrijk¶ßz 91Þml¶á|n p i9Ã

9

f ® 6@¯« ~Á ¯"8G¯ À 8 ®=~® ±1« ±0¯ ð « ® ¯ GÁ ¯±@8 ® ¯ ÀwpËáJlÞ »Å±18Ç Á2®=:Á~Á ¯U8~¯ ¯Ñ ®ÀGÀ.¬2« ®Û ¬2¯°¯S. À q¯8GÍ28G.1«<04´4±0¯° µ ¸»c.d«<0 ± ²@8~±@¬2«<0Æ. ±@´ À . À ¯Ñ2.1´ ³ °¯ À » ~Á2®À®À¯ Û ¬ ® 67.1°¯U« ±8~¯ Û ¬ ® 8 ® «2² GÁ . .1°°u¯Ñ2.d´ ³ °¯ ÀÊ® «ßvu .18~¯ 8G¬2¯ ® « µ Þml¶á ¸¹ º@¼C¿NÃî É 8G¯°.CÑ ® «2² GÁ ¯4«2± G® ±@«Ý±1Å¯Ñ ³ °.d«. ~® ±1« ±.d°°±CÇ ® « = ±1´ ³ °¯ ¯4.1«<0 ® « = ±1« À~®À ¯U«

~Á ¯U±18 ® ¯ À4~Á . À . ~®À Å É À ±1´4¯± GÁ ¯8. == ¯ ³ .d« = ¯ = 8 ®e ¯8 ® .Äµ ³ 8G¯U0 ®=G® 6@¯. =U= ¬28:. =É » À~® ²1Ó« ®eð\= .d« = ¯1» = ±1´ ³ 8G¯ ÀGÀ~® ±@«¸» ~Á ¯ ³ 8G¯U0 ®=G® 6@¯ªE35· À ¯ ~® «2² = .d«ÒÍ¯¯Ñ ¯«0¯U0 ± ® « = °¬<02¯°¯S.d8G« ® «<²Ý±1Å = °. À~À~®eð\= . G® ±@« .1«<0 ³ 8~¯S0 ®=~® ±1« 8~¬<°¯ À Å8~±@´ ® ´ ³ ¯8~Å¯ = 02. .2». À Ç¯°°. Àçã:è V ê<éIêe @Wç@Ye1é*ZGèdç(X@ã Zéëé*@Cê ? V ã:ãë ¹e¾¿LÃªb«Î.Í28G±@.@0¯8 À ¯« À ¯1» ³ 8~¯S0 ®=~® 61¯WªE35· ® « = ±18 ³ ±1Ó8G. ¯ À .d° À ± V ë`?g @ V X1ã VV ã e V ãë:ëé*@Cê ¹ ½@½7¿.d«<0 Z @Cê<ë`? V èdéIê&?ÊéIêPXhd&Z?LéI[Cãç@e1é*Z/æ V @e V è _ _ éIêLe¹ C¿uÅ±18)Ç Á<®=:Á .d²9. ® «¶0 ®E ¯8G¯« . =U= ¯ ³2 .d« = ¯ = 8 ®e ¯8 ® .4. ³<³ ° É Ã

),K0h!$1<"5.1*;7) ãëMZ V é æ?éI[Cã \^ T ®ÀÀ ±1´4¯ ~® ´¯ À 8G¯Å¯U8~8G¯U0 ± . À Z @Cê V`_ è$?b@ V`céIêPX$d&ZM?Lé*@Cê » ê@Cê&g _ @Cê@$?b@Cêé*Z\^ T » X@ãëMZ V é æP?Lé*@Cê çãGè V ê<éIêe »±18 ã:è \Y^ T Ã!Ù/¯ ÀG= 8 ®³G® 6@¯ªE3· ®À ¬ À ¬<.1°° É . ® ´4¯U0. °¯U.d8G« ® «2²±dÅ = °.1¬ À .d° ~Á ¯±18 ® ¯ À ¹ Q7¿LÃdò Á<®À ªE35· À ¯ ~~® «2² EÉc³2®= .d°° É8~¯ ÀE 8 ®=:À)Þ ±Æ. À ¯ ±1Å02¯ ð « ® ¯ = °.1¬ À ¯ À » á ±4. À ¯ ±dÅ)µ²@¯«2¯U8G.1°¸ = °.1¬ À ¯ À ».1«<0ß ±³ ± À®~® 61¯¯Ñ2.1´ ³ °¯ À Ã2ò Á ¯ À 8 ®= «2± G® ±@«¶±1Å5¯Ñ ³ °.1«<. G® ±@«Â¬ À ¯U0 ® « GÁ2®ÀÀ ¯ ~® «2²48G¯ Û ¬ ® 8~¯ À~Á . .d°° = °.1¬ À ¯ ÀÊ® « á .18~¯ 8~¬<¯ ® « À ±1´4¯ ³ 8G¯Å¯U8~8G¯U0´4±0¯°±dÅ mp Þ l ß»@Ç Á ¯8G¯ ~Á ¯³ 8~¯Å¯8G8~¯S0W´4±0¯°9±dÅ ´4. É Í\¯1»SÅ±18 ® « ÀE .d« = ¯1» ~Á ¯°¯U. À q¯8GÍ28:.d«<0W´4±0¯U° µ ¸Ãcµ/«<¯´4. É .d° À ±Ð8~¯ Û ¬ ® 8G¯ ~Á ¯ = ±1´ ³ °¯ ¯«2¯ À~À .1«<0 ´ ® « ® ´.d° ®eEÉ ±1Å á »!Ç Á ¯8G¯ = ±@´ ³ °¯ ¯«<¯ ÀGÀ´¯S.d« À~Á . .Â´.CÑ ® ´.1°° É ²@¯«2¯U8G.1° ÁcÉc³ ± ~Á ¯ À®Àá ®À Å±1¬2«0u»\.1«<0´ ® « ® ´.1° ®EÉ ´¯S.d« À~Á . )~Á ¯ ÁcÉc³ ± ~Á ¯ À®À 0±c¯ À «<± = ±@« . ® «¶8~¯S0¬2«<02.1« )= °.1¬ À ¯ À Ã ¸î É 8G¯°.CÑ ® «2² GÁ ¯ À 8 ®= «2± G® ±@«Î±dÅ¯Ñ ³ °.1«<. G® ±@«Î¬ À ¯U0 ® « = °.d¬ À .d°0 ®À~= ±C61¯8 É ¹ Q7¿ ±

.d°°±CÇ Å±18 GÁ ¯±@8 ® ¯ À!~Á . À . ~®À Å É4À ±1´4¯)± GÁ ¯8. == ¯ ³ .d« = ¯ = 8 ®e ¯8 ® .µ À® ´ ® °.d8 ®eEÉ »1. ÀGÀ ± =® .dÓ~® 6 ®EÉ » ® « ¯U8~¯ ÀEG® «<²1«2¯ À~À ¸»0¯ À~= 8 ®³~® 61¯ªE35· = .d«¶Í¯W¯Ñ ¯«<02¯U0 ± ® « = ±@8 ³ ±18:. ¯ çã:è V êéIêLe@è7ë:ë[email protected]éè$?Lé*@Cê V dçãë ¹ ½7¿N» V ë`?g @ V X1ã V Zç dcë`?bã V éIêe ¹ ¼<»¾Sº2»~½1ºC¿N» X@è$?bè í è7ëã V ãë ? V d&ZM?d V éIêe¹e¾S¼2»c½7¿ ë`d í e V @hdSæ XdéëMZY@h[7ã V`c ¹ 9Ú7¿N» çã:è V êéIêLeMd<èCçPéI?bè$?LéI[7ã _ @KX1ãçë ¹ ½¾¿.d«0 ãMd<è$?Lé*@Cê XCéë`gZY@h[7ã V`c ¹e¾C¿NÃ

6 1'*'O,.5M!$45K1<;7) ) 4+ ('*) Ö ±1« À~® 0¯8Î. ³ 8G±1Í2°¯´ ±dÅ4°¯S.d8G« ® «<² Å.1´ ® ° É 8G¯°. ~® ±1« ÀÇ Á ¯8G¯ GÁ ¯ ³ 8~¯S0 ®=~® 61¯È9«<±CÇ)°¯S0²1¯Â0 ®ÀG= ±C6@¯8 ÉÎ . À È ®À ±Ð02¯ ð «2¯ ~Á ¯ .18~²@¯ 8~¯U°. ~® ±1«! #"$%&')(*,+.-/10 »1Ç Á<®=:ÁÎÀ . ¯ À)~Á . ³ ¯U8 À ±1« + ®À .Æ02.1¬2² Á9 ¯U8)±dÅ ³ ¯8 À ±@« / » ® « ¯8G´ À±dÅ8G¯°. G® ±@« À 0¯ ð «2¯U0 ® «Í. = Èc²18G±1¬2«0WÈc«2±CÇ)°¯S0²1¯ Þ Ã@3Ü¯ GÁ ¯ 8:. ® « ® «2² À ¯ ß = ±@« À~®À ±dÅ³ ± À®~® 61¯W.d«0Â«2¯U²@. ~® 61¯¯Ñ2.d´ ³ °¯ À Å±@8 ~Á ¯ .18~²@¯ )³ 8G¯U0 ®= . ¯ 2"$%&3'#(546 Ã 8 ³ ± À~®eG® 6@¯¯Ñ2.d´ ³ °¯i kÎßu ³ 8~±C6 ® 0¯ ÀÊ® «Å±@8~´. G® ±@«È9«<±CÇ)« ±ÆÍ\¯ 8G¬2¯.d«0 ÀÁ ±@¬2°0Í\¯¯U« . ® °¯S0Í ÉÐGÁ ¯ ® «<0¬ = ¯U0 ÁcÉc³ ± ~Á ¯ À®À Ã 8 «2¯²9. ~® 61¯Â¯Ñ2.d´ ³ °¯ i k ßtz ³ 8G±C6 ® 02¯ À"® «Å±@8~´. ~® ±1«~Á . +®À Èc«2±CÇ)«¶«2± ) ±4Í¯ 8~¬2¯.1«<0 ÀÁ ±@¬2°0«<± Í¯W¯U« . ® °¯U0uÃ

ßu p87 #"$)%&3')(*91 )(!:.-; #<!<=0 » #"!$%&3')(>*')?3'>-&3@A9B0#Cß z p87 #"$%!&3')(*,&@A9D-E F<!<10 » ! #"$%&')(*')?'>-E #<!<=0CÞ pG7 91@)&%5'#(* #<!<.-91 )(!:H0 » 91@)&)%5')(* 2<!<I-;&3@9B0 »2J #&%5')(*;&3@A9D-E'#?3'H0 »

J )&%5'#(*,&3@A9K-ML2 #<N0 »2J 'A91 !O'P* #<<N0 »#J 'A91 !O'P*Q91 #(!:10 »#J 'A9= O'P*'#?3'H0 »R )(3'#<!&*,+.-;/=0TSU91@#&%5')(*;+.-;/10 » R )('#<&*,+-;/10VS J )&%5'#(*,+.-;/=0 »91 O'P* R )&10 » 91 O)'P*,&3@A9W0C

10

ªNÅ ~Á ¯ ÁcÉc³ ± ~Á ¯ À®À °.1«2²1¬.d²1¯ ñ = ±1« . ® « À .1°°u0¯ ð « ® ¯ = °.d¬ À ¯ À ¬ À~® «<² GÁ ¯ ³ 8~¯S0 ®= . ¯.d«<0ÝÅ¬2« = ±@8 ÀÉ ´"Í±@° À . ³2³ ¯U.d8 ® «2² ® « GÁ ¯¯Ñ2.d´ ³ °¯ À .d«<0ÐÍ<. = È9²@8~±@¬2«<0È9«<±CÇ)°¯S0²1¯@»Ü.³ 8~¯S0 ®=~® 61¯ªE35· À~ÉÀE ¯´ = .d« ® «<0¬ = ¯ ~Á ¯Å±@°°±CÇ ® «<² = °.1¬ À ¯Å8G±1´ôßvuÊ»<ßtzâ.d«<0 Þ 9

! #"$%&')(*,+.-/10 S J 'A91 O' *,+10W- R )(3'#<!&*,/.-;+=0 Ã

8 ° ¯8G«<. ~® 61¯° É »2.°¯U.18~«2¯U8 = ±1¬<°0 Á .761¯ ® «<0¬ = ¯U0. À ¯ ±dÅ = °.d¬ À ¯ À.9! #"$%&')(*,+.-/10 S J 'A91 O' *,+10W- 91@)&%5'#(*,/.-;+=0 Ã! #"$%&')(*,+.-/10 S J 'A91 O' *,+10W- J )&%5'#(*,/.-;+=0 Ã

ªb«0¯ À~= 8 ®³~® 61¯/Èc«2±CÇ)°¯S0²1¯0 ®À~= ±C61¯8 É »@² ® 61¯U«Âß u .1«<0 Þ ±1«2° É ».1« ® «0¬ = ¯S0 GÁ ¯±@8 É= ±1¬<°0 = ±@« . ® « GÁ ¯Å±1°°±CÇ ® «2² = °.1¬ À ¯ ÀM9

S ! #"$%&')(*,+.-/10W-91@)&%5'#(*,+.-;/=0 ÃJ 'A91 O' *,+10S ! #"$%&')(*,+.-/10 Ã9=@)&%5')(>*,+.-;/10 J #&%5')(*;+.-;/10VS R )(3'#<!&*,+.-;/=0 Ã

/«2¯ = .d« À ¯¯ ~Á . +® « GÁ ¯ ³ 8G¯U0 ®=G® 6@¯È9«<±CÇ)°¯S0²1¯0 ®À~= ±C61¯8 ÉÆÀ ¯ G® «<² = °. À~À~®ð= . ~® ±1«8~¬2°¯ À .d8G¯W²1¯«<¯8:. ¯S0u»2Ç Á ¯U8~¯S. À)® « ~Á ¯ 0¯ ÀG= 8 ®³~® 61¯ À ¯ ~~® «2²Â02. .dÍ. À ¯W8~¯U²1¬2°.d8 ®eG® ¯ À .d8G¯0¯8 ® 6@¯U0uÃ

5")L! w,.)L5M5K1*6"N7, ò Á ¯8G¯ Á . À Í\¯¯U«". À ¬2²1²@¯ À~® ±1«Æ¹ C¿9±dÅ Á ±CÇ ± ® « ¯²@8G. ¯ ~Á ¯ Ç±´4. ® « À ¯ ~~® «2² À ±1Å ³ 8G¯U0 ®=~® 61¯+.d«0"0¯ À~= 8 ®³G® 6@¯ªE35·Ãdªb« ~Á2®À!® « ¯²@8G. ¯S0Å8:.d´4¯Ç±18GÈ ~Á ¯°¯S.d8G«2¯U0 ~Á ¯±18 É¶®À . = ±1´"Í ® «<. G® ±@«±dÅ+µ ³ 8G¯U0 ®=~® 61¯S¸)8~¬2°¯ À .d«<0Ôµ0¯ ÀG= 8 ®³2~® 61¯S¸ ® « ¯U²18 ®eEÉ= ±1« ÀE 8G. ® « GÀGÁ . 8G¯ À 8 ®=Ê~Á ¯ = ±1« À ¯ Û ¬2¯U« = ¯ À ±1Å GÁ ¯ À ¯W8G¬2°¯ À Ã

~Á ¯U8 ªE3· À ¯ G® «2² ÀÁ .76@¯Â.d° À ±ÎÍ¯U¯« ® «c61¯ ÀEG® ²9. ¯U0u» ~Á ¯´± ÀE® ´ ³ ±18 .1« Í¯ ® «<²V ãçè$?Lé*@Cê\èdçWéIê2ë`?bèCêPZ:ãMg í è7ëã X×çãGè V ê<éIêe ¹e¾d¿NÃÑ = ¯°°¯« ¶³ 8~¯S0 ®=~® 61¯Ð8~¯ À ¬<° :ÀÂÁ .76@¯ÝÍ\¯¯U«. =:Á2® ¯6@¯U0ÐÍ ÉGÁ ¯Â8G¯°. G® ±@«<.d° ® « À .1« = ¯ÓLÍ<. À ¯S0Ð°¯S.d8G«2¯8+ªEî3 ¹e¾d¿ ® «â«c¬2´4¯8G±1¬ À = °. À ÓÀ®ð= . ~® ±1«.1«<0 ³ 8~¯S0 ®=~® ±1« . À È À Ã)¯ = ¯« ° É » V ë ? g @ V X1ã VvV ãéIê@ V Z:ã _ ãê7?+çã:è V êéIêLe ¹¾1¾¿.d«<0 ð 8 ÀE ÓL±@8G02¯8WîÊ. É ¯ À~® .d« = °. À~À~®ð ¯8¹¾C¿ Á .761¯.d° À ±Í\¯¯U« À ¬0 ® ¯S0uÃ ® « = ¯ ~Á ¯ À ¯ªE35·À ¯ ~® «2² À 0±/«<± !® «c6@±1°61¯ ÁcÉc³ ± ~Á ¯ À~®À Å±@8~´. G® ±@« ® «"¯Ñ ³ ° ®=®ÀÉ ´"Í±@° ®= Å±@8~´» GÁ ¯)0¯6@¯°Ó± ³ ¯U0 ¯ =:Á « ®Û ¬2¯ À 02± «2± ÊÛ ¬<.1° ® Å É . À ¯ =:Á « ®Û ¬<¯ À Å±188~¯U°. ~® ±1«<.1°<Èc«2±CÇ)°¯U0²@¯/0 ®ÀG= ±C61¯U8 É Ã

5£2©¥¢ (£ ©¢/©¢ ¨&/¢©§]

ò Á2®ÀÊÀ ¯ =~® ±1«Â8G¯6 ® ¯Ç À~Á ¯ À . ¯ÓN±dÅIÓ ~Á ¯ÓN.18 8~¯U°. ~® ±1«<.1°\0<. . ´ ® « ® «2² ¯ =:Á « ®Û ¬<¯ À ´4± À±dÅÜÇ Á2®=:ÁÆÁ .76@¯.d°8~¯S.10 É4À~Á ±CÇ)« ~Á ¯ ® 8 ³ ± ¯« ~® .d°Å±18¬ À ¯ ® «Æ8~¯S.d°ÓL° ® Å¯/. ³<³ ° ®= . ~® ±1« À Ã@ò Á ¯±C61¯8G6 ® ¯UÇ ®À ° ® ´ ® ¯U0 ±Ð8~¯ = ¯U« )¯°. G® ±@«<.d°)Ù. . ó® « ® «2²Ð02¯61¯U°± ³ ´4¯« GÀ ». ® ´¯S0×. ~Á ¯.d«<.1° ÉÀ~®À ±dÅÜ8G¯U.d°ÓN° ® Å¯/02. .1Í<. À ¯ À ¹ ½Q»¾S½C¿LÃcò Á ¯ À ¯02¯61¯U°± ³ ´4¯« GÀÁ .76@¯.W´.d8GÈ1¯ ~® «2²³ ± ¯« G® .1° ® « ~Á ¯ ³ 8G± À~³ ¯8G±1¬ À «<¯Ç .d8G¯U. À ±dÅÙ. . ó® « ® «2²".1«<0 Ø «2±CÇ)°¯U0²@¯Ù ®ÀG= ±C6@¯8 É® «¶Ù/. .dÍ. À ¯ À Ã1ª ®À Ç±@8 GÁ Ç Á2® °¯+«2± ~®=® «2² GÁ . «2±@«2¯±1Å ~Á ¯8~¯U6 ® ¯UÇ¯S0 ¯ =:Á « ®Û ¬2¯ À Í\¯Ó°±@«2² À ± ³ 8G±1²18:.d´4´ ® «2²W. ÀGÀ®ÀE .d« GÀ Ç Á<®=:ÁÁ .76@¯).´ ¬ =:ÁÆÀ ´.d°°¯U8´.d8GÈ1¯ G® «<² ³ ± ¯« ~® .d°.d«<0Ò.° ® ´ ® ¯U0Ð¬ À ¯Å¬<°«2¯ À~À Å±18 À ±1°6 ® «2²8G¯U.1°eÓN° ® Å¯ ³ 8G±1Í2°¯´ À® « = ±1´ ³ .18 ®À ±1«Ç ®~Á ªE35·02. ."´ ® « ® «2² ±c±1° À .d«0 ¯ =:Á « ®Û ¬2¯ À Ã

11

R!$)BO10h5K1<;7) 5.)0 "6O1NO),/)4!$6O16ON # 02'*4&,.,K1<3(0245K1*#P6 !FO'),2 ò Á2®ÀW®À~Á ¯ ÀE .d«<02.18G0ÝªE3· À ¯ ~~® «2² ~Á . Á . ÀÍ¯U¯«Ô¬ À ¯U0 ® «Ð«c¬2´4¯U8~±@¬ ÀWÀ ¬ == ¯ À~À Å¬2° ³ 8~¯S0 ®=~® 61¯4Èc«2±CÇ)°¯U0²@¯40 ®ÀG= ±C6@¯8 É . ³2³ ° ®= . G® ±@« À Ãò Á ¯Ç¯°°eÓNÈc«2±CÇ)« À~ÉcÀ ¯U´ À Å±18 = °. ÀGÀ®ð= . ~® ±1«W8G¬2°¯ ® «<0¬ =~® ±1« ® « = °¬<02¯Ì2± ® °¹ ºC¿S» f ±1°¯´¹ º9Ú7¿u.d«<0·8G±1²1±@°¹ º2d¿NÃcÌ2± ® ° ®À ¯ =® ¯U« .1«<0ÆÍ\¯ À ¬<«<0¯8 À ±c±0Æ0¬2¯ ± ®e:ÀÊÀ® ´ ® °.d8 ®eEÉ ±Ö °.18~ÈÂ.1«<0¶- ® Í2°¯ ÀÖ -/½ÃH/« ~Á ¯W± ~Á ¯8 Á .1«<0u» f ±@°¯U´ .d«0·8~±@²1±1°.d8G¯ =:Á .d´ ³2® ±1« À= ±1« = ¯8G« ® «2² À ¬ =U= ¯ ÀGÀ Å¬2°!ªE3· . ³2³ ° ®= . ~® ±1« À »\02¯ À~³2®e ¯ ~Á ¯Å. =~Á . /GÁ ¯ É .18~¯ À ¬2Í À .d«2Ó~® .d°° É °¯ ÀGÀ ¯ =® ¯« ÃÌ<± ® ° ®À . ± ³ Ób0±CÇ)«â°¯U.18~«2¯U8U» f ±1°¯´ ®À .ÎÍ\± ~ ±1´ÓN¬ ³ °¯U.18~«<¯8S».d«<0·8G±1²1±@°¬ À ¯ À . = ±@´ Í ® «2¯U0 À ¯S.d8 =:Á À 8:. ¯U² É Ã 8 °°W.18~¯Ð´. ® «2° É = ±1« = ¯8G«2¯U0 Ç ®eGÁÀ® «2²1°¯ ³ 8G¯U0 ®= . ¯°¯S.d8G« ® «<²ÂÅ8G±1´ ³ ± À~®~® 61¯.d«<0Ý«2¯U²@. G® 6@¯"¯Ñ2.d´ ³ °¯ À .d«<0ÝÍ<. = È9²@8~±@¬2«<0È9«<±CÇ)°¯S0²1¯ ® «.@020 ®~® ±1«Ü»·8G±1²@±1° = .d«¶.d° À ±"Í¯¬ À ¯U0 ±°¯S.d8G«Å8~±@´ ³ ± À®~® 61¯¯Ñ2.1´ ³ °¯ À±1«2° É ÃCò Á ¯ É ¬ À ¯Ê0 ®GE ¯8G¯« . == ¯ ³ .d« = ¯ = 8 ® ¯U8 ® . 9C= ±1´ ³ 8G¯ ÀGÀ® ±1«Ü» = ±C6@¯8:.d²@¯HC. == ¬28G. =É .1«<0´ ® « ® ´.d°502¯ ÀG= 8 ®³G® ±@«°¯«<² ~Á »<8G¯ À~³ ¯ =~® 61¯U° É Ã2ò Á ¯ À~ÉcÀ ¯U´ 3ªE- Ò¹ ½1Ú2»~½d¼d¿L»202¯61¯U°± ³ ¯S0Å8~±@´ô.4°¯S.d8G« ® «2² = ±1´ ³ ±1«2¯U« ±1Å/¬ ó 8 Ð¹ º9½7¿N» ® « 8~±0¬ = ¯U0 ~Á ¯ ³ 8~± ³ ± À®~® ±1«<.1° ® . ~® ±1«³ .d8:.10 ® ²1´ Í ÉÂ 8:.d« À Å±18G´ ® «2²4.d«¶ªE3· ³ 8G±1Í<°¯U´ ® « ±Æ. ³ 8~± ³ ± À®~® ±1«<.1°°¯S.d8G« ® «<² . À È\Ã6OB "0h5.1#76 #'#7NP1*04'(BO)021,.1#76 5.!)), ò Á ¯ ÀÉÀ ¯U´ ò ® °02¯ ¹e¾¿<Í\¯°±1«2² À ± ò5± ³ Ó

0±CÇ)« ® «<0¬ =~® ±1«¶±dÅ50¯ =®À® ±1« 8G¯¯/.1°²@±18 ®eGÁ ´ À Ã9ª )= .d«Í¯6 ® ¯UÇ¯S0Æ. À . ð 8 À ÓN±18:0¯U8¬ ³ Ó²18:.10¯W±1Å/¬ ® «2°.d« ÀÖ <ÃÚ»2¯U´ ³ °± Éc® «2²°±@² ®= .1° Û ¬2¯8 ® ¯ À)® « 8G¯¯«<±c02¯ À Ç Á<®=:Á® «c61±1°61¯ À. ³2³ 8G± ³ 8 ® . ¯ Á .d«<02° ® «2²Æ±1Å6C.18 ® .dÍ2°¯ À Ã<ò Á ¯"´. ® «Î.106C.d« .d²1¯W±1Åò ® °0¯ ®À/®e:À ¯ =® ¯U« =É.d«<0 = . ³ .dÍ ® ° ®EÉ ±dÅ/02¯U.d° ® «2²Ç ®~Á °.18~²@¯Æ«c¬2´ Í\¯8 À ±dÅ 8G. ® « ® «2²Ý¯Ñ2.d´ ³ °¯ À »(Ç Á2®=:Á .d8G¯~Á ¯ÆÇ¯°°ÓLÈc«2±CÇ)« ³ 8G± ³ ¯8 ~® ¯ À ±1Å+ò ® °0¯ À³ 8~± ³ ± À®~® ±1«<.1°.d« = ¯ À ±@8 À Ã q+¯U« = ¯ò ® °0¯ = ¬28~Ó8~¯U« ° É 8G¯ ³ 8G¯ À ¯« GÀ ±1«2¯±1Å ~Á ¯"´± ÀE . ³2³ 8G± ³ 8 ® . ¯ ÀÉÀ ¯´ À Å±18 ³ 8G¯U0 ®=~® 61¯Èc«2±CÇ)°¯U0²@¯0 ®ÀG= ±C6@¯8 É Ãcî¯ À~® 0¯ À~Á ¯°.d«2²@¬<.d²@¯Í ® . À »ò ® °0¯W.d°°±CÇ À Å±18Ê°±c±1ÈC. Á ¯S.10Æ.1«<0 ³ 8~¯ ³ 8G¬2« ® «2²µ. =U= ±@8G0 ® «2² ± GÁ ¯´ ® « ® ´.d°«9¬<´ Í\¯8W±dÅÊ¯Ñ2.1´ ³ °¯ ÀW= ±C61¯U8~¯S0<¸0¯ ð «2¯U0ÐÍ ÉÎ³ .18G.1´¯ ¯U8À ¯ ~® «2²ÃA 1<!$,M5Y#7!$BO)L!]!$)N7!),K,.1#76 ò Á ¯8~¯U°. ~® ±1«<.1°18G¯²@8~¯ À~À~® ±@« . À È = .d« Í¯Ê0¯ ð «2¯U0". À Å±@°eÓ

°±CÇ À.9 f ® 61¯U« 8:. ® « ® «2²Æ¯Ñ.1´ ³ °¯ À . À+³ ± À~®eG® 6@¯ ²@8~±@¬2«<0Å. =GÀ Å±18 ~Á ¯ .18~²@¯ ³ 8~¯S0 ®= . ¯ µ ! " # "%$2¸»ÜÇ Á ¯U8~¯ GÁ ¯Æ6C.d8 ® .dÍ2°¯% Á . À 8G¯U.1°!6C.d°¬2¯ À ».d«0ÝÍ. = Èc²18G±1¬2«0ÎÈc«2±CÇ)°Ó¯U0²@¯ ³ 8G¯U0 ®= . ¯0¯ ð « ®~® ±1« À » ð «<0Ð.0¯ ð « ®eG® ±@«Å±@8 µ ! " # "%$2¸» À ¬ =:ÁGÁ . ¯S. =:Á= °.d¬ À ¯ Á . À .° ® ¯8:.d°)Í ® «<0 ® «2²&ìµ. À~À ¬2´ ® «<² GÁ . ! " # "%$×.d8G¯ÂÍ\±1¬<«<0<¸Ãò Éc³2®= .1°Í<. = Èc²18G±1¬<«<0¶È9«<±CÇ)°¯S0²1¯ ³ 8~¯S0 ®= . ¯ À+® « = °¬<02¯ °¯ ÀGÀ ÓN±18~ÓL¯ Û ¬<.d° ¯ ÀGÀ »\.@020 ®~® ±1«Ü» À ¬<Í 8:. = Ó~® ±1«.d«<0´ ¬<° G®³ ° ®= . G® ±@«ÜÃ 8 «Î. ³2³ 8G±@. =:Á ±8G¯°. G® ±@«<.d°8G¯²@8~¯ À~À~® ±@« ®ÀÊ® ´ ³ °¯´4¯U« ¯U0 ® «~Á ¯ ÀÉÀ ¯´ ÌP µÌ ® 8 À /8:0¯U8)¯²@8~¯ À~À~® ±1« ÉÀE ¯´Æ¸¹ ½@½7¿@Ç Á<®=:Á³ ¯8~Å±18G´ Àu ± ³ Ób0±CÇ)«À ¯S.d8 =:Á ±1Å.8G¯ ð «2¯U´4¯« ²@8G. ³2Á Ã2ªb«Î¯S. =:ÁÎ= °.1¬ À ¯@»ÌP = .1« ³ 8G¯U0 ®= .6C.d°¬2¯WÅ±18 ~Á ¯ .d8G²1¯ 6C.d8 ® .1Í2°¯'. À~Á ¯Â±1¬ ~³ ¬ 67.1°¬<¯Æ±dÅ).Í<. = È9²@8~±@¬2«<0È9«<±CÇ)°¯S0²1¯° ®e ¯8:.d°L». À .= ±1« ÀE .d« »2±@8. À .4° ® «2¯S.d8 = ±1´"Í ® «. ~® ±1«±1Å6C.18 ® .dÍ2°¯ À . ³2³ ¯U.d8 ® «2² ® « GÁ ¯ = °.1¬ À ¯Æµ¬ À~® «<²° ® «2¯U.18)8~¯U²18G¯ ÀGÀ® ±1«\¸Ã6OB "0h5.1*;7)&( #76",M5.!$41*65 #7NP1*0 R!#PN7!4&+ + 1*6"NO ª ®À Ç¯U°°5È9«<±CÇ)« ~Á . WÖ ±1«2Ó

ÀE 8G. ® « 3±1² ®= ·8~±@²18:.d´4´ ® «<² µ Ö 3·¸ = .d« À ¬ =U= ¯ ÀGÀ Å¬2°° É 0¯U.1°/Ç ®~Á «c¬2´4¯8 ®= .d° = ±1«2ÓÀE 8G. ® « GÀ Ãò Á ¯ ® 0¯S.±dÅªb«<02¬ =G® 6@¯ Ö ±@« À 8:. ® « 3Ü±@² ®= ·8~±@²18:.d´4´ ® «2²µª Ö 3·¸¹ C¿ ®À ±Í¯U«2¯ ð2 Å8G±1´ ~Á ¯/«c¬2´ Í\¯8~Ó Á .d«<0° ® «2² = . ³ .dÍ ® ° ®eG® ¯ À ±1Å Ö 3·»9.d«<0 ± ¬ À ¯ GÁ ¯ = ±1« À 8:. ® « À ±@°6@¯8W±dÅ Ö 35· ±Î0± ³ .d8 ±dÅ GÁ ¯ À ¯U.d8 =:ÁÐ® «c61±@°6@¯U0 ® « ® «<02¬ =G® 6@¯Æ°¯S.d8G« ® «2²<Ãò5± GÁ2®À¯«<0Ý.Â´.CÑ ® ´.d°° É 0 ®À~= 8 ® ´ ® «.d« ²@¯«2¯U8G.1° ® . ~® ±1« ³ 8G±1Í2°¯´ ® «ªE35· ®À/ 8:.d« À Å±18G´4¯U0 ±)* MbQ9:GD:MbMbVU!VF,+@VSOPaIRCBE9DMN@MNBEDGT-+\FVSOPaIR@:HUJ MbQ9::DGMbMNFQ9aPa`DQ9MbD|FV!OPJc|1Q9:OPJ9)ED:aeHBEOeVUJcHa|@D.cJ9OBEOPVSJ@M!VUFuFQ9J9~BEOeVUJ9M:?

12

.d«Î¯ Û ¬ ® 6C.d°¯« /= ±1« À 8:. ® « /À . ~®À Å. =G® ±@« ³ 8G±1Í2°¯´%µ Ö ·¸Ãuò Á ¯ À ±1°¬ G® ±@« À ±1Å ~Á ¯±18 ® ²1Ó® «<.1°!ªE3· ³ 8G±1Í2°¯´ = .d«ÝÍ¯ = ±@« À 8G¬ = ¯U0Å8~±@´ ~Á ¯ À ±1°¬ ~® ±1« À ±1Å Ö ·»Ç Á2®=:ÁÒ= .d«ÐÍ\¯±1Í . ® «2¯U0Í É 8G¬2«2« ® «2². = ±1« ÀE 8G. ® « +À ±@°6@¯8±1« Ö ·Ã

I ),.02!1*O5K1<;P) 5K)L0 "6O1NO),/)4!$6O16ON # 0'*43",.4&'S5 O)#7!$1*), 46OB 4,.,K# 01*45K1*#P6 !F"'*), ªb«0 ®ÀG= ±C6@¯8 ® «2²ÒÅ¬<°°= °.d¬ À .d° ~Á ¯U±18 ® ¯ À »9. À 02±1«2¯ ® « GÁ ¯ À~ÉÀE ¯´ Ö °.1¬<0 ® ¯«¹ QS¿N»1¯S. =:Á ¯Ñ.1´ ³ °¯ ®À .vq¯8GÍ28G.1«<0´±0¯U°»2.1«<0 ~Á ¯ À~ÉÀE ¯´ À ¯U.18 =:Á ¯ À Å±@8 ~Á ¯W´4± ÀE ²1¯U«2¯8:.d° = °.1¬ À ¯ À~Á . .18~¯ 8~¬2¯ ® «¶.1°°~Á ¯´±0¯U° À Ã Ö °.d¬ À ¯ À .d8G¯Æ0 ®À~= ±C61¯U8~¯S0 ® «<0¯ ³ ¯U«<0¯« ° É Å8G±1´ ¯U. =:Á ± ~Á ¯U8U»5Ç Á2®=:ÁÔ®À .À ¬2Í ÀE .d« ~® .d°!.10267.1« .d²@¯ Å±@802. .Â´ ® « ® «2²<»Ü. ÀW= ±1´ ³ .18~¯S0 ± GÁ ¯4°¯S.d8G« ® «2²±1Å = °. À~À~®eð Ó= . G® ±@«8G¬2°¯ À µ ³ .d8 ~®= ¬2°.d8G° É °¯S.d8G« ® «<²±1Å5´"¬ ¬<.1°° É 02¯ ³ ¯«<0¯U« +³ 8~¯S0 ®= . ¯ ÀÊ® «´"¬2° ~®³ °¯³ 8~¯S0 ®= . ¯°¯S.d8G« ® «<²9¸Ãªb« Ö °.1¬<0 ® ¯«Ü» À ¯U.18 =:Á ±dÅ = °.d¬ À ¯ À®À ° ® ´ ®e ¯U0 Í ÉÔ~Á ¯°.d«<²1¬<.1²1¯Í ® . À Ã9ª GÀ . =U= ¯ ³ .1« = ¯ = 8 ® ¯U8 ® ±1« = .d«ÆÍ\¯´±0 ®ð ¯S0Í É4À ¯ ~~® «2² Ç± ³ .d8:.d´4¯ ¯8 ÀM9CGÁ ¯8~¯ÓÛ ¬<¯ À ¯U0Æ´ ® « ® ´.d°. == ¬28:. =É .d«<0Æ´ ® « ® ´.d°\«9¬<´ Í\¯8±1ÅÜ¯Ñ2.d´ ³ °¯ À= ±C6@¯8G¯U0uÃdªb«.d«2± ~Á ¯U8= °.d¬ À .d°0 ®ÀG= ±C6@¯8 ÉÆÀÉÀ ¯´»\·8 ® ´"¬ À ¹¾QS¿N» GÁ ¯Í\¯ À Ó ð 8 ÀÀ ¯U.d8 =:Á Å±18 = °.1¬ À ¯ ÀÊ®À ²1¬ ® 0¯S0Í ÉÁ ¯U¬28 ®À~®=À ´4¯U. À ¬<8 ® «2² ~Á ¯ = ±1« ð 8~´. G® ±@«±dÅ = °.d¬ À ¯ À Ã1ò Á ¯ Ö °.1¬<0 ® ¯« À~ÉÀE ¯´ Ç. ÀÅ¬28 GÁ ¯8"¯Ñ ¯U«<0¯S0 ± FÐ.18~´48Â¹ ½7¿ ~Á . ¯U«<.dÍ2°¯ À °¯S.d8G« ® «2²±dÅ+. ÀGÀ ± =® . ~® ±1«Ò8G¬2°¯ À Å8G±1´´ ¬2° ~®³ °¯W8~¯U°. ~® ±1« À ÃA 1<!$,M5Y#7!$BO)L! 0' ",M5K)L!16ONO ò5± ³ Ób0±CÇ)« ® «<0¬ =~® ±1«"±dÅ0¯ =®À® ±1« 8G¯¯ À(= .1«Í\¯6 ® ¯UÇ¯S0

. À . = °¬ À ¯U8 ® «2²´4¯ GÁ ±0 À® « = ¯Â«2±0¯ ÀW® « ~Á ¯ 8G¯¯ = ±@8~8G¯ À~³ ±1«<0 ± À ¯ :À ±dÅ)¯Ñ2.1´ ³ °¯ ÀÇ ®eGÁ4À~® ´ ® °.d8 ³ 8G± ³ ¯8 ~® ¯ À » GÁ ¬ À Å±18G´ ® «2² = ±@« = ¯ ³!Á2® ¯8:.d8 =:Á2® ¯ À Ã7ò Á2®À 6 ® ¯UÇ×ÇÊ. À .10± ³ ¯S0® « Ö <Ã Úâ¹ ¼d¿L».d«Ô¬ ³ ²18:.10¯Æ±1Å ~Á ¯¶ò ® °0¯°±@² ®= .d°0¯ =®À~® ±1« 8G¯¯°¯S.d8G«2¯8SÃ 8 8~¯U°. ~® ±1«<.1°0 ®À .1« = ¯ÓLÍ. À ¯U0 = °¬ À ¯U8 ® «2² ®À³ 8G¯ À ¯« ¯S0Æ.d° À ± ® «Ð¹ ½1ºC¿NÃ 8 «Â¯U.18~° É . ³2³ 8G±@. =:ÁÆ= ±@´ Í ® « ® «2²°¯S.d8G« ® «<².d«<0 = ±1« = ¯ ³ ¬<.d° = °¬ À ¯8 ® «2² ¯ =:Á « ®Û ¬2¯ À Ç. À® ´ ³ °¯U´¯U« ¯U0 ® « GÁ ¯ À~ÉÀE ¯´Ö ±1°.¶¹¾Uºd¿LÃ f ® 61¯U«. À ´.d°°µ À~³ .d8 À ¯7¸ À ¯ ±1Å = °. À~À~®ð ¯U0 8:. ® « ® «<² ® « À .1« = ¯ À .d«<0¶. À ¯ ±dÅ¬2« = °. ÀGÀ~®eð ¯S0 ® « ÀE .d« = ¯ À » Ö ±1°.4¬ À ¯ À î ®ÀGÀ ±@« À/= ±@« = ¯ ³ ¬<.1° = °¬ À ¯U8 ® «2²Â.1°²@±18 ®eGÁ ´ Ø î f±1« ~Á ¯4¯« G® 8G¯ À ¯ ±1Å ® « À .d« = ¯ À » = ° ® ´ Í ÀGÁ ¯ Á2® ¯8:.d8 =:Á9É¶ 8~¯U¯4.1«<0Î¬ À ¯ À~Á ¯ = °. À~À~®eð ¯U0® « À .1« = ¯ À ± ® 0¯« G® Å É µ À~® «2²@°¯±@8+0 ®À ïE¬2« =~® 61¯7¸ = °. ÀGÀ 0¯ ÀG= 8 ®³2~® ±1« À Ã

45K4%(4,.) !$)L,.5M!F"0h5y!$1*6"NO ò Á ¯ À~ÉcÀ ¯U´ Ì2¯«0¯8Ô¹ 9½7¿ À ¯U.d8 =:Á ¯ À Å±18 = ±1´4´4±1«³ .d8 GÀ ±dÅ)8~¬<°¯ À 0¯ À~= 8 ® Í ® «2². = ±1« = ¯ ³ » ~Á ¬ À Å±18G´ ® «<² À ¬<Í = ±1« = ¯ ³ 0¯ ð « ®eG® ±@« ÀW ±Í\¯¬ À ¯S0 ® « GÁ ¯8G¯Å¬28G´ ¬2°. G® ±@«±dÅ±18 ® ² ® «<.1°u8~¬<°¯ À Ã<ò Á ¯8~¯ À ¬<° +®À .4Èc«2±CÇ)°¯U0²@¯WÍ<. À ¯Ç ®eGÁ«2¯Ç ® « ¯U8~´4¯U0 ® . ¯ = ±1« = ¯ ³2GÀ .d«<0"0¯¯ ³ ¯U8 ® «2Å¯8G¯« ~® .d° À 8G¬ = ¬28~¯ ~Á .1« GÁ ¯ ® « ®~® .d° Õ. 8~¬2°¯Í. À ¯1Ãuò Á ¯ ÀÉÀ ¯U´äªb«<0¯Ñ×¹e¾S¼C¿ ®À= ±1« = ¯8G«2¯U0Ç ®eGÁÐ~Á ¯ ³ 8G±1Í2°¯´±1Å0¯ ¯U8~´ ® « ® «<²Ç Á2®=:Á . ~ 8 ® Í<¬ ¯Æ0¯ ³ ¯U«<0¯U« =® ¯ À µÅ¬2« =G® ±@«<.d°±18´ ¬2° ~® 6C.d°¬2¯U0¸ Á ±1°0 ® « ~Á ¯4² ® 6@¯«Ð8~¯Ó°. ~® ±1«<.1°502. .dÍ. À ¯1Ãò Á ¯ ® «<0¬ = ¯U0Î. 8 ® Í2¬ ¯0¯ ³ ¯«0¯« =® ¯ À/= .d«ÎÍ\¯ ¬ À ¯S0 ±±1Í . ® «Î.´±@8~¯ ÀE 8~¬ = ¬28G¯U0Ô0<. .1Í<. À ¯1Ã(î± ~Á . ³2³ 8G±@. =:Á ¯ À= .d«ÔÍ\¯Â6 ® ¯Ç¯U0Ô. À 0± ® «2² ³ 8~¯S0 ®= . ¯® «c61¯U« ~® ±1«Ü»cÇ Á ¯U8~¯4µ¬ À ¯8 À ¯°¯ = ¯U0¸ ® «c61¯U« ¯U0 ³ 8~¯S0 ®= . ¯ À .d8G¯¬ À ¯S0Å±@8 GÁ ¯±@8 É 8G¯ À 8G¬ = Ó ¬28 ® «<²<Ã

O%"N7!$#HO B"1*,K02#;7)L!& ò Á ¯ À ¬2Í2²@8~±@¬ ³ 0 ®ÀG= ±C61¯U8 ÉÐ . À È ®À 0¯ ð «2¯S0 . À Å±@°°±CÇ À.9² ® 61¯«. ³ ± ³ ¬2°. G® ±@« ±1Å ® «<0 ® 6 ® 02¬<.d° À .d«0 . ³ 8G± ³ ¯8 EÉ ±1Å GÁ ± À ¯ ® «0 ® 6 ® 0¬<.1° À Ç¯Î.d8G¯® « ¯U8~¯ ÀE ¯U0 ® «Ü» ð «<0 GÁ ¯ À ¬2Í<²18G±1¬ ³<À ±dÅ ~Á ¯ ³ ± ³ ¬<°. ~® ±1« GÁ . .d8G¯ ÀE . ~®À~®= .d°° É ´4± À® « ¯U8~¯ ÀEG® «<²2» ® Ã ¯1Ã»u.d8G¯". À °.d8G²1¯ . À³ ± ÀGÀ~® Í2°¯".1«<0 Á .761¯ GÁ ¯"´4± ÀE ¬2«c¬ À ¬.d° ÀE . ~®À~®= .d°µ0 ®ÀE 8 ® Í2¬ G® ±@«<.d°I¸ =:Á .d8:. = ¯8 ®À~®=À Ç ®eGÁ 8G¯ À~³ ¯ = ± GÁ ¯ ³ 8~± ³ ¯U8 EÉ ±dÅ ® « ¯8G¯ À Ãcò Á ¯ ÀÉÀ Ó ¯´ ó® 0± À ¹ cÚ7¿²@¬ ® 02¯ ÀÊGÁ ¯ ± ³ Ób0±CÇ)« À ¯S.d8 =:Á ±1Å ³ ± ¯« ~® .d°° É¶® « ¯8G¯ À~® «2² À ¬<Í2²18G±1¬ ³À¬ À® «2²4«c¬2´4¯8G±1¬ À ¬ À ¯U8Ób0¯ ð «2¯U0 ³ .d8:.d´4¯ ¯8 À Ã

13

/)4!$6O16ON N"4'*1*5.45K1<;P)+ # B")L', #"B 6"4+ 1*0 , ,M5K)L+ ,2 ò Á ¯+.d¬ ±1´. ¯U0 = ±1«2ÓÀE 8~¬ =~® ±1«±1Å´±0¯U° À ±dÅ0 É «<.1´ ®=À~ÉÀE ¯´ ´. É Í\¯". ® ´4¯S0. Û ¬<.1° ® . ~® 61¯´4±0¯U°50 ®À Ó= ±C61¯U8 É Ã 8 8~¯ = ¯U« !Û ¬<.d° ® . G® 6@¯´4±0¯°0 ®ÀG= ±C61¯U8 ÉÀ~ÉcÀ ¯U´ ¹ ½2¾¿N»U¬ À~® «2². À~® ´ÓN° ® È1¯8G¯ ³ 8~¯ÓÀ ¯U« . G® ±@«Ü» ®À Í<. À ¯U04±@« Ö ± ® ¯8:. À f ¯U«2´4±c02¯° ± Ç Á<®=:Á¶À® ²1«<.1° ³ 8G± = ¯ À~À~® «<² = . ³ .dÍ ® ° ®eG® ¯ ÀÁ .761¯Í¯U¯«.10<0¯U0uÃ

N"45.1#76wB"1*,K02#;P)!7 ò Á ¯ À~ÉÀE ¯´ä3 8 f 8 - f ¹¾C¿0 ®ÀG= ±C61¯U8 À . À ¯ ±dÅ0 ® ÅIÓÅ¯8G¯« ~® .d°)¯ Û ¬. ~® ±1« À Å8G±1´à.d« ¯Ñ2.d´ ³ °¯Í¯ Á .76 ® ±@8±1Å.Ò0 É «<.1´ ®=À~ÉcÀ ¯U´Ã !Ñ2.d´ ³ °¯Í¯ Á .76 ® ±@8 À .d8G¯ À³ ¯ =®eð ¯S0×Í É ° ®ÀE:À ±dÅ/´4¯S. À ¬28~¯U´4¯« GÀ ±dÅ. À ¯ ±1Å À~ÉcÀ ¯U´ 6C.d8 ® .1Í2°¯ À ».d«<0Í. = Èc²18G±1¬2«0Èc«2±CÇ)°¯U02²1¯ ³ 8G¯U0 ®= . ¯ À ¯«.dÍ2°¯ GÁ ¯ ® « 8G±c02¬ =G® ±@«Î±dÅ!«2¯UÇ 67.18 ® .dÍ<°¯ À. ÀG® ´4¯Â0¯U8 ® 6C. G® 6@¯ À » À® «2¯ À ±18 = ± À~® «2¯ À ±dÅ ÀÉÀ ¯U´6C.d8 ® .dÍ2°¯ À ÃÜ-+¯UÇ 6C.d8 ® .dÍ2°¯ À= .d«ÒÍ\¯Å¬28 GÁ ¯8 ® « 8G±0¬ = ¯U0Í É ´ ¬2° ~®³ ° ®= . ~® ±1«Ã6OB "0h5.1*;7) B"45.4%(4,K), Ã 8 G® ² Á9 ¯8 = ±@«2«2¯ =~® ±1«Ç ®~Á 0¯U0¬ =~® 61¯02. .dÍ<. À ¯ ¯ =:Á Ó

«2±1°±1² ÉWÁ . À Í\¯¯«48G¯ = ¯« ° É .106@± = . ¯U0 Í É 3Ü¬ = Ù/¯+.d¯S0 ¹ »GÚ7¿ ® « 8G±c02¬ =® «2²W.d« ® «0¬ = Ó~® 61¯Æ02. .dÍ. À ¯´ ® « ® «2² Û ¬2¯8 É °.d«<²1¬<.1²1¯ ~Á . ® « ¯²@8G. ¯ ÀW= ±1« = ¯ ³:À Å8G±1´ ªE35·» Ö 3·»0¯U0¬ =~® 61¯+0<. .1Í<. À ¯ À .1«<0´4¯ .dÓ ³ 8~±@²18:.d´4´ ® «2² ® « ±W.Õ<¯Ñ ® Í2°¯)¯U«c6 ® 8G±1«2´4¯« Å±@8!8~¯U°.dÓ~® ±1«<.1°9Èc«2±CÇ)°¯U02²1¯0 ®ÀG= ±C6@¯8 É/® «"02. .1Í<. À ¯ À Ã ® « = ¯ ~Á ¯ ³ 8 ® ´ ®~® 61¯ À ±1Å ~Á ¯Ê°.d«2²@¬<.d²@¯ = .1«¯U. À~® ° É Í\¯ = ±@´ Í ® «2¯U0"Ç ®~Á ·8G±1°±1²<» = ±@´ ³ °¯Ñ À~ÉcÀ ¯U´ À .d«<0"Í¯ Á .76 ® ±@¬28 = .d«Í\¯ À³ ¯ =®eð ¯S00¯ = °.18G. ~® 61¯° É Ãò Á2®ÀEÉc³ ¯4±dÅ ® « ¯²18:. G® ±@«±1Å = ±@« = ¯ ³GÀ Å8G±1´ 0 ®GE ¯8G¯« .d8G¯U. À ±1Å = ±@´Ó³ ¬ . ~® ±1«<.1°<°±1² ®=)= .d« ³ 8G±C61¯)¯Ñ 8~¯U´¯U° É Í¯U«2¯ ð\=® .d°2Å±@8 +Ù ó « GÁ ¯)Å¬ ¬28G¯1Ã1ª Ê= .1«°¯U.@0 ±". «2±C61¯U°<ªE35· ³ .18G.@0 ® ²@´ ±1Å ® «0¬ =G® 6@¯°±1² ®=³ 8~±@²18:.d´4´ ® «<² Û ¬<¯8 É °.d«2²@¬<.d²@¯ À Ç Á ± À ¯¬ À ¯Å¬2°«2¯ ÀGÀ ´. É Í¯ ³ 8~±C6@¯U0 ±WÍ¯ À~® ´ ® °.d8 ± GÁ ± À ¯Ê±dÅ = ±1« ÀE 8G. ® « °±@² ®=³ 8~±@²18:.d´4´ ® «2²Ã

#7+ )S0 "4'*')L6"NP)L,t16 !$),.)4!$0 ªE3· Á . À .1°8G¯U.@0 É 0¯U61¯°± ³ ¯U0«c¬2´4¯8G±1¬ À ¬ À ¯Å¬<° ¯ =:Á « ®Û ¬<¯ À Å±18)8G¯°. G® ±@«<.d°Èc«2±CÇ)°¯U0²@¯0 ®ÀG= ±C6@¯8 É Ã 8 8G¯ = ¯U« 8G¯ À ¯U.d8 =:Á 8G¯«<0 ® «ªE3· ®À ±"02¯61¯U°± ³ .d°²1±@8 ®~Á ´ À® ´ ³ °¯´4¯U« G® «<².d°° GÁ ¯"´4± ÀE³ ± ³ ¬2°.d8´. =:Á2® «2¯ °¯U.18~« ® «2² ¯ =:Á « ®Û ¬<¯ À/® « ~Á ¯ ð 8 ÀE ÓL±@8G0¯U8+Å8:.d´4¯Ç±18GÈÃ8 °8~¯S.10 É 0¯U61¯°± ³ ¯U0 ¯ =:Á « ®Û ¬2¯ À ¬ ³ ²18:.10 ® «2² ³ 8~± ³ ± À®~® ±1«.d°°¯U.18~« ® «2²Ð.d°²1±@8 ®~Á ´ À ® «Ó= °¬<0¯ ð 8 À ÓN±18:0¯802¯ =®À~® ±@« 8G¯¯+°¯U.18~« ® «2²Â¹¾¿N» ð 8 À ÓN±18:0¯8 = °¬ ÀE ¯8 ® «2²¹ ¼<»G½dºd¿L»18~¯U°. ~® ±1«<.1°²1¯«<¯ ~®= .1°²@±18 ®eGÁ ´ À ¹ ½#C¿N» ð 8 ÀE ÓL±@8G02¯8 ® « À .d« = ¯ÓNÍ<. À ¯S0°¯U.18~« ® «2²Î¹¾M1¿L» ð 8 À ÓN±18:0¯U8+8~¯ ® «2ÓÅ±18 = ¯´4¯U« °¯U.18~« ® «2²4¹e¾@¾¿2.d«0 ð 8 À ÓN±18:0¯8!îÊ. É ¯ À® .d« = °. À~À~®ð ¯8)¹¾C¿NÃ7ª ®À ¯Ñ ³ ¯ = ¯S0 ~Á . ~Á ¯¶.102. ³ . ~® ±1«Ô±dÅ ³ 8G± ³ ± À~®~® ±1«<.1°´. =:Á2® «2¯Æ°¯U.18~« ® «2²Î.1°²@±18 ®eGÁ ´ À ± ~Á ¯ ð 8 ÀE ÓL±@8G0¯U8Å8G.1´¯UÇ±@8~ÈÂÇ ® °° = ±1« ~® «c¬2¯ .d° À ± ® « GÁ ¯ .d8G¯U. À Å±18Ç Á2®=:Áð 8 À ÓN±18:0¯8 ® ´ ³ °¯´4¯« . ~® ±1« ÀÀEG® °°+0±Ò«2± ¯Ñ ®À Ãò Á2®ÀÀÁ ±1¬2°0 ³ 8~±C6 ® 0¯¶.ÝÅ¬2°° ÀG= .1°¯¶´4¯ ~Á ±c02±1°±1² É Å±@848~¯U°. ~® ±1«<.1°02. .´ ® « ® «2²Í<. À ¯S0Ý±@«ÐÅ¬ ¬28G¯ÆªE35· ® ´ ³ °¯´4¯« . ~® ±1« À ±dÅ ð 8 ÀE ÓL±@8G0¯U8îÊ. É ¯ À~® .d««<¯ ÓÇ±@8~È À » ð 8 ÀE ÓL±@8G02¯8+«2¯U¬28:.d°(«2¯ Ç±18GÈ À » ³ ± À~À~® Í<° É¶ð 8 ÀE ÓL±@8G02¯8)Å¬ .ÉÀ~ÉcÀ ¯U´ À .1«<0Î± ~Á ¯U8ªE3· ¬ ³ ²@8G.@0¯ À ±dÅ ³ 8G± ³ ± À~®eG® ±@«<.d°´. =:Á<® «2¯°¯U.18~« ® «2² ¯ =:Á « ®Û ¬2¯ À Ã

5£2©¥¢ Ê§ ¤2¥§ © c¨¥/Ü¤ ¢/¦ (£¦à¥¤

y S)L'45.1#76"4' O%"N7!$#HO 1*,K02#;7)L!ªb« ~Á ¯«<¯Ç)° É ¯U´¯U8~² ® «<² ð ¯°0±dÅ À ¬2Í2²18G±1¬ ³ 0 ®ÀG= ±C61¯U8 ÉÆ Ç±´4± À® ´ ³ ±@8 .d« ÀÉÀ ¯U´ ÀÅ±180 ®À~= ±C61¯8 ® «2² À ¬<Í2²18G±1¬ ³À .18~¯ !Ñ ³ °±18:. ¹ ½2d¿W.d«<0 ó® 0± À ¹ 9Ú7¿NÃÊò Á ¯ ð 8 ÀÀ~ÉÀE ¯´ 8G¯U. :ÀGÁ ¯Â°¯U.18~« ® «2² . À ÈÔ. À . À~® «<²1°¯Â8G¯°. ~® ±1« ³ 8~±@Í2°¯U´» ® Ã ¯@Ã»!.1°° ~Á ¯02. ..18~¯. À ÓÀ ¬2´4¯S0 ±Í¯.76C. ® °.dÍ2°¯ ® «±1«<¯ .dÍ2°¯Æµ8G¯°. G® ±@«¸»<Ç Á2® °¯ ~Á ¯ À ¯ = ±1«0Â±@«2¯¯Ñ ¯U«<0 ÀÊGÁ2®À

14

. À È ±Æ´"¬2° G® ÓN8G¯°. ~® ±1«.d°02. .1Í<. À ¯ À »2Ç Á2®=:ÁÎ®À 8G¯°. ¯S0 ±Â.Æ«9¬<´ Í\¯8±1Å!± ~Á ¯8°¯U.18~«2Ó® «2² . À È À ¹ Qc»º)2»9¼C¿N»d´4± À ° É ® « ~Á ¯/ªE3·Ä¹¾S½»G½d¼C¿NÃdªb«Í\± GÁÆÀÉÀ ¯´ ÀGÁ ¯ ³ 8G± ³ ± À~®~® ±1«<.1°µ. ~ 8 ® Í<¬ ¯ÓL6C.d°¬2¯7¸)°.d«2²@¬<.d²@¯ ®À ¬ À ¯U0 ±0¯ À~= 8 ® Í\¯ ~Á ¯ ® «<0¬ = ¯U0 ÁcÉc³ ± ~Á ¯ À ¯ À » ® Ã ¯@Ã»0 ®À Ó= ±C61¯U8~¯S0 À ¬2Í2²18G±1¬ ³<À .d8G¯¶0¯ ð «2¯S0 . À= ±1«dïE¬2« =G® ±@« À ±dÅ/Å¯U. ¬28G¯ À µ. 8 ® Í2¬ ¯ À 6C.d°¬2¯ À ¸Ãò Á ¯+´4± À® ´ ³ ±18 .1« Å¯U. ¬28G¯ À ±dÅ !Ñ ³ °±18:..d«<0 ó® 0± À= ±@« = ¯U8~« GÁ ¯+¬ À ¯+±1Å Á ¯¬28 ®À~®=ÀÅ±18 À ¬2Í2²18G±1¬ ³ 0 ®ÀG= ±C61¯U8 É ~Á ¯ Á ¯¬28 ®À~®=À .18~¯±@¬ ° ® «<¯U0¶Í¯U°±CÇWÃFÝ¯ Á .761¯)0¯6@¯°± ³ ¯U04.W8G¯°. G® ±@«<.d° À ¬2Í2²@8~±@¬ ³ 0 ®ÀG= ±C6@¯8 ÉÀÉÀ ¯U´ 2Ù ¹ ½ C¿±1« ³ 8 ® «Ó

=®³ °¯ À~Á . ¯´ ³ °± É4~Á ¯Å±@°°±CÇ ® «<² ´. ® « ® «2²@8~¯S0 ® ¯U« GÀ.9 ¯Ñ Á .d¬ À~® 61¯ ð 8 ÀE ÓL±@8G02¯8Å¯U. ¬28G¯= ±1« ÀE 8~¬ =~® ±1«Ü»¯° ® ´ ® «<. ~® ±1« ±1Å ® 8G8G¯°¯6C.d« Å¯S. ¬<8~¯ À »).d« ® ´ ³ °¯´4¯« . ~® ±1«±dÅ .×8~¯U°.dÓ~® ±1«<.1°18G¬2°¯°¯S.d8G«2¯8S» GÁ ¯Ç¯ ® ² Á9 ¯S0 = ±C61¯U8 ® «2²+.1°²@±18 ®eGÁ ´.1«<0 ® « = ±18 ³ ±18:. ~® ±1«±1Åc¯Ñ2.d´ ³ °¯Ç¯ ® ² Á@:À!® « ± ~Á ¯Ç¯ ® ² Á9 ¯S0"8G¯°. G® 6@¯+. =U= ¬28:. =É Á ¯U¬28 ®À~®= » ³ 8~±@Í<.dÍ ® ° ®À~®=)= °. ÀGÀ®ð= . ~® ±1«».d«<0¶.d8G¯U.¬2«<02¯8 Ö 8G¬2°¯ À ¯ ¯U6C.d°¬<. G® ±@«ÜÃ8 ÀGÁ ¯ ® « ³ ¬ » 2Ùâ¯Ñ ³ ¯ =GÀ µ.@¸u.Ê8G¯°. G® ±@«<.d°d02. .1Í<. À ¯ = ±@« . ® « ® «2²Ê±1«2¯´. ® « .1Í2°¯

µ8G¯°. ~® ±1«\¸Ç Á ¯8G¯W¯U. =:Á 8G±CÇ = ±18G8G¯ À~³ ±@«<0 À ±Â.4¬<« ®Û ¬2¯ éIêPXCéI[Sé*X$d<èdç .1«<0±1«2¯. 8 ® Í2¬ ¯±dÅ ~Á ¯Â´4. ® « .dÍ2°¯ ®À À~³ ¯ =®ð ¯U0×. ÀGÁ ¯ Zçè7ë:ë . ~ 8 ® Í2¬ ¯1».d«0ÄµÍ¸W.Î´±0¯ÓL°.d«<²1¬<.1²1¯0¯ ð « ®eG® ±@«¶¬ À ¯S0 ± = ±1« ÀE 8~¬ =Êð 8 À ÓN±18:0¯U8Å¯S. ¬28~¯ À Ãò Á ¯Ê´4. ® «"±1¬ G³ ¬ ±1Å Ù ®À . À ¯ ±dÅ À ¬2Í2²@8~±@¬ ³<À Ç Á ± À ¯ = °. ÀGÀ Ób0 ®ÀE 8 ® Í2¬ G® ±@« À 0 ®GE ¯8

À ¬2Í ÀE .d« ~® .d°° É Å8~±@´ GÁ ± À ¯±dÅ GÁ ¯ = ±@´ ³ °¯ ¯ 0<. .dÓ À ¯ Ã<ò Á ¯ À ¬2Í2²@8~±@¬ ³<À .18~¯ ® 0¯« G®eð ¯S0Í É= ±@«CïE¬2« =~® ±1« À ±dÅ À~É ´ Í\±1° À ±dÅ ³ 8G¯ÓN²1¯U«2¯8:. ¯U0 ð 8 ÀE ÓL±@8G02¯8Å¯U. ¬28G¯ À Ã 8 À .+Í É Ó ³ 8G±c02¬ = » Ù .d° À ± ³ 8G±C6 ® 0¯ À . ð °¯ = ±1« . ® « ® «2² ~Á ¯4´4¯« ~® ±1«<¯U0 À ¯ ±dÅÅ¯U. ¬28G¯ À .d«<0± E ¯8 À+ ±¯Ñ ³ ±@8 . À~® «2²@°¯8~¯U°. ~® ±1«Æµ. À . ¯Ñ 5ð °¯7¸uÇ ®~Á 8~±CÇ ÀÜ= ±18G8~¯ À³ ±1«0 ® «<² ± ® «<0 ® 6 ® 0¬.d° À .1«<0ð ¯°0 À)= ±@« . ® « ® «2² ~Á ¯ 8~¬ GÁ 6C.1°¬2¯ À ±dÅ(8~¯ À³ ¯ =~® 61¯/Å¯U. ¬28G¯ À Å±18 GÁ ¯W² ® 61¯U« ® «<0 ® 6 ® 0¬<.1°Ãò Á2®À .1Í2°¯ ®À+GÁ ¬ À . ³ 8G± ³ ± À~®eG® ±@«<.d° ®À ¯U08G¯ ³ 8~¯ À ¯U« . G® ±@«¶±1Å ~Á ¯ ® « ³ ¬ 02. .Â.1«<0 = .1«Í¯W¬ À ¯U0. À .d« ® « ³ ¬ ) ±6C.d8 ® ±1¬ À . ~ 8 ® Í<¬ ¯ÓL6C.d°¬2¯°¯U.d8G«2¯U8 À Ã8 « ® ´ ³ ±18 .1« Å¯U. ¬28G¯±dÅ ~Á ¯ Ù .d°²1±18 ®~Á ´ ®ÀuGÁ ¯¬ À ¯!±dÅ GÁ ¯Ç¯ ® ² Á9 ¯S0 = ±C61¯U8 ® «2²

.d°²1±18 ®~Á ´Ãdªb« GÁ ¯ = °. ÀGÀ~®= .d° = ±C61¯8 ® «2².1°²@±18 ®eGÁ ´ ±1Åu8~¬2°¯Ó À ¯ ® «<0¬ =~® ±1«Ü»@±1«2° É GÁ ¯ ð 8 ÀÅ¯Ç ® «<0¬ = ¯U0Æ8G¬2°¯ À ´4. É Í\¯±1Å ® « ¯U8~¯ ÀE . ÀÀ ¬2Í<²18G±1¬ ³ 0¯ À~= 8 ®³ ±18 À Ç ®~ÁÂÀ ¬ =® ¯« = ±C69Ó¯8:.d²1¯@» À~® « = ¯ À ¬2Í À ¯ Û ¬<¯« ° É® «<02¬ = ¯S0 8G¬2°¯ À .d8G¯ ® «<0¬ = ¯U0WÅ8G±1´ Í ® . À ¯U0 ¯Ñ2.d´ ³ °¯ À ¬2Í À ¯ :À »® Ã ¯1Ã» À ¬<Í À ¯ GÀ® « = °¬<0 ® «2²4±1«2° ÉÆ³ ± À~®eG® 6@¯/¯Ñ2.d´ ³ °¯ À «2± += ±C61¯8G¯U0ÆÍ É³ 8~¯U6 ® ±1¬ À ° É® «0¬ = ¯S08~¬2°¯ À Ã<ò Á2®À Í ® . À+= ±@« ÀE 8G. ® « À~Á ¯ ³ ± ³ ¬2°. ~® ±1«¶Å±18 À ¬2Í<²18G±1¬ ³ 0 ®À~= ±C61¯8 ÉÆ® «.4Ç. ÉÂ~Á . ®À ¬<«2«<. ¬28:.d°!Å±@8 ~Á ¯ À ¬2Í2²@8~±@¬ ³ 0 ®ÀG= ±C6@¯8 É³ 8G± = ¯ ÀGÀ Ç Á2®=:ÁÐ®À » ® «Ò²@¯«2¯U8G.1°»Ü. ® ´4¯U0Ò. 0 ®ÀG= ±C6@¯8 ® «<² ® « ¯8G¯ À~® «2² ³ 8G± ³ ¯8 ~® ¯ À ±dÅ À ¬2Í2²18G±1¬ ³<À ±dÅ GÁ ¯W¯« G® 8G¯ ³ ± ³ ¬2°. ~® ±1«Ã2ªb« = ±1«2Ó 8:. À » GÁ ¯ À ¬2Í À ¯ Û ¬2¯U« 8~¬2°¯ À® «<02¬ = ¯S0Í É GÁ ¯+Ç¯ ® ² Á9 ¯S0 = ±C61¯U8 ® «2²W.d°²1±@8 ®~Á ´ .d°°±CÇ×Å±@80 ®ÀG= ±C6@¯8 ® «<² ® « ¯8G¯ À~® «2² À ¬<Í2²18G±1¬ ³¶³ 8G± ³ ¯8 ~® ¯ À ±dÅ ~Á ¯W¯« ~® 8~¯ ³ ± ³ ¬<°. ~® ±1«ÜÃò Á ¯/Ç¯ ® ² Á9 ¯U0 = ±C61¯U8 ® «2²".1°²@±18 ®eGÁ ´ ®À® ´ ³ °¯´4¯« ¯S0 ® « À ¬ =:Á ."Ç. É4GÁ . += ±C6@¯8G¯U0

³ ± À®~® 61¯/¯Ñ2.d´ ³ °¯ À .d8G¯«2± 02¯°¯ ¯S0Å8G±1´ ~Á ¯ = ¬28G8~¯U« 8:. ® « ® «<² À ¯ Ãcªb« À ¯S.10u» ® «Â¯S. =:Á8~¬2«Ð±dÅ ~Á ¯ = ±C61¯U8 ® «2²Â°±c± ³ » GÁ ¯.d°²1±@8 ®~Á ´ À ±@8~¯ À Ç ®eGÁ ¯S. =:Á ¯Ñ2.d´ ³ °¯. = ±1¬2« WÁ ±CÇ´4.1« ÉW~® ´4¯ À µÇ ®eGÁ4Á ±CÇ×´.d« É 8~¬2°¯ À® «<0¬ = ¯U0 À ±Å.d8¸ ~Á ¯+¯Ñ2.d´ ³ °¯ Á . À Í\¯¯« = ±C61¯U8~¯S0uÃFÝ¯ ® ² Á@:À ±dÅ = ±C6@¯8G¯U0¯Ñ.1´ ³ °¯ À 02¯ = 8G¯U. À ¯. == ±18:0 ® «2² ± ~Á ¯Å±@8~´"¬2°. icµE¸ p

u »7Ç Á ¯8G¯i9µE¸ ®ÀGÁ ¯WÇ¯ ® ² Á9 ±1Å(.1«¶¯Ñ2.d´ ³ °¯Í\¯ ® «2² = ±C61¯8G¯U0 ~® ´4¯ À Ã8 67.18 ® .d« ±dÅ ~Á ¯Ç¯ ® ² Á9 ¯U0 = ±C6@¯8 ® «<².1°²@±18 ®eGÁ ´ Á . À Í¯U¯«¬ À ¯U0.d° À ± ® « ~Á ¯ = ±1«2Ó

¯Ñ ±dÅ ~Á ¯2Ù À ¬2Í<²18G±1¬ ³ 0 ®ÀG= ±C6@¯8 É .d°²1±18 ®~Á ´ä¹¾!d¿L».d«0 ® « GÁ ¯ Ö -/½CÓ Ù À ¬2Í2²@8~±@¬ ³0 ®ÀG= ±C6@¯8 É .d°²1±18 ®~Á ´ ¹ ½ d¿LÃ

15

yI )4,O!),t# 65.)L!$),M5K1*6"N76"),.,.d8 ® ±@¬ À 8~¬2°¯¯6C.d°¬<. G® ±@«Â´4¯U. À ¬<8~¯ À .1«<0 Á ¯U¬28 ®À~®=ÀÊÁ .76@¯Í\¯¯U« À ¬<0 ® ¯S0ÆÅ±@8 À ¬2Í2²@8~±@¬ ³0 ®ÀG= ±C6@¯8 É ». ® ´4¯U0 . Í<.1°.1« =® «2² GÁ ¯ À~® ¯±1ÅW.â²18G±1¬ ³ µ8G¯Å¯U8~8G¯U0 ±×. À Å. = ±18 ® «¹ ½2d¿¸Ç ®eGÁâ®GÀ 0 ®À 8 ® Í<¬ ~® ±1«<.1°¬2«c¬ À ¬<.1°«2¯ À~À µ8G¯Å¯8G8G¯U0 ±Ð. À Å. = ±@8¸Ã5ò Á ¯ ³ 8G± ³ ¯8~Ó~® ¯ À ±1ÅÅ¬2« =~® ±1« ÀW~Á . = ±1´"Í ® «2¯ ~Á ¯ À ¯ Ç±¶Å. = ±18 ÀÁ .761¯4Í\¯¯U«Ð¯Ñ ¯« À® 61¯U° ÉÀ ¬0 ® ¯S0µ ~Á ¯ 2Ó@Ó À~³ . = ¯ @¸Ã ® ´ ® °.d8G° É » ~Á ¯Ç¯ ® ² Á9 ¯S0Æ8G¯°. ~® 61¯. =U= ¬28:. =ÉÁ ¯¬28 ®À~®= »0¯ ð «2¯S0. À µ S 9¸ p (µ ç 5µ n ç(µ 7¸¸.d«0×¬ À ¯U0 ® «¹ )d¿N» 8G.@0¯ À ± E ²1¯U«2¯8:.d° ®EÉ ±1Å ~Á ¯"8G¬2°¯¶µ5µ ç» ® Ã ¯1Ã»\8G¬2°¯ = ±C61¯U8G.1²1¯S¸Ê.1«<0Î8G¯°. G® 6@¯. == ¬28:. =É µ5µ n 9¸(µ 7¸¸Ãdî¯ À® 0¯ À5À ¬ =:Á ±1Í2ïE¯ =G® 6@¯ ´4¯U. À ¬28G¯ À ±dÅ ® « ¯8~Ó¯ À~® «2²1«2¯ À~À » À ±1´4¯ À ¬2ÍïE¯ =~® 61¯ ´4¯U. À ¬<8~¯W±1Å ® « ¯U8~¯ ÀEG® «<²1«2¯ À~À ±1Å.Æ0 ®À~= ±C61¯U8~¯S0 ³ . ¯8G«= .d«Í\¯ .dÈ1¯U« ® « ± ~Á ¯". == ±1¬2« » À ¬ =:Á . À . =G® ±@«<.dÍ ® ° ®eEÉ µ . ³ . ~ ¯U8~« ®À+® « ¯8G¯ À~® «2² ® Å~Á ¯¬ À ¯8 = .1«0± À ±@´4¯ ~Á<® «2²Ç ®~Á¶®e ± Á2®À ±18 Á ¯8).10267.1« .d²@¯ ¸.1«<0Æ¬2«<¯Ñ ³ ¯ = ¯U0«2¯ À~Àµ . ³ . ¯8G« ®À)® « ¯8G¯ À~® «2² ± ~Á ¯W¬ À ¯U8 ® Å ®+®À+À ¬28 ³ 8 ®À~® «<² ± ~Á ¯¬ À ¯8 @¸¹ <¾¿NÃ

y y%(N7!#=O ;4'O45K1*#P6 )4,y!$),6C.d°¬<. G® ±@«±dÅ ® «<0¬ = ¯U0 À ¬2Í2²@8~±@¬ ³<À)® « ~Á ¯ Ö À~³ . = ¯Â¹ º d¿ À~Á ±CÇ À+= °. ÀGÀ~®eð ¯U8 ³ ¯8~Å±18~Ó´4.1« = ¯ ® « ¯8G´ À ±dÅÅ.d° À ¯.1°.18~´ ±@8 èCçëãæ@Cë:éI?LéI[Cã V è$?bã! #" p $%&(' u $%

µ ³ °± ~ ¯S0±1« ~Á ¯ ÎÓN.dÑ ®À ¸ GÁ . «2¯U¯U0 À) ±Í\¯W´ ® « ® ´ ® ¯U0u»<.1«<0 À ¯U« À~®eG® 6 ®EÉ ±18 ? V d<ãÊæ&@7ëéI?LéI[7ã V èh?Nã" p & %& % u $ '

µ ³ °± ¯S0±1« ~Á ¯ Ób.CÑ ®À ¸ ~Á . «2¯U¯U0 À+ ±Í\¯´.CÑ ® ´ ® ¯S0uÃ<ò Á ¯ ÖÀ³ . = ¯ ®À . ³2³ 8G± ³ 8 ® . ¯Å±18Ê´4¯U. À ¬28 ® «2² GÁ ¯ À ¬ == ¯ À~À ±1Å À ¬<Í2²18G±1¬ ³ 0 ®À~= ±C61¯8 É » À~® « = ¯ À ¬2ÍÓ²18G±1¬ ³<À Ç Á ± À ¯ " ) #" 8G.@0¯± E ®À= °± À ¯ ± ~Á ¯0 ® .d²1±@«<.d° = .d«×Í¯¶0 ®ÀG= .18G02¯U0â. À® « À~® ²1« ®ð= .1« Ãò Á ¯ À .1«<02.18G0. ³2³ 8G±@. =:Á®À ±Â¬ À ¯ ~Á ¯.d8G¯U.¬<«<0¯8 GÁ ¯ Ö= ±@«c61¯ÑÁ ¬<°°\0¯ ð «<¯U0Í É"À ¬2Í2²@8~±@¬ ³<À Ç ®~ÁÆ~Á ¯+Í\¯ À." ) #" 8G.@0¯± E . À . Û ¬<.d° ®eEÉ ´4¯U. À ¬28G¯Å±18 = ±@´ ³ .d8 ® «<² GÁ ¯ À ¬ == ¯ À~À ±1Å(0 ®E ¯8G¯« °¯S.d8G«2¯8 À Ã

* ¨ )¢+¥,+ 5¦.- ¢£

ò Á2®À Ç±18GÈÇÊ. À/À ¬ ³2³ ±@8 ¯U0Í É¶GÁ ¯ °±C6@¯« ® .d« ó® « ®ÀE 8 É ±1Å 0¬ = . G® ±@«Ü» =® ¯« = ¯".1«<0 ³ ±@8 »<.d«0 GÁ ¯Å¬2«0¯U0 ³ 8G±dïE¯ = Ù. . ó® « ® «2²Æ.d«0¶Ù/¯ =®À~® ±1« c¬ ³2³ ±18 Å±18î¬ À~® Ó«2¯ ÀGÀÖ ±@´ ³ ¯ G®eG® 6@¯«2¯ À~À.9 8 ¬28G± ³ ¯U.1« ® 8 ¬<.d°« ¯U8 ³ 8 ®À ¯Îµª òÓ:¾!CÓ:¾1¾@Ú@¸ÃªW.d´²18:. ¯Å¬2° ±Î3¬ = Ù/¯ +.d¯S0 » 2.c;À ±¶Ù; ¯8G± À È ® »·¯ ¯U8Ì(°. =:Á .1«<0ÒÙ8:.d²9.d« f .d´ Í\¯8G²1¯U8Å±18ïE± ® « 8~¯ À ¯S.d8 =:ÁÆ® «¶ªE35·.d«<0 À ¬2Í2²18G±1¬ ³ 0 ®À~= ±C61¯U8 É Ã

yY¤¢¨

[S?0/?21(aPVC7D:D:aHUJ9|Æ^?23ÊD4HD|1B?)\VS1L|1VJÂOPJ9|@Q9~BEOeVUJÆVUFaPVUSOPHUa5|@DG:OPMbOeVUJÂBbEDGD:M:?A1Q@`@TWOBbBED|WBEV6587.9!:<;>=?A@CBEDGFH:<IJI<KL@MON!:<NPOKABEQRDTSO@VU(MJWODGBYX[ZCMJWOZ\FO]]^DTN_WUR[ZUZ`1?

a ?+^Ü?3ÊD:9HUMbDSR5^?3Db4HUD|dB?dcOPJ@OeJ@HMbMbVC:OeHBEOPVSJÐEQ@aPD:MWOPJÝT/Q9aBEOP9aPD EDGaHBEOPVSJ9M:?[X[ZCMeBgf=@hSO@hNLQjik:<NLQfmlbMOZJnoI\iAM<?pMONq:<NrPOKABEQRDTSO@bU(MJWODRB!X[ZCMJWOZCFo]]DjN_WURAd@EOPJ9UDGR^ts * K"[ a ZuCR@?2[ aOvew [Y a Ru[ZUZud?

Y1?+^Ü?3Dx4HD|1B? * DGEMbDGGBEOPX7D!VSJOPJ9|@Q9~BEOeXSDaPVSSOP@EVUUHUTWTWOPJ@@?KLJdXdOBED|/aPD:GBEQ1EDHB^yziL@lbMOZCnoI\iAMh?qMON>K_ZhZC@hNLQFONrP6|tK_QRKZC@yLZC@hNPOI DjNU(MJWODGBXxZ\MJWOZCFO]^]DTN~WURuAdcHUSDGbBEVJR<BEVH9DHOeJ4A11EOeJ@SDG^ts0!A2R<[ZUZSZd?* XSHOPaH`9aPD+HBe ibm7q\iG1jOCkUfiNgCmi_~jo@mh7l~dqg1gg9ofÜi9q ?

16

@?+^Ü?~3ÊD84HUD|dB? * ED:aeHBEOPVSJ9HUa<|9HBHU`9HUMbDTWOPJ@OeJ@+CQ9DGbaHJ9UQcHUUDS?dKLJ6X[ZCMeBgf|zMOK_ZhQji lbMOZCnI\iAM<? MON 7.Z<QRD BEDGFOH:<NLQ@EHTH D W~@hNB<@ FONrP=O]<MOH DGB^ MO]V?rK_QFOQRDGMON@RdA11EOPJ9SD~(^ts * K~Rd[ZUZ`WIOPJ1ED:MbM~?

v ?+^Ü?~3ÊD84HUD|dB? * JWOPJ9|@Q9~BEOeXSDaPVUSOP1EVSUHTWTWOPJ9+CQ9DGbaeHUJ9UQcHSDFVU|9HBH`cHMbDTWOPJ9OPJ9ID1BEDGJc|@D:|"H`9MNBbHU~B~~?cKLJ!XxZ\MeBgfO:o>=~U(X 0?MOIJQB<MoNe@hZC@hNB<@ ,MOZCnoI\iAMh?k MO] ?rKH [email protected]@CF 9 @J@hQRDTN~WmMON MO]V?rK_QFOQRDGMONrFH2U(MJWODRBFONrP 9FB<i_DTN2@VU @CFOZ<NLDTN~WURc9<? a u w Y1R[ZUZ`1?

?+^Ü?~3ÊD.4HD|1BR//?~15aeVC7D:DGa?MbOPJ9+aPVSUOPHUa<|@DG:OPMbOeVUJ BbEDGD:M(FV:aPQ@MNBEDGEOPJ9@?LX[ZCMeBgft=@hSO@hNLQji:<NAQf lbMOZJnoI\iAM<?6MON^:<NPoKLBEQRDjSO@ UzMJWODGB XxZ\MJWOZCFO]^]DTN~WUR7Ad@EOPJ9UDGR^ts * K[ a ZudR9<?7[YUY w [7]dR[:ZSZud?

ud?+^Ü?3ÊD4HUD|dBR2^?L3ÊD:cHMbDS?Y(aHQ9MEHaÜ|@OPMb:VX7DGbC?z9FB<i_DTN@YU @CFOZ<NLDTN~WUR a a Y7h<ZUZ w [ R[:ZSZud?

`1?/? 3OeTWVUVSQ@aeVUM:R!A? 3 DGEVSMbdOHUJc| * ? ?"!+HSHM:?(KJCBED:UHBEOeJ@$#1@aHJcHBEVUbÐHUJc|k3ÊDGMb~EOe1BEOPX7DÊKLJ9|@Q@GBEOPVSJOPJ"KN^\?@KLJ!XxZ\MeBgfM%QjiL@'&(OQji6:<NLQ@hZ<NFOQRDRMONFH_MODTNLQV MON @hZJ@hNB<@MoN7.Z<QRD ,BEDGFHr:<NLQ@EHTH DW~@hNrB<@*)T:r(7V:+-,/.Rc9<?@ZS]S] w ZU]udR\[ZSZud?

Z1?A?3 DGEVSMbdOIRd^Ü?3ÊD8/ÊHMbDSR~1?~4Q@~R0?0HUaPaPDGC?t(aeHUMbMbO .9HBEOPVSJ VUF2EOPX7DG!HBEDG(dQ9HUaPOBL|@HBHÂQ9MbOPJ9THU9OPJ9D"aeD:HEJ9OPJ@@?ÔKJdX[ZCMeBgf0M1mQjiA@m|,D <Qji :<NLQf^ MONe@hZJ@hNrBh@MONdQjiA@b5@+SO@EH M<?r] @hNAQFoNP7 ??H DGBJFOQRDRMONM1 MO] ?LKQ@hZ y2@CB<i_NMOH MJWOD@hImQM32xNASeDTZCMONA] @hNAQFHx=LQRKAPOD@hIJ;4AMHfr:5X.MHjH KQRDRMONm9MeP~@EHjH DTN_WU?PRd9<?[ a Z w [:YuC?2(VSTW@Q@BHBEOeVUJcHa(cDG~9HUJ9OP:MÜQ9`@aPOe:HBEOPVSJ@M:RAdVSQ1BEcHUTW1BEVSJ<R[ZSZ@?

[:]1?A?L3 DGEVSMbdOIR^Ü?cVd|@VUEVXdMbdOI? 3ÊOPMb:VX7D~EOeJ@|11JcHTWOP:M +@EVSTËOPJc|1Q9GBEOPX7DÊaPVUSOP+@EVSHUTTWOPJ@ÊBEVTH~@OPJ9D|@OPMb:VX7DGbC?_XxZ\MeBgf y2@hNLQji#:<NAQf MON @hZJ@hNB<@0MON 9FB<i_DTN@8U @CFOZhNADTN_WURd@?Zu w []SYdRtcVUESHUJ6!+HQ@FTHJ9J<R<[:ZSZUY1?

[U[S?A?(3 DGEVSMbdOIRÜ^Ü? 3ÊD#4HUD|dBR /? 1(aPVC7D:D:aI?4D:aeHBEOPVSJ9HUa(EDGOeJ1FVE:D:TWD:JCBaPDHEJ@OeJ@@?uKJ 3?uHSDWID|2? xXxZ\MeBgf72[DWgi_Qji6:<NLQf MONe@hZJ@hNrB<@MONb:<NPoKLBEQRDjSO@ U(MJWODGBVX[ZCMJWOZCFo]]DjN_WURc@?2[S[ waOa RcA11EOeJ@SDGR9^zs * K[ R[ZSZO`1?

[ a ?A?3 DGEVUMb1OIRLs/?@^\HXCH8WID|1M:? :98@EH FOQRDGMoNFH5FOQF9DTNADTN_WU?9A11EOPJ9SD~R a ]S]1[U?[:Y1?0 ?;#ÜT|@DS?5^<DHEJ@OeJ@VUFÜcHHUGBED~EOeMNBEOP:VSJ@:D:1B|1D:MbGEOP@BEOPVUJ9MFEVSTËMbTHaea<MbDGBEMBEV:aeHUMN

MbO .9D|D+@HUTW9aPD:M:?2XxZ\MeBgf =2@hSO@hNAQji<2[K_ZCMh?A@CFONq MONe@hZC@hNB<@#MON9FB<i_DTN@0U @CFOZhNADTN_WUR^ts * Kuo`1Rc9<?2[]SY w [ a [SRAd@EOPJ9UDGR<[ZSZ@?

[+@?0 ?#ÜT|@DUR23?=0DGBbBEMb9DGEDG~?24D:aeHBEOPVSJ9HUaOeJ@MNBHUJ@:DG`cHMbD|aPDHEJ9OPJ91?>X[ZCMeBgfYyzi_DTZ<Q@J@hNLQji:<NAQf MONe@hZC@hNB<@MON!9FB<[email protected] @CFOZ<NLDTN~WURc@?<[ aa w [:YS]1R2cVUE7HJ6!+HUQ1FTHUJ@JR2[ZUZ ?

[ v ?>?+cH:ddHS|2R@?/?ÜOeHBEDGBEMbCdLA1cH9OEV@R?AdT/dBER 4?"BEdQ@EQ@MEHUT/CRD|1M:? 7.POSoFoNB<@hIDjNA0NM8H@CPeW~@ 50DTIhBJMOSO@hZ/bFoNP5FOQF9DTNADTN_WU?99DcKL uED:MbM:R[ZSZ v ?

[ ?+? * ? +aH~<?duED|@OPHBEDOPJ1XSD:JCBEOeVUJ+OeJ+OPJc|1Q9GBEOPX7D5|9HBH!D:J9UOeJ@D:DGEOPJ91?XxZ\MeBgfr=ADCBOQji2[KZ\Mh?A@CFON MoNe@hZC@hNB<@MONb9FB<i_DTN2@VU @CFOZ<NLDTN~WUR2A11EOPJ9SD~Rc^ts * K udRc@?A`SYZ@R[ZUZSY1?

[eud?+?,+aH~<R>s/?\^\H~@OP~@DS?(VCVSdOPJ9ÆQ9OPJCBED:EOPB¶GVSJ9MNBbHOPJdBEMOBEÎÜEOPTQ@M:?u\DG~@J9OPHUaEDGVUbBRJ9OPX7D~EMbOPB VUF15EOPMNBEVUaR\[ZSZO`1?

[ `1?+? * ?+aeH~"HUJ9| s?1^\H~@Oe9DU?([e1[V * .9EMNBbV|@DG[1H:CD:MbOeHUJ":aeHUMbMbO .cDG?KLJmXxZ\MeBgf M1 QjiL@oQji:<NAQ@hZhNrFOQRDGMONrFHl!MOZCnoI\iAMh?mMON^:<NrPOKABEQRDTSO@ U(MJWODGB,XxZ\MJWOZCFO]^]DTN~W )T:\U(X D.R@?Z aew []SYdRAd@EOPJ9UDG^zs * K[ Y@R[ZSZUZ1?

[:Z1?E?ÊHT`DGEUDGR23?=F^\HXCH8SRts?u a ]U] a 3D:MbGEOP@BEOPX7DOPJc|@Q@GBEOPVSJÆBE@EVUQ9SÆMbQ@`9UEVUQ9Â|@OPMNGVXSDGb2HHMbDMNBEQc|d)OPJHTWD|@OPHad|@VSTHOPJ?KLJX[ZCMeBgfM1G&OQji:<NLQ@hZ<NFoQRDGMONFOHz MONe@hZJ@hNrBh@MoNb9FB<i_DTN@ U @CFOZ<NLDjN_WH)T:o(9bUJI8KDKI.RcVUE7HJ6!+HUQ1FTHUJ@JRdOPJ41ED:MbM:?

a ]1? * ?J?OPV|9HUJ9H1R>?ÜA@HaPDS?u^<DHEJ@OeJ@MNBbEQ9GBEQ1ED|GVSJ9GD:@BEM+Q9MbOPJ@SD:J@DGBEOPHUaPSVEOPBE@TWM:?uKLJX[ZCMeBgfL DTNLQji :<NLQfYlbMOZJnoI\iAM<? MON!9FB<i_DTN2@VU @CFOZ<NLDTN~WURc9<?2[ Z w [euO`1Ru[:ZSZ a ?

a [S?03/? ? /ÊHQâHUJc|M#? 0?.(VUOPDGH1?^D:HEJ9OPJ@ÎdQ9HUaPOBHBEOPX7DÆTWVd|@DGaeM VUF/|d@J9HUTWOP4MN1MNBED:TWM:?9FB<i_DTN@ U @CFOZ<NLDTN~WUR a a Y7h[guu wa [ a Ru[:ZSZud?

aOa ? * ?N!+HHaeOOSR1K?15HBE7V1? +OEMNB!VU|1DG!ED:ED:MbMbOPVSJ?9FB<i~DTN2@8U @CFOZ<NLDTN~WUR a a Y7h<[u w [gu R[:ZSZud?

17

a Y1?c?!)OEMNBED:JR7A2?0¶EVU`D:aI?r4D:aeHBEOPVSJ9HUa9|@OPMNBHUJ@:DG`cHMbD|:aPQ9MNBEDGEOPJ@@?7KJ3?SuHUUD+ID|<? >X[ZCMeBgf2[DWgi_Qji :<NLQf MoNe@hZC@hNB<@ MON :<NrPOKABEQRDTSO@#U(MJWODGB X[ZCMJWOZ\FO]]^DTN_WUR@? a [ wa u]1R!A1@EOPJ@SDGR^ts * K[+ R[ZUZ`d?

a @?>!)aVUMbSD:J<R 0?N[ZUZ #J19aPVUH_ * TQ9aBEOPcHBbBEDGEJÝHUJc|TQ@aPBEOPMNBbHBED:UÝ|@OPMb:VX7D~bÒHMbMbOeMNBHJCB?9KJ3/? c? +9HddHU|<R;?/?cÜOeHBEDGBEMbdOLA19HU9OEV1R2?2A1TdBEHJc|b4+? BEdQ@EQ9MEHT/Æ#|@M:? 78POSoFONrBh@hI0DjN ANM8H@CPeW~@Y50DTIhBJMOSO@hZ/6FONP5FOQF 9DTNADTN_WUR a SZ w_a ud[U?zcKNâuED:MbM:?

aOv ?0s?!^HXCH8SRA?[3 D~EVSMbdO)HUJ9| c?@?EVU`D:aPJ9OP?^DHEJ9OPJ9ÎJ@VSJ@EDG:Q@EMbOPX7DÂ|@D.cJ9OBEOPVSJ@M"VFEDGaHBEOPVSJ9MOBEÆ^Ks )A?1KLJX[ZCMeBgfz|,D hQji*2xK_ZCM<?A@JFoN lbMOZCnoDTN_W6=@hIJI<DRMON MoN!U(@JFoZhNADTN_WUR@?av wa `d[S?<Ad@EOPJ9UDG?

a ?0s?2^\HXCH8HUJc|ÂA?23 DGEVUMb1OI? :<NPoKLBEQRDjSO@UzMJWODGBXxZ\MJWOZCFO]^]DTN~W-5 y2@CB<i_NLD KL@hI FONP67 ??H DOBJFOQRDRMONLIE? #ÜaPaPOeM./VbVCVd|2R\[ZUZ1?a ud?0s?d^\HXCH8URcA?_3 DGEVUMbdO2HUJc|mc?_sQ@THUV1?CKLJ9|@Q9~BEOeXSDaPVSUOPÊ@EVUUHUTWTWOPJ@+FVU!ED:aeHBEOeVUJcHa

dJ@VaPD|1SD+|1OeMbGVXSDGbC?;@8@hN2@hZCFoQRDGMON MO] ?rK_QRDTN~WW[eu"N[h9Y w_a YdR[ZUZSZ1?a `1?0s?<^\HXCH8SR+(? D:aPD J2R? +aeHU?z4D:aeHBEOeVUJcHa5MbQ@`9UEVUQ9|@OPMb:VX7DGb2 * @EVUVSMbOBEOPVSJcHa

O HBEOPVSJHU@@EVSHUBE@EVUQ9S .9EMNBbVU|1DGÜFDHBEQ@ED!:VUJ9MNBbEQ9~BEOeVUJ?~XxZ\MeBgfr:<NLQf MONe@hZJ@hNrBh@YMoN:<NrPOKABEQRDTSO@ UzMJWODGBYX[ZCMJWOZ\FO]]^DTN_WO;>:\U(X INKDKI7?A1@EOPJ@SDGR a ]U] a IOPJ1ED:MbM~?a Z1?0s?U^\HXCHSRS?+aeH~<R1?!+HX;MbD:RS^Ü?U\Vd|@VEVXdMbdOI?4Q9aPDOPJ9|@Q9~BEOeVUJ/FVUMbQ9`@UEVSQ@|1OeMbGVXC

D~b/OBEm s a NA_3?XxZ\MeBgf:<NLQ@GWOZCFoQRDTN_W7 I?A@CBEQRI M1x5FOQF9DjNLDTN~WO;25@JBEDjI<DRMONm=AKg??MOZ<Q FoNP9@hQFTU(@JFoZhNADTN_WUR;0¶VUEdMb9VU4HBBE@D>#,xc4^ !03Y3 a ]S] a (VUJ@FD~ED:J9:DURcOPJ4@EDGMbM:R a ]S] a ?

YU]1?0/?Ac4HUJ@J9OPaeH1RA//?@\VSOPX7VUJ9D:J<?JHJ4HUaPUVUEOBE9T FVU .cJ9|@OPJ9/HUaPa<OPJdBED~ED:MNBEOPJ9MbDGJdBEDGJ9:D:M:?@KLJ4+?@HU@9aIR@D|<?PRrXxZ\MeBgfxJ<@hZhN@hQRDGBEI^FONP#= oIJQ@h]I8 R99<?cZuUY w Zuo`1R[:ZSZ ?

Yd[S?03/?_cOP~@OPDS?(c4HU9OPJ9DaPDHEJ@OeJ@+OeJWBE@DJ@D+dB .cX7D!CDHEM:?zXxZ\MeBgf yzi_DTZ\P2[KZ\Mh?A@CFONlbMOZJnoDjN_W=@hIJI<DRMONMONbU @CFOZhNADTN_WUR99<?<[:]u w [ aOa ?<ÜOBETHUJR2[:Z``d?

Y a ?)K?VcV D~BEOU?)^D:HEJ9OPJ@ÐVUFCQcHUaPOBHBEOPX7D¶TWVd|@D:aPM:?KLJ×K~?.15HBE7VÒHJc| s/?^\HXCH8ÐID|@M:? X[ZCMJWOZC@hIJIDjNb9FB<i_DTN@ U @CFOZ<NLDjN_WURc@? a ]d[ w_a [guC?A1OPSTH/ÜEDGMbM:R\[ZO`uC?

YUY1?A?LcQ9SUaPDGBEVSJ<RcD|2?2:<NrPOKABEQRDTSO@YU(MJWODGB XxZ\MJWOZCFO]^]DTN~WU? * HS|1D:TWOP)ÜEDGMbM:R\[ZUZ a ?Y@?A?~cQ@SUaeD~BEVSJ?7KLJdX7DGEMbD@#ÜJCBHUOPaPTWD:JCBHUJc|ÜEVUSVSaI? @.@hN@hZCFOQRDGMoN MO] ?LKQRDjN_WO;(=O?A@CBEDGFH

DjI<IJKL@^MoN :<NrPOKABEQRDTSO@ UzMJWODGBYX[ZCMJWOZ\FO]]^DTN_WUR[Y@Y C~R\[:ZSZ v ?Y v ?A?cQ@SUaeD~BEVSJR? +9DGJ9@? # W:OPD:JCBOeJ9|@Q@GBEOPVSJ4VFÜaeVUSOP@EVSHUTWM:?[XxZ\MeBgfz|,DjZhIJQY MONe@hZ

@hNrB<@^MON!7VH WMOZ<DTQji_]^DGBYU @CFOZ<NLDjN_WyziA@JMoZRc@?9Y ` w YO`1[S?9TWMbcHdR@\VSCCV1R[:ZSZS]d?Y ?A?(cQ9SUaPDGBEVSJ<RÜ^?z3ÊD#4HUD|dB?KJc|1Q9GBEOPX7DaPVUSOP@EVSHUTWTWOPJ9@D:VUb¶HUJ9|TWDGBE@Vd|@M:?

_MoKZ<NFOH M1.U(MJWODGB XxZ\MJWOZCFO]^]DTN~WW[:Z Oa ] a Z wN uUZ1R\[ZSZ@?Yud?A? //? sOPD:J@1Q11MN\(@D:J91R4?5|1D 0VUaPF?x|zMOK_NPFOQRDRMONLIM%bDTNPOKABEQRDTSO@ H MJWODGB ?rZ\MJWOZCFO]^]DTN~WU?

Ad@EOPJ9UDG^zs * K[ aa `1R[ZSZud?YO`1?+uEVX7VSMNBR +(?NF +9H(:DGBbBR1? a ]S]d[,4VS`@Q9MNB5:aeHUMbMbO .cHBEOPVSJ"FVU(OeTW1ED::OPMbDD:JdX1OEVUJ9TWD:JCBEM:?

9FB<i_DTN@ U @CFOZ<NLDTN~WUR; a Y7~R a ]UY w_a Y1[U?YUZ1?+>1? 4+?Q9OPJ9aeHUJ<?@^DHEJ9OPJ9)aPVSUOPHUa9|@D .9J9OBEOPVSJ9MuFEVSTp4D:aeHBEOPVSJ@M:?9FB<i~DTN2@,U @CFOZhNADTN_W v a YSZ

a R\[ZUZS]d?S]1?c?SAdD:`cH@R?O4VSQ@X7D:OEVSaI?r(VSJ@MNBbHUOPJCBÜKLJc|1Q9GBEOPX7D(^VUSOPÜEVUUHTWTWOeJ@@?1KLJ/^?o3ÊD 4HD|1BR

D:|<?PR278POSoFONrBh@hI0DjNb:<NPOKABEQRDTSO@YUzMJWODGBYX[ZCMJWOZ\FO]]^DTN_WUR99? a uu w_a Z@R2K)AuED:MbM:R\[:ZSZ ?1[S? * ?2AdOeaP`DGEMbcHB R * ?\Q 9OPaPOPJ?JÆAdQ9`d_ND:~BEOeXSDcD:HUMbQ@EDVUFKJdBED~ED:MNBEOPJ9SJ@D:MbMOPJ6!)J@Va

D:|@SD3ÊOPMb:VX7DGbC?2KLJ X[ZCMeBgf|,DTZ<I<Qx:<NLQ@hZ<NFoQRDGMONFOH. MONe@hZJ@hNrB<@mMON A0NM8H@CPeW~@^50DTIhBJMOSO@hZ/FoNP5FOQF9DTNADTN_W3) AY5 5 .R@9<? a u v wa `1[SRu[:ZSZ v ?

a ? #!?A1VUTWTWDG?Y4Q9aPD:`9HUMbD/MNBbHBEO .cHBEOPVSJ9M * JÂH9@EVSHUÂBEVBE9DGVUb4ED:MNBbEQ@GBEQ@EOPJ@@?zX[ZCMeBgf|zMOK_ZhQji :<NLQfmlbMOZCnoI\iAMh? MONk:<NPoKLBEQRDjSO@mU(MJWODRBbX[ZCMJWOZ\FO]]^DTN_WUR"?0cb3LAdBEQ9|@OPD:J a YuCR(@?YuuLYUZS]dR[ZUZ@?

SY1? * ?\AdEOPJ9OPXUHUMEHUJ<Rt4? 3?7!)OPJ91RA?(cQ9SUaPDGBEVSJ<Rc? >@? #?uAdBEDGEJd`DGE1?9DWuED|@OPGBEOPX7D\V@OGVSaPVSH#XUHUaPQcHBEOPVSJ (cHUaPaPD:J@SDS?<KLJX[ZCMeBgf(|,D hQ@J@hNLQji :<NAQf _MODTNAQ0 MONef MON7.Z<QRD ,BEDGFHz:<N;Q@EHjH DW~@hNB<@~R@9? w Z1RcVUE7HJ6!+HUQ1FTHUJ@JR2[ZUZuC?

18

@?+^Ü?\Vd|1VUEVXdMbdOR<? +aH~<R2s/?c^\HXCHS?2uED|1Oe~BEOeXSDDGbFVUETHJ9:DVUF0DGOeUCBED|!4DGaHBEOPX7D* ::Q1HUGC?2KJ OPS9D:|<Rz3? * ?PR7!)VSTWVUEVMbdOIR>@?HJc| dBE7V)R>@?PRD|@OBEVUEM:R X[ZCMeBgfxM1QjiL@ Qji 2[K_ZCMh?A@CFONk MONe@hZC@hNB<@MONbX[ZhDjNBED ?H@hI0M1Y5FOQF9DTNLDjN_W FONPANM8H@CPeW~@ 50DTIhBJMOSO@hZ/)jX AY5 5I8KDKDK/.R99<? aOvv wa @RAd@EOPJ9UDGR a ]S]U]1?

v ?A?;0¶EVU`D:aI? * JÂHaeUVUEOBE9T FVUÊTQ9aBEOED:aeHBEOPVSJ9HUaÜ|1OeMbGVXSDGbVUF(MbQ9`@UEVSQ@9M:?X[ZCMeBgf(|,DTZhIJQ2[K_ZCMh?A@CFONb= o] ?MOIJDTK_] MON#XxZ<DTNrBED ?H@hIVM185FOQF9DTNADTN_W^FONPEANrM8H@[email protected]@hZ/URd@?uo` w `ud?2A1@EOPJ@SDGR2[ZUZuC?

?A?0¶EVS`DGaR!A2? 3 D~EVSMbdOI?!9DKL^\ |@D:Mb~EOe1BEOPVSJÒaPDHEJ9OPJ91EVS`9aPD:T!5\VH|@M"H¶UD:J@D~HUaÜTWVd|@D:aaPD:X7D:aÜ|1D .cJ@OBEOeVUJÆVUF!|@HBH TWOeJ@OPJ9OPJ¶KN^?2KLJ<!?zcVUEOPÂHUJ9|>@?t/DGbETHUJ@JRD:|@M:RAX[ZCMeBgfr|(FB<iWOZhKg??A@hNAQRZJ@V@hN9FoIEB<i~DTN2@EHTH@hI[U @hZ<N2@hN )j| 9bU D(+.R8 aa [Y3ÊVUbBET/Q9Jc|2RJ9OPX?3ÊVUbBET/Q9Jc|2Rc[ZUZ v ?

19

Ontology Enrichment with Texts from the WWW Andreas Faatz, Ralf Steinmetz

KOM - Multimedia Communications Lab Darmstadt University of Technology, Merckstr. 25, 64283 Darmstadt, Germany afaatz, [email protected]

1 Introduction

Automatic thesaurus and ontology construction dates back from the last three decades[1]. Our approach is a further development of methods to construct the whole ontologyautomatically. In contrast to these approaches our algorithm can only be applied, if weenrich an existing ontology instead of fully constructing the ontology.The following paper focuses on requirements for the semi-automatic enrichment ofmedical ontologies based on the statistical information of word usage. An ontology is astructured network of concepts from an knowledge domain and interconnects the con-cepts by semantic relations and inheritance. [2] gives a precise technical definition ofan ontology, that we will refer to throughout this paper:

Definition 1: An ontology is a 4-tuple := (C, is_a, R, ), where C is a set we call con-cepts, is_a is a partial order relation on C, R is a set of relation names and

: R is a function [2].Throughout this paper we assume that a concept has a character string as a descriptor.This character string may be a word or a phrase.

For our purposes we will neglect R as well as and focus on is_a as the particular re-lation, which is responsible for superconcept-subconcept dependencies. For examplebacteria is a superconcept of the concept pathogenic bacteria. Whenever we talk of ’re-lations’ or ’relational paths’ in the following sections, we refer to the is_a relation. Wealso define

Ω σ

σ ℘ CXC( )→

σ

Abstract. The following paper explains, how we can enrich an existing ontology bymining the WWW. The use of such an ontology may be manifold, for example as acomponent of information systems or multimedial repositories. The enrichment proc-ess is based on the comparison between statistical information of word usage in alarge text collection, a so called text corpus, and the structure of the ontology itself.The text corpus will be constructed by using the vocabulary from the ontology andquerying the WWW via Google.We define similarity measures by optimising their parametrisation and examine thecentral properties of the enrichment approach - along with the presentation and eval-uation of experimental results. Parametrisation of a similarity measure means assign-ing weights to each word collocation feature we first check in the text corpus andthereafter integrate into the representation of a word or a concept.

20

tologyomaina by

in do- the

hose

con-

(hu-

aturalr the

ally,ctiont cor-cuss-

andany

that a oftenotionht to by ancetion

Definition 2 : We call the restriction of an ontology := (C, is_a, R, ) to (C, is_a) thehierarchical backbone of .

Ontologies give a formal representation and conceptualisation of a knowledge domain,which is useful for the administration of large multimedial resource collections: if on-tologies reflect an agreement of a group of experts and are rich enough in the sense ofa sufficient number of concepts, ontologies are able to handle information exchangeacross the borders of one expert’s vocabulary. For example, one could ask an onto return the names of all bacteria causing diarrhoea and in this way access dknowledge without the barrier of finding out the names of the particular bacterireading texts from the domain of infectiology.Naturally the construction of an ontology is hard and expensive, as one has to tramain experts in formal knowledge representation. This is the motivation behindidea, that for a given ontology we focus on finding new concepts automatically. Tnew concepts are propositions, which extend the given ontology. For this we

• use a special text corpus derived from WWW search results• detect a set of candidate concepts from the corpus • finally select a subset of those candidate concepts ranking their similarity to

cepts already existing in the given ontology.

The final selection ends up in new concepts for the ontology to be proposed to aman) ontology engineer.The concepts have one or more descriptors, which are words or phrases from nlanguage. This implies that we develop our method finding suitable definitions fosemantic similarity of words or ordered sets of words.The paper is organised as follows: in section 2.1 we describe our approach informwhereas in 2.2 we give a survey on formalisation, which is explained in detail in se2.3. Section 3 deals with the experimental results on two very different kinds of texpora and especially on mining propositions from Google search hits. Section 4 dises related work. At last section 5 summarises and points to future work.

2 Enrichment Approach

2.1 Overview

Similarity between words is a topic from the theory of word clustering algorithms requires statistical information about the context, in which the words are used. Mapproaches check collocation features of the words in large text corpora, such word is represented by a large vector. The vector has entries communicating, howa collocation feature was fulfilled in the corpus. The vectors are sparse [3]. The nof similarity definitions by vector representations normally does not assign a weigevery single dimension of the vectors. In this paper we argue, that this is possiblesoft method using the information already defined in the given ontology. The influeof the ontological structure on the word (-vector) similarities results in an optimisa

Ω σΩ

21

e

pathseas-

cators

radict

ic dis-onal in-

n-evel in

nd the

a-suresd en-ty ofs of en-

Let-

problem, which determines which dimension in the word (-vector) representation is in-fluential for the similarity computation. The following definition should clarify, what acollocator is.

Definition 3: Let a word w be given. A collocator of w is a word, which occurs togetherwith w due to a predefined rule in a text collection (text corpus).

Thus for example in the phrase ’Medical ontology enrichment in the k-med project’ ,’enrichment’ and ’medical’ would be collocators for predefined rules like for instanc’maximal distance 5 words’ or also ’occurrence in the same sentence’.The way we include the ontological information from := (C, is_a, R, ) may beguided by different heuristics on a numerical interpretation of is_a, R and . For exam-ple the abstraction level of a concept, the interconnection by relations, relational and their lengths or the local granularity of the modelling can establish distance mures on a given ontology . Our goal is a comparison of this distance measures to the information about colloin a text corpus.

2.2 Enrichment as an optimisation problem

The core idea of our approach is computing enrichment rules, which do not contthe distance information already given by the ontology we want to enrich.We first have to state a basic assumption.

Assumption 1: There exists a consistent distance measure d expressing semanttances between the concepts in . The distance measure is based on the relatiterconnections between the concepts in the hierarchical backbone of .

By consistency we mean, that d underlies some characteristic heuristics: a long relatioal path between two concepts rises the distance between them, the abstraction lthe hierarchical backbone influences the distance measure d as well as the number ofconcepts, which are subconcepts to the same superconcept. Both abstraction anumber of siblings rise the distance. The distance measure d, which we will from nowon denote by d(x,y) for concepts x and y from also differentiates between generalistion and specialisation in an ontology. We showed in [4] that such distance meaexist indeed. Thereby our notion of ’consistent’ does not necessarily imply a goorichment quality, but just means, that the above heuristics are fulfilled. The qualithe heuristics and the resulting distance measures have to be judged by the resultrichment experiments.We also assume a text corpus to be given and determine the ordered set K of the nmost frequent collocators which cooccur with at least two concepts from .

be a vector; the i-th entry of this vector v(x) expresses, how often the descriptor of the concept was a collocator in with the i-th element of K.

Ω σσ

Ω

ΩΩ

Ω

ζΩ

v x( ) Rn∈

x Ω∈ ζ

22

e

n con-

cate-

lied

chical For

n- a set

theing in dur-

rmscol-

uteson- to be

Let now be a component-wise monotonic function of the dissimilarity betweentwo concepts x and y from . For the dissimilarity we postulate, that it must be com-putable from the vectors v(x) with concepts x from .The parametrisation just weighs each collocation feature positively, itindicates, how strong the j-th dimension is involved in the dissimilarity computation.The optimisation process consequently fits the average of the ( (v(x), v(y))) forpossible with each , to the distances d(x,y) for each pair of con-cepts from .To sum it up briefly, the optimisation establishes a dissimilarity measure, which is asnear as possible to the distance measure d in . In the next section we present a formal-isation of the algorithm. We decided to explain the details of this formalisation to makea repetition of experiments with the algorithm possible.

2.3 Formalisation of the algorithm

A distance measure on is a function d: (C X C) [0,1]. Examples of distancemeasures are:

1) d(x,y)= , where e is Euler’s constant and s denotes the number of steps along thshortest relational path between the concepts x and y. This distance definition corre-sponds to the heuristics, that long relational paths rise the distance between givecepts.2) d(x,y)=1, if there exists a relation between the concepts x and y and d(x,y)=0, if theredoes not exist a relation between the concepts x and y. This definition has only a rea-sonable application to ontologies, if the transitivity of the is_a-Relation and the connation of different relations from R is clearly stated in the a of axioms of . 3) [5] defined criteria for similarity measures in thesauri, which in turn can be appto distances in the hierarchical backbone of the ontology := (C, is_a, R , ). [4] showed, that there is an infinite number of distance measures on the hierarbackbone of an ontology fulfilling more restrictive characteristics than 1) and 2).further details we have to refer to [4].

A text corpus is a collection of text documents written in exactly one natural laguage. We assume to be electronically available. From a text corpus we defineof words or phrases to be the candidate concepts. A proposition for the ontological en-richment is a word or a phrase from , which is used similarly to the concepts fromgiven ontology. Candidates are to be predefined, for example as all nouns occurr

. Note that might also be extended by additional text material. This may happening or after the application of the enrichment algorithm.A rule set is a finite set of linguistic properties, each of which can be tested in teof its fulfilment frequency in the text corpus. In our case we will always consider location properties for the rule set .The entries of a representation matrix list, how often the j-th propertyfrom was fulfilled in for the descriptor the i-th concept from C.The enrichment algorithm processes information available from and . It compan optimal solution for the problem of fitting the distance information among the ccepts expressed by and the dissimilarity information between words or phrasesextracted from the word usage statistics considering .

fk DΩ D

Ωk k1 kn, , ,( )=

fk Dk k1 kn, , ,( )= ki 0≥

Ω

Ω

Ω →

es

Ω

Ω σ

ζζ

ζ

ζ ζ

ρ

ρmij

M C ρ ζ, ,( )ρ ζ

ζ Ω

Ωζ

23

Let us assume a given . We search for a set of non-negativereals with , which will be called configuration of the rule set . Each corre-sponds to a rule .The configuration k decides about the quantities of dissimilarity we derive from

. The Kullback-Leibler divergence generally measures the dissimilarity between twoprobability mass functions [6] and was applied successfully to statistical language mod-elling and prediction problems [7]. The Kullback-Leibler D(x,y) divergence for twowords x, y is defined as

In the basic version of the Kullback-Leibler divergence, which is expressed by formula(1), w is a linguistic property and is the probability of this property being ful-filled for the word x. In the sum indicated by formula (1), w ranges over all linguisticproperties one includes in a corpus analysis. In our case the frequencies of observingthe collocation properties are denoted by . We change (1) in such a way, thatk weighs the influence of each property w:

with in our caseConsidering our representation matrix notation we obtain

Let us clarify the notation of formula (3): denotes the i-th concept from C. Correspondingly in (3) the are the matrix en-

tries in in the row expressing the collocation properties of . With this no-tation holds. In that sense, we will be able to determine an optimal

.Taking the distances from the ontology as an input, which should be approximatedby the as well as possible, the question of finding an optimal configuration kreduces to the question:which configuration k minimises the average squared error expressed by the differences

?

M C ρ ζ, ,( ) k k1·

kn, , =k ρ= ρ ki

ρi

M C ρ ζ, ,( )

D x y,( ) P w x( )P w x( )P w y( )-----------------log

w∑= (1)

P w x( )

M C ρ ζ, ,( )

Dk x y,( ) k w( )P w x( )P w x( )P w y( )-----------------log

w∑= (2)

k w( ) k∈M C ρ ζ, ,( )

P w xi( )mil

min

n 1=

ρ∑

--------------------

l 1=

ρ∑= (3)

xi milM C ρ ζ, ,( ) xik xi( ) ki=

k k1·

kn, , =Ω

Dk x y,( )

d x y,( ) Dk x y,( )–( )2

24

cepts

owl-dialources

chicalain are

’rgeruring

outf in- main

Finally we present a formulation of this question in terms of a quadratic optimisationformula. Searching for an optimal k means searching for a minimum of the followingfitness expression:

where and for all . Note that we minimise over the setof all configurations, that means over all possible k. We now explain, which wordsphrases are propositions for the ontological enrichment.Once we optimised formula (4) we obtain the configuration in need to compute all thedistance measures between all the concepts from and the candidates. We apply anenrichment step starting with the optimal similarity measures . We only take into concern the with and a candidate y. If such a distancebetween a formerly known concept (i.e. its descriptor) and a candidate (i.e. a word fromthe corpus) formerly unknown to is lower than a predefined threshold, y propositionto enrich . A suitable threshold can for example be defined from the average of thedistances d(x,y) where x ~ y holds for some . Additionally the with and a candidate y carry even more information,namely an optimal placement of the candidate concepts. The candidate concepts and theconcepts from can be presented together, if a candidate turns out to be a proposition.This simplifies the knowledge engineer’s understanding of how the candidate conevolved and in which semantic area of they might belong.

3 Experimental results

3.1 Basic input: ontology and corpus

The ontology we enriched in our experiments is a modelling carried out by a medicalexpert during the first phase of the k-med project. K-med is an abbreviation for ’knedge based multimedia medical education’. This project tries to collect multimemedical learning resources and for the sake of reuse describe the educational resby applying a metadata scheme and a common medical ontology [8]. The ontology contains the most abstract concept disease OR symptom along with thesubconcepts measles, German measles, diarrhoea, intestinal infection. Diarrhoea itselfhas the subconcepts aqueous diarrhoea and sanguinary diarrhoea. The ontology maybe viewed as an incomplete test ontology for several reasons: it only uses hierarrelations, the superconcept and subconcept relations and also the knowledge domnot fully clarified (such that a construction like ’disease OR symptom’ with a logical ORbecomes necessary) and at least one additional abstraction level (a concept likeinfec-tions’) could make the modelling clearer. In fact, this ontology is just a part of a laontology under construction. It is based on the subjects, which have to be tought dthe first semesters of medical education in Germany.To sum it up, we enriched this intuitively modelled ontology in the sense of [9], withdeductive or inductive rules or axioms. From our point of view this enrichment ocomplete ontologies is, along with the extension of thesauri and catalogues, the

min d xi xj,( ) Dk xi xj,( )–( )2

i 1=

C

∑i 1=

C

∑ k

(4)

k k1 kn, , , = kl 0≥ kl k∈

ΩDk x y,( )

Dk x y,( ) x C∈

ΩΩ

∼ R∈Dk x y,( ) x C∈

Ω

Ω

Ω

25

, as

Googledocu- con-

of 70 of thee didm theary

t the

ilarity in theperi-

application area of our approach. Furthermore following [9], such a rather informal on-tology represents a situation, where machine learning techniques should support theknowledge engineer.The size of the ontology was kept small for the sake of a rapid application and evalua-tion. For the experiments presented in the remainder of this paper, it was easier to re-duce the possible interdependencies by referring to this ontology chunk containingseven concepts.As larger ontologies can be segmented to smaller ones - for instance to speed up thecomputation - we consider the enrichment of the ontological chunk a good starting pointfor experiments.The computation of the represenation matrix and the derivation of the optimisationproblem was carried out by an implementation of our own. For the sake of a possiblelater connection of this component to other existing ontology tools and the linguisticworkbench TATOE [10] we used Smalltalk as the language to implement the algorithm.The quadratic optimisation (4) itself was carried out by the ampl-solver [11].For our first experiments, we used a general, but very large (about 28.700.000 sentenc-es) newspaper corpus available at [12]. The corpus can be queried by on-line querytools, which also provide stemming and lemmatisation techniques. All queries are col-location queries determining, whether two words were used in at a distance of prede-fined size. Although these experiments produced bad enrichment results from themedical expert’s point of view, very important meta-properties of the algorithmcompression of the rule set and a stability of the algorithm were found.A second experiment was based on the search results of the web search engine [13]. We passed each descriptor of a concept to Google [13] and converted the ments belonging to the 10 search hits with the highest ranking into a text corpus. Intrast to the newspaper corpus this corpus is more specialised, consisting documents with 135.166 words and 15.570 sentences. Because of a restrictionconcordancer freeware in use (Wconcord, Darmstadt University of Technology) wnot apply stemming and lemmatisation to the specialised corpus we gained froWWW. In our Smalltalk implementation we included a stop list consisting of auxiliverbs, conjunctions, personal pronomina and prepositions.For the rule set we always tested, how often a collocation at maximum distance fivetokens, but in the same sentence took place.All of our enrichment results can be found in table 1. In the column at the very lefreader will find the concept from , in the middle column the concepts from the gen-eral newspaper corpus proposed to the concept from . At the right we find the prop-ositions to from the special corpus.In both experiments a candidate became a proposition, if we computed a dissimbelow 0.5. The reason for the choice of this threshold was, that a path of length 2hierarchical backbone of the test ontology led to a distance average of 0.5. All exments were carried out in German, so we show translations here.

ζ

ζ

ρ

ΩΩ

Ω

26

s
With ’stomach ache’ and ’medical doctor’ in German we did not propose word groupbut composita (’Bauchschmerzen’ and ’Arzt’ in German).
Table 1: enrichment results

concept from general corpus (28.700.000 sentences)

special (www) corpus (15.500 sentences)

disease OR symptom loss illness infections bodylegwound animals vaccinationfevercombat

intestinal infection medical doctorcause

diarrhoea ailmentepidemiccoughfevervaccinationinfections

vomitstomach achenauseafevermedical doctor

measles

German measles

sanguinary diarrhoea vomitstomach achenauseafevercan

aqueousdiarrhoea

vomitstomach achenauseafevercan

Ω

27

verlyni-rop-

ed theix ofencess,

.

ing inries

3.2 Results for general corpora We refer to the enrichment results depicted in table 1. As the rule set we used collo-cation in the same sentence at maximal distance of five tokens in .Although we find some concepts like ’infection’ which is a missing abstraction level inthe ontology, the enrichment results from general corpora are poor. We retrieve ogeneral propositions like ’body’ and even flaws like the proposition of ’wound’, ’amal’ or ’leg’. These flaws occured especially with candidates, which only had one perty from . Another problem occurs with special concepts - as expected, even a very large newspa-per corpus does not contain enough information to get propositions for the subconceptsof diarrhoea.Although we faced these problems, the experiments for general corpora were worth-while, because we identified two considerable meta-properties of the approach: stabilityand compression.

a) Stability

Before dealing with the core of the enrichment - the proposal of concepts - we testinner stability of the approach and pruned , which was a square matrsize 102, in two different ways. If contained not enough information, thdifferent pruning strategies would bare the danger of destroying the enrichment prowhich means leading to inconsistent optimal configurations or enrichment results

The first pruning strategy was keeping only the ten largest entries per row, resulta 102 X 34 matrix . The second pruning strategy only kept the ent

with >10, resulting in a 102 X 63 matrix. With both ma-

Table 2: stability

first strrategy

second strategy

suffer suffered

diseases fever

illness measles

hepatitis

die died

pregnancy pregnancy

percent percent

stomach

ρζ

ρ

M C ρ ζ, ,( )M C ρ ζ, ,( )

M C ρfirst ζ, ,( )cij M C ρ ondsec ζ, ,( )∈ cij

28

’

th-nze-tation

ine and the

m theo thehe 12

trices we set up the optimization procedure, solved expression (4) respectively and de-rived two optimal configurations and . The collocators belonging torules with nonzero weights are listed in table 2. Both strategies mean collocation atmaximal distance five words with a descriptor from the respective column of table 2.Let us comment the collocators remaining from the two pruning strategies and the op-timization. In table 2 we listed the collocators in such a way, that we immediatly see therelation between the two sets. Some of the collocators do not differ at all, some of themare only different in terms of the grammatical context they stem from (for example suf-fer’ and ’suffered’), some of them obviously carry a semantic relation (’diseases’ and’ illness’ on the one hand and their more specialised pendants ’feaver’, ’ measles’ and’hepatitis’ on the other hand). The only collocator without any direct relative in the oer set is ’stomach’, so we state, that the analysis of the resulting collocator sets of noro weight does not show any inner contradiction in the approach, the represenmatrix in our case seems to be stable and even carrying redundant information.

b) Compression of

With both pruning strategies only a few properties from achived a correspondingweight from k, with . This compression of the rule set also occured for differ-ent definitions of the ontological distance d. We assume, that this compression is closelyrelated to the sparsity of our representation matrix and to the structure of the optimisa-tion formula (4). The reason for this assumption are further experiments with artificiallyand randomly generated matrices, which we used as pseudo-representation matriceswith sparsity structures similar to the ones of and .As these experiments ended in a similar compression, we will search for a proper math-ematical reason why this reduction of influential features with nonzero weight takesplace.

3.3 Results for a special corpus based on Google hits

As we mentioned, we passed each descriptor of a concept to a web search engconverted the first 10 hits of the Google search result into text files, removingHTML-specific tagging. The results can be found in table 1.As the rule set we used again collocation at maximal distance of five words in . Forpruning reasons from our representation matrix we kept only properties from , whichwere fulfilled for at least two concepts from the small test ontology. This resulted in arepresentation matrix with 292 columns. This means, that - for the special corpus - weinitially found significantly more rules than with the common corpus, but after the so-lution of (4) we obtained 12 properties from with a nonzero weight. These were dis-tance five in the same sentence with chronic, infection, because of, seldom, diarrhoea,vaccinate, pneunomia, virus.The choice of the candidates for the enrichment was driven by the observation froprevious experiments: propositons with only one nonzero property induced flaws tenrichment. Consequently we accepted candidates, which at least fulfilled two of tremaining properties from .

kfirst k ondsec

ρ

ρi ρki ki 0>

M C ρfirst ζ, ,( ) M C ρ ondsec ζ, ,( )

ρ ζρ

ρ

ρ

29

ist

eval- con-tionalgeprop-

akingvalua-

hichwhich

orpus-

asiere more

r is not

lloca-

oreanced,

usflec-

ied to

sub-ective

The main results of the experiment with the special corpus crafted from web search hitsare• enrichment of the special concepts sanguinary diarrhoea and aqueous diarrhoea• identification of a group of symptoms (vomit, stomach ache, nausea, fever) as

propositions• lack of propositions for measles and German measles

The flaws in this enrichment are can which should actually be a member of the stopland medical doctor which is at least too overgeneral.

3.4 Discussion of the results

As a second general observation we state, that the main flaws identified during theuation in this section come from candidates, which share only one feature with acept from . We conclude, that a possible way to handle this may be an addituning of the definition of d. A potential technique for this is generating artificial usaprofiles, which do not at all reflect real words, and searching for a measure, which erly discriminates between real words and randomly constructed feature sets. There exist subjective and objective ways of evaluating the results. Roughly spethe objective evaluation methods base on reference ontologies, the subjective etions are based on expert interviews [14]. The question posed in the subjective evaluation was: Consider the ontology and the table of propositions from table 1 to be given. Wstrategy performs better? Which aspects of the enrichment results are positive, ones are negative? The subjective evaluation clearly showed, that the enrichment with the special cperforms better, as there are less flaws like wounds or leg and also less overgeneralisations like illness. In addition to this, the results of the specialised enrichment are eto perceive, as there are not too many propositions and the good propositions werprecise.The fact that our specialised enrichment with a web-based corpus performs betteas trivial as it seems. For instance, the symptom group vomit, stomach ache, nausea, fe-ver was also on the candidate list for the general corpus experiments - but the cotion information was insufficient, although the corpus was very large. Comparing the experiments, possible causes for lacking propositions for measles andGerman measles may be found in the structure of out test ontology. It contains minformation about diarrhetic diseases. At least the special text corpus must be bala preprocessing step we will include in future experiments.Other good candidates (like ’virus’) did not become propositions in the special corpexperiment, as we did not use stemming and lemmatisation. This introduces all intions of verbs to the candidate set. Instead of this, the inflections should be unifone candidate. The main goal of a series of evaluations should be finding a correlation betweenjective and objective measures. If such a correlation exists, even imprecise obj

Ω

Ω

30

evaluation measures can give answers to the quality of parameter tuning or candidatestrategies. The reason for this is, that the objective evaluation measures have to correlateordinally but not cardinally with the subjective ones. Objective evaluation measures we are developing are guided by the notion of precisionin document retrieval [1]. Naturally we need a larger series of experiments to detect apossible correlation between objective and subjective evaluation measures.

4 Related work

Two main branches of automated ontology construction by natural language processingin general and checking collocators in our special case may be identified: those, whichbase the similarity of concepts or their descriptors on syntactic criteria or collocationdirectly (we will refer to those as first type) and those, which take statistic samples ofthe features of a concept or its descriptor. For example the first type declares conceptsor their descriptors as similar, when they often occur together in one sentence. Theworks of [15]and [16] are examples of this type of enrichment. The second type declares words as similar, if they are used in an similar context. Forexample, if a word w is used with a word v in the same sentence very often, and also aword u is used with a word v in the same sentence very often, then w and u would besimilar according to the assumptions of the second type approaches, even if w and unever appear in the same sentence all over the text corpus. Note the significant differ-ence between the approaches: the first type would state a high similarity between w andv, also between u and v. Representatives of the second type are the works of [17], [18],[19] and also [20]. The latter ones use syntactic parsing instead of pure collocation in-formation. [20] moreover does not assume identity between concepts and their descrip-tors as we did in this paper.If we try to group our approach among the first or second type approaches, we can clear-ly state that the way we define word similarities and dissimilarities is due to the secondtype as we pointed out in the previous section. Nevertheless the way we restrict ourview to certain candidate concepts and their respective features also refers to the firsttype: a candidate concept is a concept, of which we determine similarities or dissimilar-ities to all ontological concepts. As for almost natural computational restrictions we arenot able to compute the dissimilarities for all the words from a corpus, we must restrictour observations. Our candidate strategies in section 3 explain possible restrictions indetail. They are influenced by ideas of the first type.Although - except [21] - none of the related works mentioned here directly focuses onontology enrichment from the WWW, all these works have in common, that their auto-mated construction features can be extended to our enrichment goals of propositionidentification and placement. [15] used pure collocation information for gaining new concepts, but also focuses onqualitative issues of the collocations to derive statements about relations and the behav-iour of the relations.[16] focused on the so-called salient words, which are able to disambiguate word mean-ing very well. In a way [16] is also a extending special form of ontology, namely a the-saurus. In contrast to our approach his focus is on disambiguation, which was further

31

omnt, andpinionn ex-

easensederallyura-.iningorm-m].

euralent basis

ologyal cor-omainm the of a

thm. treat-

thodon-mp-tion.

them from,ale forhe op-

developed by [21], but with web resources. In our approach a concept can be proposedto two different concepts, but may also disambiguate, depending on our ontological dis-tance measures. Especially the identification of the symptom group vomit, stomachache, nausea, fever propositions for several concepts from would be simply impos-sible with the approaches of [16] or [21], as they found on discriminating descriptors.They achieve this by a test which detects the mostly diverging contexts of concepts ofa given ontology (or thesaurus).[19] designed the Mo’K workbench for word clustering and building hierarchies frthe clusters in a second step. We also implemented a feature based environmeour goal is even more specialised than word clustering. However we share the owith [19], that collecting many features and assigning weights to the features is acellent basis for similarity definitions. Our implementation is in Smalltalk, wherMo’K uses C. Smalltalk remains a possible language, as we end up in very condrepresentations with a few features of nonzero weight (compare section 3.2). Genspoken, in contrast to our work (which systematically computes an optimal configtion k for a rule set ) [19] do not explicitly offer a strategy for choosing the weightsAn approach related to [19], but more founded on collocation networks and determartificially specialised corpora can be found in [18] and a comparison of the perfance of specialised and common corpora in word clustering can be found in [AsiuFinally [17] experimented with word similarities expressed by an unsupervised nnetwork algorithm, the Kohonen map [23], but also for clustering, not for enrichmgoals. Our evaluation methods result from the enrichment goal and could also be afor an evaluation of the Kohonen maps in word clustering, a task desired by [22].

5 Conclusion and future work

We presented a method for ontology enrichment and applied it to a medical ontchunk. Our evaluation shows, that the strategy to derive propositions from a specipus seems to end up in clearer and more error prone conceptual propositions to a dexpert. The experiments have to be repeated with other specialised corpora froweb, the major task for future work. Instead of Google one could refer to the hitsmedical search engine like Medivista.From all our experiments we identified very important meta-properties of the algoriThese are a possible basis for future extension of the algorithm: a more systematicment of the initial rule set by gradually extending the word distances can be achieved,if we are able to keep the compression property of the algorithm. From our point of view the integration of the presented algorithm in a Delphi me[9] for k-med-like projects or the evolving semantic web is very promising. In the ctext of the task of the k-med ontology, automatically identifying a whole group of sytoms is especially helpful for managing documents for case-based medical educaA number of other interesting questions comes along with the approach. Amongare the following ones: how do we construct and balance a suitable corpus to learnwhich linguistic preprocessing is necessary or helpful, how does the approach sclarger ontologies. The latter question is again closely related to our observations: t

Ω

ρ

ρ

32

ce,

wl-

inki,

)

g

-

ttle,

timisation problems end up in an identification of a few relevant collocation featuresand the representation matrix can stand a pruning preprocessing.Also the question of evaluating the results is interesting for related areas such as Ko-honen maps for documents and word clustering algorithms [17].

References

1. Spark-Jones K.: Readings in Information Retrieval, Morgan Kaufmann, 19972. Stumme, G., Mädche A.: FCA-Merge: A Bottom-Up Approach for Merging Ontol-ogies, Proceedings of the 17th International Joint Conference on Artificial IntelligenSeattle, USA, August, 1-6, 2001, San Francisco/CA: Morgan Kaufmann, 2001 3. Sahlgren,M.: Vector-Based Semantic Analysis: Representing Word Meanings Basedon Random Labels, Proceedings of the ESSLLI 2001 Workshop on Semantic Knoedge Acquisition and Categorisation, Helsinki, Finland, 20014. Faatz, A., Hoermann, S., Seeberg, C., Steinmetz, R.: Conceptual Enrichment of On-tologies by means of a generic and configurable approach, workshop notes of the ES-SLLI 2001-workshop on semantic knowledge acquisition and categorisation, HelsFinland, August 20015. Resnik, P.: Semantic Similarity in a Taxonomy: An Information-Based Measure andits Application to Problems of Ambiguity in Natural Language, Journal of Artificial In-telligence Research, 11, 19996. Kullback, S. and Leibler, R.A.: On Information and Sufficiency, Annals of MAthe-matical Statistics 22, 19517. Dagan I., Perreira F., Lee L.: Similarity-based Estimation of Word Co-occurence Probabili-ties, Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics,ACL’94, New Mexico State University, June 19948. Stefan Hörmann and Ralf Steinmetz: Creating courses with learning object metada-ta, Multimedia Systems, Springer, Berlin/Heidelberg, to appear 20029. Clyde W. Holsapple and K.D. Joshi: A collaborative approach to ontology design,Communications of the ACM, Vol. 45, 2, February 200210. Alexa, M.: Text type analysis with TATOE. Storrer, A. & B. Harriehausen (eds.(1998): Hypermedia für Lexikon und Grammatik. Gunter Narr Verlag, Tübingen.[Asium]: D. Faure and C. Nedellec: ASIUM: Learning subcategorization frames andrestrictions of selection, in Y. Kodratoff, editor, 10th Conference on Machine Learnin(ECML 98) -- Workshop on Text Mining, Chemnitz, Germany, April 1998 11. ampl optimisation software, http://www.ampl.com12. Institut für deutsche Sprache, www.ids-mannheim.de13. the Google search engine, www.google.de14. Andreas Faatz: Enrichment Evaluation, technical report TR-AF-01-02 at Darmstadt University of Technology15. Heyer, G.; Läuter, M.;Quasthoff, U.; Wittig, Th.; Wolff, Chr.: Learning Relationsusing Collocations, Proceedings of the IJCAI Workshop on Ontology Learning, SeaUSA, August 2001

33

ngi-tober

ec

uda-

a

16. David Yarowsky: Word-Sense Disambiguation using Statistical Models of Roget’sCategories Trained on Large Corpora, Proceedings of COLING-92, Nantes, France,199217. Lagus, K.: Studying similarities in term usage with self-organizing maps. Proceedings of NordTerm 2001, 13-16 June, Tuusula, Finland. pp. 34-45. ,200118. Gaël de Chalendar and Brigitte Grau. "SVETLAN’ or how to Classify Words usingtheir Context", proceedings of the 12th International Conference on Knowledge Eneering and Knowledge Management, EKAW 2000, Juan-les-Pins, France, Oc2000, pages 203-216 Rose Dieng and Olivier Corby (Eds.), Springer, 2000,19. Bisson, G. and Nédellec, C. and Cañamero L.: Designing clustering methods forontology building - The Mo’K workbench, in Staab, S. and Maedche, A. and NedellC., editors, Ontology Learning ECAI-2000 Workshop, Berlin, August 200020. Hahn author Hahn, U. and Schnattinger, Ontology Engineering via Text Under-standing, IFIP'98 --- Proceedings of the 15th World Computer Congress, Vienna/Bpest, 199821. Eneko Aguirre, Mikel Lersundi: Extracción de relaciones léxico-semánticas partir de palabras derivadas usando patrones de definición. Procesamiento del Len-guaje Natural 27: 165-172 (2001)22. Lagus, K.: Text Mining with the WEBSOM. Acta Polytechnica Scandinavica, Math-ematics and Computing Series no. 110, Espoo, Finland 2000 23. Teuvo Kohonen, Self Organising Maps, 3rd Edition, Springer Series in InformationSciences, Vol. 30, Springer, Berlin

34

Using Ontologies to Discover Domain-Level

Web Usage Profiles

Honghua (Kathy) Dai and Bamshad Mobasherhdai,[email protected]

School of Computer Science, Telecommunication, and Information Systems,DePaul University, Chicago, Illinois, USA

Abstract. Usage patterns discovered through Web usage mining are ef-fective in capturing item-to-item and user-to-user relationships and sim-ilarities at the level of user sessions. Without the benefit of deeper do-main knowledge, such patterns provide little insight into the underlyingreasons for which such items or users are grouped together. This canlead to a number of important shortcomings in personalization systemsbased on Web usage mining or collaborative filtering. For example, ifa new item is recently added to the Web site, it is not likely that thepages associated with the item would be a part of any of the discoveredpatterns, and thus these pages cannot be recommended. Keyword-basedcontent-filtering approaches have been used to enhance the effectivenessof collaborative filtering systems by focusing on content similarity amongitems or pages. These approaches, however, are incapable of capturingmore complex relationships at a deeper semantic level based on differenttypes of attributes associated with structured objects. This paper rep-resents work-in-progress towards creating a general framework for usingdomain ontologies to automatically characterize usage profiles containinga set of structured Web objects. Our motivation is to use this frameworkin the context of Web personalization, going beyond page- or item-levelconstructs, and using the full semantic power of the underlying ontology.

1 Introduction and Problem Statement

The goal of Web usage mining [24] is to capture and model the behavioralpatterns and profiles of users interacting with aWeb site. The discovered patternsare usually represented as collections of pages that are frequently accessed bygroups of users with common needs or interests. Such patterns can be used tobetter understand behavioral characteristics of visitor or user segments, improvethe organization and structure of the site, and create a personalized experiencefor visitors by providing dynamic recommendations [25, 7, 3, 15, 9, 10, 22].

At a conceptual level, there may be many different kinds of objects withina given site that are accessible to users. At the physical level, these objectsmay be represented by one or more Web pages. For example, consider a sitecontaining information about movies. This site may contain pages related tothe movies themselves, actors appearing in the movies, directors, studios, etc.

35

Conceptually, each of these entities represents a different type of semantic object.During a visit to this site, a user may access several of these objects togetherduring a session.

Given the session-based Web usage data, a variety of mining techniques canbe used to discover patterns, including clustering, association-rule or sequentialpattern discovery. For example, clustering of user sessions may result in thediscovery of a group of similar sessions based on pages commonly accessed withinthose sessions. In order to find an aggregate representation of user interestscaptured by such a cluster, one might derive the cluster centroid containing aset (or a vector) of pages that are common among cluster elements. In [18,17], we called these aggregate representations of session clusters aggregate usageprofiles and used these profiles for Web personalization.

Specifically, each user session s can be viewed as an n-dimensional vectorover the space of all pages, i.e.,

t = 〈w(p1, s), w(p2, s), . . . , w(pn, s)〉 ,where w(pi, t) is a weight, in session s, associated with the page pi. The weightscan be binary representing the existence or non-existence of a page access in thesession, or they can be some value representing the user’s interest on the page(e.g., the page stay time). Applying data mining techniques, such as clustering,to this space may result in a set CL = cl1, cl2, . . . , clk of session clusters, whereeach cli is a subset of the set of sessions.

Given a session cluster cl, we can construct a usage profile prcl as a set ofpageview-weight pairs by computing the centroid of cl:

prcl = 〈p, weight(p, prcl)〉 | weight(p, prcl) ≥ µwhere

– the significance weight, weight(p, prcl), of the page p within the usage profileprcl is given by

weight(p, prcl) =1|cl| ·

∑s∈cl

w(p, s);

– w(p, s) is the weight of page p in session s ∈ cl; and– the threshold µ is used to focus only on those pages in the cluster that appearin a sufficient number of sessions in that cluster.

Each such profile, in turn, can be represented as a vector in the original n-dimensional space. The aggregate representation of common usage patterns assets or vectors of pages makes such usage profiles quite useful for personalizationand collaborative filtering [12]: given a new user, u who has accessed a set ofpages, Pu, so far, we can measure the similarity of Pu to the discovered profiles,and recommend to the user those pages in matching profiles which have not yetbeen accessed by the user.

While in the above discussion we have focused on clustering as the primarydata mining technique for the discovery of usage profiles, it should be noted that

36

a variety of other techniques can also be used. For example, frequent itemsetsdiscovered as part of association rule mining [1] on the usage data, would alsolead to sets of items or pages representing usage profiles [16].

One problem with the above approach is that the profiles only capture com-mon usage patterns at the page level. They do not reflect the underlying reasonswhy the group of users represented by the profile are interested in the accessedpages. This can lead to a number of important shortcomings in personalizationsystems based on Web usage mining or collaborative filtering. For example, ifa new item is recently added to the Web site, it is not likely that the pagesassociated with the item would be a part of any of the discovered patterns (dueto a lack of sufficient usage level data). Yet, a user who fits one of the profilesmay indeed be interested in the new item, because the underlying domain char-acteristics of the item might correspond to those of items in one or more of theprofiles.

In the movie site example mentioned above, consider the situation where adiscovered usage profile may contain pages related to a number of movies manyof which are from the same genre and have been directed by the same director.Using the above standard approach for personalization, pages appearing in thisprofile may be recommended to a user who has accessed some of the otherpages in that profile. However, if a new movie is added to the site with similarproperties, it will not appear in any profile until a sufficient number of users haveaccessed this movie together with other similar movies. This problem is oftencoined as the “new item problem” in collaborative filtering.

A common approach to resolving this problem has been to integrate contentcharacteristics of pages with the usage patterns [6, 20, 21, 19]. Generally, in theseapproaches, keywords are extracted from the content on the Web site and areused to either index pages by content or classify pages into various content cat-egories. In the context of personalization, this approach would allow the systemto recommend pages to a user, not only based on a matching usage profile, butalso (or alternatively) based on the content similarity of these pages to the pagesuser has already visited.

Keyword-based approaches, however, are incapable to capturing more com-plex relationships among objects at a deeper semantic level based on the inherentproperties associated with these objects. In our movie site example, the abovecontent-based filtering approach would allow the system to recommend othermovies based on similarities in their textual descriptions or other content char-acterisitcs. But, the system would have considerable difficulty in recommending,for example, unrelated movies from the same genre, having the same main actorsas those already accessed by the user, etc. To be able recommend different typesof complex objects using their underlying properties and attributes, the sys-tem must be able to rely on the characterization of user segments and objects,not just based on keywords, but at a deeper semantic level using the domainontologies for the objects.

This paper represents work-in-progress towards creating a general frameworkfor using domain ontologies to automatically characterize usage profiles contain-

37

ing a set of structured Web objects. Our motivation is to use this frameworkin the context of Web personalization, going beyond page-level constructs, andusing the full semantic power of the underlying ontology. This effort involves thefollowing tasks:

1. Given a page in the Web site, we must extract domain-level structuredobjects as semantic entities contained within this page. This task involvesthe automatic extraction and classification of objects of different types intoclasses based on the underlying domain ontologies. Our goal is to createa representation for a usage profile discovered through Web usage miningprocess (e.g., through clustering or association rule mining), not simply as aset of pages, but as a set of structured objects embedded in those pages.

2. Given a set of structured objects representing a usage profile, we must createan aggregated representation as a set of pseudo objects each characterizingobjects of different types occurring commonly across the user sessions. Wecall such a set of aggregate pseudo objects a Domain-level Aggregate Profile.Thus, a domain-level aggregate profile characterizes a collection of similarusers based on the common properties of objects in the domain ontologythat were accessed by these users.

We begin by providing a formal framework for the representation of the do-main ontology and for creating aggregate representations of domain-level objects.We then discuss how the resulting aggregate profiles can be used in the contextof personalization.

2 Representing Domain Ontologies for Web Objects

There has been much recent work in designing ontology languages to formallyrepresent knowledge on the Web, such as RDFS [2] and DAML+OIL [13]. Theontology language DAML+OIL has been extended to include formal semanticsand reasoning support by mapping to the description logic SHOQ(D) [14].

Generally speaking, in our current work we adopt the syntax and semantics ofSHOQ(D) to represent domain ontologies. In SHOQ(D), the notion of concretedatatype is used to specify literal values and individuals which represent realobjects in the domain ontology. Moreover, concepts can be viewed as sets ofindividuals, and roles are binary relations between a pair of concepts or betweenconcepts and datatypes. The detailed formal definitions for concepts and rolesare given in [11, 14]. Because our current work does not focus on reasoning taskssuch as deciding subsumption and membership, we do not focus our discussion onthese operations. The reasoning apparatus in SHOQ(D) can be used to providemore intelligent data mining services as part of our future work.

In SHOQ(D), a concept has the meaning of a set of individuals, which inour context are called “objects”. The notion of a concept is quite general andmay encompass a heterogeneous set of objects with different properties (roles)and structures. In the present work, we are interested in deriving aggregaterepresentations of a group of objects that have a homogenous concept structure

38

(i.e., have similar properties and data types). For example, we may be interestedin a group of movie objects, each of which has specific values for properties suchas “year”, “genre”, “actors”, etc. For the purpose of presentation, we call such agroup of objects a class. Thus, in our framework, the notion of a class representsa restriction of the notion of a concept in SHOQ(D).

More specifically, our notion of class can be defined in the context of SHOQ(D)as follows.

Definition 1 A class C I (I is the set of all individuals in the domainontology) is a set of objects where there exists a set of roles R, such that,∀r ∈ R,Dr = v2 | (v1, v2) ∈ r, and C ∀R.Dr.

We call the roles that characterize the class C attributes. These attributestogether define the internal properties of the objects in C or relationships withother concepts that involve the objects in C. And we denote the domain of valuesof the attribute r as Dr. Furthermore, because we are specifically interested inaggregating objects at the attribute level, we extend the notion of a role toinclude a domain-specific combination function and an ordering relation.

More formally, a class C is characterized by a finite set of attributes AC ,where each attribute a ∈ AC is defined as follows.

Definition 2 Let C be a class in the domain ontology. An attribute a ∈ AC isa 4-tuple, a = 〈Ta, Da, ψa,a 〉, where– Ta is the type for the values for the attribute a.– Da is the domain of the values for a;– a is an ordering relation among the values in Da; and– ψa is a combination function for the attribute a.

The “type” of an attribute in the above definition may be a concrete datatypeor it may be a set of objects (individuals) belonging to another class. Given atype Ta for an attribute a, we have Da dom(Ta).

In the context of data mining, comparing and aggregating values are essentialtasks. Therefore, ordering relations among values are necessary properties forattributes. We associate an ordering relation a with elements in Da for eachattribute a. The ordering relation a can be null (if no ordering is specifiedin the domain of values), or it can define a partial or a total order among thedomain values. For standard types such as values from a continuous range, weassume the usual ordering. In cases when an attribute a represents a concepthierarchy, the domain values of a are a set of labels, and a is a partial orderrepresenting the “is-a” relation.

Furthermore, we associate a data mining operator, called the combinationfunction ψa, with each attribute a. The combination function ψa defines an ag-gregation operation among the corresponding attribute values of a set of objectsbelonging to the same class. This function is essentially a generalization of the“mean” or “average” function applied to corresponding dimension values of setof vectors when computing the centroid vector. In this context, we assume that

39

the combination function is specified as part of the domain ontology for each at-tribute of a class. An interesting extension would be to automatically learn thecombination function for each attribute based on a set of positive and negativeexamples.

Classes in the ontology define the structural and semantic properties of ob-jects in the domain which are “instances” of that class. Specifically, each objecto in the domain is also characterized by a set of attributes Ao corresponding tothe attributes of a class in the ontology. In order to more precisely define thenotion of an object as an instance of a class, we first define the notion of aninstance of an attribute.

Definition 3 Given an attribute a = 〈Ta, Da, ψa,a 〉 and an attribute b =〈Tb, Db, ψb,b 〉, b is an instance of a, if Db ⊆ Da, Tb = Ta, ψb = ψa, and b is arestriction of a to Db. The attribute b is a null instance of a, if Db = ∅.Definition 4 Given a class C with attribute set AC = aC

1 , aC2 , . . . , a

Cn , we say

that an object o is instance of class C, if o has attributes Ao = ao1, a

o2, . . . , a

on

such that, aoi is a, possibly null, instance of aC

i , for 1 ≤ i ≤ n.In the context of SHOQ(D), we can use the concept inclusion axiom o C

to denote the “instance-of” relation between o and C. In general, the domainontology can be represented as a set of assertions on classes and attributes.

Based on the definitions of attribute and object instances, we can provide amore formal representation of the combination function ψa. Let c be a class ando1, o2, . . . , on a set of object instances of c. Let a ∈ AC be an attribute of classc. The combination function ψa can be represented by:

ψa(〈ao1 , w1〉, 〈ao2 , w2〉, . . . , 〈aon , wn〉) = 〈aagg, wagg〉,where each aoi belonging to object oi is an instance of the attribute a, andeach wi is a weight associated with the attribute instance aoi representing thesignificance of that attribute relative to the other instances. Furthermore, aagg isa pseudo instance of a meaning that it is an instance of a which does not belongto a real object in the underlying domain. The weight wagg of aagg is a functionof w1, w2, . . . , wn.

Given a set of object instances, o1, o2, . . . , on, of class C, a domain-levelaggregate profile for these instances is obtained by applying the combinationfunction for each attribute in c to all of the corresponding attribute instancesacross all objects o1, o2, . . . , on.

An Example

As an example, let us come back to the movie Web site discussed in the previoussection. The Web site includes collections of pages containing information aboutmovies, actors, directors, etc. A collection of pages describing a specific moviesmight include information such as the movie title, genre, starring actors, director,etc. An actor or director’s information may include name, filmography (a set of

40

Fig. 1. The Ontology for a movie Web site

movies), gender, nationality, etc. The portion of domain ontology for this site,as described, contains the classes Movie, Actor and Director (see Figure 1).The collection of Web pages in the site represent a group of embedded objectsthat are the instances of these classes.

The class Movie has attributes Year, Actor (representing the role “actedby”), Genre, Director, etc. The Actor, and Director attributes have values thatare other objects in the ontology, specifically, object instances of classes Actorand Director, respectively. The attribute Year is an example of an attributewhose datatype is positive integers with the usual ordering. The attribute Genrehas a concrete datatype whose domain values in DGenre are a set of labels (e.g.,“Romance” and “Comedy”). The ordering relation for Genre defines a partialorder based on the “is-a” relation among these labels (resulting in a concepthierarchy of Genre’s a portion of which is shown in Figure 1).

An attribute a of an object o has a domain Da. In cases when the attributehas unique value for an object,Da is a singleton. For example, consider an objectinstance of class Movie, “About a Boy” (see Figure 2). The attribute Actorcontains three objects (H. Grant, R. Weisz and T. Collette) that are instancesof the class Actor (for the sake of presentation we use the Actor’s name tostand for the object of Actor). Therefore, DActor = H. Grant, R. Weisz and T.Collette. Also, A real object may have values for only some of the attributes. Inthis case the other attributes have empty domains. For instance, the attributeDirector in the example has an empty domain, and is thus not depicted in thefigure.

We may, optionally, associate a weight with each value in the attribute do-main Da (usually in [0, 1]). This may be useful in capturing the relative im-portance of each attribute value. For example, in a given movie the main actorsshould have higher weights than other actors. In our example, the object actor H.

41

Fig. 2. An Example of an Object in Class Movie

Grant has weight 0.6 and the object Actor Rachel Weisz has weight 0.1. Unlessotherwise specified, we assume that the weight associated with each attributevalue is 1.

In the object o shown in Figure 2, the domain for the attribute Genreis the set “Genre-All”, “Comedy”, “Romantic Comedy”, “Kid & Family”.The ordering relation o

Genre is a restriction of Genre to Genre-All, Comedy,Romantic Comedy, Kid & Family.

Let us now define the combination functions for some of the attributes in classMovie. Note that the combination functions are only applicable when creatingan aggregate representation of a set of objects. For the attribute Name, we areinterested in all the movie names appearing in the instances. Thus we can defineψName to be the union operation performed on all the singleton Name attributesof all movie objects.

The attribute Actor contains a weighted set of objects in class Actor. Insuch cases we can use a vector-based weighted mean operation. The domain ofthe resulting aggregate attribute instance D′

Actor can be viewed as the union ofthe domains of the Actor attributes of the individual Movie objects: D′

Actor =∪iDActori, and the weight for an object o in D′

Actor is determined by

w′o =

∑i wi · wo∑

iwi.

For example, applying ψActor to 〈S, 0.7; T, 0.2; U, 0.1, 1〉, 〈S, 0.5; T,0.5, 0.7〉, 〈W, 0.6; S, 0.4, 0.3〉 will result in the aggregate domain D′

Actor of〈S, 0.58〉, 〈T, 0.27〉, 〈W, 0.09〉, 〈U, 0.05〉.

As for the attribute Year, the combination function may create a range ofall the Year values appearing in the objects. Another possible solution is to dis-cretize the full Year range into decades and find the most common decades thatare in the domains of the attribute. For example, applying ψY ear to 〈2002,1>, 〈1999, 0.7〉, 〈2001, 0.3〉 may result in an aggregate instance Y ear′ ofattribute Year with D′

Y ear = [1999, 2002].The attribute Genre of class Movie contains a partial ordering relation

Genre which represents a concept hierarchy among different Genre labels. Thus,for each instance object p of Genre, the relation a

Genre specifies the restriction

42

of the partial ordering relation to Dp. The combination function, in this case,can perform tree (or graph) matching to extract the common parts of the con-ceptual hierarchies among all instances [23]. Given a set of objects p1, . . . , pnof Genre, we can define ψGenre(〈p1, w1〉, 〈p2, w2〉, . . . , 〈pn, wn〉) = 〈∩iDpi , w

′〉,where w′ is the average of the weights of instance values in the intersection.The relation ′

Genre is a restriction of Genre to ∩iDpi . This function pro-vides us the common segment of conceptual hierarchy for Genre in the setof movie instances. For example, applying ψGenre to “Romantic Comedy”,“Kids&Family”, “Romance”, “Comedy”, “Romance”, and assuming thatall weights are 1, will result in an aggregate instance Genre′ of attribute Genrewith the value “Romance”.

3 Creating an Aggregated Representation of a UsageProfile

We now return to our original problem: how to create an aggregate represen-tation of a discovered usage profile at the domain-level. As noted earlier, a us-age profile at the page or item level can be viewed as a weighted set (or vec-tor) of pages. The problem of extracting instances of the ontology classes fromthese pages is an interesting problem in its own right and has been studiedextensively (see, for example, [5]). Here we assume that, either using manualrules, or through supervised learning methods, we can extract various objectinstances represented by the pages in the original page- or item-level usage pro-file. Thus, the usage profile can be transformed into a weighted set of objects:pr = 〈o1, wo1〉, 〈o2, wo2〉, . . . , 〈on, won〉 in which each oi is an object in the un-derlying domain ontology and wi represents oi’s significance in the profile pr.The profile represents a set of objects accessed together frequently by a group ofusers (as determined through Web usage mining). Our goal is to create an aggre-gate representation of this weighted set of objects to characterize the commoninterests of the user segment captures by the profile at the domain level.

Given the representation of a profile pr as a weighted set of objects, theobjects in pr may be instances of different classes c1, c2, . . . , ck in the ontology.The process of creating a domain-level aggregate profile begins by partitioningpr into collections of objects with each collection containing all objects that areinstances of a specified class (in other words, the process of classifying the objectinstances in pr). Let gi = 〈oci

1 , woci1〉, . . . , 〈oci

m, wocim〉 be the elements of pr that

are instances of the class ci.Having partitioned pr into k groups of homogeneous objects, g1, . . . , gk, the

problem is reduced to creating aggregate representation of each partition gi.This task is accomplished with the help of the combination functions for each ofthe attributes of ci some of whose object instances are contained in gi. Once therepresentatives for every partition of objects are created, we assign a significanceweight to each representative to mark the importance of this group of objectsin the profile. In our current implementation the significance weight for eachrepresentative is computed as the sum of all the object weights in the partition.

43

Input: A weighted set of objects: O = 〈o1, w1〉, . . . , 〈on, wn〉Output: The domain-level aggregate of O: 〈o′1, w′

1〉, . . . , 〈o′k, w′k〉

Main Procedure:Result = ∅Partition objects in O into g1, g2, . . . , gk,

such that ∀gi,∃ a class cgi , with objects in gi being instances of cgi

For each gi(i = 1, . . . , k) doBuild a pseudo object o′i to be an instance of cgi :

For each attribute a in the attribute set of class cgi do

Let ao′i be an instance of a in the object o′i:

ao′i = ψa(gi)

EndforResult = Result ∪ o′iw′

i =∑

ok∈giwk

EndforReturn Result

Fig. 3. The Algorithm DPA for Creating and Aggregate Representation of a WeightedSet of Objects

However, significance weight can be computed using other numeric aggregationfunctions. Figure 3 summarizes the algorithm DPA for creating domain-levelaggregate profiles.

Examples Continued: Generating Domain-Level Aggregate Profiles

We present two examples of transforming usage profiles into domain-level aggre-gates using ontological information. The first example is based on our hypothet-ical movie Web site (Figure 1), and the second is based on the results of ourexperiments with a real Web site containing real estate property listings.

Suppose we have discovered a Web usage profile and transformed it into aweighted set of 3 movie objects in the class Movie (see Figure 4). We cangenerate a domain-level aggregate representation of these movies by applyingthe DPA algorithm to this profile. Because the objects are the instances of thesame class, the DPA algorithm does not have to partition the objects. Thus, wecan directly apply combination functions on each attribute in the class Movie.The details of combination functions used in this example were described in theprevious section. Figure 5 shows the pseudo object instance representing thedomain-level aggregation of these objects.

Note that the original item-level profile gives us little information about thereasons why these objects were commonly accesed together. However, after wecharacterize this profile at the domain-level, we find some interesting patterns:they all belong to Genre “Romance”, and the actor S has a high score comparedwith other actors. This might tell us that this group of users are interestedparticularly in the movies belonging to “Romance” and are particularly fond ofthe actor S.

44

Fig. 4. A Weighted Set of Objects in a Usage Profile from a Movie Web Site

Fig. 5. An Example of Domain-Level Aggregate Profile from a Movie Web site

Our next example is based on a real usage profile discovered from the usagedata belonging to a real estate Web site containing various property listings.The ontology of the Web site has a single class property. An object which isthe instance of class property represents a real estate property that has beenlisted for sale. This ontology is depicted in Figure 6.

In this case, all attributes have atomic values. For example, the attributenumber of bedrooms has a range of values 1, 2, 3, 4, 5, 6 and style of buildinghas values among 1 Story, 2 Story, Town Home, Ranch. Figure 6 also showsan example of a property object instance with the value sets for each attributein dashed boxes.

In our experiment, we began by clustering similar user sessions. Then wegenerated the usage profiles obtaining the cluster centroids in which each item’sweight represents the percentage of cluster sessions containing that item. Forexample, one of the usage profiles obtained from our experiment is: 〈Property1,1〉,〈Property 2,0.7〉,〈Property 3,0.18〉,〈Property 4,0.18〉.

45

Fig. 6. The Ontology of a Real Estate Web site Containing Property Listings

We used the “weighted average” as our combination function for the at-tributes containing numeric data. Let wai be the weight associated with instanceai of attribute a. Because all the attribute instances contain singletons as theirvalue sets, we can use vai to represent the only value for the instance ai . Thecombination function ψa for each of the numeric attributes (Price, Size andRooms) gives the value set D′

a of the aggregate instance, as follows:

D′a =

∑i vai × wai∑

iwai

For the attribute Location, the combination function computes the union ofall the values in this attribute. In this example, this function returns Chicago,Evanston. For the attribute Style, the combination function computes the mostfrequent style values in this attribute. In this case, the function returns 2 Story.Finally, we used the same combination function for the attribute Year as theone used in the class movie of the previous example. Figures 7 and 8 show theoriginal usage profiles discovered through clustering and the resulting domain-level aggregate profile, respectively.

After generating domain-level usage profile, we have the information aboutthe average price, size, number of rooms, as well as the locations, major styles,and year range. Such information not only enriches the profile with more knowl-edge about the listed items, but also provides us more possibilities for Web per-sonalization. In the following section we discuss how we can leverage domain-levelaggregate profiles for Web personalization.

46

Fig. 7. A Weighted Set of Objects in a Usage Profile from a Real Estate Web Site

Fig. 8. An Example of Domain-Level Aggregate Profile from a Real Estate Web site

4 Web Personalization Based on Domain-Level UsageProfiles

Domain-level aggregate profiles present users’ interests not just as a set of itemsor pages, but in terms of the common underlying properties and relationshipsamong those items that are captured by object attributes. This fine-graineddomain knowledge enables more powerful approaches to personalization.

We consider the browsing history of the current user to be a weighted set ofitems that the user has visited. The weight can be based on the amount of timethe user spends on each item or other measures that represent the significance ofthe items in the current browsing session. The algorithm DPA can be applied tothis weighted set to create an aggregate domain-level representation of the userbrowsing history. We call this aggregate representation the current user model.Given the current user model, there are two possible approaches to generatingpersonalized recommendations for the user. These two approaches are depictedin Figure 9.

The first approach searches the domain ontologies to recommend the itemsthat have similar features with the current user model. We call the recommen-dationed objects the instantiations of the current user model. This approach isessentially a generalization of the keyword-based content filtering approach anddoes not rely on any discovered usage profiles. While this approach does notencounter the “New Item” problem, it does not have the advantages of collabo-rative filtering systems, namely, considering item-to-item relationships that arebased on how items are used or accessed together rather than based on contentsimilarity. In addition, this approach requires the definition of possibly complexsimilarity or distance functions for each of the attributes of the underlying classesin the ontology.

47

Fig. 9. Personalization Framework utilizing Domain Knowledge

The second method generates recommendations based on the discovereddomain-level aggregate profiles. The recommendation engine matches the cur-rent user model with all the discovered usage profiles. The usage profiles withmatching score greater than some pre-specified threshold are considered to rep-resent this user’s potential interests. A successful match implies that the currentuser shares common interests with the group of users this usage profile repre-sents. The recommendation engine also matches the items in the domain withthe user’s potential interests and will recommend to the user the matching items.This approach is more complex than the pure content-based approach in thatit involves two matching processes: matching the current user with the usageprofiles and matching the item candidates with the user’s potential interests.

The second approach has a number of advantages. First, it retains the user-to-user relationships that can be captured by the discovered usage profiles. Sec-ondly, in contrast to standard collaborative filtering, it provides more flexibilityin matching usage profiles with current user model because in this case match-ing involves comparison of features and relationships, not exact item identities.Furthermore, the items do not have to appear in any usage profiles in order tobe recommended, since fine-grained domain relationships are considered duringthe instantiation process. The following example shows that this approach canalso be used to solve the “New Item” problem.

Let us consider the real estate Web site discussed in the previous section.We again consider the real usage profile containing property 1, property 2, prop-erty 3 and property 4 depicted in Figure 7. Suppose that there is a new itemproperty 5 (recently added to the site) which is not included in any usage pro-files. Assume that property 5 has attributes 〈Price = 370000, Location =Chicago, size = 4500, rooms = 4, style = 2story, Y earbuilt = 1993〉.

48

Furthermore, assume that a user has browsed pages related to property 1 andproperty 2. If the recommendation engine were based on item-level usage profiles,property 5 would not be recommended. The recommendation engine would in-stead recommend property 3, property 4. However, the current user model (prop-erty 1 and property 2) is indeed more similar to property 5 than to property3 or property 4 in terms of their domain-level attributes. If the recommenda-tion engine uses domain-level aggregate profile of Figure 8, it would be able torecommend property 5.

5 Conclusion and Future Work

In this paper we have presented a general framework for using domain ontologiesto automatically characterize usage profiles containing a set of structured Webobjects. Our motivation has been to use this framework in the context of Webpersonalization, going beyond page- or item-level constrcuts, and using the fullsemantic power of the underlying ontology.

In our approach we consider a Web site as a collection of objects belonging tocertain classes. Given a collection of similar user sessions (e.g., obtained throughclustering) each containing a set of objects, we have shown how to create anaggregate representation of for the whole collection based on the attributes ofeach object as defined in the domain ontology. This aggregate representation isa set of pseudo objects each characterizing objects of different types commonlyoccurring across the user sessions. We have also presented a framework for Webpersonalization based on domain-level aggregate profiles.

Currently we assume that we have the predefined combination functions foreach class attribute specified as part of the domain ontology. One area of futurework involves the study of machine learning techniques in order to discover thebest way to summarize the attribute automatically. Another area of future workinvolves the creation of general as well as domain-specific distance ro similarityfunctions allowing for the comparison of objects in different classes with thedomain-level profiles. Finally another interesting area of work will be to exploreuse of discoverd domain-level aggregates from Web usage mining to enrich theexisting domain ontology for a Web site.

References

1. R. Agrawal and R. Srikant. Fast Algorithms for Mining Association Rules. Proc.20th Int. Conf. Very Large Data Bases, VLDB, 1994.

2. D. Brickley, and R.V. Guha, Resource Description Framework (RDF)Schema Specification 1.0, World Wide Web Consortium, 2000.http://www.w3.org/TR/rdf-schema/

3. B. Berendt and M. Spiliopoulou. Analysing navigation behaviour in web sites in-tegrating multiple information systems. In VLDB Journal, Special Issue on Data-bases and the Web. 9(1):56-75, 2000.

49

4. J. Broekstra, M. Klein, S. Decker, D. Fensel, F. van Harmelen, and I. Horrocks.Enabling knowledge representation on the web by extending RDF schema. In Pro-ceedings of the 10th World Wide Web conference, Hong Kong, China, May 1–5,2001.

5. M. Craven, D. DiPasquo, D. Freitag, A. McCallum, T. Mitchell, K. Nigam, S. Slat-tery. Learning to construct knowledge bases from the world wide web. In ArtificialIntelligence, 118:69-113, 2000.

6. M. Claypool, A. Gokhale, T. Miranda, P. Murnikov, D. Netes, and M. Sartin.Combining Content-based and Collaborative Filters in an Online Newspaper. InProceedings of the ACM SIGIR ’99 Workshop on Recommender Systems: Algo-rithms and Evaluation. University of California, Berkeley, Aug. 1999.

7. R. Cooley, B. Mobasher and J. Srivastava. Data preparation for mining WorldWide Web browsing patterns. Journal of Knowledge and Information Systems, (1)1, 1999.

8. M. Deshpande and G. Karypis. Selective markov models for predicting web-pageaccesses. In First International SIAM Conference on Data Mining, 2001.

9. E. Damiani, B. Oliboni, E. Quintarelli, and L. Tanca. Modeling users’ navigationhistory. In Workshop on Intelligent Techniques for Web Personalisation at the17th International Joint Conference on Articial Intelligence (IJCAI01), Seattle,Washington, (USA), 2001.

10. X. Fu, J. Budzik, and K. J. Hammond. Mining navigation history for recommenda-tion. In Proc. 2000 International Conf. Intelligent User Interfaces, New Orleans,LA, January 2000. ACM.

11. R. Giugno and T. Lukasiewicz. P-SHOQ(D): A Probabilistic Extension ofSHOQ(D) for Probabilistic Ontologies in the Semantic Web. Accepted for pub-lication in Proceedings of the 8th European Conference on Logics in Artificial In-telligence (JELIA’02), Cosenza, Italy, September 2002. Lecture Notes in ArtificialIntelligence, Springer, 2002.

12. J. Herlocker, J. Konstan, A. Borchers, and J. Riedl. An algorithmic framework forperforming collaborative filtering. In Proceedings of ACM SIGIR’99, 1999.

13. Ian Horrocks. DAML+OIL: a reason-able web ontology language. In Proc. ofEDBT 2002, March 2002. To appear.

14. I. Horrocks and U. Sattler. Ontology reasoning in the SHOQ(D) description logic.In B. Nebel, editor, Proc. of the 17th Int. Joint Conf. on Artificial Intelligence(IJCAI-01), pages 199-204. Morgan Kaufmann, 2001.

15. B. Mobasher, R. Cooley and J. Srivastava. Automatic personalization based onWeb usage mining. In Communications of the ACM, (43) 8, August 2000.

16. B. Mobasher, H. Dai, T. Luo, and M. Nakagawa. Effective Personalization Based onAssociation Rule Discovery from Web Usage Data. In Proceedings of the 3rd ACMWorkshop on Web Information and Data Management (WIDM01), held in con-junction with the International Conference on Information and Knowledge Man-agement (CIKM 2001), Atlanta, Georgia, November 2001.

17. B. Mobasher, H. Dai, T. Luo, and M. Nakagawa. Improving the effectiveness ofCollaborative Filtering on Anonymous Web Usage Data. In Proceedings of the IJ-CAI 2001 Workshop on Intelligent Techniques for Web Personalization (ITWP01),August 2001, Seattle.

18. B. Mobasher, H. Dai, T. Luo, and M. Nakagawa. Discovery of Aggregate Usage Pro-files for Web Personalization. In Proceedings of the Web Mining for E-CommerceWorkshop (WebKDD’2000), held in conjunction with the ACM-SIGKDD Confer-ence on Knowledge Discovery in Databases (KDD’2000), August 2000, Boston.

50

19. B. Mobasher, H. Dai, T. Luo, Y. Sun, and J. Zhu. Combining web usage and con-tent mining for more effective personalization. In Proceedings of the InternationalConference on ECommerce and Web Technologies (ECWeb), 2000.

20. D. Mladenic. Text learning and related intelligent agents: a survey. IEEE IntelligentSystems, 14(4):44–54, 1999.

21. M. Pazzani. A Framework for Collaborative, Content-Based and Demographic Fil-tering. Artificial Intelligence Review, Dec. 1999, pp. 393-408.

22. Pitkow J. and Pirolli P. Mining Longest Repeating Subsequences to Predict WWWSurfing. In Proceedings of the 1999 USENIX Annual Technical Conference, 1999.

23. R. Ramesh, L. V. Ramakrishnan. Nonlinear pattern matching in trees. In Journalof the ACM, 39(2):295-316, 1992.

24. J. Srivastava, R. Cooley, M. Deshpande, P-T. Tan. Web usage mining: discoveryand applications of usage patterns from Web data. SIGKDD Explorations, (1) 2,2000.

25. Myra Spiliopoulou and Lukas C. Faulstich. WUM: A Tool for Web UtilizationAnalysis. In Proceedings of EDBT Workshop WebDB’98, Valencia, Spain, Mar.1998.

51

Multi-faceted Learning for Web Taxonomies

Wray Buntine and Henry Tirri

Helsinki Inst. of Information TechnologyHIIT, P.O. Box 9800

FIN-02015 HUT, Finlandfwray.buntine,[email protected]

Abstract. A standard problem for internet commerce is the task ofbuilding a product taxonomy from web pages, without access to corpo-rate databases. However, a nasty aspect of the real world is that mostweb-pages have multiple facets. A web page might contain informationabout both cameras and computers, as well as having both specicationand sale data. We are interested in methods for supervised and unsu-pervised learning of multiple faceted models. Here we present results formulti-faceted clustering of bigram words data.

1 Introduction

A recognized problem for internet commerce is the task of building a producttaxonomy from web pages, without access to corporate databases, and thenpopulating a database with link information about service, repair, spare parts,reviews, product specications, product family and company home pages, andpurchase information from retailers. A key precursor for this task is the ability tobuild classication hierarchies in a supervised or unsupervised manner. However,a nasty aspect of the real world is that most web-pages have multiple facets. Aweb page might contain information about cameras and computers, as well ashaving both specication and sale data. Whereas another page might mix aproduct index with partial specication and sales data. We are interested inmethods for supervised and unsupervised learning of multi-faceted models.

Clustering or unsupervised learning is now a standard method for analysingdiscrete data such as documents, and is now being used in industry to create tax-onomies from web pages. A rich variety of methods exist borrowing theory andalgorithms from a broad spectrum of computer science: spectral (eigenvector)methods [1], kd-trees [2], using existing high-performance graph partitioning al-gorithms from CAD [3], hierarchical algorithms [4] and data merging algorithms[5], etc.

All these methods, however, have one signicant drawback for typical appli-cation in areas such as document or image analysis: each item/document is to beclassied exclusively to one class. Their models make no allowance for instance,for a product page to have 60% digital camera content and 40% laptop com-puter content. It is 100% one way or another, and any uncertainty is only about

52

2

whether to place the 100% into one or the other class. In practice documentsinvariable mix a few topics, readily seen by inspection of the human-classiedReuters newswire, so the automated construction of topic hierarchies needs tore ect this. One alternative is to make clustering multi-faceted whereby a docu-ment can be assigned proportionally (i.e., using a convex combination) across anumber of clusters rather than uniquely to one cluster.

Authors have recently proposed discrete analogues to principle componentsanalysis (PCA) intended to handle discrete or positive only count data of thekind used in the bag-of-words representation of web pages. Methods includenon-negative matrix factorization [6], probabilistic latent semantic analysis [7]latent Dirichlet allocation [8], multinomial PCA [9]. A good discussion of themotivation for these techniques can be found in [7], and an analysis of relatedreduced dimension models and some of the earlier statistical literature here canbe found in [10], and theory and algorithms are presented in [9].

Multinomial PCA by itself does not perform multi-faceted clustering becauseon average a page/item/document might be composed of up to 50 componentsin some of our experiments, and this does not re ect the behaviour of a \fewdierent topics" we were looking for. We have recently made modications tothe standard algorithms so that multinomial PCA performs multi-faceted clus-tering. That is, it performs a clustering whereby some items/documents/pagesare assigned with proportion to a few dierent classes.

In this paper, we rst expand more on what we mean by a multi-facetedmodel, and then we give some examples from our working system. Our experi-ments are conducted on bigram data for words collected o a good fraction ofGoogle's database of web pages for August 2001. The data sizes allowed us toexperiment to understand how many components might be produced.

2 Contrasting Clustering and Multi-Faceted Clustering

For concreteness, consider the problem in terms of the usual \bag of words"representation for a document [11]. Here the items making up the sample aredocuments and the features are the counts of words in the document. A doc-ument is represented as a sparse vector of words and their occurrence counts.All positional information is lost. With J dierent words, the dimensionality forwords/features, each document becomes a vector x 2 ZJ , where the total

Pj xj

might be known. Traditional clustering becomes the problem of forming a map-ping ZJ 7! f1; : : : ;Kg, where K is the number of clusters. Whereas techniquessuch as PCA form a mapping ZJ 7! RK where K is considerably less than J .

The problem we consider, however, is to represent the document as a convexcombination, thus to form a mapping ZI 7! CK where CK denotes the subspaceof RK where every entry is non-negative and the entries sum to 1 (m 2 CK

implies 0 mk 1 andP

kmk = 1). Callm the reduced image of a document.For instance, suppose we are performing a coarse clustering of newswires

into topics: the topics found might be \sports", \business", \travel", \interna-tional", \politics", \domestic", and \cultural". Consider a document about a

53

3

major sport-star and the overlap of his honeymoon with a big game. Then tradi-tional clustering might output the following: \the document is about sports". Amore rened clustering system that represents uncertainties as well might out-put: \with 90% probability it is about sports, with 7% probability it is aboutcultural, and 3% probability about something else". General multinomial PCAconsidered in this paper might output: \50% of the document is about sports,35% of the document is about cultural, 7% about business, 5% about interna-tional". The supposed business content is really a discussion of the hotel for thehoneymoon and the supposed international content comes from the location ofthe honeymoon. Note here general multinomial PCA plays the role of dimen-sionality reduction, and places similar kinds of words into the same bucket forcompression purposes rather than any real topic identication.

The problem we consider is also to perform multi-faceted clustering, whichserves the purpose of extracting multiple mutually occurring topics from a doc-ument. Suppose m 2 CK is the reduced image of a particular document. Formulti-faceted clustering, m should have most entries zero, and only a few en-tries signicantly depart from zero. A measure we shall use for this is entropy,H(m) =

Pj mj log(1=mj). Thus multi-faceted clustering prefers low entropy

reduced images from CK . In the limit, when the average entropy of the reducedimages is 0, the mapping becomes equivalent to standard clustering. With thesimple honeymoon example above, the output could be reduced to: \70% ofthe document is about sports, 30% of the document is about cultural". Thismakes the document have 2H(m) = 1:85 eective topics, as opposed to the orig-inal PCA example above with more proportions (0.5,0.35,0.07,0.05) which had2H(m) = 3:17 topics.

Note that clustering is usually varied using the dimensionality K, their num-ber of components, not the nature of their decomposition, H(m). Approachessuch as the Information Bottleneck method [12] and hierarchical approaches ingeneral yield clustering at dierent scales and do not relax the assumption ofmutual exclusivity. We make the distinction here between the term componentwhich is a derived feature discovered for dimensionality reduction, and a facet,which is similar but is intended instead to be a relaxation of a topic. Documentsshould have a few facets for good multi-faceted clustering but many componentsfor eective dimensionality reduction.

3 Theory

We brie y review the theory of the multinomial version of PCA, and discuss theextensions. More details of the basic theory appear in [9].

Given a document, we rst to sample a K-dimensional probability vectormthat represents the proportional weighting of components, and then to mix itwith a K J matrix whose k-th row represents a word probability vectorfor the k-th component. For a document with a total count of L words in itsbag-of-words representation x, this is modelled as:

m Dirichlet() ;

54

4

x Multinomial(m; L) ;

where is a vector ofK-dimensional parameters to the Dirichlet. Thus, the meanof each entry xj is a convex combination of a column of , the probabilities forthe j-th word for dierent components.

This probability model does not readily yield an algorithm. The proportionvector m is a hidden variable but it cannot be treated with the standard EMalgorithm for hidden variables. However, if we can introduce a second hiddenvariable for each document w which is the word counts now broken out by wordindex j and topic index k as a JK matrix. This matrix has row totals equal thebag of words data x. An algorithm can be derived which iteratively recomputesan expected value for both the topic proportionsm and the word counts brokenout by topic w.

The following iterative algorithm can be derived using variational methods[9].

Theorem 1. Given the hidden variable model above, and the priors: m Dirichlet() and kl; Dirichlet(2f), where f is an empirical word proba-bility vector and is some other vector giving priors for the rst Dirichlet.The following updates converge to a lower bound of log p(x;;) that is op-timal for all product approximations q(m)q(w) for the hidden value posteriorp(m;w j;;x). The subscript [i] indicates values from the i-th document.For this the K-dimensional vector is an intermediate variable representinga Dirichlet approximation to the posterior distribution for the i-th documentsproportions m[i] and the J K I-dimensional array is an intermediatevariable representing the multinomial probabilities for an approximation to theposterior distribution for the rows of the i-th documents word matrix w[i].

j;k;[i] 1

Z1;j;[i]k;j exp

0(k;[i]) 0

Xk

k;[i]

!!;

k;[i] k +Xj

rj;[i] j;k;[i] ;

k;j 1

Z2;k

2fj +

Xi

rj;[i] j;k;[i]

!;

where 0() is the digamma function, and Z1;j;[i] and Z2;k are some normalizingconstants.

The exponential in the rst rewrite rule is an estimate ofmk;[i]) as exp(Eqflogmk;[i])g)which tends to reduce the component entropyH(m[i]). Note the last rewrite ruleis the standard MAP estimate for a multinomial parameter vector.

To reduces the entropies of the component proportionsm[i] even further, weuse the additional updates:

j;k;[i] 1

Z1;j;[i]k;jmk;[i]

55

5

mk;[i] nk;[i]P

k nk;[i] H(m[i]) log 1=mk;[i]

;

where is a Lagrange multiplier that can be increased to decrease the entropyH(m[i]). Note this replaces the above update for j;k;[i]. These modied updates

correspond to using an entropic prior pr(m[i]) / expH(m[i])

, which is a

weighted version of Brand's [13].

4 Experimental Setup

Data was collected about word occurrences from a signicant portion of theEnglish language documents in Google's August 2001 crawl. After HTML andother tokens are removed, the basic text is processed to determine the mostfrequent 5000 words consisting only of letters `a'-`z' ignoring case. Their co-occurrence data, the so-called bigram data was also collected. Bigrams are onlycounted for contiguous words in the same phrase: not broken by punctuation(excepting `-'), line breaks or other formatting tokens. The large number ofdocuments used allows the bigram data to be 17% non-zero for bigrams of thetop 5000 words. Note, some web pages contain seemingly random text and morethan enough jargon. The top word \to" has 139; 597; 023 occurrences and the5; 000-th word \charity" has 920; 343 occurrences. The most frequent bigramis \to be" with 20; 971; 200 occurrences, while the 1; 000-th most frequent is\included in" at 2; 333; 447 occurrences.

In this case, the role of document in the theory is played by a word, and therole of word, is played by the words appearing after this word.

The code for our system is 1300 documented lines of C, with error checking,input parsing, diagnostic reporting, and component display. The code runs com-parably to a PCA algorithm, converging in maybe 10-30 iterations, depending onthe accuracy required. It outputs a HTML page with internal links representingthe dierent aspects of the multi-faceted model constructed.

To measure the component entropies, we use 2H(m j d) where H(m j d) isthe mean of the individual entropies H(m[i]) for each document (which is aconditional entropy, hence the notation).

5 Experimental Results

We conducted a number of experiments as listed below.

5.1 Basic Illustration

We clustered the 5000 most frequent words on the web into 1500 dierent multi-faceted classes based on their occurrence in bigrams. The average eective num-ber of components per word (measured by 2hH(m j d)i) using standard multino-mial PCA is about 30 and the distinctions are diÆcult to interpret in many cases.Using the modied version of multinomial PCA which reduces the entropy of

56

6

the component vector m, we got this to the more manageable gure of about3 eective components per word. Many words had a majority component withprobability over 0.7, while a few had up to 20 dierent components.

Below we give some examples of the dierent facets for dierent words. Theseare presented here to illustrate the method. Note the clusters here representword use which is subtlety dierent to word meaning. Look at the examples for\wedding" below to see this. Our interpretation of each facet is given in italicprior to the list of words included in the facet,

\wedding": occasion: birthday, christmas, romantic, holiday, holidays,vacation, wedding, anniversary, happy,;

jewelry: rush, mini, gold, silver, jewelry, diamond, bell, wedding,\love": aection: kiss, love

preference: prefer, expect, recommend, need, like, probably, suggest,love, want, mean, say, require, never, think

emotion: confusion, pride, stress, pleasure, danger, fear, depression,honor, pain, comfort, britain, suering, passion, joy, glory, concern, de-sire, wealth, beauty, strength, escape, feeling, insight, promise, satisfac-tion, peace, respect, love

\four": number: seven, eight, ve, nine, six, twelve, ten, four, three, twenty,two, one

some/few: few, several, ve, ten, six, four, couple, three, half\scene": event: universe, situation, era, meal, scene, incident, lesson,

province, instance, issue, game, case, event, series, state, class, mission,project, school, sale, unit

play/performance: festival, scene\eorts": group work: initiative, initiatives, projects, proposals, collabo-

ration, eorts, eort, programs, activities, strategies, work attempts: attempts, aim, attempt, eorts, eort relationships: minds, hearts, lives, voices, bodies, families, eorts, com-mitment, attention, relationship, original, work

For instance, \love" is broken into an aection term, an preference term, and anemotion term, whereas \eorts" is broken into a group work term, an attemptsterm, and a relationship term.

5.2 Explaining Component Dimensionality

Why does standard multinomial PCA produce dierent eective number of com-ponents? We varied the input in a number of ways to explore this question.

First, we ran the system with dierent starting dimensions for the numberof components allowed. Given I documents (i.e., words in the Google data) andJ words/features per document (again, words in the Google data), then with Kstarting components, we are attempting to reduce the I J word count matrixto the product of a I K document topic matrix and a K J topic to wordmapping. Some of the K topics may be rarely used and contribute little, thus

57

7

the eective number of total components can be much less. We measure this as2H(p) where p is the K-dimensional vector of mean proportion for a topic inall documents (i.e., the mean of the rows of the I K document topic matrix).This contrasts with the eective number of topics/components per document2H(m j d) which is the conditional entropy on the I K document topic matrix,or the mean of the entropies of the rows of the I K document topic matrix.

Second, we down-sampled the data. We sampled without replacement fromthe data vector to reduce the total size of each \document" by subsamplingfactors of 100, 1; 000, 10; 000, 100; 000 respectively. Note the bigram word datahad huge starting counts. This was done to produce \documents" of dierentsizes. The characteristics of the data sets produced are given in Figure 1. The

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

1 10 100 1000 10000 100000 10

100

1000

10000

100000

1e+06

1e+07

Spa

rsity

(%

non

zero

)

Mea

n co

unt

Subsampling factor

Data characteristics

sparsitymean count

Fig. 1. Data characteristics for subsampled bigram data

X-axis is the subsampling factor. The solid line (axis on left) represents theproportion of non-zeros in the resultant data, and the dotted line (axis on right)represents the count of words in the resultant \document".

The results from these experiments are reported in Figure 2. Each curve rep-resents the results for one subsampling factor, i.e., an X value on Figure 1. Sothe top curve, which has no subsampling and where documents are 8; 000; 000words on average, the mean eective number of components per document in-creases upto about 40. The second from the bottom curve (/10000) has about 80words per document and typically 3-4 mean eective number of components perdocument, no matter how many components are inferred from the data (from20 upto about 800). The third from the bottom curve (/1000) has about 800

58

8

0

5

10

15

20

25

30

35

40

0 200 400 600 800 1000 1200

Mea

n ef

fect

ive

num

ber

of c

ompo

nent

s pe

r do

cum

ent

Effective number of components

Bigram data for 5000 words

baseline/100

/1000/10000

/100000

Fig. 2. Component dimensions

words per document and typically 10 mean eective number of components perdocument, no matter how many components are inferred from the data (from20 upto about 1000).

From these experiments, we can conclude that the mean eective number ofcomponents is largely in uenced by the document size. This would be result ofthe statistical capacity of the document to support a number of components.Small newswires and web pages can be 100 words, and thus could support a fewtopics statistically. Larger ones are about 1000 words and could support uptoabout 10 topics statistically. Web pages larger again that correspond to specsheets, long details of factual data, etc., could support even more topics.

6 Conclusion

We have demonstrated that recent extensions to PCA for multinomial data areinadequate for multi-faceted clustering and given results for a modication to thebasic algorithm that performs better in this regard. We argued that multinomialPCA is really a dimensionality reduction algorithm, and not designed for multi-faceted clustering.

It remains to be seen how the modied algorithm will perform on the sug-gested task of performing multi-faceted clustering of web-pages as a pre-processingstep to data mining for the semantic web. For this, we would need at least theability to generate full topic hierarchies, and to perform automatic labelling/namingof topics in the hierarchy. We expect that by doing this for multiple companiesat once, useful hierarchies could be established.

59

9

Another research direction is to modify these algorithms to create supervisedversions of them, whereby each item/document/page is tagged with multipletopics (as the Reuters and AP news-wires can be), and the task is to learn amodel for the component topics and the proportions for each.

References

1. Shi, J., Malik, J.: Normalized cuts and image segmentation. IEEE Transactionson Pattern Analysis and Machine Intelligence 22 (2000) 888905

2. Moore, A.: Very fast EM-based mixture model clustering using multiresolutionkd-tree. In: Neural Information Processing Systems, Denver (1998)

3. Han, E.H., Karypis, G., Kumar, V., Mobasher, B.: Clustering based on associationrule hypergraphs. In: SIGMOD'97 Workshop on Research Issues on Data Miningand Knowledge. (1997)

4. Vaithyanathan, S., Dom, B.: Model-based hierarchical clustering. In: UAI-2000,Stanford (2000)

5. Bradley, P., Fayyad, U., Reina, C.: Scaling clustering algorithms to large databases.In: Proc. KDD'98. (1998)

6. Lee, D., Seung, H.: Learning the parts of objects by non-negative matrix factor-ization. Nature 401 (1999) 788791

7. Hofmann, T.: Probabilistic latent semantic indexing. In: Research and Develop-ment in Information Retrieval. (1999) 5057

8. Blei, D., Ng, A., Jordan, M.: Latent Dirichlet allocation. In: NIPS*14. (2002) toappear.

9. Buntine, W.L.: Variational extensions to EM and multinomial PCA. In: ECML2002. (2002)

10. Hall, K., Hofmann, T.: Learning curved multinomial subfamilies for natural lan-guage processing and information retrieval. In: ICML 2000. (2000)

11. Baeza-Yates, R., Ribeiro-Neto, B.: Modern Information Retrieval. Addison Wesley(1999)

12. Tishby, N., Pereira, F., Bialek, W.: The information bottleneck method. In: 37-thAnnual Allerton Conference on Communication, Control and Computing. (1999)368377

13. Brand, M.: Structure learning in conditional probability models via an entropicprior and parameter extinction. Neural Computation 11 (1999) 11551182

60

#ESWKUKVKQP QH 9GD 5GOCPVKEU DCUGF QP C /WNVKNKPIWCN

6GZV/KPKPI 6GEJPKSWG

%JWPI*QPI .GG CPF *UKP%JCPI ;CPI

&GRCTVOGPV QH 'NGEVTKECN 'PIKPGGTKPI

0CVKQPCN -CQJUKWPI 7PKXGTUKV[ QH #RRNKGF 5EKGPEGU

%JKGP -WPI 4QCF -CQJUKWPI [email protected]

&GRCTVOGPV QH +PHQTOCVKQP /CPCIGOGPV %JCPI ,WPI 7PKXGTUKV[

6CKPCP [email protected]

#DUVTCEV 6JKU RCRGT FGUETKDGU JQY UGOCPVKE YGD KUUWGU UWEJ CU OCEJKPG

WPFGTUVCPFCDNG UGOCPVKE TGRTGUGPVCVKQP KP OWNVKRNG NCPIWCIGU CPF CESWKUKVKQP

QH YGD UGOCPVKEU CTG VCEMGF D[ OGCPU QH C OWNVKNKPIWCN VGZVOKPKPI CRRTQCEJ

6JKU RCRGT RTGUGPVU CNIQTKVJOU VJCV GPCDNG OWNVKNKPIWCN VGZV OKPKPI DCUGF QP

PGWTCN PGVYQTM VGEJPKSWG PCOGN[ VJG UGNHQTICPK\KPI OCRU 51/ HQT CWVQ

OCVKECNN[ ITQWRKPI UKOKNCT OWNVKNKPIWCN VGZVU KG %JKPGUG CPF 'PINKUJ VGZVU

9G VTGCV VJG 9QTNF 9KFG 9GD CU QWT OWNVKNKPIWCN EQTRWU CPF WVKNK\G VJG VGZV

OKPKPI OQFGN HQT ITQWRKPI UGOCPVKECNN[ UKOKNCT FQEWOGPVU 9JGP GZRQUG VQ

TCY OWNVKNKPIWCN VGZV VJG OQFGN ENWUVGTU YQTFU CPF FQEWOGPVU CEEQTFKPI VQ

VJGKT UGOCPVKE TGNCVGFPGUU VQ HQTO C UGOCPVKE PGVYQTM 6JG OQFGN CNUQ UWI

IGUVU C PQXGN GZRNCPCVKQP QH OWNVKNKPIWCN VGZV ENWUVGTKPI DCUGF QP VJG UGOCPVKE

TGNCVGFPGUU +P CFFKVKQP YG RTQRQUG C HTCOGYQTM DCUGF QP VJG UGOCPVKEU GZ

VTCEVGF HTQO VJG TGUWNVKPI FQEWOGPV ENWUVGTU CNQPI YKVJ C :/. FQEWOGPV

VGEJPKSWG VQ ECTT[ QWV VJG VCUM QH TGECVGIQTK\CVKQP QH YGD RCIGU 6JG YGD

RCIGU CTG TGECVGIQTK\GF DCUGF QP VJG UGOCPVKE TGNCVGFPGUU KP VJG EQTRWU (QT

KORNGOGPVCVKQP YG CRRNKGF VJG FGTKXGF OWNVKNKPIWCN VGZV OKPKPI CRRTQCEJ KP

VJG RTQEGUU QH CWVQOCVKE EQPUVTWEVKQP QH C PGY VGZV FKTGEVQT[ KP QTFGT VQ C

:/. FQEWOGPV FCVCDCUG VJCV CEJKGXGU VJG CKOU QH C UGOCPVKEUDCUGF MPQYN

GFIG ECVGIQTK\CVKQP

+PVTQFWEVKQP

6JG 5GOCPVKE 9GD KU WUGF VQ FGPQVG VJG PGZV GXQNWVKQP UVGR QH VJG 9GD YJKEJ

GUVCDNKUJGU C NC[GT QH OCEJKPG WPFGTUVCPFCDNG FCVC 6JG FCVC KU UWKVCDNG HQT CWVQOCVGF

CIGPVU UQRJKUVKECVGF UGCTEJ GPIKPGU CPF KPVGTQRGTCDKNKV[ UGTXKEGU YJKEJ RTQXKFG C

RTGXKQWUN[ PQV TGCEJCDNG ITCFG QH CWVQOCVKQP 6JG WNVKOCVG IQCN QH VJG 5GOCPVKE 9GD

KU VQ CNNQY OCEJKPGU VJG UJCTKPI CPF GZRNQKVCVKQP QH MPQYNGFIG KP VJG 9GD YC[ KG

YKVJQWV EGPVTCN CWVJQTKV[ YKVJ HGY DCUKE TWNGU KP C UECNCDNG CFCRVCDNG GZVGPUKDNG

OCPPGT

61

5GOCPVKEU KP 9GD %QPVGPV

4GEGPV YQTM KP EQORWVCVKQPCN NKPIWKUVKEU UWIIGUVU VJCV NCTIG COQWPVU QH UGOCPVKE

KPHQTOCVKQP ECP DG GZVTCEVGF CWVQOCVKECNN[ HTQO NCTIG VGZV EQTRQTC QP VJG DCUKU QH

NGZKECN EQQEEWTTGPEG KPHQTOCVKQP 5WEJ UGOCPVKEU JCU DGGP DGEQOKPI KORQTVCPV CPF

WUGHWN HQT DGKPI CU CP GUUGPVKCN TGRTGUGPVCVKQP QH EQPVGPV KP GCEJ YGD RCIG RCTVKEW

NCTN[ YKVJ VJG KPETGCUKPI CXCKNCDKNKV[ QH FKIKVCN FQEWOGPVU KP XCTKQWU NCPIWCIGU HTQO

CNN CTQWPF VJG YQTNF +P VJKU YQTM YG CVVGORV VQ FGXGNQR C PQXGN CNIQTKVJOKE CRRTQCEJ

HQT OWNVKNKPIWCN VGZV OKPKPI VGEJPKSWG VQ GZVTCEV UGOCPVKE KPHQTOCVKQP HTQO VJG YGD

VGZV EQTRQTC 7UKPI C XCTKCVKQP QH CWVQOCVKE ENWUVGTKPI VGEJPKSWGU YJKEJ CRRN[ VJG

5GNH1TICPK\KPI /CRU 51/ YG JCXG EQPFWEVGF UGXGTCN GZRGTKOGPVU VQ WPEQXGT

CUUQEKCVGF FQEWOGPVU DCUGF QP %JKPGUG'PINKUJ DKNKPIWCN RCTCNNGN EQTRQTC CPF C

J[DTKF %JKPGUG'PINKUJ EQTRWU 9JGP GZRQUGF VQ TCY OWNVKNKPIWCN VGZV VJG OQFGN

ENWUVGTU YQTFU CPF FQEWOGPVU CEEQTFKPI VQ VJGKT UGOCPVKE TGNCVGFPGUU VQ HQTO C UG

OCPVKE PGVYQTM CU C UQWTEG QH OGVCFCVC HQT VJG EQPUVTWEVKQP QH VJG 5GOCPVKE 9GD

/WNVKNKPIWCN 6GZV/KPKPI HQT %TGCVKPI /GVCFCVC

/WNVKNKPIWCN VGZV OKPKPI KU VJG GZVTCEVKQP QH KORNKEKV RTGXKQWUN[ WPMPQYP CPF UG

OCPVKECNN[ UKOKNCT VGTOU CPF FQEWOGPVU HTQO OWNVKNKPIWCN EQTRQTC 6JG QDLGEVKXG KU

VQ GUVCDNKUJ EQORWVGT RTQITCOU VJCV CWVQOCVKECNN[ ENWUVGT EQPEGRVWCNN[ TGNCVGF VGTOU

CPF VGZVU KP XCTKQWU NCPIWCIGU

+P VJKU YQTM VJG RTKOCT[ VJTWUV KU VQ OCMG VJG VGZV OKPKPI VGEJPKSWGU HWNN[ OWNVK

NKPIWCN OKPKPI QRGPUQWTEG VGZVU KP FKHHGTGPV NCPIWCIGU *QYGXGT QWT IQCN KP VJGUG

GZRGTKOGPVU KU PQV VQ GUVCDNKUJ VJG KFGCN HWNNHWPEVKQPCN OWNVKNKPIWCN KPHQTOCVKQP FKU

EQXGT[ U[UVGO CU VJG VKOG CPF TGUQWTEGU TGSWKTGF HQT VJKU VCUM CTG EQPUKFGTCDNG

4CVJGT YG TGUVTKEV QWT CVVGPVKQP VQ DKNKPIWCN EQTRQTC CPF VT[ VQ WPFGTUVCPF VJG DCUKE

TGSWKTGOGPVU HQT GHHGEVKXG OWNVKNKPIWCN KPHQTOCVKQP FKUEQXGT[ CPF VJG RTQDNGOU VJCV

CTKUG HTQO C UKORNG KORNGOGPVCVKQP QH UWEJ C U[UVGO 6JKU TGUGCTEJ YQTM KU OCKPN[

ECTTKGF QWV D[ GZRGTKOGPVKPI YKVJ VGZV GZVTCEVKQP HTQO RCTCNNGN EQTRQTC C XCTKCVKQP QH

DKNKPIWCN EQTRQTC $KNKPIWCN EQTRQTC ECP DG WUGF KP OCP[ YC[U (QT OWNVKNKPIWCN

KPHQTOCVKQP CEEGUU DKNKPIWCN EQTRQTC OCMG KV RQUUKDNG VQ KPXGUVKICVG U[PVCEVKE UGOCP

VKE CPF NGZKECN TGNCVKQPUJKRU DGVYGGP NCPIWCIGU CPF CTG CNUQ KORQTVCPV UQWTEGU QH

EQPVTCUVKXG GXKFGPEG QH NCPIWCIG KP WUCIG )GPGTCNN[ URGCMKPI VJGTG CTG VYQ FKHHGTGPV

V[RGU QH DKNKPIWCN EQTRQTC RCTCNNGN CPF EQORCTCDNG EQTRQTC 2CTCNNGN VGZV EQTRQTC CTG

UGVU QH VTCPUNCVKQPGSWKXCNGPV VGZVU KP YJKEJ IGPGTCNN[ QPG VGZV KU VJG UQWTEG VGZV CPF

VJG QVJGTU CTG VTCPUNCVKQPU %QORCTCDNG VGZV EQTRQTC CTG UGVU QH VGZVU HTQO RCKTU QT

OWNVKRNGU QH NCPIWCIGU YJKEJ ECP DG EQPVTCUVGF CPF EQORCTGF DGECWUG QH VJGKT EQO

OQP HGCVWTGU *QYGXGT WPNKMG RCTCNNGN KG VTCPUNCVKQPGSWKXCNGPV EQTRQTC VJG[

CNYC[U EQPEGTP C TGUVTKEVGF UWDNCPIWCIG 6JG[ QHHGT C UQWTEG QH FCVC QP PCVWTCN NCP

IWCIG NGZKECN GSWKXCNGPVU YKVJKP C IKXGP FQOCKP =? (QT OWNVKNKPIWCN VGZV OKPKPI KP

VJKU YQTM YG GZRGEV KV VQ DG CEVGF CU C UVCTVKPI RQKPV HQT GZRNQTKPI VJG KORCEVU QP

NKPIWKUVKEU KUUWGU YKVJ VJG OCEJKPG NGCTPKPI CRRTQCEJ VQ OKPKPI UGPUKDNG NKPIWKUVKEU

GNGOGPVU HTQO OWNVKNKPIWCN VGZV EQNNGEVKQPU +V KU DGNKGXGF VJCV KH VJG EQTRWU KU NCTIG

GPQWIJ VJGP UVCVKUVKECN QT QVJGT CNIQTKVJODCUGF VGEJPKSWGU ECP DG WUGF VQ RTQFWEG

62

DKNKPIWCN VGTO GSWKXCNGPVU QT CUUQEKCVKQPU D[ EQORCTKPI YJKEJ UVTKPIU EQQEEWT KP VJG

UCOG UGPVGPEGU QXGT VJG YJQNG EQTRWU 6JGTGHQTG KP VJKU YQTM YG GORNQ[GF C RCTCNNGN

EQTRWU VJG DKNKPIWCN OCIC\KPG CTVKENGU HTQO VJG 5KPQTCOC /CIC\KPG VQ ETGCVG C UGV QH

FWCNNCPIWCIG VTCKPKPI FQEWOGPVU HQT FKUEQXGTKPI ETQUUNCPIWCIG VGTO CUUQEKCVKQPU

CPF FQEWOGPV CUUQEKCVKQPU

51/ DCUGF 6GZV/KPKPI #RRTQCEJ

&QEWOGPV 2TGRTQEGUUKPI

6JG RTQRQUGF U[UVGO HQEWUGU QP VJG VCUM QH HKPFKPI CUUQEKCVKQPU KP EQNNGEVKQPU QH VGZV

$CUGF QP CUUQEKCVKQP UKOKNCT FQEWOGPVU VJTQWIJ VJG RTQRQUGF VGZV OKPKPI RTQEGUU

ECP DG ICVJGTGF KP C ENWUVGT (QT CP 'PINKUJ EQTRWU VJG FQEWOGPV RTGRTQEGUUKPI KU

SWKVG UVTCKIJVHQTYCTF 1WT CRRTQCEJ DGIKPU YKVJ C UVCPFCTF RTCEVKEG KP KPHQTOCVKQP

TGVTKGXCN +4 =? VQ GPEQFG FQEWOGPVU YKVJ XGEVQTU KP YJKEJ GCEJ EQORQPGPV EQTTG

URQPFU VQ C FKHHGTGPV YQTF CPF VJG XCNWG QH VJG EQORQPGPV TGHNGEVU VJG HTGSWGPE[ QH

YQTF QEEWTTGPEG KP VJG FQEWOGPV +P RTCEVKEG VJG TGUWNVKPI FKOGPUKQPCNKV[ QH VJG

URCEG KU QHVGP VTGOGPFQWUN[ JWIG UKPEG VJG PWODGT QH FKOGPUKQPU KU FGVGTOKPGF D[

VJG PWODGT QH FKUVKPEV KPFGZGF VGTOU KP VJG EQTRWU #U C TGUWNV VGEJPKSWGU HQT EQP

VTQNNKPI VJG FKOGPUKQPCNKV[ QH VJG XGEVQT URCEG CTG TGSWKTGF 5WEJ C RTQDNGO EQWNF DG

UQNXGF D[ GNKOKPCVKPI UQOG QH VJG OQUV EQOOQP CPF UQOG QH VJG TCTGUV YQTFU CPF

D[ CRRN[KPI C PWOGTKECN CNIQTKVJO UWEJ CU .CVGPV 5GOCPVKE +PFGZKPI .5+ =? OGVJQF

(QT C %JKPGUG EQTRWU VJG FQEWOGPV RTGRTQEGUUKPI KU TGNCVKXGN[ EQORNKECVGF 5KPEG C

%JKPGUG UGPVGPEG KU EQORQUGF QH EJCTCEVGTU YKVJQWV DQWPFCTKGU UGIOGPVCVKQP KU

KPFKURGPUCDNG 9G GORNQ[ C FKEVKQPCT[ UQOG OQTRJQNQIKECN TWNGU CPF CP CODKIWKV[

TGUQNWVKQP OGEJCPKUO HQT UGIOGPVCVKQP +P CFFKVKQP YG CNUQ GZVTCEV PCOGF QTICPK\C

VKQPU RGQRNG CPF NQECVKQPU CNQPI YKVJ FCVGVKOG GZRTGUUKQPU CPF OQPGVCT[ CPF RGT

EGPVCIG GZRTGUUKQPU 6JG TGUV QH VJG RTQEGUU KU VJG UCOG CU VJCV QH VGZV OKPKPI KP CP

'PINKUJ EQTRWU

5GNH1TICPK\KPI /CRU

6JG UGNHQTICPK\KPI OCR 51/ =?=? KU QPG QH VJG OCLQT WPUWRGTXKUGF CTVKHKEKCN

PGWTCN PGVYQTM OQFGNU +V DCUKECNN[ RTQXKFGU C YC[ HQT ENWUVGT CPCN[UKU D[ RTQFWEKPI C

OCRRKPI QH JKIJ FKOGPUKQPCN KPRWV XGEVQTU QPVQ C VYQFKOGPUKQPCN QWVRWV URCEG

YJKNG RTGUGTXKPI VQRQNQIKECN TGNCVKQPU CU HCKVJHWNN[ CU RQUUKDNG #HVGT CRRTQRTKCVG

VTCKPKPI KVGTCVKQPU VJG UKOKNCT KPRWV KVGOU CTG ITQWRGF URCVKCNN[ ENQUG VQ QPG CPQVJGT

#U UWEJ VJG TGUWNVKPI OCR KU ECRCDNG QH RGTHQTOKPI VJG ENWUVGTKPI VCUM KP C EQO

RNGVGN[ WPUWRGTXKUGF HCUJKQP +P VJKU YQTM YG GORNQ[ VJG 51/ OGVJQF VQ RTQFWEG

VYQ OCRU HQT VGZV OKPKPI PCOGN[ VJG YQTF ENWUVGT OCR CPF VJG FQEWOGPV ENWUVGT

OCR

63

6JG 9QTF %NWUVGT /CR CPF &QEWOGPV %NWUVGT /CR

6JG YQTF ENWUVGT OCR VJCV KU GORNQ[GF HQT FQEWOGPV GPEQFKPI KU RTQFWEGF CEEQTFKPI

VQ YQTF UKOKNCTKVKGU OGCUWTGF D[ VJG UKOKNCTKV[ QH VJG EQQEEWTTGPEG QH VJG YQTFU

%QPEGRVWCNN[ TGNCVGF YQTFU VGPF VQ HCNN KPVQ VJG UCOG QT PGKIJDQTKPI OCR PQFGU $[

OGCPU QH VJG 51/ CNIQTKVJO YQTF ENWUVGTU ECP DG QTFGTGF CPF QTICPK\GF CU PQFGU QP

VJG OCR .GV ZK∈ℜ0 ≤ K ≤ / DG VJG HGCVWTG XGEVQT QH VJG K

VJFQEWOGPV KP VJG EQTRWU

YJGTG 0 KU VJG PWODGT QH KPFGZGF VGTOU CPF / KU VJG PWODGT QH FQEWOGPVU 9G WUGF

VJGUG XGEVQTU CU VJG VTCKPKPI KPRWVU VQ VJG OCR 6JG OCR EQPUKUVU QH C TGIWNCT ITKF QH

RTQEGUUKPI WPKVU ECNNGF PGWTQPU 'CEJ PGWTQP KP VJG OCR JCU 0 U[PCRUGU .GV

YL]YLP^ ≤ P ≤ 0_ ≤ L ≤ , DG VJG U[PCRVKE YGKIJV XGEVQT QH VJG LVJ PGWTQP KP VJG

OCR YJGTG , KU VJG PWODGT QH PGWTQPU QP VJG OCR 9G VTCKPGF VJG OCR D[ VJG 51/CNIQTKVJO5VGR 4CPFQON[ UGNGEV C VTCKPKPI XGEVQT ZK HTQO VJG EQTRWU

5VGR (KPF VJG PGWTQP L YKVJ U[PCRVKE YGKIJVU YL YJKEJ KU ENQUGUV VQ ZK KG

5VGR (QT GXGT[ PGWTQP N KP VJG PGKIJDQT QH PQFG L WRFCVG KVU U[PCRVKE YGKIJVU D[

YJGTG αV KU VJG VTCKPKPI ICKP CV VKOG UVCOR V

5VGR +PETGCUG VKOG UVCOR V +H V TGCEJGU VJG RTGUGV OCZKOWO VTCKPKPI VKOG 6 JCNV

VJG VTCKPKPI RTQEGUU QVJGTYKUG FGETGCUG αV CPF VJG PGKIJDQTJQQF UK\G CPF IQ VQ

5VGR

6JG VTCKPKPI RTQEGUU UVQRU CHVGT VKOG 6 YJKEJ KU UWHHKEKGPVN[ NCTIG VJCV GXGT[ HGC

VWTG XGEVQT OC[ DG UGNGEVGF CU VTCKPKPI KPRWV HQT EGTVCKP VKOGU 6JG VTCKPKPI ICKP CPF

PGKIJDQTJQQF UK\G DQVJ FGETGCUG YJGP V KPETGCUGU

#HVGT VJG VTCKPKPI RTQEGUU VJG OCR HQTOU C 9QTF %NWUVGT /CR D[ NCDGNKPI GCEJ

PGWTQP YKVJ EGTVCKP YQTFU (QT VJG PVJ

YQTF KP VJG EQTRWU YG EQPUVTWEV CP 0

FKOGPUKQPCN XGEVQT XPKP YJKEJ QPN[ VJG P

VJGNGOGPV KU PQP\GTQ 6Q NCDGN VJG PGWTQPU

YG RTGUGPV GCEJ XPVQ VJG OCR CPF HKPF VJG DGUV OCVEJKPI PGWTQP 5KPEG VJG PWODGT QH

PGWTQPU KU IGPGTCNN[ OWEJ UOCNNGT VJCP VJG PWODGT QH YQTFU GCEJ PGWTQP KP VJG OCR

OC[ JCXG OWNVKRNG NCDGNU 9G OC[ UC[ VJCV C PGWTQP HQTOU C YQTF ENWUVGT DGECWUG VJG

ENQUGN[ TGNCVGF YQTFU YKNN OCR VQ VJG UCOG PGWTQP 6JG YQTF ENWUVGT OCR CWVQPQ

OQWUN[ ENWUVGTU YQTFU CEEQTFKPI VQ VJGKT UKOKNCTKV[ QH EQQEEWTTGPEG 9QTFU VJCV VGPF

VQ DG HQWPF KP VJG UCOG FQEWOGPV YKNN DG OCRRGF VQ ENQUG PGWTQPU KP VJG OCR (QT

GZCORNG VJG %JKPGUG YQTFU HQT PGWTCN CPF PGVYQTM QHVGP QEEWT UKOWNVCPGQWUN[ KP

C FQEWOGPV 6JG[ YKNN OCR VQ VJG UCOG PGWTQP QT PGKIJDQTKPI PGWTQPU QP VJG OCR

9QTFU VJCV FQ PQV QEEWT KP VJG UCOG FQEWOGPV YKNN OCR VQ FKUVCPV PGWTQPU QP VJG

OCR #EEQTFKPIN[ YG ECP FGHKPG VJG TGNCVKQPUJKR DGVYGGP VYQ YQTFU CEEQTFKPI VQ

VJGKT EQTTGURQPFKPI PGWTQPU KP VJG YQTF ENWUVGT OCR CPF VJG OKPKPI VCUM YKNN DG

PLQ NLN

ML Z[Z[ −=−

ROGROGQHZ

OLOOW Z[ZZ −+= α

64

RGTHQTOGF DCUGF QP UWEJ TGNCVKQPUJKRU 6JG VTCKPGF OCR CNUQ HQTOU C &QEWOGPV %NWU

VGT /CR D[ NCDGNKPI GCEJ PGWTQP YKVJ EGTVCKP FQEWOGPVU 6JG FQEWOGPV HGCVWTG XGE

VQTU ZK CTG RTGUGPVGF VQ VJG OCR VQ NCDGN VJG PGWTQPU &QEWOGPVU YKVJ UKOKNCT MG[

YQTFU YKNN OCR VQ VJG UCOG QT PGKIJDQTKPI PGWTQPU 6JG UKOKNCTKV[ DGVYGGP VYQ

FQEWOGPVU OC[ DG ECNEWNCVGF D[ OGCUWTKPI VJG 'WENKFGCP FKUVCPEG DGVYGGP VJGKT

OCRRGF PGWTQPU KP VJG OCR 5KPEG VJG PWODGT QH VJG PGWTQPU KU OWEJ NGUU VJCP VJG

PWODGT QH VJG FQEWOGPVU KP VJG EQTRWU OWNVKRNG FQEWOGPVU OC[ OCR VQ VJG UCOG

PGWTQP 6JWU C PGWTQP HQTOU C FQEWOGPV ENWUVGT $GUKFGU PGKIJDQTKPI PGWTQPU TGRTG

UGPV FQEWOGPV ENWUVGTU QH UKOKNCT OGCPKPI KG JKIJ MG[YQTF EQQEEWTTGPEG HTG

SWGPE[

&KUEQXGT[ #NIQTKVJO

+P VJKU UGEVKQP YG GZRNCKP JQY VJG YQTF ENWUVGT OCR CPF FQEWOGPV ENWUVGT OCR GHHGE

VKXGN[ OQFGN VJG TGNCVKQPUJKR DGVYGGP VJG YQTFU CPF FQEWOGPVU 9G VTCPUHQTO C

FQEWOGPV VQ C XGEVQT QH YQTF QEEWTTGPEG #HVGT VJG UGNHQTICPK\KPI RTQEGUU VYQ

FQEWOGPVU YKNN OCR VQ PGCT PGWTQPU KH VJG[ EQPVCKP UKOKNCT YQTF QEEWTTGPEGU 9JGP

FKHHGTGPV YQTFU CTG NCDGNGF QP VJG UCOG PGWTQP QT PGCT PGWTQPU QP VJG YQTF ENWUVGT

OCR VJG[ VGPF VQ QEEWT KP C TGUVTKEVGF UGV QH FQEWOGPVU 1P VJG QVJGT JCPF KH VYQ

YQTFU UGNFQO EQQEEWT KP CP[ FQEWOGPV VJG[ UJQWNF PQV DG NCDGNGF QP PGCT PGWTQPU

6JKU KU DGECWUG VJG PGWTQP OC[ DG XKGYGF CU TGRTGUGPVKPI C XKTVWCN FQEWOGPV EQPVCKP

KPI VJQUG YQTFU NCDGNGF QP KV 6YQ YQTFU YKNN DG OCRRGF VQ VJG UCOG PGWTQP KH CPF

QPN[ KH VJG[ QHVGP EQQEEWT KP VJG UCOG FQEWOGPV QVJGTYKUG VJG XKTVWCN FQEWOGPV

OC[ PQV EQPVCKP VJGUG YQTFU UKOWNVCPGQWUN[ 0GKIJDQTKPI PGWTQPU KP VJG YQTF ENWUVGT

OCR TGRTGUGPV YQTF ENWUVGTU EQPVCKPKPI UKOKNCT YQTFU KG YQTFU VGPF VQ EQQEEWT KP

VJG UCOG FQEWOGPV *GPEG VJG UGNHQTICPK\KPI OCR OC[ OGCUWTG VJG YQTF EQ

QEEWTTGPEG UKOKNCTKV[ COQPI FQEWOGPVU 9G FGHKPG VJG UKOKNCTKV[ DGVYGGP VJG RVJ

YQTF CPF VJG SVJ YQTF CU HQNNQYU

−1−

+ TS 1*1*

=TS'

YJGTG 0R KU VJG PGWTQP NCDGNGF D[ VJG RVJ YQTF CPF )0R KU VJG VYQFKOGPUKQPCN

ITKF NQECVKQP QH 0R 5WEJ UKOKNCTKV[ OGCUWTGU VJG NKMGNKJQQF QH VJG EQQEEWTTGPEG QH

YQTFU .CTIG UKOKNCTKV[ TGXGCNU VJCV VJG VYQ YQTFU QHVGP EQQEEWT KP VJG UCOG UGV QH

FQEWOGPVU YJKEJ OC[ DG EQPUKFGTGF CU C MKPF QH CUUQEKCVKQP RCVVGTP COQPI VJG

YQTFU

1P VJG FQEWOGPV ENWUVGT OCR C PGWTQP TGRTGUGPVU C FQEWOGPV ENWUVGT VJCV EQPVCKPU

FQEWOGPVU QH UKOKNCT OGCPKPI YJKEJ KU FGHKPGF D[ VJG UGV QH JKIJN[ EQQEEWTTKPI

YQTFU VJG[ EQPVCKP 5KPEG YG VTCKP VJG OCR WUKPI VJG GPEQFGF FQEWOGPV HGCVWTG XGE

VQTU CU KPRWV VJG YGKIJV XGEVQT QH C PGWTQP TGRTGUGPVU VJG QEEWTTGPEG QH VJG YQTFU KP

C XKTVWCN FQEWOGPV 5WEJ C XKTVWCN FQEWOGPV OC[ DG WUGF CU VJG EGPVTQKF QH VJG

FQEWOGPV ENWUVGT CUUQEKCVGF YKVJ VJCV PGWTQP 1P VJG FQEWOGPV ENWUVGT OCR QPN[

VJQUG FQEWOGPVU EQPVCKPKPI QXGTNCRRKPI YQTFU OC[ OCR VQ VJG UCOG PGWTQP &QEW

OGPVU EQPVCKPKPI PQPQXGTNCRRKPI YQTFU OC[ OCR VQ FKUVCPV PGWTQPU 0GKIJDQTKPI

PGWTQPU TGRTGUGPV FQEWOGPV ENWUVGTU YKVJ UKOKNCT QXGTNCRRKPI UGVU QH YQTFU VJWU

VJG EQQEEWTTGPEG QH YQTFU OC[ DG FGVGTOKPGF D[ VJG PGKIJDQTJQQF URCVKCNN[ (QT

CP[ FQEWOGPV YG ECP HKPF KVU UKOKNCT FQEWOGPVU D[ GZCOKPKPI KVU OCRRGF PGWTQP

65

CPF PGKIJDQTKPI PGWTQPU KP VJG FQEWOGPV ENWUVGT OCR 6JG UKOKNCTKV[ DGVYGGP VJG KVJ

CPF MVJ FQEWOGPV KU FGHKPGF CU HQNNQYU

−−+= NL 1*1*

NL'

YJGTG 0K JGTG KU VJG PGWTQP NCDGNGF D[ VJG KVJ FQEWOGPV CPF )0K KU ITKF NQECVKQP

QH 0K CU KP 'S

&QEWOGPVU YJKEJ EQPVCKP VJG UCOG UGVU QH YQTFU YKNN FGHKPKVGN[ OCR VQ VJG UCOG

PGWTQP TGUWNVKPI KP JKIJ UKOKNCTKV[ FGHKPGF KP 'S /QTGQXGT GXGP KH VJGUG FQEW

OGPVU FQ PQV EQPVCKP GZCEVN[ VJG UCOG UGV QH YQTFU YG OC[ UVKNN UC[ VJCV VJG[ CTG

EQPEGRVWCNN[ UKOKNCT DGECWUG VJG[ UVKNN EQPVCKP EQOOQP YQTFU VJCV QHVGP EQQEEWT

KP VJGUG FQEWOGPVU CPF VJG FKUUKOKNCT YQTFU CTG NKMGN[ QEEWTTKPI KP FQEWOGPVU

OCRRGF VQ VJG PGCTD[ PGWTQPU &QEWOGPV ENWUVGT OCRU RTQXKFG CP GHHGEVKXGN[ YC[ VQ

HQTO FQEWOGPV ENWUVGTU # PGWTQP QP VJG OCR TGRTGUGPVU C FQEWOGPV ENWUVGT 9G ECP

CNUQ FGHKPG VJG UKOKNCTKV[ DGVYGGP VYQ ENWUVGTU D[ VJG FKUVCPEG QH VJGKT EQTTGURQPFKPI

PGWTQPU

−−+= OM 1*1*

OM'

YJGTG L CPF N CTG VJG PGWTQP KPFKEGU QH VJG VYQ ENWUVGTU

6JG UKOKNCTKVKGU FGHKPGF KP 'S CPF RTQXKFG UQOG MPQYNGFIG CDQWV VJG

FQEWOGPVU KP VJG EQTRWU 9G WUG 'S VQ FKUETKOKPCVG YQTFU DCUGF QP VJG MPQYNGFIG

FKUEQXGTGF HTQO VJG EQTRWU 6JKU KU CNUQ VTWG HQT 'S CPF 'S VQ FKUETKOKPCVG FQEW

OGPVU CPF ENWUVGTU TGURGEVKXGN[ 9QTF CUUQEKCVKQPU CU YGNN CU FQEWOGPV CUUQEKCVKQPU

CTG ENGCTN[ FGHKPGF D[ UWEJ UKOKNCTKVKGU +V KU PCVWTCN VQ CRRN[ UWEJ CUUQEKCVKQPU VQ

CRRNKECVKQPU QH FQEWOGPV TGVTKGXCN KPFGZKPI CPF ENWUVGTKPI # SWGT[ YJKEJ KU HQTOW

NCVGF CU C MG[YQTF QT C EQODKPCVKQP QH MG[YQTFU CFLCEGPV D[ $QQNGCP QRGTCVQTU KU

UGPV D[ VJG WUGT VQ TGVTKGXG TGNGXCPV FQEWOGPVU 9G RGTHQTO C UGCTEJ VJTQWIJ VJG

YQTF ENWUVGT OCR VQ QDVCKP VJG NCDGNGF PGWTQPU QH VJG MG[YQTF QT EQODKPCVKQPU QH

MG[YQTFU 6JG FQEWOGPVU NCDGNGF QP VJG EQTTGURQPFKPI PGWTQPU QP VJG FQEWOGPV

ENWUVGT OCR CTG UGNGEVGF CPF RTGUGPVGF VQ VJG WUGT /QTGQXGT CNN FQEWOGPVU OC[ CNUQ

DG RTGUGPVGF D[ FQEWOGPV UKOKNCTKV[ FGHKPGF KP 'S &QEWOGPVU YKVJ JKIJGT UKOKNCTK

VKGU CRRGCT GCTNKGT KP VJG SWGT[ TGUWNV 6JG WUGT IGVU C NKUV QH TGNGXCPV FQEWOGPVU GXGP

YJGP VJG SWGT[ YQTFU OC[ PQV QEEWT KP VJG FQEWOGPVU +ORQTVCPV KPHQTOCVKQP CPF

PGY MPQYNGFIG TGNCVGF VQ VJG WUGTU SWGT[ OC[ DG TGXGCNGF D[ QWT OGVJQF

/WNVKNKPIWCN 6GZV /KPKPI HTQO C *[DTKF %JKPGUG'PINKUJ

%QTRWU

(WTVJGTOQTG YG VGUVGF VJG FGXGNQRGF FKUEQXGT[ CNIQTKVJO CICKPUV CTVKENGU QH C J[DTKF

%JKPGUG'PINKUJ EQTRWU IGPGTCVGF D[ OKZKPI VJG %JKPGUG YGDFQEWOGPVU YKVJ VJG

'PINKUJ YGDFQEWOGPVU HTQO VJG UGNGEVGF DKNKPIWCN OCIC\KPG CTVKENGU KP VJG 5KPQTCOC

/CIC\KPG +P VJG EQTRWU GCEJ DKNKPIWCN FQEWOGPV RCKT KU URNKV KPVQ VYQ FQEWOGPVU CP

'PINKUJ CPF C %JKPGUG FQEWOGPV +P VJKU ECUG GCEJ VTCKPKPI FQEWOGPV KP GKVJGT %JK

PGUG QT 'PINKUJ KP VJG EQTRWU KU FGEQPUVTWEVGF KPVQ C DCI QH %JKPGUG CPF 'PINKUJ

66

VGTOU TGURGEVKXGN[ #NVJQWIJ VJG FQEWOGPV RCKTU EQPVCKPGF KP VJG RCTCNNGN EQTRWU CTG

VTCPUNCVKQPGSWKXCNGPV YG TGICTF GCEJ QPG QH VJG FQEWOGPV RCKT CU CP KPFGRGPFGPV

VTCKPKPI FQEWOGPV 6JG FGXGNQRGF 51/ OGVJQF VTGCVU VJKU FQEWOGPV CU C DCI QH

HTGGN[ KPVGTOKPINGF %JKPGUG CPF 'PINKUJ VGTOU 6JG %JKPGUGVGZV RCTV QH VJG EQTRWU

YCU HKTUV RTGRTQEGUUGF D[ VJG YQTF UGIOGPVCVKQP RTQITCO CU KP RTGXKQWU GZRGTKOGPVU

(QT 'PINKUJ FQEWOGPVU VJG FQEWOGPV RTGRTQEGUUKPI UVKNN DGIKPU YKVJ VGTO KPFGZKPI

GUVCDNKUJKPI UVQR NKUVU UVGOOKPI CPF VJGP GPEQFKPI FQEWOGPVU YKVJ DKPCT[ XGEVQTU

KP YJKEJ GCEJ EQORQPGPV EQTTGURQPFU VQ C FKHHGTGPV YQTF CPF VJG XCNWG QH VJG EQO

RQPGPV TGHNGEVU VJG RTGUGPEG QH YQTF KP VJG FQEWOGPV

9G EQPUVTWEVGF UGNHQTICPK\KPI OCRU VJCV EQPVCKP PGWTQPU KP ª ITKF HQTOCV

VQ EQPFWEV GZRGTKOGPVU DCUGF QP C UOCNN EQNNGEVKQP QH VJG EQTRQTC %JKPGUG CTVKENGU

CPF 'PINKUJ CTVKENGU 6JG TGUWNVKPI OCRU IKXG VJG KORTGUUKQP VJCV VJG GHHGEVKXGPGUU

QH VGZV ENWUVGTKPI KU SWKVG UKIPKHKECPV KP C OKZVWTG QH 'PINKUJ CPF %JKPGUG GZRGTKOGP

VCN CTVKENGU #P GZCORNG QH TGUWNVKPI YQTF ENWUVGTU HTQO VJG VTCKPGF YQTF ENWUVGT OCR

KU UJQYP KP 6CDNG

6CDNG #P GZCORNG QH TGUWNVKPI YQTF ENWUVGTU HTQO VJG VTCKPGF YQTF ENWUVGT OCR

5[UVGO (TCOGYQTM

6JG U[UVGO HTCOGYQTM KU UJQYP KP (KI $[ CRRN[KPI VJG 51/ PGWTCNPGVYQTM

VGEJPKSWG YG GZVTCEV UGOCPVKEU HTQO VJG EQPVGPVU QH C JWIG *6/.FQEWOGPV EQNNGE

VKQP WUKPI VJG FGXGNQRGF OWNVKNKPIWCN VGZVOKPKPI CNIQTKVJO =?=? 6JG GZVTCEVGF

UGOCPVKEU ECP DG WUGF VQ RTQFWEG OGVCFCVC CPF QPVQNQI[ VQ FGUETKDG VJG KPHQTOCVKQP

KP VJG QTKIKPCN YGD FQEWOGPVU 6JGTGHQTG VJG QTKIKPCN YGD FQEWOGPVU ECP DG VTCPU

HQTOGF KPVQ :/. FQEWOGPVU CPF UVQTGF KP VJG :/. FQEWOGPV FCVCDCUG YKVJ VJG

FGXGNQRGF QPVQNQI[ HQT VGZV ECVGIQTK\CVKQP #U UWEJ VJG GZKUVKPI YGD RCIGU ECP DG

CPCN[\GF VQ IGPGTCVG C UGV QH OGVCFCVC VQ FGUETKDG VJGKT EQPVGPV CPF RTQFWEG CP QPVQN

QI[ HQT VJG :/. FQEWOGPV FCVCDCUG VJTQWIJ C VGZVOKPKPI VGEJPKSWG DCUGF QP VJGKT

UGOCPVKE TGNCVGFPGUU

VLQRUDPD ` ` DCF DTKFI ECWIJV EJKNFNKM

EJKPILW EJTKUUK EQOOWP EQORTGJGPU EQPVTKDWV EQWPVT[Y

EWNVKX EWTKQU FTQXG GCUKNK GPFNGUUNK GXGPV GXGPVW HWNHKN

IQCN ITGCVGUV JKIJGUV KORCTV KPGZJCWUV KPHGTKQT LKCHQPI

LQK OCIC\KP OKUUKQP OQFGN OQFGUVK RNCPV RQVGPVK

RTQDNGO RTQHGUUKQP RWTUW TGEQTF TGRTGU TGURQPU UEJQNCT

UGPU UGTX URCTM URGEKCNKUV VTCPUOKV YCK YCPI YQPFGT 3

m33lýlÍl°ptÚ] ! 3ª >1

G3«Þ¥ÿÏzçu Nâ !Ô a

IØWõõ^¥Å@³£Ú

F1nEïÇá¦üZ<)

W¢SªëÊi7êh:W C Z «

E¯Zé"³@PíRA

67

(KI 5[UVGO (TCOGYQTM

4GNCVGF 9QTM

6GZV OKPKPI KU C PGY KPVGTFKUEKRNKPCT[ HKGNF +V EQODKPGU VJG FKUEKRNKPGU QH FCVC OKP

KPI KPHQTOCVKQP GZVTCEVKQP KPHQTOCVKQP TGVTKGXCN VGZV ECVGIQTK\CVKQP OCEJKPG NGCTP

KPI CPF EQORWVCVKQPCN NKPIWKUVKEU VQ FKUEQXGT UVTWEVWTG RCVVGTPU CPF MPQYNGFIG KP

NCTIG VGZVWCN EQTRQTC 9KVJ VJG JWIG COQWPV QH KPHQTOCVKQP CXCKNCDNG QPNKPG VJG

9QTNF 9KFG 9GD JCU DGGP DGEQOKPI C HGTVKNG CTGC HQT VGZV OKPKPI TGUGCTEJ 6JG 9GD

EQPVGPV FCVC KPENWFG WPUVTWEVWTGF FCVC UWEJ CU HTGG VGZVU UGOKUVTWEVWTGF FCVC UWEJ CU

*6/. FQEWOGPVU CPF C OQTG UVTWEVWTGF FCVC UWEJ CU VCDWNCT FCVC KP VJG FCVCDCUGU

*QYGXGT OWEJ QH VJG 9GD EQPVGPV FCVC KU WPUVTWEVWTGF VGZV FCVC 6JG TGUGCTEJ

CTQWPF CRRN[KPI MPQYNGFIG FKUEQXGT[ VGEJPKSWGU VQ WPUVTWEVWTGF VGZV KU VGTOGF

MPQYNGFIG FKUEQXGT[ KP VGZVU -&6 =? QT VGZV FCVC OKPKPI =? QT VGZV OKPKPI #F

XCPEGU KP EQORWVCVKQPCN TGUQWTEGU CPF PGY UVCVKUVKECN CNIQTKVJOU HQT VGZV CPCN[UKU

JCXG JGNRGF VGZV OKPKPI FGXGNQR CU C HKGNF 4GEGPVN[ VJGTG JCXG DGGP UQOG KPPQXCVKXG

VGEJPKSWGU FGXGNQRGF HQT VGZV OKPKPI (QT GZCORNG (GNFOCP WUGU VGZV ECVGIQT[ NCDGNU

CUUQEKCVGF YKVJ 4GWVGTU PGYUYKTG VQ HKPF WPGZRGEVGF RCVVGTPU COQPI VGZV CTVKENGU

=?=?=?=? 6GZV OKPKPI D[ WUKPI UGNHQTICPK\KPI OCR 51/ VGEJPKSWGU JCU CNTGCF[

ICKPGF UQOG CVVGPVKQPU KP VJG MPQYNGFIG FKUEQXGT[ TGUGCTEJ CPF KPHQTOCVKQP TGVTKGXCN

HKGNF 6JG RCRGT QH =? RGTJCRU OCTMU VJG HKTUV CVVGORV VQ WVKNK\G 51/ WPUWRGTXKUGF

PGWTCN PGVYQTMU HQT CP KPHQTOCVKQP TGVTKGXCN YQTM +P VJKU RCRGT JQYGXGT VJG FQEW

OGPV TGRTGUGPVCVKQP KU OCFG HTQO OCPWCNN[ UGNGEVGF KPFGZGF VGTOU CPF KU VJWU PQV

TGCNN[ TGCNKUVKE +P CFFKVKQP COQPI VJG OQUV KPHNWGPVKCN YQTM YG EGTVCKPN[ JCXG VQ

OGPVKQP 9'$51/ =?=?=? 6JGKT YQTM CKOU CV EQPUVTWEVKPI OGVJQFU HQT GZRNQT

KPI HWNNVGZV FQEWOGPV EQNNGEVKQPU VJG 9'$51/ UVCTVGF HTQO *QPMGNCU UWIIGUVKQP

QH WUKPI VJG UGNHQTICPK\KPI UGOCPVKE OCRU =? CU C RTGRTQEGUUKPI UVCIG HQT GPEQFKPI

FQEWOGPVU 5WEJ OCRU CTG KP VWTP WUGF VQ CWVQOCVKECNN[ QTICPK\G KG ENWUVGT

$KNKPIWCNEQTRQTC

6GZVOKPKPI

5GOCPVKEU DCUGFVGZV ECVGIQTK\CVKQP

XML parser

XML document repository

68

FQEWOGPVU CEEQTFKPI VQ VJG YQTFU VJCV VJG[ EQPVCKP 9JGP VJG FQEWOGPVU CTG QTICP

K\GF HQNNQYKPI VJG UVGRU KP VJG RTGRTQEGUUKPI UVCIG QP C OCR KP UWEJ C YC[ VJCV

PGCTD[ NQECVKQPU EQPVCKP UKOKNCT FQEWOGPVU GZRNQTCVKQP QH VJG EQNNGEVKQP KU HCEKNKVCVGF

D[ VJG KPVWKVKXG PGKIJDQTJQQF TGNCVKQPU 6JWU WUGTU ECP GCUKN[ PCXKICVG C YQTF ECVG

IQT[ OCR CPF \QQO KP QP ITQWRU QH FQEWOGPVU TGNCVGF VQ C URGEKHKE ITQWR QH YQTFU

&KHHGTKPI HTQO CDQXG VTCFKVKQPCN WVKNK\CVKQP QH 5GNH1TICPK\KPI /CRU KP VGZV OKPKPI

KP VJKU RTQLGEV KPENWFKPI RTGXKQWU YQTM =?=?=?=?=? YG FKF PQV GORNQ[ VJG

OCR TGRTGUGPVCVKQP HQT XKUWCNK\CVKQP 4CVJGT YG RTQXKFG C EQPEGRVDCUGF SWGT[

OGVJQF HQT FQEWOGPV TGVTKGXCN HTQO VJG FCVCDCUG KP QTFGT VQ GHHGEVKXGN[ KORTQXG VJG

FKHHKEWNVKGU KP PCXKICVKPI VJG OCRU HQT VGZV UGCTEJ 1WT U[UVGO CNNQYU WUGT VQ RGTHQTO

C MG[YQTFDCUGF UGCTEJ UKOKNCT VQ VJCV QH QVJGT UGCTEJ GPIKPGU 1PEG WUGT KPRWV C

MG[YQTF VJG KPRWV VGTO YCU WUGF CU CP GPVT[ RQKPV KPVQ VJG IGPGTCVGF YQTF ENWUVGT

CPF VJG UGCTEJKPI VQQN YQWNF TGVWTP C NKUV QH UWIIGUVGF FQEWOGPVU VJCV YGTG TGNCVGF VQ

VJG KPRWV VGTO CEEQTFKPI VQ VJG 51/ ENWUVGTKPI CNIQTKVJO 9KVJ RCTVKEWNCT GORJCUGU

QP RTQEGUUKPI OWNVKNKPIWCN KPHQTOCVKQP UQWTEGU VJG FGXGNQRGF 51/ VGEJPKSWG RGT

HQTOU C NCPIWCIGPGWVTCN OGVJQF VQ VCEMNG VJG NKPIWKUVKEU FKHHKEWNVKGU KP VJG VGZV OKP

KPI RTQEGUU 6JG CNIQTKVJO YCU RTGUGPVGF KP FGVCKN KP QWT RTGXKQWU YQTM

=?=?=?=?=?

%QPENWUKQPU

5VKNN KP VJKU ECUG VGZV OKPKPI HTQO VJG QJ[DTKF %JKPGUG'PINKUJ EQTRWUR VGZV EQNNGE

VKQP CEVWCNN[ TGRTGUGPVU V[RKECN UKOKNCTKV[ TCVJGT VJCP VTCPUNCVKQP GSWKXCNGPEG *QY

GXGT KP VJKU ECUG VGTOU KPENWFKPI DQVJ %JKPGUG CPF 'PINKUJ QPGU YJKEJ CTG EQPEGRVW

CNN[ UKOKNCT CTG ITQWRGF KP C PGWTQP CPF PGKIJDQTKPI PGWTQPU +P CFFKVKQP CP KPVGTGUV

KPI QWVEQOG KU VJCV VJG TGUWNVKPI ENWUVGT ITQWRU YQTFU YKVJ UKOKNCT EQPEGRVU YJKEJ CTG

TGRTGUGPVGF GKVJGT KP 'PINKUJ QT %JKPGUG 6JKU KORNKGU VJG FGXGNQRGF OKPKPI CNIQ

TKVJO KU PGWVTCN VQ NCPIWCIGU KG 'PINKUJ CPF %JKPGUG WUGF KP VJG GZRGTKOGPVKPI

FQEWOGPVU 6JKU RQVGPVKCN RTQXKFGU RQUUKDKNKVKGU VQ GUVCDNKUJ CP GHHGEVKXG YC[ HQT EQP

EGRV GZVTCEVKQP HTQO OWNVKNKPIWCN KPHQTOCVKQP UQWTEGU (WTVJGTOQTG VJG TGUWNVKPI

VGTO ENWUVGTU ECP DG WUGF VQ GUVCDNKUJ C UVTWEVWTG QH EQPEGRVU QT CP QPVQNQI[ 6JG

QPVQNQI[ KU QHVGP NCPIWCIGKPFGRGPFGPV QT CV NGCUV NCPIWCIGPGWVTCN UQ VJCV KV ECP DG

WUGF KP OWNVKNKPIWCN 5GOCPVKE 9GD CRRNKECVKQPU

#EMPQYNGFIGOGPV 6JKU TGUGCTEJ JCU DGGP RCTVKCNN[ UWRRQTVGF D[ VJG 0CVKQPCN 5EK

GPEG %QWPEKN KP 6CKYCP 05% '

4GHGTGPEGU

'DJDQ , )HOGPDQ 5 DQG +LUVK +³.H\ZRUG%DVHG %URZVLQJ DQG $QDO\VLV RI /DUJH

'RFXPHQW 6HWV´ 3URF RI WKH )LIWK $QQXDO 6\PSRVLXP RQ 'RFXPHQW $QDO\VLV DQG ,QIRUPD

WLRQ 5HWULHYDO 6'$,5 /DV 9HJDV 19

&GGTYGUVGT 5 &WOCKU 5 (WTPCU ) CPF .CPFCWGT - Q+PFGZKPI D[ .CVGPV 5GOCPVKE

#PCN[UKUR ,QWTPCN QH #OGTKECP 5QEKGV[ HQT +PHQTOCVKQP 5EKGPEG

69

)HOGPDQ 5 DQG 'DJDQ , ³.'7 .QRZOHGJH 'LVFRYHU\ LQ 7H[WV´ 3URF RI WKH )LUVW $Q

QXDO &RQIHUHQFH RQ .QRZOHGJH 'LVFRYHU\ DQG 'DWD 0LQLQJ .'' 0RQWUHDO

)HOGPDQ 5 .ORVJHQ : DQG =LOEHUVWHLQ $ ³9LVXDOL]DWLRQ 7HFKQLTXHV WR ([SORUH 'DWD

0LQLQJ 5HVXOWV IRU 'RFXPHQW &ROOHFWLRQV´ 3URF RI WKH 7KLUG $QQXDO &RQIHUHQFH RQ

.QRZOHGJH 'LVFRYHU\ DQG 'DWD 0LQLQJ .'' 1HZSRUW %HDFK

(GNFOCP 4 &CICP + CPF *KTUJ * Q/KPKPI 6GZV 7UKPI -G[YQTF &KUVTKDWVKQPUR , QH

+PVGNNKIGPV +PHQTOCVKQP 5[UVGOU 8QN

*GCTUV /# Q7PVCPINKPI 6GZV &CVC /KPKPIR 2TQEGGFKPIU QH #%.O VJG VJ #PPWCN

/GGVKPI QH #UUQEKCVKQP HQT %QORWVCVKQPCN .KPIWKUVKEU

+RQNHOD 7 .DVNL 6 /DJXV . DQG .RKRQHQ 7 ³1HZVJURXS ([SORUDWLRQ ZLWK :(%620

0HWKRG DQG %URZVLQJ ,QWHUIDFH´ 7HFKQLFDO 5HSRUW $ +HOVLQNL 8QLYHUVLW\ RI 7HFKQRO

RJ\ /DERUDWRU\ RI &RPSXWHU DQG ,QIRUPDWLRQ 6FLHQFH (VSRR )LQODQG

.DVNL 6 +RQNHOD 7 /DJXV . DQG .RKRQHQ 7 ³:(%6206HOI2UJDQL]LQJ 0DSV RI

'RFXPHQW &ROOHFWLRQV 1HXURFRPSXWLQJ´ 9RO

.RKRQHQ 7 ³6HOI2UJDQL]LQJ )RUPDWLRQ RI 7RSRORJLFDOO\ &RUUHFW )HDWXUH 0DSV´ %LRORJL

FDO &\EHUQHWLFV 9RO SS

-QJQPGP 6 5GNH1TICPK\KPI /CRU 5RTKPIGT8GTNCI $GTNKP

-QJQPGP 6 Q5GNH1TICPK\CVKQP QH 8GT[ .CTIG &QEWOGPV %QNNGEVKQPU 5VCVG QH VJG #TVR +P

0KMNCUUQP . $QFGP / CPF <KGOMG 6 GFKVQTU 2TQE QH +%#00 VJG VJ +PVGTPCVKQPCN

%QPHGTGPEG QP #TVKHKEKCN 0GWTCN 0GVYQTMU 8QN .QPFQP 5RTKPIGT

.GG %* CPF ;CPI *% Q6QYCTFU /WNVKNKPIWCN +PHQTOCVKQP &KUEQXGT[ VJTQWIJ C 51/

DCUGF 6GZV /KPKPI #RRTQCEJR +P 2TQEGGFKPIU QH +PVGTPCVKQPCN 9QTMUJQR QP 6GZV CPF 9GD

/KPKPI 6JG 5KZVJ 2CEKHKE 4KO +PVGTPCVKQPCN %QPHGTGPEG QP #TVKHKEKCN +PVGNNKIGPEG 24+%#+

/GNDQWTPG #WUVTCNKC #WIWUV 5GRVGODGT

.GG %* CPF ;CPI *% Q# /WNVKNKPIWCN 6GZV /KPKPI #RRTQCEJ $CUGF QP 5GNH

1TICPK\KPI /CRUR +P #RRNKGF +PVGNNKIGPEG 6JG +PVGTPCVKQPCN ,QWTPCN QH #TVKHKEKCN +PVGNNK

IGPEG 0GWTCN 0GVYQTMU CPF %QORNGZ 2TQDNGO5QNXKPI 6GEJPQNQIKGU 5RGEKCN +UUWG QP

6GZV CPF 9GD /KPKPI 5%+'+ +P 2TGUU

/HH &+ DQG <DQJ +& ³$ :HE 7H[W 0LQLQJ $SSURDFK %DVHG RQ 6HOI2UJDQL]LQJ 0DS´

3URF RI WKH $&0 &,.0 QG :RUNVKRS RQ :HE ,QIRUPDWLRQ DQG 'DWD 0DQDJHPHQW

:,'0 .DQVDV &LW\ 0LVVRXUL 86$

/HH &+ DQG <DQJ +& ³$ 7H[W 'DWD 0LQLQJ $SSURDFK 8VLQJ D &KLQHVH &RUSXV %DVHG

RQ 6HOI2UJDQL]LQJ 0DS´ 3URF RI WKH )RXUWK ,QWHUQDWLRQDO :RUNVKRS RQ ,QIRUPDWLRQ 5H

WULHYDO ZLWK $VLDQ /DQJXDJH ,5$/ 7DLSHL 7DLZDQ

.KP : 5QGTIGN & CPF /CTEJKQPKPK ) Q# 5GNH1TICPK\KPI 5GOCPVKE /CR HQT +PHQTOCVKQP

4GVTKGXCNR 2TQE QH VJG #%/ 5+)+4 +PVON %QPH QP 4GUGCTEJ CPF &GXGNQROGPV KP +PHQTOC

VKQP 4GVTKGXCN 5+)+4O %JKECIQ +.

2KEEJK ' CPF 2GVGTU % Q%TQUU.CPIWCIG +PHQTOCVKQP 4GVTKGXCN # 5[UVGO HQT %QORCTCDNG

%QTRWU 3WGT[KPIR +P %TQUU.CPIWCIG +PHQTOCVKQP 4GVTKGXCN GFU D[ )TGHGPUVGVVG )

-NWYGT #ECFGOKE 2WDNKUJGTU

4KVVGT * CPF -QJQPGP 6 Q5GNH1TICPK\KPI 5GOCPVKE /CRUR $KQNQIKECN %[DGTPGVKEU 8QN

5CNVQP ) CPF /E)KNN /, +PVTQFWEVKQP VQ /QFGTP +PHQTOCVKQP 4GVTKGXCN /E)TCY*KNN

$QQM %QORCP[ 0GY ;QTM

<DQJ +& DQG /HH &+ ³$XWRPDWLF &DWHJRU\ *HQHUDWLRQ IRU 7H[W 'RFXPHQWV E\

6HOIRUJDQL]LQJ 0DSV´ ,Q 3URFHHGLQJV RI WKH ,(((,116(116 ,QWHUQDWLRQDO -RLQW &RQ

IHUHQFH RQ 1HXUDO 1HWZRUNV ,-&11 &RPR ,WDO\ -XO\ 9RO ,,,

;CPI *% CPF .GG %* Q#WVQOCVKE %CVGIQT[ 5VTWEVWTG )GPGTCVKQP CPF %CVGIQTK

\CVKQP QH %JKPGUG 6GZV &QEWOGPVUR +P 6JG (QWTVJ 'WTQRGCP %QPHGTGPEG QP 2TKPEKRNGU CPF

2TCEVKEG QH -PQYNGFIG &KUEQXGT[ KP &CVCDCUGU 2-&& .[QP (TCPEG 5GR

70

!"#$%$!&(')* + ,&-#$ .+/ 012#43 05

67897;:<>=@?BA-:C=ED;FEG"HEIEIJAK:L0L0DM89:CA-IN=+G":PORQ2IJ7;D9SKFJAKIJ=+LTRHEI6 :P77;U7

VBWPXZY[]\_^*WP`-\a;bdceaf^*Xhgh\_ij`lknmKoCijWC`hoCWpij`hkhq rsceaftjtjWCkfW

uv`hijwxWP[_r]iy\z!a;b|s~@WP[MKWCWC`|>~@WP[MhWCWP`Nh|v#!uv

m-oCa;\_tY;`l ;x9f;ff;;

-;fK;x-Nxx-PK--J]f-;-;x ]¡xfd ¢£ ¤x¥

¦§¨C©Cª«Z¬;©f*®¯ Wm-WC^Y;`\_ijon°+WC~ijr*Y±w-ijr]ijaf`²a;bvY±^Y;o ¯ ij`lW*[_WY;lY;~htjW!°³WC~abJ[_WPr]afgh[_oPWCrCij`\_WP[_tji`h´xWnY`Z"oPaf`l`hWCoP\_W0\ ¯ [_afghk ¯ ^*W9\MYµ¶lY\MYs· iy\ ¯ oCaf^*^*a;à;`-\_a;tjafkfijWCrÇd¹`\ ¯ irºXZYX@WP["·eW2WP»KXltja[_W\ ¯ W!ij^*XZYoP\"r]gho ¯ Y¼mKWC^Y`\_io*°³WC~·eafglt ¯ YwfWaf`¾½Y;o ¯ ij`hWº¿JWY[_`hij`lk!Y;tjkfa[_iy\ ¯ ^*r>glr]WbRa;[vglr]WP[sXh[_aÀZtjij`lkY;`lX@WP[_r]a;`ZYtijr_Y\_ia;`N¸JÁvgK[ ¯ zKX@a;\ ¯ WPr]ireir#\ ¯ Y\#tjWY[_`lij`lk"bÂ[_af^Ã\ ¯ WmKWP^Y;`\_ijo>°³WC~r ¯ afgltafgK\_X@WP[]bÄa[_^Å\][MY;hiy\_ijaf`ZYt#tjWY[_`hi`hk¾bÂ[_af^Æ\_a-lYCzEq r0°+a[_tÂ+°ÇiKW!°³WC~bRa;[È~@a;\ ¯ X@WP[]bRa;[_^Y`loCW¼Y;`lÉYoCoCgK[MY;oPz¸ ¹¶`)\ ¯ ijrXZYX@WP[2·eW³Xh[_WPr]WC`\2[_WCr]ghtj\_ra;~h\MY;ij`hW)· iy\ ¯ \·eaÊKiyËEWP[_WC`\±hY\MY;r]W9\_r^Y[_´xWCKµÌglXÊ· iy\ ¯ r]WC^Y;`\_ijo¾^*WP\MYµhY\MYKÍKghr]i`hkº\ ¯ WCr]Wv·W ¯ YwxWvij`-wxWCr¢\_ijkfY\_WhiyËEWP[_WC`\#ij`lr¢\MY`loCWs[_WCXh[_WCr]WP`-\MY\_ijaf`lrY`Znw;Y[_ijafghretjWY[_`hij`lk\_WCo ¯ `liÎglWPrÇÁvgK[eij`liy\_iY;th[_WCr]glty\_rd· iy\ ¯ \ ¯ W>ÏvYPÐ ÑjwxWsYCzWCrY`ZpBµ¶ÏsÏ)Y;tjkfa[_iy\ ¯ ^*re·eWP[_WBKir_YXlX@a;i`\_ij`lkKKtWCYfhij`hkºghre\_aºW9»lY^*ij`lW#\ ¯ Wsglr]Wab\ ¯ W0Ò[_a;kfaftÓY;tjkfa[_ij\ ¯ ^¸@uvr]ij`hk*¹¢¿£ÒÔ\_WCo ¯ `liÎglWCr ·W·eWP[_W"Y~ltjW\_ahijr]oCawxWP[^*WY`lij`lkbÄght£Y`Z!·eW~@WCtjijWCwxWflX@a;\_WC`\_iY;tjtyz[_WPglr_Y;~htjW´K`ha· tWChkfW;¸

Õ |sgK\ ¯ a;[#\_aº· ¯ af^ÖYtjt£oCa[][_WCr]X@af`lhWC`hoCWr ¯ a;glt~@WY;lK[_WPr]r]WJ¸

71

Ø ÙlÚ>Û@ÜEÝÞºß0à£ÛEá_Ý>Ú

â£ãJä*åäæ±çèNé¶êÌëì+äfí4î ïðñ0OÄDBA!òZOÄDMORóKIÈORI³?ôEORUCô¾89ó@=EAxõÓö Ds÷Ê7ø³?ORTÂTÓø£7"7ùZ897IN=@7;=¾?O8PôQAhUCôEORIE7¼:P7;Ah=EA-øJTÂ7³UóKIl8P7Il8;F#A-IN=)?ôE7;:97³7òh7:PõÊ:P7;D9óKHJ:PU7È?OÂTRTø£7¼Q±A-:PúK7f=Zû¢HEü)HJD9OÂIJýQAhUCôEORIE7:P7;AK=JA-øETR7*Q78PA-û¢=JA8PAJþJÿBôE7ORIl897;Ih8POÂóhI²OÄD89ôNA8º=@ó@UHEQ7Il8PDºóKI+8PôE7Z7;QAKIl89OÄU÷ 7;ø±?OÂTRT£UóKIZòK7;õ:P7;A-TJQ27fA-IEORIEý*øZõ2HJDMORIEýDM89:PHJU8PHE:P7;=±=EA-8PAûóK:PQ±A8CD#A-IJ=øZõ2:977:P:9ORIEý89óUóhQQ2óhIóKIl89óhTÂóhýKOR7;D;þ÷ 7ø£7TROÂ7;òK78PôJA8 ORIEOÂ89OÄA-TRTÂõ89ôJ7Z7Q±A-Il8PORU÷Ê7ø±?ORTRTJUóhIJDMOÄDM8 óôJA-IJ=@û¢U:PA 897;=2üJAKýK7fD#Q!HJUCô±TRORúK78PôE70÷Ê7ø±?v7úZIEó?89ó@=EAxõKFhüE:PóòlOÄ=@ORIEýn89ôJ70D9AKQ7OÂIóK:9ûQA-89ORóKIFhøEH@8BOÂI±Q±AhUCôEOÂIJ7:97fAK=EAKøETR7óK:PQ+þEóh:>7ù@AKQüETÂ7hFK?v70UóhHETÄ=±7IZòZORDPA-ýh7DM7;Q±A-Il89OÄUQAK:9úZHEüÊAKU;UóKQüJAKIZõlORIEý¾A³UóhIZòK7Il8POÂóhIJA-T0ÿ üJAKýK7!ýhOÂòZORIEý³OÂIóh:9Q±A-89ORóKI D9HJUCôÊAKDâ£ã@ê*êºéãEä>äfíJçlä @èNè£ç!"!ê æ2è£ä#%$dãJä&'!)('*#'!,+níCä#!%-KäCäè/.eèJê10ä#!2ê é34$dãZêé¢ä5Âä6Eã-è£ä¼è7 @æ±íCä#!ê98'::;=<'>@?BA>CÇäé]ëD÷Ê7ÈøN7;TÂOR7òh7289ôJA-8!DMHJUCôÇDM8PA-89OÄUOÂIóh:9Q±A-89ORóKI=@7QóKIND_8P:PA-897;D2óhIETÂõÇüNA-:982óB089ôJ7³ü£ó-8P7Il89OÄA-T"óh:EZ7Q±A-Il8PORU³÷Ê7ø897;UCôJIEóKTRóKýhOÂ7fDFd89ôE7;OÂ:=@7üETRóõZQ7Il8D9ôEóhHETR=A-TRTÂó?Góh:±AK=@òA-INU7;=ÉüE:PóHJTROÂIJýÔQ78PôEó@=ED±UAKüJA-øETR7¾óBºAKUIlHEOÂ:PORIEýúlIJó?TÂ7f=@ýK7D9HJUCô±AKDJ" @èJèÓç'!K5êL(Kä#"íCçèM-4dã7!PäPë2!)-Kä2-æN'é*0éãEäê1!0æ±çé]ä!êÌç'542!)æ8O@<C+éPQ8BORSA'$0ç4>ä515ç'2çèT3³íCçèM-UdãV WäXç9YZS)³å[3èJéãEä#êä!DÇ÷ ôJ7IÔA8P:9HJ7Z7Q±A-Il8PORU±÷Ê7øÉ7ù@ORDM8PDORI 89ôEOÄD\óh:9Q OÂ8ø£7;UóhQ7;DHJDM7HETFeAKIJ=ÇDMôJóKHETÄ=F]óh:!ORIJDM8PA-INU7KFA-TRTÂó?^óK:Èø£789897;:¾Q±A8CUCôEOÂIJýÇóDM7;Q±A-Il89OÄUAKTÂTRõÉ7;IE:9OÄUCôE7f=üJ:9ó@=@HJU8¼=@7fD9U:9ORü@89ORóKIND?O8PôDM7;QAKIl89OÄUºHNDM7;:üE:9óBHJTR7;D;þ

Y²çhëPã@ê è£ä,5ÂäCç'!èNê è@¾897fUCôEIEóhTÂóhýKOR7;DdôJAxòK7søN7;7IA-üEüETROR7;=ORI8PôE7UóhIh8P7ùZ8óE8PóZ=JAxõ£ö Dd÷ óh:9TÄ=÷ OR=@7¼÷ 7;øÇ89óÔôE7TRüÇHJD97:CDHJIJ=Ç89ôE7;OÂ:?vAxõÔ89ôE:PóKHEýhô 89ôE7¾HEIEQ±A-IJAKýK7fA-øETR7ÈA-QóKHJIh8óOÂIóh:9Q±A-89ORóKI89ôJA-8È7ù@ORDM8PD;þ>L 8_õZüEOÄUAKTD9U7IJAK:9ORó ORIZòKóhTÂòh7;DAKUIhHJORD9O8POÂóhIóBnAÊQó@=@7T0óAHJD97:fö D!OÂIl897;:97fD_8CD?ôEOÄUCôÉU;A-I 89ôJ7I)ø£7¼HNDM7f=Ê89óÔQ±AKúK7È:P7;UóKQQ7IJ=EA-89ORóKIJD;F7hþ ýNþséãZê5jê èW(E9ã 5L-¼íCä_*ê èNé¢ä!9ä#éÇóK:!ë2è7Cê`-hä#!!é ã@ê"[!)S-' Jëé`$ ê éBêaê æê15Rç'!é6±éãZê èW'\3B cb 0xäí% Bfã@éM[!Pä#0fê`' WX5d3±þ@L òxAK:9OR78_õóBdA-üEüJ:9ólAKUCôE7fD#7ù@OÄD_8óK:vTÂ7fA-:PIEORIEý\:PóKQ ÷Ê7ø@û]UóhIh8P7Il8F89ôE7fDM7:PAKIEýK7a:PóKQ Q789ôJóZ=JD0?ôEOÄUCô²UCôEóZóhD97n89ó¼OÂýhIEóK:P7A-TRTeºÿ )Q±AK:9úZHEü²AKIJ=³89:P7;A-87òK7;:9õl8PôEOÂIJý AKDüETÄA-ORI)8P7ùZ8;Fe89ó 89ôJóhD97¾?ôEOÄUCôÉQ±A-úh7¾HJD97³ó8PôE7¼TRORQ2OÂ897f=ÉDM89:PHJU8PHE:P7¼ORI0ÿ ²AKIJ=289:P7;A-8 89ôJ789OÂ89TR7KFlôE7fAK=@ORIEýJFKTÂORIEúlû¶8P7ùZ8PDs=@OLf£7;:97;Ih8PTÂõhþgºINU7UóhIl897Il8>ôNAKD ø£77;I7ùZ89:CAKU897;=U:PóKQ =@ó@UHEQ7Il8CDF@8PôE7IE7ùZ80D_8P7üOÄDB89óÈAKüEüETRõ¾ORIóK:PQ±A89ORóKI+:P78P:9OR7òA-TÓ8P7;UCô@ûIEOhIlHE7fDFNDMHJUCô+AKDDM89óhüZ?sóh:P=¾:97;Q2óòAKT¶FEDM897;Q2QORIEýJFE897:PQÆ?v7ORýKôl8POÂIEý±AKIJ=¼D9óóKIþNLøJA-ýKûó û¢?sóh:P=ED:97;üE:97fDM7;Il8PA8POÂóhI¼OÄD089ôE7;I²HJD97;=³89óEóK:PQÆ8PôE78P:PAKOÂIEORIEýÈORIJDM8PAKIJU7fD:P7IlHEOR:97f=¼øZõ89ôE7*TR7;AK:9IEORIEý±A-TRýKóh:9OÂ89ôJQ³þeORýKHE:P7Ni"üJ:9óòZOÄ=@7;DBA±DPUCôE7Q±A8PORU"òZOR7?(óB8PôEORDüE:9ó@U7;DPDþ

×]In89ôE7 ?vóK:PúüE:P7;D97Il8P7;="ôE7:P7 ?v7#?ORD9ôn89ó07ù@üETRóK:P789ôE7 ORQüJAKU8óZ8PôE7jZ7;QAKIl89OÄU÷ 7;øóKI2HJDM7;:üE:9óBHJTROÂIEý*AKIJ=!ü£7:CD9óKIJAKTÂOÄD9A-89ORóKI]FfQóK:P7vD9ü£7;UOLHNU;A-TRTÂõhF?s7BôJAxòh7sORIZòK7;DM89ORýhA-897f=ôEó?QAhUCôEORIE7û¢TÂ7fA-:PIEORIEý*897fUCôEIEOIlHE7;DvUóhHETR=ÈøN7"HJD97;=ÈOL?v70ôJAh=¾AhUU7fD9D 89óDM7;Q±A-Il89OÄUQ±A-:PúZHEüóK:n7òh7:Põ÷ 7;ø :P7;D9óKHE:CU7hþkgºHJ:*ôlõZü£ó-8PôE7;D9ORD"OÄD"89ôJA-8n89ôE7EZ7;QAKIl89OÄU÷Ê7ø D9ôEóKHETÄ= ôE7TRüDMóhTÂòh7NHEIJ=EAKQ27;Il8PA-TvüE:PóKøETR7Q±D89ôJA-8!Q±A-úK7¾QAhUCôEORIE7ÈTR7;A-:PIEORIEýl:PóKQ 89ôE7¼÷ 7;øÇ89ó@=EAxõ=@OnmÈUHET8fFøZõ*üE:9óòZOÄ=@ORIEýºDM89:PHJU8PHE:P7;=ORIóh:9Q±A8POÂóhIFf:97f=@HJUORIEýnA-QøEOÂýhHEOÂ8_õKFxA-IJ=!üE:9óòZOÄ=@ORIEýHJDM7HETZ:977:P7IJU7;DÓ8Pó0øJAhUCúZýK:PóKHEIJ=nORIóh:9Q±A8POÂóhI*OÂI89ôE7jóK:PQ óBJóKIl8PóKTRóKýKOR7;D;þ÷Ê7sD9HEýKýh7;DM889ôJA-8TÂ7fA-:PIEORIEý":9óhQÃD97Q±AKIh8PORU;A-TRTÂõºQ±A-:PúK7;=@ûHEü*=EA-8PAD9ôEóKHJTR=nóhH@89ü£7:)óK:PQ(TR7;AK:9IEORIEý":PóKQHEIJDM89:PHJU8PHE:97f= óh:³D97QOÂû¢DM89:PHJU8PHE:P7;= 8P7ùZ8;FB?OÂ89ôÃ:97;ýhAK:P= 8PóORIJU:P7;AhDM7f= AKU;UHE:CAKUõKFvO¶þ 7KþQ2óh:97nQ7;AKIEORIEýHETdA-IJ=³QóK:P7nHJD9AKøETR7*:97fDMHETÂ8PD;FEAKDB?v7TRTAhDA±=@7;U:97fAKD97ºORI³89ôE7*8POÂQ7AKIJ=:97fDMóhHE:PU7;DvIE7;7;=@7f=È8Pó±7ù@7;UH@897"89ôJ7*TÂ7fA-:PIEOÂIJýAKTÂýhóK:PO8PôEQ+þo °+a[_tÂ ° iÂKW °³WC~ cea;`lr]a;[]\_ijgh^ mKWP^Y;`\_ijo °+WC~ ¹¶`liy\_iY\_ijwxWf¯ \]\_XTp qSq·#·#·¸ ·&r-¸ a;[_k'q;;;K#qr]·jq

i

72

Remove HTML

Plain text documents

HTML Pages

IR

Bags-of-words

InstanceRepresentation

Binary Term Vector

Learning Algorithm

Learning outcome

scthu 'vZ ¿JWY[_`hi`hknbR[_a;^\ ¯ Wv°³WC~wm-o ¯ WC^Y\_ijo"xsijWP·

yTzy |~@Se][`[7

ÿBôE7A-üEüJ:9ólAKUCôÈ?v7nôJAxòK7º8PAKúK7I³OÄDBøJAKD97;=¼óKI+A8_õlüJORU;A-TQ±AKUCôEORIE7nTR7;A-:PIEORIEý±A-üEüJTÂOÄUA-89ORóKIOÂI¾89ôE7nUóhIh8P7ùZ8Bó89ôJ7º÷ÊóK:PTR=¾÷ OR=E7º÷Ê7ø]ZHJD97:9û¢üE:9óBHJTROÂIEýNþhÿBôEOÄDBD9U7IJAK:9ORóüE:PóòZOR=E7;D>AKIóKüEü£óK:989HEIJO8_õ 89ó²7ù@üETÂóh:97È89ôE7¼øN7;ôJAxòZOÂóhHE:óAIZHEQøN7;:óBTÂ7fA-:PIEOÂIJý²A-TRýKóK:POÂ89ôEQ±D;þ÷Ê7AKDPDMHEQ7!89ôJA-8nA¼HJD97:"ôJAKD"ORIh8P7:CAKU897;=²?O8Pô A³DMõ@DM897Q óhI D97òK7;:PAKTó@UU;AKD9OÂóhIJD;FNü£7:PôJA-üNDølõ:PA-89ORIEýÇA ÷ 7;ø üJA-ýh7+óK:±Q±A-úZORIEý)AÊüEHE:CUCôJAKD97A8¾A-I 7û¢UóKQQ7:CU7DMOÂ897hþ&Eóh:ÈóKHJ:üEHE:PüNólDM7fD89ôJ7!7ùEAhU8ºDPU7;IJA-:POÂóÈ=Eól7fDIEóK8ºQ±A89897;:;þ£<>AKUCôóB89ôE7fDM7OÂIl897;:PAhU8POÂóhIJD"óK:PQ A89:CA-ORIEOÂIJý³ORIJDM8PAKIJU7hFTRAKøN7;TÂTR7;= ?O8PôÇDMóhQ27±UTRAhD9Dn=@7;üN7;IJ=@ORIEýóKI 89ôE7ÈAhU8POÂóhIÔü£7:)óK:PQ27f=ølõn8PôE7BHJD97:fþ4JóK:7ùEAKQ2üJTÂ7ÌD2]ôJ7vQORýKôl8ôJAxòK7v:PA-897f=!AºøNóZóKúl)#7:Põ*ýKóZó@=EFD9ó8PôE7=EA8CAA-ø£óKH@8s89ôJA-8vø£óZóKú289ôE7;IÈø£7;UóhQ7;D>89ôE7"OÂIND_8CA-IJU7ºA-IN=±O8CDsUTRAhD9D>?sóhHETÄ=±øN7V#7;:9õ2ýKóZó@=JþÿBôE7!UCôJAKTÂTR7IEýh7OÄD89ó¾HJDM72AÈD9780óB#D9HJUCôORIJDM8PA-INU7;D89ó³AKUIlHEOR:P7AÈUTÄAKDPDMOLHNU;A89ORóKI+Qó@=@7T?ôEORUCô UAKI ø£7HJD97;=Ô89ó+üE:P7;=@OÄU8!UTRAhD9D"TÄA-ø£7TÄDóK:H@89HJ:97±ORIJDM8PA-INU7;D;þkEóh:*7ùEA-QüETR7KFORIA-I¾7ûøJHJDMORIE7fD9DvUóhIh8P7ùZ8;F@D9HJUCô¼A!QóZ=E7TUóhHETÄ=±89ôJ7I¾ø£7nHJD97;=±8Pó:P7;UóKQQ7IJ=ÈüE:9ó@=@HNU8PD89ó²A³HJD97:fþÿó7ùZüJTÂóh:97289ôJ7±OÂQüJAhU8óBDM7;Q±A-Il89OÄUQ±A-:PúlHJüF?s7:P7IlHEOR:97±A³IZHEQ!ø£7:!ó=EA8CAKD978PDdóNOÂIl897;:PAhU8POÂóhIJD?ôJORUCô7ù@ORDM8OÂI2A0D97QOû]DM89:PHJU8PHE:97f=*8P7ùZ8eóh:9Q±A-8;FxAKDe?s7;TÂTZAKDeORIAa@7Q±A-Il89OÄUB÷Ê7øTRAKIEýKHNA-ýK7hFDMHJUCôAKD*,hþ÷ 7?vóKHETÄ=89ôE7;IøN7AKøETÂ7v8PónUóKQüJAK:97v89ôJ7üN7;:óh:9Q±A-INU70óBTR7;A-:PIEORIEý\:PóKQ 8PôE7"üETRAKOÂI¾897ùl8"óh:9Q±A8v?O8Pô¼TR7;AK:9IJOÂIEý:9óhQ DM7;Q±A-Il89OÄUQ278PA-û¢=EA-8PAJþ

¯ \]\_XTp qSq·#·#·¸ ·&r-¸ a;[_k'q#sV"]q

ï

73

×]I)89ôJ7¼IE7ùZ8D97;U89ORóKIF?v7³=@7fD9U:9ORø£7È89ôJ7+=JA8PAhDM78PD?s7¼ôJAxòK7¼HJD97;=óh:óhHE:7ù@ü£7:9ûOÂQ7Il8CDFs89ôE7;I(ORI(D97;U8POÂóhI)?v7Ô=@OÄD9UHJDPD¾óhHE:¾7ùZü£7:PORQ27;Il8PDÈ?OÂ89ô úZIEó?TÂ7f=@ýK7²D9üJAK:PD97TÂ7fA-:PIEOÂIJýJFd89ôE7+A-TRýKóh:9OÂ89ôEQ±D!HJDM7f=FOÂIND_8CA-IJU7¾:P7üE:P7;D97Il8PA-89ORóKIJDA-IJ=):97fDMHETÂ8PD;þ×]IÉD97;U8POÂóhI¼?v7=@OÄDPUHJDPD0úZIEó?TR7;=EýK7!ORIl897;IJDMORòK7TÂ7fA-:PIEORIEýJF£HJD9OÂIEý¼89ôJ7×]IJ=EHJU8POÂòh7óKýKOÄU!6>:9óhýK:CA-Q2ûQ2ORIEýÇA-TRýKóh:9OÂ89ôJQ 6 :PóKýKóhT¶FA-ýlA-ORI=@OÄDPUHJDPDMORIEý A-TRýKóh:9OÂ89ôEQ±D;Fe=EóZUHEQ7Il8È:P7üE:P7;D97Il8PA-89ORóKIA-IJ=¼:97fDMHJT8CDþ7eORIJA-TRTRõKF@?v7n=EORDPUHJDPDD9óKQ7n:97;TRA-897f=¾?vóK:PúÈA-IJ=¼üE:P7;D97Il8óKHE:UóKIJUTÂHNDMORóKIJD;þ

=eÛ*@Û

÷ ôE7I!UóKQQ7IJUOÂIJý89ôEOÄDd?sóh:9úº?s7 úZIE7;? 89ôJA-88PôE7Z7;Q±A-Il89OÄU#÷Ê7ø?BAKDdD_8POÂTRThòh7:Põ0QHJUCôOÂI²OÂ8PDºORIÌA-IJUõKFÓøEH@8"?s72D_8POÂTRTeôEóKü£7;=+8PôJA8nOÂ8º?vóKHETÄ=ø£7!ü£óhDPDMORøETR789ó9HJIJ=²D97Q±AKIh8PORU;A-TRTÂõQAK:9úh7;=Zû¢HEüÇ=EA-8PAHEü£óKIÉ?ôEORUCô 8PóÔøJAhDM7¾óKHE:7ù@üN7;:9ORQ7Il8PD;þe0IóK:989HEINA897;TÂõhFd?v7ÈôJAxòh7øN7;7I HEIJAKøETÂ78PóVHJIJ= A-IZõÔD9HEøJDM8PAKIh8PORAKT A-QóKHJIh8*óBv=EA-8PAUóhIl8PA-ORIEORIEý+DMHJUCôÊQ±A-:PúZHEü7-þEóK:89ôEOÄD:P7;AhDMóhI ?v7±?s7;:97óh:PU7;= 89óýh7IE7;:PA-897±óKHJ:*ó?I/ÌüN7;:9ôNA-üJD*:CA8PôE7:A-:989OLHNUOÄA-T1=EA8CAEþ×¢82ORD*óhHE:ôEóhüN7hFôEó?v7òh7:fFÓ89ôJA-889ôJORD=EA8CA?OÂTRTsD97:PòK728Pó²OÂTRTRHJD_8P:PA-8978PôE7¾OÄD9D9HE7fDAKDPDMó@UOÄA8P7;=¾?O8Pô³TR7;AK:9IJOÂIEý:PóKQ 8PôE7Z7;QAKIl89OÄUn÷ 7;øþ

kzy \ #j¡`¢]£¥¤U¡¡£S@×_ÿÿdAKTÂú@D)¦"OÄDA-I³óKIJTÂORIE7nüNóh:M8CA-TMóh:vORIóh:9Q±A8POÂóhI³AKøNóhH@8OÂIóK:PQA-89ORóKI¼897fUCôEIEóKTRóKýhõ±DM7;QûOÂIJAK:PD0ýKORòK7;I²A80HEIJOÂòh7:CDMOÂ89OR7;D0OÂI8PôE7Óþ£×¢8º?BAKD8PôE7!óhIETRõ³AKüEüETRORU;A89ORóKIóB>A9Z7;QAKIl89OÄU ²A-:PúZHEüTÄA-IJýKHJAKýK7B?s7B?v7:P7A-øETR7B89óaHJIN=?ôJORUCôôJAh=2D9HmÈUOR7Il8>AKQ2óhHEIl8PDóBÓD9HJUCô=EA8CAüEHEøETRORUTÂõAxòA-ORTÄA-øETR7KþE×¢8"HJD97;DL §gº×_J¨n8Pó¾=@7fD9U:9ORøN7!8PAKTÂú@D;FE8PAKTÂúh7:CDFJTÂó@UA-89ORóKI²AKIJ=ó-89ôJ7:ºUóKIJU7ü@8CD:P7TÄA89ORIEý±8Pó¾D97QOÂINA-:CDþJÿBôJ7!DMõ@DM897;QÅOÄDüEHJøETÂOÄU!AKIJ=A-IZõhóKIE7nOÄD:P77*8PóDMHEøJQ2OÂ8e89ôE7;OÂ:eó?I*8PAKTÂúÓþxÿBôE7 8CA-TRú@DdUóKQ7>ORID97òh7:CA-T'óh:9Q±A8CDFf7OÂ89ôJ7:AKDeAüETÄA-ORI0ÿ Z÷ 7;ø³üNA-ýK7nóh:AKD"L Z§gº×_ þTdORýKHJ:97ï2DMôJó?DvAKI+7ùEA-QüETR7nódø£ó-8PôVóK:PQA-8PD;þ

LTÂ89ôEóhHEýKô³89ôJ7;D97=@ó@UHEQ7Il8CDA-:P7*ýK7;IE7:CA8P7;=9:9óhQ 89ôJ7D9AKQ27*øNAKUCúlû7;IJ=³=EA8CA-øJAhDM7hF89ôE7;:97>AK:97 D9óKQ7>=EOnfÓ7:P7IJU7;DOÂI!UóKIl897;Il8;F7hþ ýNþ8PôE7&0ÿ ZÈòK7;:PD9ORóKInOÂINUTRHJ=@7;DdAøEORóKýh:PAKüEôZõó89ôJ7ºA-HE89ôEóh:;Fh?ôEOÂTR78PôE7,0L §gº×_òA-:PORAKIh8sOÂIJUTÂHN=@7;DsQóK:P7OÂIóK:PQA-89ORóKI±óhI89ORQ27A-IJ=±üJTRAhU7KFh78CU-þZÿBôE7×_ÿÿeA-TRúZDsDM78vUóKIl8PAKOÂIE7f=±=@7;DPU:PORü@89ORóKIJD>óJ©Bn8PA-TRú@DFl7;AhUCôó?ôEOÄUCôóK:PQ27f=¾AKI¾ORIJD_8CA-IJU70ORI¼óhHE:B=EA8CAKD978;þ@ÿBôE7a©OÂIND_8CA-IJU7;Dv?s7;:97ºQ±A-IZHJA-TRTRõUTRAhD9D9OLHJ7;=±øZõ7;AKUCô2ó£HJDK¶6><F-GnLG2FKL,º6eOÂIl8Pón8_?sónUTÄAKDPDM7fDFx7;O8PôE7:ê èNé¢ä!9ä#é¶ê èWvóK:è[-éê èNé¢ä!9ä#Céê è@øJAKD97;=óhI89ôE7089OÂ89TR7KFZA-H@8PôEóK:vA-IJ=±AKøJD_8P:PAhU8fþKÿBôE:P770UTRAhD9D9OnHN7;=òA-:PORAKIl8PD#óB£8PôE70:CAx? =EA8CA?s7;:97"89ôZHJDU:97fA8P7;=þeORýKHE:P7\ªD9HEQQ±A-:PORD97;Dv8PôE7UTÄAKDPDB=@OÄD_8P:9ORøEH@8POÂóhIJDóe7;AhUCô¼òh7:CDMORóKIþ

kz1 \¬«®SW£W@¯¥¤U¡S¡£SÿBôE7V°º<±G7;D97;A-:CUCôE×]IJ=E7ù²ÈOÄDA =@ORýKOÂ8PAKTBTÂORøE:CA-:PõÊóB0:P7;D97;A-:CUCôÇüJAKüN7;:PD2?OÂ89ôEORI=±sóhQ2ûüEH@89ORIEýl@UOR7INU7Kþ×¢8*=@óZ7;D"IEóK8"üJ:9óòZOÄ=@7=@ó@UHEQ7Il8PD"?OÂ89ô DM7;QAKIl89OÄUQ±A-:PúZHEüFÓøEH@8n89ôJ7TRAhUCúóBA-IZõóK89ôE7;:vD9óKHJ:PU7;D#óB=EA8CA?OÂ89ô¾AKüEüE:PóKüE:PORA-897Q±A-:PúZHEüZÌó-89ôJ7: 8PôJA-I¼×_ÿÿdAKTÂú@D2óK:CU7;=²HJDº89ó³TÂóZóKúVóh:"=@OLfÓ7:P7Il8"?BAxõ@DóBsAKUIhHJOÂ:POÂIJý¼D9HJUCôÔ=EA-8PAEþÿBôE7N7fDM7fA-:CUCôE×]IJ=@7ù³ °+Wv·eafghtÂ~@WhWCtjijk ¯ \_W2iybY`zaf`hWBoCafght2X@afij`\ glr>Y\>r]gho ¯ YnlY\MY;r]afgK[_oCWS´µ ¯ \]\_XTp qSq·#·#·¸ iy\]\MYt´-rC¸ a;[_k¶ ¯ \]\_XTp qSq·#·#·¸ hY;^*t ¸ a;[_k· ¯ \]\_XTp qSqoCiy\_WCr]WCWP[¸ `¸9¸ `lWCof¸ oCa;^qoCr

74

scthu ¹Jºv® ½¿+Y;`lVs|B½¿ Áv¹¢¿³»hY;^*XhtWPrbR[_af^ ¹ ®#® Y;tj´-rC¸

75

üE:9óòZOÄ=@7;D8PôE7VHETÂT897ùZ8±óº89ôE7üNA-ü£7:CDF#A-IN=AKIORIJ=@7ùZORIEýÇD9õZDM897;Q»óK:±UO8CA89ORóKINDF#ORIAK=E=@OÂ89ORóKIFEO8üJ:9óòZOÄ=@7;Ds89ôJ7UóK:P:P7;D9üNóhIJ=@ORIEý¼vORøJÿ7ù¼7;Ih8P:9OR7;D;FED977eOÂýhHE:97Jþ

½¾¿_ÀÁ%ÂÃ6Ä_Ä6Å%¾¿)Æ)Ç6È"É6Ê2ÃË)ÄPÁ%Ì_ÍÀ%ÄPÁÎ)ÂPÁ6Ï%ÐP¿2Ã6ÄÑÐPÊÒ6Ó%ÂPÁ,ÔÕ®Ö×ØÊ2ÃË)Ä6Á,Ð¿)ÅÙ4×LÚ6Û'×JÜ)Ð_ÄPÁ#ÕXÑÒ2¾Ò)Ý_ÄÔÕ`Þß)Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄKà6Ò6Ê)Å_áÂ6Îâ)ÄÏ2ÂPÁáKãÂP¿2Ç_¾_ÇÒ%Ä¿2ÃPá"â)Â_ÅÄ_ÝÇ%ÕXÑä2Â_ÂPË_Ò%¾PÒ)Ý_Ä"Ô\Õ`ß_Á)ÂÃ6ÄÄ6Å)¾¿)Æ%Ç"ÂPÎÒ6Ó2ÄåPÌPÒ_Óæ`¿Ò)Ä6Á6¿%ÐPÒ%¾_Â¿%Ð_Ýà6á6Ï_À%ÂÇ_¾Ê_ÏÂ¿,ã)ÂÏ_À_ÊÒ%ÄPÁÞ6Á%ÃÓ2¾Ò)ÄÃPÒ6ÊÁ)Ä2ÕXÑÐ_Å_Å6Á)ÄÇÇ"ÔÕ`çÂÝ6ÅãÂÐÇÒ4ÑÞ6Ê2ÇÒ_Á)ÐÝ¾6Ð2Õ2Ñá)Ä_ÐPÁÔÕPåPÌ_Ì_ÍXÕ2ÑÊ)Á)ÝÔ\ÕPÃ_¾Ò)ÄÇ_Ä_ÄPÁ×è¿2é4×è¿%ÄÃ4×hÃ6ÂÏ%ê_ÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6Ä4×dÓÒ6Ï%Ý2Õë

scthu ìE »hY;^*XhtWBeij~ ® WP»`\][]z!bÂ[_af^ÖÏBdcí>WCr]WY[_o ¯ ¹¶`ZKWP»E¸

¼sORøJÿd7ù)OÄDAÔôEORýKôJTÂõ)DM89:PHJU8PHE:P7;=óh:9Q±A-8;F#A-IJ=ÉORD8PôE7:P7#óh:97³7;AhDMORTÂõÇUóKIZòK7;:M8P7;= 8PóA-IZî øJAKD97;=óK:PQA-8;FÓDMHNUCô AKDKvþ÷ 7±UCôEólDM7K,vFAhD0OÂ8"OÄDº89ôE7÷W±ºö D"øJAhDMOÄUZ7Q±A-Il8PORU*÷Ê7ø:P7üJ:97fDM7;Ih8CA8POÂóhI¼TÄA-IEýhHJA-ýh7KþJÿBôJ7!UóhIlòh7:CDMORóKI¾89ó9, ?BAKDBü£7:)óK:PQ27f=ølõ Q±A-úZORIEýÇ8PôE7 OÄ=@7;Ih8POnHN7:³ó*89ôJ7ÔüJAKüN7;:íÌO¶þ 7Kþï Jë%(hä!#O:XJä!®#'!æ±çè£ëCä%¾89ôE7 DMHJøð_7;U8A-IJ=³7;AhUCô³A-8M8P:9ORøEH@8P7û¢òxAKTÂHJ7"TRORIE7AüE:P7;=@OÄUA-897*A-IJ=¼óKøð_7;U8ORIA-IlKÇ8P:9ORüETR7Kþ7eORýKHE:P7aOÂTRTÂHND_8P:PA-897;Dv8PôE7\K)ýh7IE7;:PA-897;=9:PóKQ 8PôE7\¼vOÂøJÿd7ù+A-üEü£7;AK:9ORIEý2OÂIleORýKHE:P7Eþ

ñòPó Ï%Ýô)ÄPÁ%Ç_¾_Â¿)ÔXÕPåS×nõXÕ ò6öñ ÁÅ6Î÷nÖ6øù ó Ï%ÝP¿2ÇPÔXÕ1Ó)Ò_Ò6ÀB÷LêêPú_ú_ú×Ã_ÇPÅ'×LÐPä)ÅP¿B×LÐ)ÃS×èÊË)ê%ûÆ_Æ6Á%¾Ï_¿%ÄÇ6êÄ ó À%ê_ü_ýXÕó Ï%ÝP¿2ÇS÷dÁÅ_ÎÔXÕ1ÓÒÒ6ÀB÷Lê_ê6ú_ú_ú×dú%þ4×LÂPÁÆ)ê)åPÌ_Ì_Ìê6õÍ_ê_ÍÍ_ÚPÁÅ6Î%ÚÇPáP¿Ò%Ð ó Ú¿%Ç6ü_ýXÕ öñ ¾¿_ÀÁ%ÂÃ6Ä_Ä6Å%¾¿)Æ)ÇjÁÅ_Î÷LÐä%ÂPÊÒÔXÕ`É6Ê2ÃË)ÄPÁ%Ì_ÍÀ%ÄPÁÎ)ÂPÁ6Ï%ÐP¿2Ã6Ä2Õ öñ ÐÊ)Ò6Ó%ÂPÁ ö Ö×eØÊ2ÃPË)ÄPÁ,Ð¿%ÅÙ4×LÚ6Û×JÜ)Ð_Ä6Á ñ ê_ÐÊ)Ò6Ó%ÂPÁ öñ Ò%¾PÒ)Ý_Ä ö ÞKß)ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄàPÒ6Ê)Å_áÂPÎKâ)ÄPÏ%ÂPÁáKã)Â¿2Ç_¾_ÇPÒ)Ä¿2ÃPáKâ)Â6ÅÄ_Ý)Ç ñ êPÒ%¾PÒ)Ý_Ä öñ ä%ÂÂPË_Ò%¾Ò%Ý_Ä ö ß_Á%ÂÃ6Ä_Ä6Å%¾¿)Æ)ÇÂPÎ,Ò6Ó%Ä,åPÌ6Ò6Ó\æ`¿)Ò)ÄPÁ6¿%Ð6Ò%¾6Â¿%ÐÝà6áPÏ_À%Â)Ç_¾Ê_ÏÂP¿,ãÂÏ_ÀÊÒ)ÄPÁÞ_Á%ÃÓ2¾Ò%ÄÃÒ6ÊÁ%Ä ñ êä%ÂÂPË_Ò%¾Ò%Ý_Ä öñ Ð6ÅÅ6Á)ÄÇ_Ç ö çÂ_Ý6Åã)Â_ÐÇÒ4ÑÞPÊ2ÇÒ_Á%Ð_Ý¾6Ð ñ ê_Ð6Å_Å6Á%ÄÇ_Ç öñ áÄÐPÁ ö åPÌÌ_Í ñ ê6á)Ä_ÐPÁ öñ ÊÁ%Ý ö Ã_¾Ò%ÄÇ6Ä_ÄPÁB×è¿%é4×è¿2ÄÃS×hÃ6ÂPÏ%ê6ÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_Ä4×èÓÒ6Ï2Ý ñ êÊ_Á)Ý öñ ê¾¿_À)Á)ÂÃ6Ä_Ä_Å)¾¿)Æ)Ç öñ êPÁÅ6ÎB÷dÖ6øù ö

scthu 2ÿ£ sV»hY;^*XhtWvWC`lWP[MY\_WbÂ[_af^Öeij~ ® WP»E¸

ÿBôE7Q78PA-û¢=JA8PA ýh7IE7;:PA-897f= :PóKQ ¼sORøJÿd7ù ORDÈòK7:PõD9ORQ2üJTÂ7hF>A-IJ= TÄAKUCú@DQ±A-IZõó89ôE7üJ:9óhüN7;:M8POÂ7fD#óAn:97fA-TT@7Q±A-Il89OÄUB÷Ê7ø±:P7;D9óKHE:CU7hF-DMHJUCô±AhD#óKIl89óhTÂóhýKõ:P7#7;:97;IJU7fDþgºIJ7IEó-8CA-øETR7sUCôJAK:PAhU8P7:PORDM89OÄUORD89ôNA8dOÂ8ORDdQ!HJUCôDMôEóh:M8P7:8PôJA-I!89ôE7süETÄA-ORI*897ùl8:P7üE:P7;D97Il8PA-89ORóKIó8PôE7ÊDPA-Q7 OÂIJDM8PAKIJU7 8_õZüEOÄUAKTÂTRõaªÇ?vóK:C=ED%þ>×]IÃAK=E=@OÂ89ORóKIFs8PôE7 ¼sORøJÿd7ù =@óZ7;DOÂIJUTÂHN=@7ºD9óKQ7ORIóK:PQ±A8POÂóhI8PôJA8vORD>IEó-8s89óø£7KóhHEIJ=ORI±89ôJ7KHETRTJ8P7ùZ8;FlDMHNUCôÈAKD !è£ç5é¶ê é5ÂäAKIJ=Z3KäCç!_jM Jí#5jêÌëCç-é¶ê`-èJþ

ÿBôE7,7fDM7fA-:CUCôE×]IJ=@7ù2üJ:9óòZOÄ=@7;DsAcæ[ @é]ä!0åëêÌäè£ëCä !ê1!9äCëé6'!X3"TRORDM89ORIEý!89ôE7"QóhDM8UOÂ897;=³üJAKüN7;:PDOÂI+7;AhUCô¼óBiDMHEøð_7;U80U;A8P7ýKóh:9OR7;D;þZ×]I³89ôE7fDM7U;A897;ýKóh:9OR7;D"7fDM7fA-:CUCôE×]IJ=@7ùTÂOÄD_8CDKª©B©üNA-ü£7:CDþó?v7òh7:fFlAIZHEQ!ø£7:BóB89ôJ7;D97nAKüEü£7;A-:BHEIN=@7:BQóK:P708PôJA-I³óKIJ7nU;A89û7ýKóh:9õ+A-IN=?s7;:97!:97;Q2óòh7;=FJ89ó¼TÂ7fAxòK7lïhï±ORIJDM8PAKIJU7fDORIÔóKHJ:º=EA-8PAKD978fþkgºIJ7A-üJüE:9ólAKUCô ¯ \]\_XTp qSqoCiy\_WCr]WCWP[¸ `¸9¸ `lWCof¸ oCa;^q;Kij[_WPoP\_a;[]z¸ ¯ \_^*t

ª

76

89ó+TÂ7fA-:PIEORIEý9:PóKQ 8PôEOÄD=EA8CA¾?vóKHJTR= øN728Pó+A-8M8P7Qü@8*89ó³TR7;A-:PI A¼UTRAhD9D9OnHN7:*UAKüJA-øETR72ó=@ORDPU:PORQ2ORIJA-89ORIEý¾ø£78_?v77;I²89ôJ7Vi±U;A897;ýKóh:9OR7;D;þT0ó?s7;òK7:fFNAhDnAÈQ!HJT8POû]UTÄAKDPD"üE:9óhøETR7Q ORDDMHEøND_8CA-Il89OÄA-TRTÂõ¼ôJA-:C=@7;:B89ôJAKI²A±DMORQüETÂ7*øJOÂIJAK:9õ¼TÂORúK7=@OÄD9TÂORúK7*UTÄAKDPD9OnHNU;A8POÂóhIF@?s7AKTRD9ó±:CA-I89ôE7º7ù@üN7;:9ORQ7Il8vA-8M8P7Qü@89ORIEýA*øEORIJA-:Põ±UTÄAKDPDMOLHNU;A89ORóKIóòh7: 7fAKUCôÈóB89ôE7inU;A897;ýKóh:9OR7;D;FO¶þ 7Kþ£êºéãZê*Jç%Jä!nç-èL0ýK7;Ih8CDjEç2Eä!2$&'!ºêºê é>çèT30éãEä\-é ãJä#!8<ë5Âç'%ä#)÷ 7"UCôEóhD9789ôE7,HJòK7nUA-897;ýKóK:POR7;D#?OÂ89ô¼89ôE7"QóhDM8vORIJDM8PAKIJU7fDcóh:søEORIJAK:9õ±UTÄAKDPD9OnHNU;A8POÂóhIFhINA-Q7TRõVY²çëPã@ê è£ädäCç'!èJê èW'$]+,!é¶ê vëêÌç5èJé]ä515jênhäè£ëCä#$è#!æç-é¶ê`-è0äé!êÌä0xç5è$! \ @æç-è"c-æKM @é¢ä!èJé]ä!9çKëé¶ê`-èÊAKIJ=#2çé]çKíCç'ä#þ#÷ 7AKTRD9ó D97TR7;U897;=É89ôJ7l+jhäèJé!AKDóKIJ7¼óB89ôE7D9Q±A-TRTÂ7;:ýK:PóKHEüJD;þMdORýKHJ:97N©³üE:97fDM7;Il8PD089ôE7±UTÄAKDPDn=@OÄD_8P:9ORøEH@8POÂóhIZóh:º8PôEOÄD*=EA8CAKD978;þk°0ó-89728PôJA8n89ôJ7GnLG òxAK:9OÄA-Il8ôNAKDci$QóK:P7 ORIJD_8CA-IJU7;DeAKDQ2óh:97 8PA-TRú@Dø£7;U;A-Q7sAxòA-ORTÄA-øETR7:9óhQ×_ÿÿdAKTÂú@D;FôEó?s7;òK7;:s8PôE7;D97n?s7;:97nIE7;òK7;:BUTRAhD9D9OLHJ7;=¼øZõ¼6><óK:L,º6þ

%'& ¨C©C« & ¬)(h¨ * t,+ (K¨ - t ¨/. t0+ (h¨B|1 2; f r$3Òd r /2 rx|sVÒ r r$4

scthu )5J ¹ ®#® Ytj´Kr VY\MY;r]WP\vÒ[_a;ÀZtjWf¸

-!«l©C«l¨6(K© 78* ¦ % %:9 ;=<>% -@? ¦ u ( & ©¨Ïvgh^"~@WP[#ab¹`hr¢\MY;`hoCWCr xf r$4f r$2f rrx A4; )4x -/4

scthu )BJ >WCr]WY[_o ¯ ¹`ZKWP»!VY\MYr]WP\sÒ[_aÀZtjWf¸

C DÃÚÝFE#G_dÞ HcJILK,eÜTNM]ÜJÚ0á9Ú H

O]zy PM¯']FQSRS`TP

÷ 7=@7HJIE7 Uè['"5Âä2-häåJç'!2äVdäCç'!èNê è@BAKDd8PôE7A-üEüE:PóhAhUCôn89ôJA-8#ORDd89:CAK=EO8POÂóhIJA-TRTÂõn8CA-úh7IølõD_8CA89OÄDM89OÄUA-TZóh:#üE:9óhøJA-øJOÂTRORDM89OÄUsQAhUCôEORIE7sTR7;AK:9IJOÂIEý"Q78PôEó@=EDþK×]I!8PôE7B÷ 7;øUóKIl897ùZ8;Fx89ôJ7üE:97fDM7;IJU7!óK:nA-øJD97IJU7ó>U7:98PAKOÂIÔ=EORDPU:POÂQORIJA8POÂIJýÈúK7;õl?vóK:C=ED?OÂ89ôJOÂIÊA¾D978ºóB#89:CA-ORIEORIEýOÂIJDM8PAKIJU7fDnORDnHNDM7f=²8Pó³üJ:9ó@=@HJU7±A³UTRAhD9D9OLHNUA-89ORóKI²Qó@=@7TþeZHNUCôÊA³Qó@=@7T#OÄD"89ôE7;I HJDM7f=89ó¾üJ:97f=@ORU8º8PôE7UTÄAKDPDó&H@89HE:P7!HJIJDM7;7I OÂIJDM8PAKIJU7fDþkZHJUCô AÈQó@=@7Te8_õZüEOÄUA-TRTRõ+UóKIJD9OÄD_8CDó!AÉD978³ó?v7ORýKôl8PD¾óK:³A)üE:PóKøJAKøEORTÂOÄD_8PORU²Q2ó@=@7;Tóh:¾8P7:PQ±DÈó@UUHE:9:PORIEý)?OÂ89ôEORI(7fAKUCôUTÄAKDPDþ@LQó@=@7TTR7;AK:9IE7f=¥:PóKQÆD9HJUCô+A-I+A-üEüJ:9ólAKUCôÈOÄDBQ±A-ORIETÂõ¾HJDM7HET?OÂ89ôEORI³89ôE7*ýhOÂòh7I7ù@üN7;:9ORQ7Il8;FEAKIJ=¼?OÂTRTøN7nHJD97TR7;DPDjóK:UTRAhD9D9OnõZORIEý!ORIJDM8PAKIJU7fD:9óhQ A2=EOnfÓ7:P7Il8DMHJøð_7;U8=@óKQ±A-ORIF-?ôJ7:P7s8PôE70=@OÄDPU:POÂQORIJA8POÂIEýnúh7õZ?sóh:P=JDdQORýKôl8 ø£70=@OLf£7;:97;Ih8fF-óK:#7;òK7;I2ORIJDM8PAKIJU7fD:9óhQ 89ôE7DPA-Q7*=@óhQAKOÂI³?ôEOÄUCô+ôJAxòK7"øN7;7I+üE:P7üE:Pó@U7fD9D97;=¾OÂIA=@OLf£7;:97;Ih8?BAxõKþ

©

77

O]z1 WU`[7¯'®SFX £EóK: 89ôEOÄD>üNA-:98 óóKHE:vD_8PHJ=@õhFK?v7HJD97;=28_?vó*?s7;TÂTNúlIJó?I2Q±AhUCôEOÂIJ7TÂ7fA-:PIEOÂIJýA-TRýKóK:POÂ89ôEQ±D;FøNóK89ô2óN?ôJORUCôóKü£7:CA897vHJD9OÂIJý"A7;A-89HE:P7sòh7;U89óK::P7üE:P7;D97Il8CA89ORóKI*óBN7fAKUCô!ORIJDM8PAKIJU7hþxÿBôJ77;A8PHE:P7;DvHJD97;=¾89ó±=@7;DPU:PORøN7"OÂIND_8CA-IJU7;DBòA-:POÂ7f=Èø£78_?v77;I¼óhHE:=@OnfÓ7:P7Il8A-üJüE:9ólAKUCôE7fD¶DM7;7øN7;TÂó?þ°ûóKTÄ=¼U:9ólD9DsòA-TROR=EA-89ORóKI î ïið?BAKDvHJD97;=¼?ôE7I+AKDPDM7fD9D9ORIEý8PôE7nüN7;:óh:9Q±A-INU7"ó89ôE7*Q78PôEó@=ED;þJÿBôE7A-TRýKóh:9OÂ89ôJQDv?ORTÂTIEó?(ø£7=@OÄD9UHJD9D97;=¼øE:POÂ7AYJõKþ

Z!ç\[] 0ä_^0ç3KäXîniið>OÄD*ADMORQüETROnHJ7f=ÊòK7:CD9OÂóhI óBs8PôE7E¼BAxõK7;DnUTRAhD9D9OnHN7:fF?ôEORUCô 8PAKúK7fD*AüE:9óhøJA-øJOÂTRORDM89OÄU±A-üJüE:9ólAKUCô 89ó²TR7;AK:9IJOÂIEýNþJ°0A$` aRòK7E¼BAxõK7fDn:97f=@HJU7fD*8PôE7¾UóhQüETÂ7ù@O8_õÊóB89ôJ7IEóK:PQ±A-TB¼BAxõK7fDUTÄAKDPD9OnHJ7;:øZõºQ±AKúlORIEýB8PôE7sAhD9D9HEQü@89ORóKI"8PôJA8d89ôE7&7;A8PHE:P7;DóB@AKI*OÂIND_8CA-IJU7A-:P7UóKIJ=EO8POÂóhIJA-TRTÂõÔOÂIJ=E7ü£7IJ=@7;Il8;þ×]I üE:CAKU89OÄU7hFÓ89ôEOÄDAKDPD9HEQü@89ORóKI DM7;TR=EóKQôJóKTÄ=ED"89:PHE7KFøEH@8È8PôE7ÔA-TRýKóh:9OÂ89ôJQ D_8POÂTRT0ü£7:)óK:PQ±D±?v7TRTóKI Q±A-IZõÉ:P7;AKTû¢?vóK:PTR=UTÄAKDPDMOLHNUA-89ORóKI8PAhDMú@D;FôJAxòZOÂIEýøN7;7I³DMôEó?I8Póø£7"7IlHJA-T£ORI¼AKU;UHE:CAKUõ!8Pó!IE7;HE:CA-T£IE78_?vóK:Pú@D>A-IN=È=@7;UORD9ORóKI±8P:97;7TÂ7fA-:PIEOÂIJý+îLi)-ðþ

UbcZäCç'!9ä#éLZ!äênxãEí2' !ed°°,¾î Kð>OÄDA-IÇA-TRýKóh:9OÂ89ôJQ?ôEOÄUCô ?OÂTRT>TÄA-ø£7T HJIEúZIEó?I OÂI@ûD_8CA-IJU7;D?OÂ89ôn8PôE7#TRAKøN7;TKóBh8PôE7 Q±A'ð_óh:9OÂ8_õól89ôJ71dÉIJ7;A-:P7;DM8IE7ORýKôZø£óKHE:CDORI°û]=@OÂQ7;IJDMORóKINA-TDMüJAhU7hF?ôE7:P7K° ORD#89ôE7IZHEQøN7;: ó[7;A-89HE:P7;DHJD97;=óh:>=@7;DPU:PORøEOÂIJý"7fAKUCô2OÂIND_8CA-IJU7Kþd°,°ORD#AºTRAfõ!AKTÂýhóK:PO8PôEQ+F;Q7fA-IEORIEý"89ôJA-8#O8#?ORTÂT@IEóK8#ýK7IJ7:CA897BAºQ2ó@=@7;T@øJAKD97;=óKI!89ôE7B8P:PAKOÂIEûOÂIEýOÂIND_8CA-IJU7;D;FhøEHE8sóKIJTÂõ?ôE7;I¼AKD9úK7f=289ó2UTÄAKDPDMOLõ2A!IE7? ORIJDM8PAKIJU7º?OÂTRTNOÂ8BüN7;:óh:9Q 89ôJ7UóKQüEHE8PA8POÂóhIJD*IE7;7;=@7f=Ê89óÔUTRAhD9D9OLõ²89ôJ7ÈýKORòK7;IÊORIJDM8PAKIJU7hþÿBôE7Èòh7:CDMORóKIÊó d°° HJDM7f=OÂI óKHJ:7ù@ü£7:POÂQ7Il8CD*ORD!A+òxAK:9OÄA8POÂóhI óBs8PôE7ÈDM8PA-IN=EA-:C= A-TRýKóK:POÂ89ôEQ+FA-øJTÂ78Pó²=@7;AKT#?O8PôDMõZQ!ø£óKTROÄUºORIEüEHE8PDî ð¢þ

O]z,O P£¡PSRBhgUjie¯'W£PMS¡\P£÷ 7ôJAxòh7sORIZòK7fD_8POÂýlA897f=8PôE:P77=@OLf£7;:97;Ih8 ?BAxõ@DeóN:P7üJ:97fDM7;Ih8POÂIJýºORIJDM8PA-INU7;DORI289ôEOÄD DM89HJ=Eõ[óKIE7#øJAhDM7f="óKIn89ôJ7#897ùl8eUóKIl8P7Il8ó@óKHJ:=@ó@UHEQ7Il8CDFfA-IJ="8_?vó?ôEOÄUCô"Q±AKúK7#HJD97KORI!D9óKQ7?vAxõóB89ôE7nDM7;Q±A-Il89OÄUQ78CAû]=EA8CAEþK÷Ê7"?OÂTRTNIJó? =E7;DPU:POÂø£707fAKUCô±ó8PôE7;D97ºAKüEüE:PóhAhUCôE7;D;þ

yTz!k`¡ePjjlk LºD>AnøJAhDM7;TÂORIE7A-üJüE:9ólAKUCô?s7HJD97;=±A*D9ORQ2üJTÂ7Q78PôEó@=289óU:97fA897B8P:PAKOÂIEûOÂIEý)=JA8PA îmZFið¢þ ÿBôEORDÈ?vAhDøJAKD97;=HEü£óKI 89ôE7¬0ÿ òh7:CDMORóKIó"89ôE7×_ÿÿeA-TRú@D±=EA8CAA-IJ=Ê8PôE7¥HETRT>8P7ùZ8!óBv8PôE797fDM7fA-:CUCôE×]IJ=@7ù A-:989OÄUTR7;D;þJJóK:7fAKUCô OÂIND_8CA-IJU7È?s7¾:97;Q2óòh7;=A-TRTIZHEQ!ø£7:CDA-IJ=+AKTÂT?vóK:C=EDvóTR7IEýK89ô+TÂ7fD9Dv8PôJA-IJF@øN7óK:P7A-üEüJTÂõZORIEý±A±D_8PóKüZ?sóh:P=¾TÂOÄDM8?ôEORUCôÇ:P7Qóòh7;DnIEóhI@û¢UóKIl897;Il8?sóh:P=ED*D9HJUCô)AhD±ê é`$ºé ãJäX$"é ãJäê1!2$"äé]ëDe÷Ê7¾A-TÄD9ó7ù@üETRóK:P7;=A-üEüETROÄUA8POÂóhIóB£AnDM897QQORIEýA-TRýKóh:9OÂ89ôEQ îLi)n-ðF:P7;=EHJUORIEýn?sóh:P=JDdTRORúK7ë2æ[ @é]ä#!2$dë2æ[ @éoê è@$#ë2æ[ @é]ä!2e8Pó¼ë2æ[ @éþlÿBôJ70OÄ=@7;A*?BAKD 89óQAKúK7ýh7IE7;:PAKTÂOÄD9OÂIEý*óòh7:>UTRAhD9D97;D 7;AKD9OR7:fFøEH@8>?v7"óKHEIN=2DM897;QQ2ORIEýn8Pó*QAKúK7TRO8989TR7óK: IJón=EOnfÓ7:P7IJU7ORI2ü£7:)óK:PQ±A-IJU7KF-AKIJ=UCôEóhD97IEó-8"89ó+HJD97OÂ8;þegºINU789ôEOÄDnOÂIJO8PORAKTeüJ:97ûüE:Pó@U7fD9D9OÂIJý¾ôJAh=²ø£77I UóhQüETÂ7897f=FÓ?s72ôJAK=Ô89ôJ7óKü@8POÂóhI²ó>7;O8PôE7:nU:P7;A-89ORIEý¼A¾øEORIJA-:Põ+òK7;U89óh:0?OÂ89ôÊ7;AKUCô7;TÂ7;Q7Il8nUóK:P:97fDMü£óKIN=@OÂIJý±89ó³A897:PQÆORI³89ôE7=EóZUHEQ7Il80òhóZU;A-øEHJTRAK:9õhFlóh:DMóhQ7,óh:9QÆó?s7;OÂýhôh8P7;=³òK7fU8PóK:øJAhDM7f=³óKI+ADMHEøNDM78ó#AxòxAKOÂTÄA-øJTÂ7º897;:9Q±D2þ7ºHE7*8Pó89ôE7D9Opf;7nóe89ôE7òhó@UA-øJHETRAK:9õiSªEF 897;:9Q±D"óh:89ôE7\7fDM7fA-:CUCôE×]IJ=@7ù¾=EA8CAKD978X>?v7=@7fUOÄ=@7;=¾89óÈAh=@óKü@8B8PôE7*TRA-8M8P7:A-üJüE:9ólAKUCôþ

ÿq×_,)?s7;OÂýhôh8CDnî ©-ðÓ?v7:P7nUA-TÄUHJTRA-897;=EóK:B7;AhUCô¾óB89ôJ7º8P7:PQ±DA-IJ=È89ôE7¥i4ª?O8Pô89ôE72ôEORýKôE7fD_8*ÿrx×_ :PAKIEúZOÂIEý¾?s7;:972DM7;TÂ7fU897f=þÓÿBôE7üJ:97fDM7;IJU7óK:nA-øJD97IJU7!óB 7;AKUCôÔó

78

89ôE7fDM7±8P7:PQ±D!?BAKDn8PôE7IÉHJDM7f=Ê89óÊU:P7;A8P7ÈAøEORIJA-:Põ 897:PQ òK7;U89óh:;þJeORýKHE:P7s+D9ôEó?DAKI7ùEA-QüETR7nóe89ôE7*üE:Pó@U7fD9D9OÂIJý2DM8PAKýK7;DvORIZòKóKTRòK7f=¾AKIJ=¼89ôE7aHJINA-T:P7üE:P7;D97Il8PA-89ORóKIþ

tuwv x:v y)z:|!T~>)/Aqy6ñ ÓÒ6Ï%Ý öñ Ó%Ä_Ð6Å ö_ñ Ò%¾Ò%Ý_Ä ö â)Ð)ÃÓ2¾¿%ÄÛÄÐPÁ6¿2¾¿%ÆÎ_Á)ÂÏÒ6Ó%ÄKàÄÏ%Ð¿Ò2¾_ÃL)Ää ñ êPÒ%¾Ò%Ý_Ä ö_ñ êPÓ%Ä_Ð6Å öñ ä%Â6Å_á öñ ÓXå ö â%ÐÃÓ2¾¿2ÄÛÄ_ÐPÁ_¿2¾¿)Æ"Î_Á%ÂÏ,Ò6Ó2ÄKà_ÄÏ%ÐP¿Ò%¾_ÃL%Ää ñ êÓ#å öñ ¾ ö ÜáçPÊ_¿_¿%Ð6ÁÞ_Þ)ÇÒÁ)Ð¿)ÅKç_Á%¾Ï_¿%Ä)Ç ñ ê¾ öñ À ö æ`¿,Ò6Ó2¾_Ç,Ç6ÄÏ2¾¿2ÐPÁKú)ÄÆ%¾ô)ÄKÅ)ÄPÒ)Ð¾6Ý)ÇÂ¿ÂÊ)ÁÁ)ÄÃ_Ä¿Ò,Ä ó À%ÄPÁ%¾Ï2Ä¿Ò%Ç"Â¿Ý_Ä_ÐPÁ_¿2¾¿)ÆÎÁ)ÂÏÒ6Ó2Ä,Ç6ÄÏ2Ð¿Ò%¾_Ã"ú)Ää

o

q:z::|TT~qz:uw/))e:6'uwA!z'yAy/ASuwo/o/ \¡¢ ¤£o$)¥F¦¤ o¤§¢,T¨T :©\/o/¢ !¡ ¤£)¢A¥S¦ª ¤§¢,\¨T :©\¥'«A$¤£F/¤¦§£¤¬¥:£1 ¦\¦¤ or¢¤£¥: r¬/ §¤¡ ¦F£ o§T o®¯$ o£ §¢¦¡¢ o¤£$)¥L¦¤ o¤§¢,F¨T ¤©

° cv :yV¤q'T)v wuwv qv yAz:cv yAxFcuwqTAwv yAxq\±j²:³,´±µ¡ ¤£)¢A¥/¶µ¦¤ o¤§¶\·¤§,·¤¡¢·e¥'¸6¶Tª¥ o§0¶¹§¤¡ º6¶\ e®ª¯) o£ o§¦¶j«A£¡µ»:»¤»¼ v yAz'u½1cuch6c:u¹¢'u¹c¾)v \v yAwcz:y):¿ ¶ ¿ ¶TÀ'¶\À'¶TÀ'¶ ¿ ¶TÀ»ª»:»

scthu $ÁE ¹¶`lr¢\MY`loCW>WCXh[_WPr]WC`\MY\_ijaf`\w½2WP\ ¯ a-¾f¸

kza&¯'W¡ScPÂg9¤sÃ ¡T£¡7£Äk`¡eP jlk gºHE:IE7ùl8±A-üJüE:9ólAKUCôÇ?BAKDD9OÂQORTRAK:!8Pó89ôE7HN:PDM8;FNøEH@80ORIJDM897;Ah=³ó#7QüETRóõZOÂIEýÈAüJTRAKOÂI³897ùZ80:97;üE:P7;D97Il8PA-89ORóKIUóK:7fAKUCô+OÂIND_8CA-IJU7?s7)Q±AKúK7ÇHJD97)ó8PôE7ÉQ±A-:PúK7f=Zû¢HEüÃ=JA8PAJþºÿBôEOÄDÔ=JA8PA ?vAhDüE:P7üE:Pó@U7;DPD97;= OÂIÖ7fD9D97IEû89OÄA-TRTÂõ 8PôE7ÇD9AKQ27 Q±A-IEIE7;:AKD¼89ôE7ÊüETÄA-ORI 8P7ùZ8;F?OÂ89ôÃóKIE7ÊORQ2ü£óK:98PAKIl8+=@OLfÓ7:P7IJU7KþB×]I0ÿ =@ó@UHEQ7Il8PD8PôE78CA-ýhDüJ:9óòZOÄ=@7VóK:PQA-8M8POÂIEýÇORIóh:9Q±A8POÂóhIF#Oþ 7hþ Å!íÆÅ897;TÂTÄD±HJD89ôJA-88PôEORD8P7ùZ82D9ôEóKHJTR=ÉøN7³üE:POÂIl897f=ÇORIÉø£óKTÄ=F?ôEOÂTR7¾ORI)8PôE7³Q278PA-û¢=EA-8PAQHNTÂ7fD!8PôE7+=@On û7:P7Il8¥î, Ç'K8PAKýhD:P7üE:P7;D97Il8DMóhQ27+ORIóK:PQ±A8POÂóhIA-ø£óKHE828PôE7+Q7;A-IJOÂIEý óB089ôJ7UóKIl8P7Il8;FNO¶þ 7KþSÅ5LfëPç-é¶ê`-è¹Æ8P7TRTRD0HJD8PôJA8nA8PAKTÂú³ôJAhD0A±TRó@UA8POÂóhIþJ÷Ê7=EOR=IJó-80?BA-Il88PóOÂýhIEóK:P7v89ôEOÄD ORIóK:PQ±A8POÂóhIFKD9ó*OÂI8PôEORD>A-üEüE:PóhAhUCô?v7B89:P7;A-8#89ôE7î, û¶8CA-ýhD>AKD Ah=E=@OÂ89ORóKIJAKT897ùZ80UóKIl897;Ih8fþEL0D?v7nHJD97ÿq×_,)89ó±D97TR7;U8ôJOÂýhôETÂõ¾:PAKIEúK7f=±897;:9Q±DB8Pó±A-üEü£7;AK:OÂI³89ôJ7OÂIJDM8PAKIJU7:P7üE:P7;D97Il8CA89ORóKIFÓUóKQQóKIETRõó@UUHE:9:PORIEý¾8CA-ýlDº?ORTRT óvóvUóKHE:CD97ø£7±OÂýhIEóK:P7;=þó?v7òK7;:;Ff8CA-ýhD?ôJORUCôó@UUHE:#OÂI:97SIhHJ7Il89TRõ!?ORTÂTND_8POÂTRTHJIJ=28PôE7OR: ?BAxõ*OÂIl8Pón89ôE7ORIJDM8PA-INU7KþdORýKHJ:97ÈnDMôJó?DA-I³7ùEA-QüETR7nóe89ôEOÄDORIJD_8CA-IJU7n:97;üE:97fDM7;Il8PA8POÂóhIþ

O]zÈÉV£ePeÊg9¤sÃÌË¥` ¡ÎÍ ÃJW¡)Qe¯' |~¡iFicP gºHE:289ôEOR:C=AKüEüE:PóhAKUCôÉ?vAhDøJAKD97;=2óKI8PôE7óhøJDM7;:9òA-89ORóKI89ôJA-8>óKI289ôJ7,Z7Q±AKIh8PORU÷ 7;øFKQ±A-:PúZHEüüE:PóòlOÄ=@7fD#D_8P:9HJU89HE:P7KFDMó*ORIJDM897;Ah=2óÓ89ôE:Pó?ORIEýAx?vAxõn8PôEOÄD>DM89:PHJU8PHE:97øZõ89:P7;A-89ORIEý*A-TRTE89ôJ7897ùl8vAKD#óhIE7HEIEOÂ8>?v7üE:9ó@U7;DPDM7f=*89ôJ7BUóKIl897;Il8óBN7fAKUCô8CA-ýnDM7;üJA-:CA8P7TRõKþ<>AhUCô!7;TÂ7;Q27;Il8ORI2óKHE:#ORIJDM8PA-INU7vòK7;U89óh:89ôE7;IÔø£7;UAKQ789ôJ7D978ºóB ?sóh:P=JD?ôEOÄUCô²ó@UUHE:9:P7;=+?OÂ89ôJOÂIÔA¼U7;:M8CA-ORI+8PA-ýTFJ8PôE7UóKIl897;Il8

n

79

tuwv x:v y)z: ´±ÈA/)qy6ñ6ó Ï%Ý öñ ÁÅ_Î öñ Ò)Ð_ÝPË¾6Å_Ô$ÏnÏ%Ý)Ç6ÄÏú)ÄPäXåÏ öñ Ò%¾Ò)ÝÄ ö â)ÐÃÓ2¾¿%Ä"ÛÄÐPÁ6¿2¾¿%ÆÎ_Á)ÂÏÒ6Ó%ÄKàÄÏ%Ð¿Ò2¾_ÃL)Ää ñ êPÒ%¾Ò%Ý_Ä öñ ÇÀ%Ä_Ð6Ë)ÄPÁ öñ ¿2ÐÏ%Ä ö ç6Ê_¿_¿%ÐPÁKÞÞ)ÇÒ_Á)ÐP¿)ÅKç6Á%¾Ï_¿%ÄÇ ñ ê¿%ÐÏ%Ä öñ Ê)Á)Ý ö ÓÒÒ6ÀB÷Lê_ê6ú_ú_ú×hÃÇPÅ'×LÐä%ÅP¿B×LÐÃ4×èÊË)ê)ûPÆ_Æ6Á%¾Ï¿%ÄÇ ñ êPÊÁ)Ý öñ Î%Ð ó ¿_Ê_Ïä%ÄPÁ öÐ¤ÑªÑ åPÍ_Í Ñ ÍªÒ_þ Ñ ÍÍ ñ êPÎ)Ð ó ¿_Ê_Ï_ä2ÄPÁ öñ êÇÀ%ÄÐPË)ÄPÁ öo

q:z::c'6:uw)!z:y)Ly/)Fuw®:r¡£¬wÓr§¤¡ º§p§¢¡¢ V/o/¢ 1¡¢ ¤£o$)¥¦¤ o¤§1¨T ¤©1¦w¯) ¤º/ £r¤ r¥:«A$¤£ /¤¦§£¤¬¥:£1 ¦r¤ «A£¡¦ª¬/©¬¤>¥¤¥:£1 ¦Ô«A£¡)Óª®:$«A©o £F¦w¯) ¤º o£

° cv :yV¤q'T)v wuwv qv yAz:cv yAxFcuwqTAwv yAxq\±j²:³,´±µ¡ ¤£)¢A¥/¶µ¦¤ o¤§¶\·¤§,·¤¡¢·e¥'¸6¶Óª®:$«A© £¶¹ª¥A o§p¶! e®ª¯) o£ o§\»¤»ª»¼ v yAz'u½1cuch6c:u¹¢'u¹c¾)v \v yAwcz:y):

¿ ¶ ¿ ¶TÀ'¶ ¿ ¶TÀ'¶\À'¶T»ª»:»scthu )ÕJ ¹¶`lr¢\MY`loCW9sWPXh[_WCr]WC`\MY\_ia;` w)½2WP\ ¯ a-É-¸ Ïsa;\_W¼hiyËEWP[_WC`hoCW¾bÂ[_af^\_WP[_^ wfWCoP\_a;[2ij`Óijk;gh[_W0f

ó>8PôEOÄD"8PA-ý+?BAKD"üE:P7û¢üE:PóZU7;DPDM7f=ORIÔ8PôE7ÈD9AKQ7Q±A-IJIE7:AhDóh:nQ2789ôEó@=EDEi2AKIJ=ÊïEFO¶þ 7KþA-üEüETRõZOÂIJý³A³D_8PóKüZ?vóK:C=+TÂOÄD_8fFOÂýhIEóK:PORIEý¼D9ôEóK:98º?vóK:C=ED;FN78PUKþÿBôE7TR7IJý-89ô ós89ôE72OÂIND_8CA-IJU7òK7;U89óh:8PôE7I2øN7fUAKQ27s89ôJ7sIZHEQøN7;:eóBJHEIEOIlHE7v8PAKýhDdOÂI89ôE7B=@ó@UHJQ27;Il8PD;FfIJó-889ôE7vIZHEQ!ø£7:ósHJIEOhIlHE7?vóK:C=ED;þÓÿBôJORD*AKüEüE:PóhAKUCôÔORDnü£óhDPDMORøETR7øN7fUA-HNDM789ôE7±òAK:9OÄA-Il8PDºóvøNóK89ô°0A$` aRòK7¼vAxõh7;D AKIJ==d"û6°7fA-:P7;DM8&°7ORýKôZø£óKHE:s?s7ºHJD97;=ÈD9HEüEü£óK:9897f=ÈDM78PD>óòA-TRHE7;DvAKD>7TR7Q7Il8CD>ORI89ôE7*ORIJDM8PA-INU7nòK7fU89óh:PD;þ

÷ ôEOÂTR789ôEOÄDnA-üEüE:PóhAhUCôQ±AK=@72DM7;IJDM7óK:"QóhDM8º8CA-ýlDFÓDMóhQ78PAKýhDºôJAK=ÔA¼UTÂ7fA-:PTÂõ+=@7ûHJIE7;=ORIh8P7:PIJA-TDM89:PHJU8PHE:P7¼?ôE7;:97³ORIóK:PQ±A89ORóKI?sóhHETÄ=)ø£7³TRóhDM8±On"üE:P7û¢üE:9ó@U7;DPDM7f=þeL0I7ùEA-QüETR7¼OÄD!89ôE7Z+ÖYâ72£êÌëHJ7;TR=Éó8PôE7³×_ÿÿdAKTÂú@D=EA8CAKD978fFd?ôEOÄUCôÉýhOÂòh7;D!89ôE7¼89óhüEOÄUó*89ôJ7²8PAKTÂú AKD³A)DM89:POÂIJýÉDMHNUCô AKDl+ÖeYâT%ÓêÌë×S*-æKM @é¢ä! åT34é¢äæ Ø"!lçèNêLïfçéê`è×T*-æKM @é¢ä! cæ2æ @èJêÌëCçéê`è Z!äé&!)(4×µèJé]ä!è£äé&!)(ê èW-þN×PeüE:P7üJ:9ó@U7fD9D97;=¾IEóK:PQ±A-TRTÂõ89ôEOÄD±?vóKHETÄ= øN7øE:PóKúh7I HEü OÂIl8Pó D97üNA-:CA897³8P7:PQD±AKIJ=89ôE7²AhUUHE:PA-897+Q7;AKIEOÂIJýÊTRóhDM8;FDMó+?v7=EOR= IEóK8*üE:97;üE:PóZU7;DPDº89ôEOÄD*8CA-ýJþ]eORýKHE:P7_Ù+DMôEó?D*AKI 7ùEA-QüETR7±ó>8PôEOÄD*OÂIND_8CA-IJU7:97;üE:97fDM7;Il8PA8POÂóhIþ

O]z0Ú gU@£$Q®£â£ãJäÛhâBâJç5 ('9!9ä#2 5é`ÇA-:P7±DMôEó?I ORI eOÂýhHE:97fD¥i)AKIJ= iBihþJE:PóKQ89ôE7fDM7È:P7;D9HETÂ8PD;FOÂ8!ORDOÂQQ7;=@OÄA8P7TRõ¾IEóK89OÄU7fA-øETR7"89ôJA-889ôE7!GnLG òA-:PORAKIl8BóBd8PôE7UTÄAKDPD9OnHJ7f=¾×_ÿÿeA-TRú@D=EA8CATÂ7f=89ó0ü£óZóK:e:97fDMHJT8CDþ;÷ 7vøN7;TÂOR7òh7 8PôJA8e89ôEOÄDdORDeA-I!AK:M8POnÌAhU8=@HE7 8Pó8PôE7sQ±AKIlHNA-ThUTRAhD9D9OLHNUA-89ORóKIó89ôE7=EA8CAF@8PôE72GnLG òA-:PORAKIl8:97AYJ7;U8PDATR7;DPD0UTÂ7fA-:PTÂõ¾=E7#HJIE7f=+OÂIl8P7:P7;DM80üE:PóHJTR7n89ôJAKI

Ù

80

tuwv x:v y)z: ´±ÈA/)qy6ñ6ó Ï%Ý öñ ÁÅ_Î öñ Ò)Ð_ÝPË¾6Å_Ô$ÏnÏ%Ý)Ç6ÄÏú)ÄPäXåÏ öñ Ò%¾Ò)ÝÄ ö â)ÐÃÓ2¾¿%Ä"ÛÄÐPÁ6¿2¾¿%ÆÎ_Á)ÂÏÒ6Ó%ÄKàÄÏ%Ð¿Ò2¾_ÃL)Ää ñ êPÒ%¾Ò%Ý_Ä öñ ÇÀ%Ä_Ð6Ë)ÄPÁ öñ ¿2ÐÏ%Ä ö ç6Ê_¿_¿%ÐPÁKÞÞ)ÇÒ_Á)ÐP¿)ÅKç6Á%¾Ï_¿%ÄÇ ñ ê¿%ÐÏ%Ä öñ Ê)Á)Ý ö ÓÒÒ6ÀB÷Lê_ê6ú_ú_ú×hÃÇPÅ'×LÐä%ÅP¿B×LÐÃ4×èÊË)ê)ûPÆ_Æ6Á%¾Ï¿%ÄÇ ñ êPÊÁ)Ý öñ êÇÀ%ÄÐPË)ÄPÁ öo

q:z::c'6:uw)oy/)Fuw)c ¢uw'hcz'xq'y6cy6ñ6ó Ï%Ý öñ ÁÅ_Î öñ Ò)Ð_ÝPË öñ Ò%¾Ò)ÝÄ ö Ï%ÐÃÓ2¾¿%ÄÝ_ÄÐPÁ6¿2¾¿%Æ,Ç6ÄÏ%ÐP¿Ò%¾_Ãú%Ää ñ êPÒ2¾Ò)Ý_Ä öñ ÇÀ%Ä_Ð6Ë)ÄPÁ öñ ¿2ÐÏ%Ä ö Æ6Ê_¿_¿%ÐPÁÐÐÇÒ_Á)ÐP¿)ÅKÆ6Á%¾Ï_¿%ÄÇ ñ ê¿%ÐÏ%Ä öñ Ê)Á)Ý ö Ã_Ç6ÅÐä)ÅP¿Æ_Æ6Á%¾Ï¿%ÄÇ ñ êPÊÁ)Ý öñ êÇÀ%ÄÐPË)ÄPÁ ö

o

Ü cv y)xÇc¾AÔ¢' Tv yAxÇcz'x:Tz'\pz¤c)uwo§,¤¡ º'¶§¢p§¢¡ ª¶µ¦w¯) ¤º/ £ª¶¤ ª¶«A£¡¹»ª»¤»³0y)cz'yA:Ý'Þ ¶ Ý /o/¢ ª¶$¡ ¤£)¢A¥/¶$¦ª ¤§¢,ª¶$¨T ¤© Þ ¶ Ý6Þ ¶ Ý ¥'«A$¤£ª¶/¤¦§£¤¬'¶$¥'£r o¦ Þ ¶ Ý ¦ª¬'¶/©¬¤¶)¥¤¥:£1 ¦ Þ»¤»ª»

scthu )ßJ ¹¶`lr¢\MY`loCW>WCXh[_WPr]WC`\MY\_ijaf`\w½2WP\ ¯ a-rK¸

à¦bà áâ ¦Ö-Öá ¦äãµ(Kª« u (vlá.R« t0&Äå (æJ© $4K¸ f)ç 2$3-¸ r)4ç 2f-¸ rf$ç 2f-¸ 3)4ç¹E 9 - sèå « u ¨"«Z¨"©'(æJ© K¸ f$ç 2x¸ r;ç 2f-¸ r;ç 2f-¸ 4)4çìE 9 - sèå « u ¨"«Z¨ s (K«h©/é£ª(K¨ rh¸ 4$ç x-¸ 4-ç fK¸ 4K/ç r$4-¸ ç

scthu 'vêN >WCr]ghtj\_r#bRa;[sÏBYPÐ ÑwfWYCzWCr#µ¹ ®#® Yt´-r>VY\MYr]WP\¸

à¦bà áâ ¦b-@á ¦bãµ(-ª« u (7ë(-©/ìTí¹î vïÔá.Â« t,&Äå (æJ© f¸j$ç rK¸ x$ç 2f-¸ rK/ç 2h¸ 2x)ç7ë(-©/ìTí¹î ¹¹ï 9 - sèå « u ¨0«Z¨ å (æN© )2-¸ 3fç 2$3K¸ r)3ç A3K¸j/4$ç 2-f¸ 4x)ç7ë(-©/ìTí¹î ì¹ï 9 - sèå « u ¨0«Z¨ s (-«l©6é£ª(h¨ ;-¸ fç 2x-¸ r-ç ;-¸jCç ;-¸ 4$ç

scthu 'v@vZ sWPr]glty\_r#bÄa;[>pBµÌÏvWCY[_WCr¢\>ÏvWPik ¯ ~@afgh[#µd¹ ®#® Y;tj´-r>VBY\MY;r]W9\¸

i$

81

78éT.R© tr< .Â«Z¨¨ 78* ¦ % %'9 ;_<>% -@? ¦ u ( & ©¨7ë(-©/ìTí¹î vjïá.R« t0&Äå (æ 4rK¸ r)4ç 4rK¸ 4;ç ;-¸ f$ç ;-¸ rx)ç )4-¸ 2$3ç 4x-¸ f$ç A4K¸ 3Sr$ç7ë(-©/ìTí¹î¹ï 9 - sèå « u ¨"«Z¨ å (æJ© -¸ rç 3-f¸ -/ç 4$3-¸ A2ç 3-f¸ $2$ç 3Sr-¸ 2x$ç 3;h¸ Srç 3f-¸ r$ç7ë(-©/ìTí¹îìï 9 - sèå « u ¨"«Z¨ s (K«h©/é£ª(K¨ -f¸j2rç 4)3K¸ 2x)ç 4$4-¸ f$ç 3;K¸ rSr$ç 3K;¸ -ç 3Kf¸ 4f$ç 3f-¸ 4x)ç

scthu 4v¹J sWCr]ghty\_r bRa;[>ÏBY9Ð ÑjwxW0YCzWCr#µJsWPr]WY[_o ¯ ¹¶`ZhW9»VY\MYr]WP\¸

78éT.R© tr< .Â«Z¨¨ 78* ¦ % %'9 ;_<>% -@? ¦ u ( & ©¨7ë(-©/ìTí¹î vjïá.R« t0&Äå (æJ© 2K¸ ;$ç 3rK¸ r$3$ç 3K;¸ A2ç 3h¸ f$ç 3Sr-¸ 3$2ç 3x-¸ x$ç 3)2K¸ x$ç7ë(-©/ìTí¹î¹ï 9 - sèå « u ¨"«Z¨ å (æJ© )2K¸ x$ç 3-f¸ r$ç 3f-¸ rç 3f-¸ r$3$ç 3Sr-¸ hç 3;h¸ ;ç 3h¸ 3x)ç7ë(-©/ìTí¹îìE 9 - sèå « u ¨"«Z¨ s (K«h©/é£ª(K¨ h¸j63ç 4)3K¸ 4Sr$ç 4$3-¸ 2)3ç 3;K¸ -/ç 3Sr-¸jç 3x-¸ 2f$ç 3)4K¸ )4$ç

s&tLu 'vBìJ >WCr]glty\_rbÄa[vpBµÌÏsWY[_WCr¢\ ÏsWCijk ¯ ~@afgK[ µe>WCr]WY[_o ¯ ¹`lhWP»!VY\MY;r]WP\¸

89ôE7+8_?vó ó-89ôJ7:ÈòA-:POÄA-Il8PD;FøNóK89ô ón?ôEOÄUCô?v7:P7U:P7;A-897;= øZõÉAhUAK=E7QORU;D:97fDM7fA-:CUCôE7:CD?O8Pô²DMü£7;UOnHNUOÂIl897;:97fD_8CDþNLIEóK89ôE7;:üEôE7;IEóKQ7IJóKI?v7UA-I+óhøJDM7;:9òh7:9óhQ 89ôJ7:97fDMHJT8CDORD89ôE7nü£ólóh:sü£7:)óK:PQ±A-IJU70óBe 78PôEó@=UUQ±A-üJüEOÂIJýN,Ê8CA-ýhDs8Pó7fA89HJ:97fD%þl÷Ê7"øN7;TÂOR7òh789ôEOÄDOÄDUA-HNDM7f=¾øZõ±8PôE7*TRAK:9ýh7ºIZHEQøN7;:ó#=@OÄD_8POÂIJU88PAKýhDv?ôEOÄUCô+AKüEü£7;A-:BORI³89ôE7*×_ÿÿeA-TRú@DQ278PA-û¢=EA-8PAJFdA-IN=Ê89ôE79ÌAKU889ôJA-8!8PôE7¾Q±A'ð_óh:9OÂ8_õÔóBB89ôE7¾897ùZ8PHJA-TBUóhIh8P7Il8!OÄDUóKIl8PAKOÂIJ7;=?O8PôEOÂIÔòK7:PõU7?8PAKýhD;FND9HJUCô²AKDa+níXé!9çKëé$q^ê`¢å(häé¢ëCãÈAKIJ=(âê é5RäþÿBôEOÄD0?vóKHJTR=U;A-HJD9789ôE7²ORIJDM8PAKIJU7:P7üE:P7;D97Il8CA89ORóKI óK:E ²789ôJóZ==Ê8PóÇôJAxòK7òh7:PõÉD9üJA-:CDM7+òh7;U8PóK:CD?O8Pô A7? 7;A8PHE:P7;DUóhIl8PA-ORIEORIEý2TRAK:9ýh7ºIZHEQøN7;:PDBóB897;:9Q±DFZ89ôE7;:97AK:97º89ôZHJDQAKIZõ±:97f=@HEIJ=EAKIl87;A8PHE:P7;D?ôEOÄUCô =EóÊIJó-8±üE:PóòlOÄ=@7+A-IZõÇORIóK:PQ±A89ORóKI89ôNA8ÈUAKIø£7+HJDM7f= óh:ÈU:P7;A-89ORIEý89ôE7±Qó@=@7TþJ 789ôEó@=)ï¼OÄD*IEó-8Af£7fU897f=ÊølõÔ89ôEOÄDAhD"89ôE7±7;Ih8POÂ:P7È=@ó@UHJQ27;Il8ORD"8P:97fA897f=AKDóKIE7ÈHJIEO8ó0897ùZ8óK:2OÂIJDM8PAKIJU7¼ýK7;IE7:CA8POÂóhIÇüEHE:PüNólDM7fDþdORIJAKTÂTRõKFe?s7¼óKøJD97:PòK7È89ôJA-8 78PôEó@= iA-IN= 789ôEó@= ï²TR7;Ah=)8Pó òK7:PõÇDMORQOÂTÄA-::P7;D9HETÂ8PD;þ#÷ 7+ø£7TROÂ7;òK7³89ôJA-889ôEOÄDORDUA-HNDM7f=)øZõ)A-IEóK89ôE7;:±A-:989OLÌAKU8ó089ôE7=EA-8PAKD978fþ#LTRTó08PôE7+OÂIND_8CA-IJU7;D2OÂI89ôJ7+=JA8PAhDM78UóKIl8CA-ORI8PôE7ÔDPA-Q7²D978¾óBaL, §gº×_Ã8CA-ýlDF 7;òK7;I OL*89ôJ7;D97Q2ORýKôl8¾ø£7²7;Q2üE8_õ óh:U7:98PAKOÂI¾OÂIJDM8PAKIJU7fDþZL0Dv?v7ºHNDM7"ÿq×_,Ê8Pó2D97TR7;U8vôJOÂýhôETÂõ:CA-IEúh7;=8P7:PQ±Dcóh:sORIJUTRHJD9OÂóhIOÂI 8PôE7¾ORIJDM8PA-INU7±òK7fU8PóK:fFA-IJ=Ê8CA-ýIJAKQ7;D!AKüEü£7;A-:!OÂI)AKTÂT 7ùEA-QüETR7;D;F89ôJ7õÊ?OÂTRT>IE7;òK7;:ýK78OÂIl89ó²89ôE7¥HNIJA-TsOÂIJDM8PAKIJU7¾:97;üE:97fDM7;Il8PA8POÂóhIþÿBôZHJD;F89ôJ7¾üETÄA-ORI 897ùl82A-IJ= Q78CAû]=EA8CAòK7:CD9OÂóhIJD*ó8PôEOÄD2=JA8PAhDM78A-:P7È7fD9D97Il89OÄA-TRTRõ 8PôE7¼DPA-Q7hþd÷Ê7¾?vóKHETÄ=)AKIh8PORUOÂüNA897È89ôJA-82ORIA±89:PHE7¥Z7Q±AKIh8PORU!÷ 7;ø =EA-8PAhDM78;FN89ôE72Q78PA-û¢=JA8PA¾?sóhHETÄ=øN72:PORUCôE7;:0ORIÔIJA-89HE:P7AKIJ=²OÂ8PDüE:97fDM7;IJU7n?vóKHETÄ=¼üE:9óòZOÄ=@7"Q2óh:97nORIóK:PQ±A8POÂóhI¾8PôJA-I³89ôJ7*üEHE:P7º8P7ùZ8ORIJD_8CA-IJU7;D;þ

â£ãJää0ä#äCç'!9ëCãè[-Kä'ðQ!9ä#X 5jé³UAKI²ø£7\óKHJIJ=ORIZeOÂýhHE:979ifï±AKIJ=iJþNÿBôE7\HJ:PDM8"UóKTRHEQIýKORòK7;D:P7;D9HET8CDóK:289ôE7+QHET8POû]UA-897;ýKóK:Põ üE:PóKøETR7Q+F#A-IJ=AKD27ù@ü£7;U897;=Fd8PôE7õA-:P7¼ü£óZóK:fþd"ûP°07;A-:P7;DM8IE7ORýKôZø£óKHE:ü£7:)óK:PQ±DQ!HNUCô³?vóK:CD97º8PôJA-Il°ºA)` aRòK7¼BAxõK7fDA889ôJ7Q!HETÂ89OÂû]UTÄAKDPDUTÄAKDPDMOLHNUA-89ORóKIF£?v7!ø£7TROÂ7;òK789ôEOÄD"ORDnUAKHJD97;=øZõhd"û¢IE7;AK:97fD_80IJ7ORýKôZøNóhHE:"øN7;OÂIEý³òK7;:9õD9HJDMûU7üE89ORøETÂ7È89ó8PôE7¾ORIJUTRHJD9OÂóhI óOÂ:P:P7TR7òA-Il8óK:!:97f=@HEIJ=EAKIl8A8989:POÂøEHE897;D;FdAhDn89ôJ7¾=@OÄD_8CA-IJU7Q2789:PORU*UóKQøEOÂIJ7"Q7fAKD9HE:97;Q7Il8PDjóK:AKTÂTó8PôE7a7;A-89HE:P7;D*îLi©-ðþ

ii

82

ñ DÃÚÝFE#G_dÞ HcÃÙhÚsÛÚZáªò&óM]ÜJÚ0á9Ú H

ÚJzy PM¯']FQSRS`TP

gºIÉ8PôE7ZZ7Q±AKIh8PORU³÷Ê7ø UóhIl897Il8±OÄD:97;üE:97fDM7;Il897;=òZORA AÊTRóKýKOÄUAKTBTRAKIEýKHJAKýK7¼OÂI ?ôEOÄUCôæäCçèNê è@ORDsUTR7;A-:PTRõ=@7HJIE7;=þh×]I±óhHE:cHJ:CDM8 ýK:PóKHEüó7ùZü£7:PORQ27;Il8PD;FK?ôEORTÂ7"DMóhQ7HJDM70?vAhDQAh=@7±ós89ôE7¼D_8P:9HJU89HE:P7üE:PóòlOÄ=@7f=²øZõÔQ±A-:PúZHEüFOÂ8PD*TRóKýhORU;A-T#IJA-89HE:P7?vAhDnOÂýhIEóK:P7;=þM¼võ7ù@üETRóKOÂ89ORIEýÔ8PôE7VHETRTüNóK897Il8PORAKTvóB89ôJ7l@7Q±A-Il89OÄU¼÷ 7;ø?v7³AK:9ýhHE7È8PôJA8±OÂ8D9ôEóhHETR=ø£7üNólD9D9OÂøJTÂ7n8Pó±TÂ7fA-:PI¼:PHETR7;DAKIJ=¼DM8PA-897;Q27;Il8PDORI²A2TÂóhýKOÄUA-T:P7üJ:97fDM7;Ih8CA8POÂóhIÈ89ôNA80OÄDD9OÂQORTRAK:89ó89ôNA8HJD97;=UóK:B8PôE7UóhIh8P7Il8;þ

ÚJz1 PSQFRBwô[hõ&[[cRöka¯'TT¯'¡X÷X ePe

÷ 7ôNAxòK7UCôEólDM7;I+89ó±HJD97*89ôJ76 :PóKýhóKT×]IJ=EHJU8POÂòh7óhýKOÄU*6 :PóKýK:CA-QQORIEýVÌ×_6DMõ@DM897;Q³þ6 :PóKýKóhT#ôJAKD!øN7;7IÉ=E7#HJIE7f=)AKD9LÆDM8PAKIJ=EAK:P= 6 :PóKTRóKý+ORIl897:PüE:P78P7:?OÂ89ô)ORIJ=@HNU89ORòK7³UA-ûüJA-øEORTRO8POÂ7fD))îLiSªðþd×¢8OÄDAKøETR7±89ó²TR7;AK:9IÇúZIEó?TR7;=EýK7ZÌ7ù@üE:P7;DPDM7f=ÊAhD!6>:9óhTÂóhýüE:97f=@OÄUA8P7;D2:9óhQ DMHEüJüETÂOR7;=Ê7ù@AKQüETÂ72ORIJD_8CA-IJU7;DnA-IN= D9HEüEü£óK:989ORIEý³øJAKUCúZýK:PóKHJIJ=ORIóK:PQ±A89ORóKIþÓÿBôJ7A-TRýKóK:POÂ89ôEQ ôNAKD*ø£77;I)DMHJU;U7fD9DHETRTÂõ HJDM7f= OÂIÇ7ù@üN7;:9ORQ7Il8PDaóh:!A-INA-TRõZD9OÄDnóQ!H@8CA-ýh7IEOÄUAKU8POÂòZOÂ8_õAKQ2óhIEýhDM8#IEOÂ89:PóhA-:PóKQ±A-89OÄU>QóKTR7;UHETR7;Dî ïð¢FK=@:PHEý*=@7fDMORýKIî ªxðNAKIJ=2üE:9óK897;OÂIDMôNA-ü£7üE:97f=@OÄU89ORóKI îni-ð¢þsÿBôE7Ô89ôE7;óK:Põø£7ôEORIJ= ×_d6ÆAKIJ= 89ôE7ÊóK:PORýKORIJA-T06>:9óhýKóhTA-TRýKóK:POÂ89ôEQ ORD=@7;DPU:POÂø£7;=¼ORI îLi-ðþJóK:óhHE:7ù@ü£7:POÂQ7Il8CDB?s7nHJD97;=±v6 :PóKýhóKThJþ Jþ ø

ÚJz,O |~@Se][`[7

÷ 7º7ù@üETRóK:P7;=8PôE7ºAKüEüETRORU;A89ORóKI±óB6 :PóKýhóKTEORI±89ôJ7"UóKIl897ùl8só89ôJ7°º<± 7fDM7fA-:CUCôE×]IJ=@7ùºA-8PAKD978fFlAhD>OÂ8BORD>89ôE7"TÄA-:PýK7:>óóKHJ:v=EA-8PAhDM78PDvA-IJ=±Q±AKüJD>7fAKD9OÂTRõ89óA6 :PóKTRóKý:P7üJ:97fDM7;I@û8PA8POÂóhIþÿBôE72×_ÿÿdAKTÂú@Dn=EA8CAKD978"ôJAKDnA¼Q!HJUCô :9OÄUCôE7:nD978nó>Q78CAû]=EA8CA¾?ôJORUCô ORD"QóK:P7=@OnmÈUHET8¾89óÇ:P7üE:P7;D97Il8±OÂI 6>:9óhTÂóhýJþ×]I óKHE:9,rùÉ6 :PóKTRóKýÊQAKüEüEORIEýJF>?s7ôJAxòh7+HJDM7f=89ôE7ÈD9OÂQüETR7EK(=EA-8PA³Qó@=@7T#óB>ú2 JícäCëé$[!9ä2--êÌëPç-é¢ä#$híwäCëécû±8P:9ORüETR7;DFÓ89ôE7±×_ÿÿeA-TRú@D=EA8CAKD978ORD7IJUóZ=E7;=³HJD9OÂIEý¥L, §gº×_>F@?ôEOÄUCôFE?ôE7;I¼8P:97fA8P7;=³AKDBüETÄA-ORIlK,vFNýK7IEû7:CA897fDQ±A-IZõ¼UóKQüETR7ù³:97;OnHNU;A8POÂóhI¾8P:9ORüETR7;D?ôJORUCô+?vóKHETÄ=+DMôE:PóKHN=¾8PôE7Q7;AKIEORIEý±ó89ôJ7=@ó@UHEQ7Il8CDFUóKQüJAK:97f=Ô8Pó+89ôE7¾OÂIl89HJO8POÂòh7±Q27fA-IEORIEýóBv8PôE7ÈD9OÂQüETR7:*8P:9ORüETR7;D\:9óhQ 89ôJ77;D97;AK:PUCôE×]IN=@7ùF@DMHNUCô³AhDné ãJäç @é ã7'!_é ã@êç!é¶êÌë5Âäê¥" @èJè£ç!N!ê æ2è£ä#þ

6 :PóKýKóhT-?BAKDd:9HEI!óKIA0:PAKIJ=@óhQ2TRõ*DM7;TÂ7fU8P7;=!D9HEøJD978ó]i$óE8PôE7s89ó-8CA-TWlïhïüNA-ü£7:CD:9óhQ 89ôJ7Z7;D97;AK:PUCôJ×]IJ=@7ùF AhD±:9HJIEIEORIEýÇ7ù@ü£7:POÂQ7Il8CD±?O8Pô 89ôE7QHJTÂTn=EA8CAKD978±?vóKHETÄ=ôJAxòK708PA-úh7I¾89óZóTRóKIJýJF@AKDv?v7º?v7:P7"UóKIl89ORIZHEóKHNDMTRõ28_?s7fA-úZOÂIJý!TR7;A-:PIEORIEý!üNA-:CA-Q78P7:CDsAKIJ=OÂIJDM8PAKIJU7B:P7üE:P7;D97Il8CA89ORóKINDþL0Dø£7#óh:97hF-?v7v:CA-IøEORIJA-:Põ7ù@ü£7:POÂQ7;Ih8CDóòK7;:d8PôE7=@OLf£7;:97;Il8UTÄAKDPDM7fDFøEHE8?OÂ89ô)6 :PóKýhóKT?v7ÈA8M8P7Qü@8P7;= 8PóTÂ7fA-:PI AUTRAhD9D9OLHNUA-89ORóKI¬óK:7fAKUCô óB89ôJ7iUTÄAKDPDM7fDþ@÷Ê7HJD97;=+A±D9OÂIJýKTR76 :PóKTRóKý2üE:97f=@OÄUA8P7*óe89ôE7aóh:9Q ê èÔ5Âç'%üÈý0ç!é¶êÌë5Âäþ08Pó:97;üE:97fDM7;Il8UTRAhD9DBQ7;Q!ø£7:CDMôEORüþJÿBôJORDø£7;U;A-Q7"89ôE78PA-:PýK78UTÄA-HJD97n?ôEORUCô6>:9óhýKóKT£?vóKHETÄ=89:Põ±89ó±TR7;AK:9Iþ

ÿ Ò[_a;kfa;t£ijr>bÂ[_WCWCtyz2Yw;Y;ijtY;~htjW0a;`ltjij`lWBbÂ[_af^ ¯ \]\_XTp qSq·#·#·¸ haof¸ ijof¸ Y;of¸ gh´@qvr ¯ ^q;mKa;bÂ\·Y[_W#q¸

ixï

83

ÚJz0Ú gU¤ Ã~¡7£Qy@£¥¯'e@¯õjT[cR

×]IEO8PORAKTÂTRõ+?s72UCôEóhD97AÈòh7:Põ³D9ORQ2üJTÂ72A-üEüE:PóhAhUCô³89ó¾Q±AKüZKÉ8Pó¼6>:9óhTÂóhýJþJLºD8PôE7K=EA8CAû¢Q2ó@=@7;T#:97;üE:P7;D97Il8PDsüX JíwäCëé`$M!9ä2-êÌëCçé]äX$aKícäPëépþ8P:9ORüETR7;D;F?s77QüETRóõK7f= AD9ORIEýKTR76 :PóKTRóKý"üE:97f=@OÄUA8P7UAKTÂTR7;=+é!ê [5ÂäþeORýKHE:P7\i"ORTRTÂHJDM89:CA8P7;D#óhHE: ORIEOÂ89OÄA-TJ:P7üJ:97fDM7;Ih8CA8POÂóhI2AKIJ=UóK:P:P7;D9üNóhIJ=EDe89ó!89ôE7,K, AKüEüN7fA-:POÂIJý*OÂI9eORýKHE:P7Jþ@°óK8970ôJó?É8PôE7,¼vORøJÿ7ù8_õZü£7Q±A-üJD89óÈANKÉUóhIJU7;ü@8?OÂ89ôEORI³89ôE7*IJAKQ7;D9üJAKU7ºóBd8PôE7;D97*7ù@ü£7:POÂQ7Il8CDþ

Ò_Á%¾À%ÝÄJÊÁ)ÝSÑÉPÊ2ÃË%ÄPÁ)Ì_ÍÀ2ÄPÁ_Î)ÂPÁ_Ï%Ð¿2Ã6ÄSÑÏ®Ã_¾Ò)ÄÇ_Ä_ÄPÁ×è¿2é4×è¿%ÄÃ4×hÃ6ÂÏ%ê_ÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6Ä4×dÓÒ6Ï%Ý)Ï×Ò_Á%¾À%ÝÄJä%Â_Â6Ë_Ò%¾Ò)ÝÄÑÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄÑÏLß_Á)ÂÃ_Ä_Ä6Å)¾¿%Æ)Ç"ÂPÎÒ_Ó%Ä,åPÌPÒ_Óæ¿Ò)ÄPÁ6¿2ÐPÒ%¾6Â¿2Ð_Ýà6áPÏ_À2ÂÇ_¾Ê_ÏaÂ¿,ãÂÏÀ_ÊÒ)ÄPÁ,Þ6Á%ÃÓ2¾PÒ)ÄÃÒ6Ê)Á)Ä)Ï×Ò_Á%¾À%ÝÄeÒáPÀ2ÄÑÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄÑÏLÓÒ_Ò6ÀB÷hê_êPú_ú_úB×hÃ_ÇPÅ'×hÐä)ÅP¿B×hÐÃS×èÊË%ê)ûÆ_Æ6Á2¾Ï_¿%ÄÇ_ê_Ä ó À%ê_ý)¾¿_ÀÁ%ÂÃPÄ_Ä6Å)¾¿)Æ)ÇAÏ×Ò_Á%¾À%ÝÄÐ6Å_Å_Á)ÄÇ_ÇÑÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏhç)Â_Ý6Åã)Â_ÐÇÒ4ÑÞPÊ2ÇÒ_Á%Ð_Ý¾6Ð)ÏS×Ò_Á%¾À%ÝÄeÒ%¾Ò%Ý_ÄÑÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏhÞß%ÄPÁ_Î)ÂPÁ_Ï%Ð¿2Ã6ÄàPÒ6Ê)Å_á,ÂPÎâ%ÄÏ%ÂPÁáãÂ¿2Ç_¾ÇÒ)Ä¿2Ã6á"â)Â6ÅÄ_Ý)ÇAÏ×Ò_Á%¾À%ÝÄáÄ_Ð6Á4ÑÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄÑÏ®åPÌ_ÌÍ)Ï×Ò_Á%¾À%ÝÄÐÊÒ_Ó%ÂPÁ4ÑÉ6Ê2ÃË)ÄPÁ%Ì_ÍÀ%ÄPÁÎ)ÂPÁ6Ï%ÐP¿2Ã6ÄÑÏLÖ×ØÊ2ÃË%ÄPÁ,Ð¿)Å,Ù4×LÚ6Û'×eÜ)Ð_ÄPÁÏ×

scthu 'v'ÿ£ sV"+`hoCa-hij`hkw¹`hij\_iYt£|>XlXh[_afY;o ¯ ¸

67:PôJA-üJD2AKD7ù@üN7fU897f=F 8PôEOÄDÈA-üEüJ:9ólAKUCôÉ=@OÄ= IEó-8ÈýhOÂòh7+òK7:PõÇýKóZó@=:97fDMHETÂ8PD¾AKD289ôJ7DM7fA-:CUCô¼D9üJAKU7óh:6 :PóKýKóhTNø£7;U;A-Q7n7ùZ89:P7Q7;TÂõÈTÄA-:PýK7hþ0HE7"8Pó6>:9óhýKóhTNóhIETÂõ¾ôJAxòZOÂIEý2óKIJ7üE:97f=@OÄUA8P7±89ó HJD97¾ORIÇ89ôE7+UóhIJDM89:PHJU8POÂóhIÇó89ôJ7¾:P7;D9HET8UTÄA-HJD97KF89ôE7+A-TRýKóh:9OÂ89ôJQ?vóKHETÄ=IhHJORUCúZTRõýK78*TRóhDM8n=Eó?I AUÌA-HETÂ8_õüJA-89ôÊó>8PôE7±DM7fA-:CUCô@û89:P77?OÂ89ôÊOÂINUóK:P:P7;U8*UóhIJDM8PA-Il8CDóK:ORIJUóh:9:P7;U8vHEIJOnHNU;A8POÂóhIJDFEAKIJ=³IE7òh7:B:P7;Uóòh7:fþ

gºHE:¥HJ:CD_8¼OÂQüE:PóòK7;Q27;Il8?BAKD±89óÉUCôJAKIEýK7+8PôE7²?BAxõÉ?v7²:P7üE:P7;D97Il897f=)8P:9ORüETR7;D;þ>×]I@ûD_8P7;AK=ÔósU;AKDM89ORIEýÈ8PôE7Q A-TRTd8Pó¾8PôE7±D9AKQ7!üE:P7;=@OÄUA-897hF£?v7U:P7;A8P7;=²A³6 :PóKTRóKý¾üE:97f=@OÄUA8P7UóK:P:P7;D9üNóhIJ=@ORIEý8Pó7fAKUCôU,ÉüE:97f=@OÄUA8P7KFZ7Kþ ýJþé!ê [5Räüç' @éã!2$ï# Jë2(hä#!#OB:XJä#!®#!æç-è£ëCäX$b WDJ Jë2(Kä!b þÆø£7;UAKQ7³ç' @éã!üï Jë%(hä!#O:XJä!®#'!æ±çè£ëCä#$9b DJ Jë2(Kä!bþþkZ7;UóKIJ=ETÂõhFJ?v7:97fUóKýhIEOÄDM7f=Ç89ôNA8¾6 :PóKýhóKTBóKü£7:CA8P7;D2óKI DM89:POÂIJýhDA-IJ= OLº8_?vó TÂOÂ897;:PAKTRD±AK:97+IEóK8Ôä:ðZçKëé5d37IlHJA-TF6>:9óhýKóKTôJAhD*IEó?BAxõ²óBýK7;IE7:CA-TROÄDMORIEý+óòK7;:º8PôE7Q+þeÿBôEORD!TÂ7f=ÊHJDn89óÔüE:97;üE:PóZU7;DPDA-TRTsDM89:POÂIEýlDD9ó+89ôJA-87;AhUCô ?sóh:P=Êø£7;UAKQ7¾ADM7;üJA-:CA8P7È6 :PóKTRóKýVÌAKU8;þEóK:!7ùEA-QüETR7KFdOÂI@ûD_8P7;AK=óeé¶ê é5Âäü5ÂäCç'!èJê èW?:SB!ê æèÓäX2$ab äCç'!èNê è@2!)-æ éãEä0åÓäæ±çèJéêÌë¾ì+äfí7b'þ?s7B?vóKHETÄ=ôJAxòK7£éê é5Âäü"5ÂäCç!èJê è@@?':4!ê æ2è£ä#2$b 5ÂäCç!èJê è@ebþ4$ éê é5Âäü"5ÂäCç'!èNê è@@?':SB!ê æ2è£ä#%$Nb äæ±çèNé¶êÌëTbþç-è[-!é¶ê é5Âäü5ÂäCç'!èJê èW?:SB!ê æèÓäX2$ab >ä;íTb/þ*þ-×]I±AK=J=@O8POÂóhI!8Pón89ôEOÄD#D9OÂQüETR7üE:P7û¢üE:9ó@U7;DPDMORIEý?s7ÈAKTRD9ó²A-üEüJTÂOR7;=ÇA+TRORDM8!óBD9õZIEóKIZõZQ±DóK:2UóKQQóKIJTÂõ HJDM7f=ÇA-øEøJ:97;òlOÄA8POÂóhIJD*A-IN= QOÄD_ûDMü£7TRTÂORIEýlDF£7Kþ ýJþ7[!)fë#$JM!);ë#A-IJ=Q[!)fëCäPä2--ê è@2A-TRTdQ±AKü89óE[!)fëCäPä2--ê è@CFê¢ðhé ãQ±A-üND8Pó<xéãZF 78PU-þjdORIJAKTÂTRõKF#?v7D_8CA-IJ=EAK:P=EORD97;=É89ôE7+:P7üE:P7;D97Il8CA89ORóKIónA-H@8PôEóK:±IJAKQ7;D289óíHJ:CDM8OÂIEOÂ89OÄA-TsüETRHJD!D9HE:PIJA-Q7KFdAKD\¼vORøJÿ7ù =Eól7fD*IEóK8DMü£7;UOnõ A²DM8PAKIJ=EA-:C=þÿBôJORD*Q7fA-IJD89ôJA-8+,5d @è !9äCäCëCäX$ !9äCäCëPäa+\D A-IN=l+,5n @è D "!9äCäPëCäAKTÂTQAKü³89ó9+\D "!9äCäPëCäþJL0D?O8Pô¼8PO8PTÂ7?sóh:P=ED?v7?vóKHETÄ=U:P7;A8P7;=2óKIJ76 :PóKTRóKýÌAKU8*óK: 7fAKUCô±A-H@8PôEóK:fþKL0I±7ùEA-QüETR7óÓóKHE:cHJIJA-T:97;üE:97fDM7;Il8PA8POÂóhI¼AKüEüN7fA-:CDvOÂIleOÂýhHE:P7i4ª@þ

iS

84

ÊÁ)ÝÉ6Ê2ÃË)ÄPÁ%Ì_ÍÀ%ÄPÁÎ)ÂPÁ6Ï%ÐP¿2Ã6ÄÑÏ1Ã¾Ò)ÄÇ6ÄÄPÁ×è¿%é'×è¿%ÄÃS×Ã6ÂÏ%ê6É6Ê2ÃË)ÄPÁ%Ì_ÍÀ%ÄPÁÎ)ÂPÁ6Ï%ÐP¿2Ã6Ä4×èÓ)Ò6Ï%Ý)ÏS×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏnÀÁ%ÂÃ6Ä_Ä6Å%¾¿)Æ)ÇAÏS×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏ®åPÌ6Ò6ÓÏ×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏ1¾¿)Ò)ÄPÁ6¿%Ð6Ò%¾6Â¿%ÐÝ)Ï×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏ1ÇPá6Ï_À%ÂÇ_¾Ê_ÏÏ×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏ1Ã6ÂPÏ_À_ÊÒ)Ä6ÁÏ×ä%Â_ÂPË_Ò2¾Ò)Ý_ÄPú%ÂPÁÅeÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏÐPÁ2ÃÓ2¾Ò)Ä)ÃÒ6ÊÁ)Ä$Ï×ÒáPÀ%ÄÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏLÓÒ_Ò6ÀB÷hê_êPú_ú_úB×hÃ_ÇPÅ'×hÐä)ÅP¿B×hÐÃS×èÊË%ê)ûÆ_Æ6Á2¾Ï_¿%ÄÇ_ê_Ä ó À%ê_ý)¾¿_ÀÁ%ÂÃ6Ä6Ä6Å)¾¿)Æ)ÇAÏ×Ð6Å_Å6Á)Ä)Ç_ÇÉPÊXÃË)ÄPÁ)ÌÍÀ%ÄPÁ_Î%ÂPÁ6Ï%Ð¿XÃ6ÄÑÏhçÂ_Ý_ÅãÂ_Ð)ÇÒ4ÑÞPÊXÇÒ_Á)Ð_Ý)¾6Ð)Ï×Ò%¾Ò)Ý_Ä6ú)ÂPÁÅÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏnÀ2ÄPÁ_Î)ÂPÁ_Ï%Ð¿2Ã6Ä$Ï×Ò%¾Ò)Ý_Ä6ú)ÂPÁÅÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏ1ÇPÒ6Ê)Å_á$Ï×Ò%¾Ò)Ý_Ä6ú)ÂPÁÅÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏnÏ2ÄÏ%ÂPÁáÏ×Ò%¾Ò)Ý_Ä6ú)ÂPÁÅÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏ1Ã_Â¿2Ç_¾_ÇPÒ)Ä¿2ÃPáÏ×Ò%¾Ò)Ý_Ä6ú)ÂPÁÅÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏnÏ2Â6ÅÄ_ÝÇ)Ï×áÄ_ÐPÁÉPÊ2ÃË)Ä6Á)Ì_ÍÀ%Ä6Á_Î)ÂPÁ6Ï2Ð¿2Ã6ÄÑÏ`åPÌ_Ì_Í)Ï×ÐÊÒ6Ó%Â6ÁÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄÑÏLÖ×ØPÊ2ÃË)ÄPÁÏ×ÐÊÒ6Ó%Â6ÁÉPÊ2ÃPË)ÄPÁ)Ì_ÍPÀ%ÄPÁ_Î)Â6Á6Ï%Ð¿2Ã_ÄÑÏÙ4×JÜ%Ð_ÄPÁÏS×

scthu 4v5J sV`loPa-hij`lk,w!mKWCoCa;`Z|sXlXK[_axYo ¯ ¸

ÚJz gU@£$Q®£

AKUCú!óDMüNAKU7BüE:P7òh7Il8PDHND*:PóKQÖüE:P7;D97Il8POÂIEýn8PôE706>:9óhýKóKTZ:P7;D9HETÂ8PDORINHETÂTþ0ó?s7;òK7;:;Fx?v7?OÂTRT =@OÄD9UHJD9DnD9óKQ77;A8PHE:P7;Dºó>8PôE7:97fDMHJT8CD"AKIJ=²?ORTRT#üE:9óòZOÄ=@7!ORTRTÂHJDM89:CA8POÂòh77ùEA-QüETR7;D;þEóK:QólD_8UTÄAKDPD97;Dv89ôJ7*QAð_óK:PO8_õ±óB8PôE7*:9HJTÂ7fD=@ORDPUóòh7:P7;=¾ølõ¼6 :PóKýKóhT£AK:97nóB89ôJ7aóK:PQVê èÔ5Âç42ü\ï# Jë2(hä#!#OB:XJä#!®#!æç-è£ëCä_þFQ7;AKIEOÂIJýÔ8PôJA8±6 :PóKýhóKT>?BAKDHEIJAKøETR7¾8PóZHJIN=AKIlõUóKQQóKIZ7;A-89HE:P7;DºøN78_?s7;7I OÂIJDM8PAKIJU7fD0óB 89ôJ72ýhOÂòh7IÔUTRAhD9D;FÓA-IJ= DMORQüETÂõ:P78PHE:PIE7;=²AKIê èÔ5Âç42)UTÄA-HNDM7v89ôNA8 TRORDM8PD AKTÂTE89ôE7ORIJDM8PA-INU7;D#=E7;UTÄA-:P7;=!89ó*ø£7ORI289ôJA-8>UTÄAKDPDþ-ÿBôEOÄD#üE:PóKøEûTÂ7;Q ORDAKTÂQólD_8U7;:M8CA-ORIETRõÈUA-HNDM7f=ÈøZõ±89ôJ7DMQ±A-TRTÓIZHEQ!ø£7:óBJ7;A-89HE:P7;DvHJD97;=¾89ó±=@7fD9U:9ORø£77;AKUCôÊORIJD_8CA-IJU7KFAKIJ=ÊóòK7;:9TÄA-ü øN78_?s7;7IÇDMóhQ7±óv89ôE7ÈUTRAhD9D97;D;FQ±A-úZOÂIJý³OÂ8=@OLmÈUHETÂ8\óh:89ôE7AKTÂýhóK:PO8PôEQ 89ó±OÄ=@7Il8POnõ³=@ORDPU:PORQ2ORIJA-89óK:Põ±ýK7;IE7:CA-TROÄD9A-89ORóKIJD;þ

07fDMüEOÂ897+8PôEOÄD±üE:PóKøETR7Q+F D9óKQ7+:9HETR7;D?v7:P7=@OÄD9UóòK7:P7;=Ç89ôJA-8¾UóòK7;:97f=)QóK:P7³89ôJAKIAÔD9OÂIEýhTÂ7¼OÂIND_8CA-IJU7KþEóK:27ùEAKQ2üJTÂ7hFeORýKHE:P7ZiS©ÔD9ôEó?D*8PôE7¼:PHETR7;DýK7;IE7:CA8P7;=í:PóKQ489ôJ7+&läèNé`ÓUTÄAKDPDF89ôE7*HN:PDM8]HJòh7#:9HETR7;DdA-:P7 D_8P:PAKOÂýhôl8óK:P?BA-:C=FA-IJ="7INUA-üNDMHETÄA8P7 óKøZòZORóKHJD[ÌAhU8PDA-ø£óKH@8*üEHEøJTÂOÄUA-89ORóKIJD*ORI 89ôE7ÈAK:97fA¾óB*çhäèJé897fUCôEIEóhTÂóhýKOR7;D;þM0ó?s7;òK7;:;F£89ôJ7ÈDMOÂùZ89ôÊ:9HETR7KFê èÔ5Âç42ü`+Lþ0né¶ê é5Âä&'!)-ü`+$¶í2-ê þBDRFeORDnQóK:P7OÂIl8P7:P7;DM89ORIEýJþk¼"0×"OÄD*A-IÇA-øEøJ:97;òlOÄA8POÂóhIóh:íPä5jêÌä%2$"-Kä#ê1!9ä#çè[-¾ê èJé]äèJéê`èFJA±UóKQQóKI+üJAK:PAh=@OÂýhQ ?OÂ89ôEORI²A-ýh7Il8PDB:P7;D97;AK:PUCôÊîni$Ùð¢þLI+AKU8POÂòh70:P7;D97;AK:PUCôE7;:>ORIÈ8PôE7*A-ýK7;Il8PD&HJ7TÄ=¾?vóKHJTR=EHJIJ=±8PôEOÄDA-TRQ2ólD_8BAKDvóKøZòZORóKHJDvAKDs89ôJ7ó-89ôJ7:*:PHETÂ7fDFøJAKD97;= óKI 89ôE7;OÂ:!úlIJó?TÂ7f=@ýK7±A-IN=Ô7ùZü£7:POR7IJU72óBs8PôE7NHJ7;TR=þeÿBôEOÄD*6 :PóKýKóhT:97fDMHETÂ8BORDs89ôZHJDBAüEOR7;U7nóBýK7;IE7:CA-T£úZIEó?TÂ7f=@ýK7hFh?ôJORUCô¼ORDvIEóK8vóhIETÂõ±HND9AKøETÂ7"ORIÈ89:PõZOÂIJý!8PóUTÄAKDPDMOLõ±IE7? :97fDM7fA-:CUCôÈüJAKüN7;:PDj:PóKQ 89ôJ7a7;D97;A-:CUCôE×]IJ=E7ùFZøEH@8UóhHETR=¼A-TÄDMó2ü£ó-897;Il89OÄA-TRTÂõøN7AKüEüETROÂ7f=¾óhH@8PD9OR=E7n89ôEOÄDB7ù@ü£7:POÂQ7Il8fþTZ7òh7:CA-TÓód89ôE7*óK89ôE7;:UTÄAKDPDMOLHNU;A89ORóKINDsAKTRD9óýK7IEû7:CA897f=2:9HETR7;D>óAD9OÂQORTRAK:#8_õZüN7hþK÷Ê7HJIJ=8PôE7;D97:P7;D9HETÂ8PD:PóKQ óKHE:s6 :PóKýKóhTZ7ù@üN7;:9ORQ7Il8PDòK7:Põ 7ùEUOÂ89ORIEýNFAKIJ=ÇüE:P7;D97Il8±AÔD97TR7;U89ORóKIÉORI eORýKHE:P7ZiZþÿBôE7³:9HETR7;D:CA-IEýh7E:PóKQ489ôJ7DMTROÂýhôl89TRõ²øEO0f;A-:P:P7KFÓDMHJUCô AKD±ç515MEç2Eä!2[ Jí5jê9ãEä2-Ôê è 045d @æ±ä98CÊç'!9äÊâÓãEä2'!X3F89ó+:9HJTÂ7fDUóKIl8CA-ORIEOÂIJýORIl897:P7;DM89ORIEý±A-IN=¾ü£ó-8P7Il89OÄA-TRTÂõ¾:97;HJD9AKøETR7ºúZIEó?TR7;=@ýh7KFED9HJUCô¼AhD ç%Jä#!2[ JíA5jê9ãEä2-+í#39Y!lçèhUç' #Cæ±çèJè ç%EäCç'!ê èW+ê èÇí%S('#ê éãl5RäPç!èJê èW³ê è éãEäéê é5Âäç!9äê èé ãJä1vä#5L-U_aYÔçhëPã@ê è£äÈäCç!èJê è@Kþ

<#ùEA-QORIJA8POÂóhI²ó 8PôE7±6 :PóKýhóKT:P7;D9HETÂ8PD"ORIJ=@OÄUA8P7!8PôJA86>:9óhýKóKTdORD"óòh7:)HE8M8POÂIEý¾89ó³89ôJ7üE:9óhøETR7Q+FÓølõ²U:97fA8POÂIEý³A¾TÄA-:PýK7IZHEQ!ø£7:nó"ê èÔ5Âç'%º:9HETR7;D;þÿBôEOÄDnORD"ø£7;UAKHJD97OÂ8*ORD"ORQû

i

85

;K

xf! #" %$ ¤&('! ) +*K',) " xf! #" %$ ¤&('! ).- +/KCK¡-(') " xf! #" %$0K f(';x! ]; " 1K f(';x! 2'¡K f " xf! #" %$03f-4! )5&f%66xx ]f ¢¡xf ]£ ¤x¥76(8x-PKK-6;9x(6:Efx) " x9f('x! ¢;; " 1- f;';x! ];; " xf! #" %$03;! <;>=== " K ;;(';x4! ];x;K " xf! #" %$0K f(';x! ¶¡K "

scthu 4vBJ »KoCWP[_XK\ a;bÒ[_afk;aftM>WCr]ghtj\_r#µd|sk;WC`\_r »KX@WP[_ij^*WC`\

K?Kfxf A@x x;h; xf! #" %$CB'¤f- ,! )ED@ FHG '¤xfK I'J'f' ÇIKKPZL-K') " xf! #" %$CB'¤f- ,! )%'CKf¤-K'K A/Kf¤hfKxxM) "

---¡Kx;-, xf! #" %$0K f(';x! ¶K;x&('¤hf " xf! #" %$0K f(';x! f;¤hKNx " xf! #" %$0K f(';x! ];xff-x " xf! #" %$0K f(';x! Ox-hf-'h "

F P&hQ/-;fK xf! #" %$x¤x¡( fP&-f! ) F 'x-SRK¤?lx) " ±¡''¥KP f;('x! ;;fh " xf! #" %$0K f(';x! ¶¡Kf " 1K f(';x! _f "

T ';x-Px@ xf! #" %$K;-M! EUUV " ¡(''¥K f(';x! _'?f;K; "

D;-¤KP3 xf! #" %$¡(''¥- f;';x! ¶KN3 " xf! #" %$0K f(';x! ¶KxfZ;¥l " xf! #" %$0K f(';x! ¢;h3fK' "

W &-';3 xf! #" %$0N' ¤h4! <;YX "

scthu 'vÁ@ Ò[_a;kfaft[sWCr]ghty\_rµ#mKY;^*XltjWjsghtWPr

78* ¦ % %'9 ;=<>% -@? ¦ u ( & ©C¨9 (K¬f«j.,.eï 2f-¸ ;$ç )4-¸ 3Srç )2K¸ 3f$ç ¸ rKç rx-¸ ç )4K¸ 3rç

scthu 'vÕJ sWPoY;tjtNbRa;[sÒ[_gl`lWC2Ò[_afkfa;tM>gltjWCrµJsWPr]WY[_o ¯ ¹¶`ZhW9»VBY\MY;r]W9\¸

i4ª

86

üNólD9D9OÂøJTÂ7óK:6>:9óhýKóhTJ8Pó2ýh7IE7;:PAKTÂOÄD97"AKIlõNHJ:M8PôE7:?OÂ89ôEóhH@80U:P7;A-89ORIEý2:9HETR7;Dv8PôJA80AK:97"IEóK8UóK:P:P7;U8óK:¥i$2ZÅóvAKTÂT8PôE7±ORIJD_8CA-IJU7;D;þÓ×¢8?sóhHETR= ø£7=E7;D9OÂ:CA-øETR7289óA-TRTRó?Ö:9HJTÂ7fD"89ôJA-8?sóhHETR=+UóK:P:97fU89TRõ¾UTRAhD9D9OnõhFJDPAxõKF¹ÙÙ#Z óB89ôE7ORIJDM8PAKIJU7fDF@8PôZHJD0AKTÂTRó?OÂIJý2Qóh:97*ýh7IE7;:PAKTÂOÂûD9A-89ORóKIJDBA-IN=ÈüE:PHEIEORIEý89ôE7*D978Bód:P7;D9HETÂ89ORIEý2:9HETR7;D;þ¼võÈ=@óKORIEý89ôEOÄDv?s7n?vóKHETÄ=ÈôEóhüN7HETRTÂõýK78A2DMQ±A-TRTR7:BDM78vóB:9HJTÂ7fDcóh:v7fAKUCô¾UTRAhD9D;FZA-IJ=¼A!QHJUCô¼D9Q±A-TRTÂ7;:vD978Bó>ê èF5Âç42vDM8PA-897ûQ27;Il8PD;þJó?v7òh7:fFAKD*6 :PóKýhóKTôJAhD*IEó+øEHEORT8OÂIÇQ78PôEó@=¬óK:!=@óhOÂIJý+89ôEOÄDF?s7ÈUCôEólDM728Pó8PA-úh7"A2òh7:PõÈDMORQüETÂ7*AKüEüE:PóhAKUCô¼AKDóKTRTRó?D@A-TRTÓ89ôJ72ê èF5Âç42s:PHETR7;DB?v7:P7"=EORDPUAK:P=@7f=¾AKIJ=óKIETRõ+89ôE7QóK:P72úlIJó?TÂ7f=@ýK7û:PORUCô:PHETR7;D":P78PAKOÂIJ7;=ZóK:n7;AhUCôÔUTÄAKDPDþÓ÷ ôE7;IÔ8PôEOÄDn:97f=@HJU7f=DM78vOÄDvHJDM7f=¥óh:vUTRAhD9D9OnH£UA8POÂóhIFh89ôE7"üE:P7;UOÄD9OÂóhIÈó8PôE7":9HETR7;DsOÄDvDM89ORTRTi$2Z AhD>IJó2AK:M8PORUTÂ7fD?OÂTRT7òK7;:0ø£7ORIJUóK:P:97fU89TRõUTÄAKDPDMOLHJ7f=FJôEó?v7òh7:º:97fUA-TRTe?OÂTRTdIJó¼TRóKIEýh7:ºøN72üN7;:7fU8fþÓÿBôJ7üN7;:PU7Il8PAKýK7"ó:97fUA-TRTR7;=¼OÂIJDM8PAKIJU7fDóK:7fAKUCô+UTÄAKDPDvAK:97nüE:P7;D97Il8P7;=¼OÂIleOÂýhHE:P7i$nEþ

[ \/ÔGdÛÞ^] Ý Ü`_

÷ 7±A-:P7Ax?vAK:97ósTROÂ8M89TR7?sóh:9ú²UóKIJU7:PIE7;=Ô?O8PôÊAKüEüETRORU;A8POÂóhIÔóvQ±AKUCôEORIE72TÂ7fA-:PIEOÂIJý¾8PóZ7Q±A-Il8PORU¼÷ 7;ø=EA-8PAJþÿBôJORD2OÄD!ORIUóKIl89:CAKDM8!8PóÊAKüEüETRORU;A89ORóKIND8Pó²89ôJ7³÷Ê7øFeóB0?ôEOÄUCô89ôE7;:97ôJAxòK7ø£77;I²Q±A-IZõKþ[Eóh:º7ù@AKQüETÂ7hF[Zõ@D9úZOÂTRTba ÷Ê7ø£7:982îLixðdHJD97;DºQ±AKUCôEORIE7TR7;AK:9IEûOÂIEý²89óÔAhUIlHEOR:97ÈA+Qó@=@7TBA-øJTÂ78Pó²üE:P7;=@OÄU8!?ôJORUCôÇTROÂIJúZD!óKI)A÷Ê7øÇüNA-ýK7¼A+HJDM7;:!?ORTÂTHJIJ= HJD97#HJT¶þs×¢8¼=@óZ7;DÈ89ôEOÄDÈøZõ A-IJAKTÂõ@D9OÂIEý)AÇD978¼ó*÷ 7;ø üJAKýK7;DÈQAKIZHJA-TRTÂõ:PA-897f=øZõAÈHJD97:fFJ?ôEOÄUCôÔAK:97*8PôE7IÔüE:9ó@U7;DPDM7f=+HJDMORIEý³D_8P:9HJU89HE:CA-Td×_897fUCôEIEOIlHE7;D;þM@õZD9úZOÂTRTca ÷Ê7ûøN7;:M8eHJDM7fDdA,°0A$` aÂòh7c¼BAxõK7;DUTÄAKDPDMOLHJ7:fFøJH@8d8PôE7sAKH@89ôJóK:CDA-TÄD9ó:97;üNóh:M8dOÂIZòK7fD_8POÂýlA8POÂóhIJDHJD9OÂIJýIE7;AK:97fD_8IE7ORýKôZø£óKHE: AKTÂýhóK:PO8PôEQ±DFx×_EFü£7:CU7;ü@89:PóKIJDAKIJ=!QHETÂ89OÂûTÄAxõK7;:eIJ7HE:CA-T@IE78_?sóh:9ú@D;þ÷ 7;øl?BA8CUCôE7:2înihF#ifïðF£TROÂúh7Zõ@DMúZORTÂTa ÷Ê7ø£7:98;F£ORD"AÈøE:Pó?D9OÂIEýÈAKOR=²?ôEORUCô A8M8P7Qü@8CD8PóA-IEIEóK8PA-897A¼÷Ê7øÔüNA-ýK7?O8Pô ORIóh:9Q±A8POÂóhI²óKI ?ôJA8nTROÂIEú@D"A¾HNDM7;:ºQORýKôl8,HNIJ=ÔHNDM7HET¶þÿBôE7!A-HE89ôEóh:PD7ù@üETÂóh:97f=AòA-:POÂ78_õ¼óBTR7;AK:9IJOÂIEýÈAKTÂýhóK:PO8PôEQ±DFNDMHJUCô²AhD÷ ORIEIEó? î Ù-ðeAKIJ=÷ óh:P=EDM8PA-8;þ[IJ=@7;:9TRõZOÂIJýÈ89ôJ7D9õZDM897;Q ?vAhD0A-IORIJDM8PAKIJU7:97;üE:97fDM7;Il8PA8POÂóhI+?ôEORUCô =@OR=ÔA8Mû897Qü@8!89ó7ù@üETRóKOÂ8QóK:P7óv89ôE7¾DM89:PHJU89HE:P7±óv89ôE7Eºÿ =@ó@UHEQ7Il8PD;FAKDnTRORIEúÔ8P7ùZ8;FôE7;Ah=@7:CDF78PUKþ?s7;:97289:P7;A-897f= =@OLfÓ7:P7Il89TRõKþJ78POpf;ORA î nð OÄDA³øE:9ó?D97:"ôE7;TÂü£7:fF=@OÄDMüJTRAxõh7;=OÂI¾A!DM7;üJA-:CA8P7?OÂIJ=Eó? IJ7ùZ8>8Pó*89ôE7ºHJD97:fö D>øE:9ó?D97:fþ-×¢8vüE:97û78PUCôE7fD>AKTÂTNóKH@8PýKóKORIEý*TROÂIJúZD:9óhQ 89ôE7±UHE:9:P7Il8nüJA-ýh72AKIJ= ?OÂTRT =@óA¾øJ:97fAK=Z8PôZHJ:CDM8nD97;AK:PUCôÔ89óAK=@òZOÄDM789ôJ72HNDM7;:"óhI?ôEORUCôÇTRORIEú@D*89óòZOÄDMOÂ8!IE7ùZ8;þe78POpf;ORAHNDM7fDÿrx×_ 8Pó²7ùZ89:CAKU8!UóhIh8P7Il8\:PóKQ üJA-ýh7;D;FA-IJ=±HNDM7fD 8PôE7º?v7ORýKôl8P7;=2897;:9Q±D>89óOR=@7;Il89OLõÈ=@ó@UHEQ7Il8CD>Q±A8CUCôEORIEý8PôE7ºHNDM7;:;ö D ORIl897:P7;DM8;þ

d e ÝsÚ0àGMß,Zá_Ý>Ú,

gºHE:È:P7;D9HET8CD±ôJAxòh7+=E7QóKIJDM89:CA8P7;=89ôNA8È8PóZ=JAxõ£ö D±AxòA-ORTRAKøETR7l@7Q±A-Il89OÄU÷Ê7ø Q±A-:PúZHEüUA-IJIEó-8ø£7>7ù@üN7fU897f=n89óºóKH@8PüN7;:óh:9QUóhIZòK7Il8POÂóhIJA-TQ±AKUCôJOÂIE7>TÂ7fA-:PIEOÂIJý0A-üEüJTÂOR7;=n8PóüETÄA-ORI897ùZ8fFE?O8Pô+:97;ýhA-:C=EDs89ó±AhUUHJ:PAhUõ±ód89ôE7*TR7;AK:9IJ7;=¼Q2ó@=@7;T¶þTó?v7òK7;:;FlOÂ8Q!HJDM8øN7*IJó-897f=89ôJA-8A-üEüJTÂõZORIEý+89ôE7¼D9AKQ27ÈAKTÂýhóK:PO8PôEQ 8Pó+89ôE7EHETÂTs897ùZ8!óA-I)AK:M8PORUTÂ7±óB©¾?vóK:C=ED;FA-IJ=28PóEi)nTROÂIJ7;D óBJK UóZ=E7KFK?ôJOÂTR7ºDM89ORTÂTNýK78M89ORIEý7SIlHJA-TRTÂõ2ýKóZó@=2üE:P7;=@OÄU8POÂòh70AKU;UHE:CAKUõ=@óZ7;D"UóKIND_8PO8PH@897A-I²ORIJU:97fAKD97OÂI üN7;:óh:9Q±AKIJU7!AKIJ=ÔDPUAKTRAKøEORTÂOÂ8_õKþN×]I²89ôJ7!÷Ê7ø UóKIl897ùZ889ôEOÄD>OÄDs7;D9üN7fUOÄA-TRTÂõ2OÂQü£óK:98PAKIh8vAKDvA-TRýKóK:POÂ89ôEQ±D#?ORTÂT£ø£707ù@ü£7;U897;=±89óDPUAKTÂ789óQORTÂTROÂóhIJDsóüJA-ýh7;D;þ4°07òh7:989ôE7;TÂ7fD9D;Fx?s7v:97;Q±A-ORI2D9óKQ7?ôNA8HEIJDPA89OÄDHJ7;=?O8PôóhHE:#UHE:P:P7Il8e:P7;D9HETÂ8PDóh:AIZHEQ!ø£7:eóE:P7;AhDMóhIJD;þLTÂ89ôEóhHEýKô!?s7vA8M8P7Qü@8P7;=*8PóKHJIN=*:97fA-TWZ7Q±A-Il8PORU>Q278PA-û¢=EA-8PAJF?v7AK=@QOÂ8 89ôJA-8s?v7A-:P7IEó-8BUóhQ2üJTÂ7897TRõ!ôNA-üEüZõ2?O8PôÈóKHE:s=EA8CAKD978PD;þKÿBôE70×_ÿÿdAKTÂú@D>=JA8PAhDM78

iS©

87

ORDn8PóZóÔDMQ±AKTÂT#8PóøN7¼A-øETR789ó²=@:CAx? A-IZõlHJ:PQ UóKIJUTÂHNDMORóKI¬:9óhQ AKIZõ²:97fDMHJT8CDFAKIJ=Ê89ôJ77;D97;AK:PUCôE×]IN=@7ù¾=EA-8PAhDM78ôJAKDBýh7IE7;:PA-897;=¼Q78CAû]=EA8CA:PóKQ ADMóhHE:PU7n?ôEORUCô+?BAKDBIE7;òK7;:Q27fA-Il808Pó¾üE:PóòZOR=E7*:97fA-T>æ±äCçèNê è@KþÓL0TRD9óJF£?ôE7;I89ôE7NZ7Q±AKIh8PORU!÷ 7;ø²ø£7;UóhQ7;Dº:97fA-TRO8_õhFQAKIZõ2D9HEüEü£óK:989ORIEý8P7;UCôEIJóKTRóKýKOR7;D>DMôEóhHETÄ=±øN7"AxòA-ORTRAKøETÂ7hFKQóhDM8sD9OÂýhIEOLHNUAKIh8PTÂõ2óKIl8PóKTRóKýKOÄUAKTDMHEüJüNóh:M8±A-IJ=Ç89ôE7AxòA-ORTRAKøEORTÂOÂ8_õ óB0ORI7:P7IJU7¼7;IEýKORIE7fDFe?ôEOÄUCôD9ôEóKHETÄ=ÉAKTÂTRó?GóK:27;AhDMõýK7IJ7:CA-TRORDPA8POÂóhIJDóJ89ôE7vúZOÂIJ=]+gf$ÓåNê æ 5Âç-é¢ä2-+ºèNè£äCç'5jê èWçèM- 2äPÓé ãµc!2éç'!9äºç'515Eé3%JäX_2åÓäCç!9ëPãÉç'5 @'!ê éãZæn?ôEORUCôÇUóKHETÄ=Ôø£7±HJD97;=Ô89óQÌAhUORTÂOÂ8PA-897±UTÄAKDPDMOLHNU;A89ORóKI 8PAhDMú@DnD9HJUCôAKD#8PôE7"óKIE7fD ?v70ôJAxòh70A89897Qü@8P7;=±ORIÈ8PôEOÄD>üJAKüN7;:;FløEH@8v?ôEOÄUCô¾A-:P7IE7fA-:PTÂõ2OÂQü£óhDPDMORøETR78Pó=@ORDPUóòh7:B?OÂ89ôEóhH@80A-IZõÈøJAhUCúZýK:PóKHEIJ=ÈOÂIóh:9Q±A-89ORóKIþ

h]zy ÃFQJ$QJ¯'ji T¯4¢

÷ 7ÊüETÄA-I 8Pó UóKIl89ORIZHE7Ô7ùZüJTÂóh:9ORIEý)OÄD9D9HE7fD¼UóKIJU7:PIE7;= ?OÂ89ôÃ6 :PóKýhóKT0AKIJ=dnIEó?TR7;=EýK7×]Ih8P7IJD9ORòK77;A-:PIEORIEýÈóhI89ôE7NZ7Q±AKIh8PORU!÷ 7;øFÓüE:POÂQ±A-:PORTÂõ+A89897Qü@8POÂIJýÈ89ó¼H@89ORTRORD97!øJAhUCúlûýK:PóKHEIJ=ORIóK:PQ±A89ORóKI89óÇôJ7TRü 6>:9óhýKóhTsýh7IE7;:PAKTÂOÄDM7³ø£78M8P7:¼óòK7:±UTRAhD9D97;D;þ÷ 7üETÄA-I 8Pó7ù@üETRóK:P7ÈýK7;IE7:CA8POÂóhIÊóB0DMHNUCô)øJAKUCúZýh:9óhHEIJ=ÊOÂIóK:PQA-89ORóKI :PóKQ4óKIl8PóKTRóKýKOR7;D:977:P7IJU7;=OÂI³89ôJ7*Q278PA-û¢=EA-8PAJFZAhDB?s7;TÂTdAKDs89ôE:PóKHJýKô¼89ôE7*HJD97nóýK7;IE7:CA-TÓøJAhUCúlýh:9óhHEIJ=¾OÂIóh:9Q±A-89ORóKIDMHJUCô+AKDAhDDMõZIEóhIZõlQ±DBóh:v?vóK:C=³DMORQOÂTÄA-:PO8_õhþ k-þ

÷ 70AK:97Bòh7:Põ!ORIh8P7:P7;DM897f=ORI89:PõZOÂIJýn89ó!AKüEüETRõ89ôJ7:97fDMHJT8POÂIEý6>:9óhTÂóhý"UTRAKHJD97;D#óhH@8PD9OR=E789ôE7nóK:PORýKORIJA-T£7ù@ü£7:POÂQ7Il8fþEL0Dv89ôJ7":P7;D9HETÂ8PDAK:97HJ:PDM8óK:C=@7:vTRóKýKOÄUºO80D9ôEóKHJTR=¾øN7nü£óhDPDMORøETR789ó¾Q±AKü²89ôJ7Q øJAhUCú¾8Pó¼AÈ:P7üJ:97fDM7;Ih8CA8POÂóhIOÂIíK, óh:ºA¼DMORQOÂTÄA-:"TÂóhýKOÄUøJAhDM7f=VóK:PQ±A8;F89ôZHJD7ùZü£óK:989ORIEý289ôJ7±æNS-Kä5l89ôJA-8?vAhDBTÂ7fA-:PIE7;=þ

0HE7B8Pón89ôE7UHE:P:97;Ih8>D9ôEóh:M8CUóKQORIEýhDeóBkZ7;Q±A-Il89OÄUB÷ 7;ø±=EA8CAn?s7üETÄA-I89ó=EóaHE:M8PôE7:7ù@üN7;:9ORQ7Il8PD?OÂ89ô óKHJ:UHE:P:97;Ih8±=EA-8PAKD978CDþ#6>:9ORQ±A-:POÂTRõÇ?s7³ORIh8P7IJ=8Pó :P7û]UTÄAKDPDMOLõ 89ôJ77;D97;AK:PUCôE×]IN=@7ù+üJA-ü£7:CD"øJAKD97;=²óKIÔü£7:CD9óKIJAKTdORIl897:P7;DM8;FN89ôZHJD"QóòZOÂIJýEHE:989ôE7;:"89ó?BA-:C=ED89ôE7¼üN7;:PD9óKIJAKTÂOÄDPA89ORóKIÇAKIJ=ÇTR7;A-:PIEORIEý²HJD97:QóZ=E7TÄDDPU7;IJA-:POÂólD*HJD97;=)AhDQó-8POÂòA-89ORóKIóh:89ôEOÄD ?vóK:PúÓþ-÷ 7º?sóhHETÄ=2AKTRD9ó*TROÂúh789ó:PHEIÈ6>:9óhýKóKT@?OÂ89ôÈ89ôE7HJTÂTM7;D97;AK:PUCôJ×]IJ=@7ù=EA8CAKD978fFIEó-8Kð_HND_8A²DMQ±AKTÂTvDMHEøNDM78!óB89ôJ7¾ORIJD_8CA-IJU7;D;FAKD!?s7;TÂTvAKDA8M8P7Qü@8POÂIEý²89ólHJIN=)AýKóZó@=?vAxõóQ±A-üEüJOÂIEýL §gº×_+89ó6 :PóKTRóKýJF-D9ó*8PôJA8>?v7Q±Axõ2:9HEIÈ6>:9óhýKóhT@7ù@üN7;:9ORQ7Il8PD?O8Pô¼8PôE7*×_ÿÿeA-TRúZD=JA8PAhDM78;þ

\ l%ÜÚ0àMe

f¸>af~@W9[]\|>[_^*r¢\][_a;`lkhEVBYzK`hWh[_WPij\MYkh ®¯ a[_r¢\_WC`gmfafY;o ¯ ij^*rCNY;`l ® a;^ ½2iy\_o ¯ WCtjt ¸"°+WP~hµ·Y\_o ¯ WP[Xp|tjWY[_`lij`lkvY;XlXK[_WC`\_ijoCWdbRa;[Ó\ ¯ W·ea;[_t0· ihW·WC~J¸¹`onnpnqcrs2tvuxwyzr7|ps(~ux|w+qvw>Ytv|uw0#tvuxwy;@XZYkfWCr12#wlÓ/3)3x-¸

-¸ ® ij^ eWP[_`lW9[_r¢µ¿JWCWf¸ ° ¯ Y\ \ ¯ W r]WC^Y`\_io ·eWC~ ijr]`Jq \ ~lgK\ oY` [_WPXh[_WCr]WC`\¸¯ \]\_X7p qq·#·#·¸ ·&rK¸ a[_k'qVvWCr]ijkf`l¹¶r]r]glWCrq#sV"Ó`la;\¸ ¯ \_^*t /3)3$4K¸

rK¸0m-oCa;\]\Óceafr¢\Y;`Zm\_WCwxWC`mhY;tC~@W9[_kh¸|³·eWCijk ¯ \_W`lWCY[_WCr¢\£`lWCijk ¯ ~@a[Y;tjk;a;[_iy\ ¯ ^ÉbÄa;[£tjWY[_`hi`hk· iy\ ¯ r¢zK^º~@aftjijovbÄWY\_gh[_WCrC¸vuxw`p4Etvw#uxwy;£p fXw-)4-/3$3rK¸

h¸0m@¸£VvgZlY`li ¸ ®¯ WKir¢\MY`loCW9µÌ·eWCijk ¯ \_W³´µÌ`lWCY[_WCr¢\]µÌ`lWPik ¯ ~@afgh[[_ghtW;¸q%c#tEw7~<uw7~wr#~v5|~ww2vtwu<~_Jml½±cdµe2K )p rx;Xwrff/3x;-¸

®¯ ijr lY\MY oPafglt bÄa[ WP»hY;^*XhtW ~@W \MY´xWC` bR[_a;^ °+a[MhÏsWP\ ¯ \]\_XTp qSq·#·#·¸ oPafkfr]oPi¸ XK[_i`hoCWP\_a;`N¸ WhgWqs· `Wqf

i

88

-¸ÒYglt° ¸kÓi`h`Ndm\_WCX ¯ WC`Ô½2glk;kftjWP\_af`JVYwKiÒYkfWfdY;`l²|sr ¯ · ij`Êm[_ij`lijwfYr_Y;`J¸²Ò ¯ Y[]µ^YoCafX ¯ a[_WÈhijr]oCawxWP[]zÊglr]ij`lk\ ¯ W±ij`lhglo9\_iwfW±ta;kfijo³XK[_afk[MY;^*^*ij`lk+r¢zKr¢\_WC^Ò]>ÁÁB¿¸1vuxw4Etvw#uxwy;r;hÌµrf)p hwKfK/3)3$4-¸

2K¸0mKY;ty\_af` Y;`legho9´-tjWPzcs¸ ® WP[_^nµ·WCijk ¯ \_ij`hk*Y;XhXh[_axYo ¯ WCr>i`ÈYgh\_a;^Y\_ijo\_WP»-\>[_W9\][_iWPwfYt¸qvwYYtv|uwt%>v~v~uxwywow2>y| w7 Ep -2r#wK;r-63$4)4K¸

-¸pWP`n¿£Y`lkK¸ZÏvW9· r¢·WCWKWP[Xp¿JWY[_`hij`lkB\_avÀltj\_W9[d`hWP\_`lW9· rÇl¹¶` t%>vvEuwy~p%bqvw#5tvwuw¡w>>tEwvow+vuxw`p4Etvw#uxwy;ZXZYkfWCrjrSr-)w'rSr$3-/3)3x-¸

4K¸ º WC`h[]z0¿JijWC~@WP[_^Y;`J¸h¿NWP\_iCiYBpf|s`"Y;k;WC`\\ ¯ Y\Yr]r]ijr¢\_r·eWC~"~K[_a· r]ij`hkh¸Z¹¶`nc ¯ [_ijrem@¸f½2WCtjtijr ¯ WChiy\_a;[MtE>vEuxwy~%¢#¢£,t5w71qvw75tw2uw2¡¤uxw7zw>>tEwvw1ntvu ¥c<u¡qvw75<¡x¡ u¦yw29£XZYkfWCr>3xw3f)3-¸d½2a;[_kxY`pY;gKbÄ^Y;`h`³Xhgl~ltjijr ¯ W9[_rº¹¶`lof¸dpÓmhY;`½Y\_WCaKc|0uBmh|#w ;/3$3f-¸

3K¸Ï0¸K¿Niy\]\_tjWCr¢\_af`hWf¸¿NWCY[_`lij`hknÎglijoM´Ktyzn· ¯ WC`2iy[][_WCtjWCw;Y;`\ Y\]\][_ij~lgh\_WPr Y;~@afgh`ZpK|`lW9·Çtji`hWY[]µ\ ¯ [_WCr ¯ a;t2Y;tjk;a;[_iy\ ¯ ^¸vuxw`p4Etw7uxwy;@K x)p A4x2wrK64K/3)4$4-¸

CK¸Vº¸N½2io ¯ ijWfJV"¸£m-XlijWCkfWPt ¯ Y;ty\_WP[NY;`l³cs¸ ® YCzKtja;[¸§¨tE¡cw+r#.ux~u¡p¡~~u ¥cvuwlo ¯ YXh\_WP[sKlXZY;k;WCrA3xXwlXr;fK¸#tjtjijr º a;[]·aa-E£63$3;K¸

;f¸ ® ¸ ½2ij\_o ¯ WPtt ¸ª©p~vuwI,tw7uxwy; o ¯ YXh\_WP[Û2K>½Yo ¯ i`hW+¿JWY[_`hi`hkh#XlY;k;WCr³w-;f-¸½2o/s[MY·#µ º itjt £63$3f-¸

-¸Vº¸d½2tÂY;hWC`hijof¸ usr]i`hk¼\_WP»-\*tjWY[_`lij`lk¼\_a ¯ WCtjX ·WP~ ~K[_a· r]ij`hkh¸Ç¹¶`JtE>vEuxwy~%#« ¬qvw75tw2uw2¡cw>>tEwvw+|w#®v|s25tqvw#5t%<uwh¸ º ce¹e¹`\_WP[_`lY\_ijaf`lY;tf-fZÏsWP· Ás[_tjWY;`hrC@¿£|Kumh|l;f-f¸

2rK¸0m@¸-½2ghkfkftjWP\_a;`N¸E¹¶`KwfWP[_r]W WC`\MY;ijtj^*WC`\eY`Z*Ò[_afk;aft ¸4§¨¯IwtEuwg|ps2uxwy4rs7E<u¡u~v~#wqvw27<ux°yu±tEyt%|o|uxwy;NXrh®rµÌ)p xXwKA4$2-63$3f-¸

Ph¸0m@¸l½2glkfk;tjWP\_af`J@¸Kpi`hkhhY;`l½¼¸lm-\_W9[_`K~@W9[_kh¸Ò[_a;\_WCij`r]WPoCaf`llY[]z!r¢\][_ghoP\_gh[_WBXK[_WhijoP\_ija;`ghr]ij`lkntjafk;ijoPµÌ~ZY;r]WC^Y;o ¯ ij`lWstWCY[_`lij`hkh¸t%5uxw+wyuxw`tvuxwy;@KÌ;9£/3)3x¸

-¸0m\_WCX ¯ WC`2½2glkfk;tjWP\_af`ÈY`Z+m;a ¯ `NÓij[]\ ¯ ¸dceXh[_a;kfaftjK¸ BpJ|É\_gK\_a;[_iY;tNi`\][_a-KgloP\_ijaf`J¸62K¸ ® WP[][]zÒYCzK`lWf¸7²±ux| w7~uw¡ ux³E7<uwńw³5s2tE~Yw7.uw´£,t§ËtE~§üyP7t

,tw7uxwy;¸dÒ ¯ V \ ¯ WCr]ijrCZuv`hijwxWP[_r]iy\z!a;b|s~@WP[MKWCWC`NE63$3$3-¸-¸0½2ijo ¯ YWCtMmh¸hÒYYY`li 7mxY;oM´½2gh[MY^Y\_r]gNKY`Z*VY`lijWCtJeijtjtr]ghrÇm-zKr]´-ijtjtZ·WC~@WP[]\XpK¹hWC`\_iybRz-µ

ij`hk*i`\_WP[_WPr¢\_i`hkn·WC~±r]iy\_WCrÇ¹`tE>vEuxwy~o%z#zn|tuw§zuw2¡bwY>tw2´wntvu ¥c<u¡2qvw#5<¡x¡ u¦yw2vµ7¡¦¶·hXZYkfWCrs#w2-fÓ63$3$2-¸

64K¸0½ ÒÓa[]\_WP[¸|>`2Ytk;a;[_iy\ ¯ ^ bRa;[ r]g7¸"»r¢\][_iXhXlij`lkK¸ ® WCo ¯ `hioCY;tsWPX@a;[]\CKhÒ[_a;k;[MY;^J/3)4f-¸63K¸|¸lmKtjaf^Y`Y;`ls¸l¿Na;kxY;`J¸d|>[_o ¯ iy\_WCoP\_gK[_WCr Y;`l\_aaftjr bÄa;[ ¯ gl^Y`hµÌtjij´xWBYkfWC`\_rC£63$3)4K¸K¸|¸@m[_ij`lijwfYr_Y;`JW¸lV"¸lpij`lkh@mE¸Z½2glk;kftjWP\_af`JEY;`l±½¼¸mK¸E ¸@m-\_WP[_`-~@WP[_kK¸#cY[_oCij`la;kfWC`hWCr]ijr

XK[_WhijoP\_ija;`lrglr]ij`hk¹¢¿NÒ ¸*¹¶`m@¸NVo¹CWP[_afr]´-idY;`l¼Ïº¸N¿NYw[MY(¹of£WChiy\_a;[_rC4t%>vvEuwy~¬%´#º qvw75tw2uw2¡p»+tE¼~E(sjw+qvw27<ux°yu±tEyt%|o|uxwy;lwxaftjgl^*W0A3x-lXZYkfWCr;rXwK)4f-¸Jm-Xh[_ij`lk;WP[]µ`xdW9[_tÂYkhJ/3)3x-¸

f¸0½¼¸em\_af`lW;¸Ô|sr¢zK^*Xh\_a\_ioPrºbÄa[!Y;`l YkxY;ij`hr¢\oP[_afr]r¢µÌw;Y;tjilY\_ija;`N¸I©u|tu½¼S2p )3Xwrf-63x;-¸

i$n

89

Semantic Methods and Tools for Information PortalsThe SemIPort Project

Jorge Gonzalez-Olalla, Gerd Stumme

Institute of Applied Informatics and Formal Description Methods AIFB,University of Karlsruhe, D–76128 Karlsruhe, Germany

http://www.aifb.uni-karlsruhe.de/WBSjgo, [email protected]

1 Objectives and Research topics.

The goal of the SemIPort project is to develop innovative methods for presenting, stor-ing and accessing scientific information within a Semantic Information Portal. Theproject will enable users to access a semantically structured information inventory,while allowing them to integrate their own information.

Several scenarios as the AIFB portal and the DBLP online bibliography will be usedto test the developed methods. Later, the results of the project will be implemented inthe form of an Internet Portal for the German Informatics Society (GI) as a competencyand service network .

The project is financed by the German bmb+f for a period of 3 years starting June2002. The participants are the Institute for Applied Informatics and Formal DescriptionMethods at the University of Karlsruhe (coordinator), the German Research Center forArtificial Intelligence, the Fraunhofer Institute for Integrated Publication and Informa-tion Systems, and the Data Bases and Information Systems Group at the University ofTrier.

Theresearch taskswithin the project are:(1) Ontology Modelling and Metadata Standards. (2) Scalable Storing, Processing

and Querying of Integrated XML and RDF Inventories (Knowledge Warehouse). (3)Web Mining and Ontology-based Knowledge Integration. (4) Visualization and Brows-ing of Complex Data Inventories. (5) Personalization and Agent-based Interaction.

The most related to the workshop is task (3). This topic will concentrate on:i) Semantic Web Content and Structure Mining. The explicit semantics of semantic

web data will be used to improve the results of ’classical’ web and Data Mining tech-niques. E.g. for ’hubs and authorities’ one could distinguish between different kinds ofauthorities.

ii) Semantic Web Usage Mining. The behavior of the portal’s visitors will be usedto optimize the portal’s structure and usability, by developing tools that will help tomanage and update the ontology according to the actual needs of the users. A ’semanticlog file’ will be created and analyzed for this purpose.

iii) Flexible techniques for knowledge integration will be developed. Existing meth-ods of schema integration for relational data bases and ontologies will be analyzed andexpanded for dynamic integration.

90

!"

#$$% &'()

*+'), -

./$$ $

0 (1 ())(

! " # " $ %" & " '(

" % ) *+

' ",-.

"/ &+ "" "/ "/ ,0. % "" " $ %/ " 1 " 2 $ $ ",3. 24 " " / " " +4 5 "/ 6 2 ) *+ 7) "" " 24 ' $ " % ""

,-. + 789 )8/ : ' "/;

/ 7/ #/ < 0==0

,0. )8 2 79/ : + 24 + " > " 8 + 6/; !!

) % ) ?% @"/ / &-A 4 B 4B & "/ > / #/ --CD-33/ #? / 78 % E F" #@/>/ < 0==0

,3. )8 2 79/ :? " "" " " /; " #$ !! #

$ %& ' ()* !!/ 4 &/ / 20==0 91

1

Semantic Web and Retrieval of Scientific Data Semantics

Goran Soldar Dan Smith School of Computing and Mathematical Sciences School of Information Systems University of Brighton University of East Anglia [email protected], [email protected]

Project Description

Introduction The results of scientific activities that include observations, experiments, interactions, deductions, etc. are stored in data sets and often made available to the scientific community. These large volumes of scientific data are typically kept and managed in an ad-hoc manner and it requires a substantial effort to discover and evaluate data quality and suitability for a particular analysis. The provision of appropriate metadata and links to related resources enables the possibility of intelligent assistance in finding and evaluating data sets for scientific research. Such tools are of increasing importance as the volume of scientific data is increasing rapidly due to the extensive deployment of automated data collection and monitoring systems. To find and process scientific data sets available on the Web is time consuming despite the fast and powerful search engines. For example to find files related to temperatures we used the Google search engine. The key word “Temperature” was entered and search restricted to the University of East Anglia Climate Research Unit (http://www.cru.uea.ac.uk). Google returned 773 pages. For a scientist browsing every single web page to find required data sets, without guarantee that they will find what they need, is simply too much time consuming exercise. It would be desirable that data files are described in a way that they could be quickly found and their semantics learnt automatically by a machine. The framework for such a semantic retrieval of scientific data is offered by W3C in the form of Resource Description Framework (RDF) [1,2]. This project seeks to address methods and provide an architecture for describing, managing, retrieving, and extracting semantic information from specific science domains. The conceptualization of the subject domains including the development of the vocabulary [3], is performed using RDF model and schema constructs. We distinguish between two levels of metadata related to RDF, Instance Metadata and Schema Metadata. The description of a particular resource using RDF constructs is known as the instance metadata, whereas Schema Metadata describes a particular ontology The notion of a semantic case is being introduced to capture semantic forms. The heterogeneity of data sources is exposed and integration of data is achieved through a mediation process [4]. The primary goal of our project is to enable machine processing of semantic information that would be utilized to query the actual information from data sets To address the problem of extracting semantic information from data files, we build an ontology for the meteorology domain which is then use to create semantic cases as file description templates. We use RDF Model Syntax and RDF Schema to create semantic cases instances. The architectural infrastructure is based on the Semantic Retrieval Model [5]. Storage and Management of Ontologies The management of RDF structures is an open issue, and W3C does not recommend any particular method for manipulating such data. By definition RDF Model is set of triples. This property can be utilised to achieve manipulation of RDF triples as a Relational Model. The storage and manipulation of RDF graph structures represents a problem. It does not matter whether the data is kept in RDF syntax or as triples, the problem is that there is no RDF data management system that provides access and manipulation of ontologies. In addition the existing RDF parsers are design to work on single documents only, so the problem is how to access documents remotely, and how to achieve the maintenance either of parts or of the whole documents in terms of modification, insertion, and deletion. Another question that arises is whether there is a need to keep whole RDF documents at all. Since modern database management systems are based on the concept of the Relational Model it is considered to use the existing RDBMS for manipulation of RDF data or alternatively to build RDF native data storage model.

92

2

Retrieval and Extraction of Semantics (RDF Triple Engine) The system is based on client-server architecture that includes specialized RDF servers. The advantage of this approach is in data management facilities of RDBMS that are utilized for the manipulation of raw RDF data. The architecture comprises the 4 layers: 1) Interface Layer, 2) Web Infrastructure Layer, 3) RDF Management Layer and 4) DBMS Layer The interface layer provides access to semantics descriptions to human users as well as to application programs. The prototype of the Semantics Retrieval Language (SRL) is currently being developed to provide semantics-orientated retrieval of DBMS managed RDF data. QBE-like Web form is used to assist users in specifying their requests although SQL-like statements are also supported. The interface handlers (access rights and authentication) are built as Java servlet modules and they run inside the Apache Tomcat servlet engine, which is seamlessly connected to the Apache Web server. All valid requests are transparently forwarded to the Semantics Support Server (SSS). Since SSS runs as a separate service, applications can establish direct connection with this server using the internal SRL protocol over TCP/IP connection. RDF Triples Engine (RTE) is a module responsible for manipulating triples and executing semantic queries. The check for namespace existence for each prefix is performed by RTE before the records are inserted into the database. A user with no knowledge of graph structures and its specific elements (containers) will find it difficult to assemble the full picture from the output. This problem is solved by adding the additional query processing to RTE, which is aware of RDF semantic and also is able to produce results suitable for human use (HTML document that conforms to XHTML structure) or for further processing (XML structure). References [1] WWW Consortium "Resource Description Framework (RDF), Model and Syntax Specification", Available at: http://w3.org./TR/REC-rdf-syntax. [2] WWW Consortium "Resource Description Framework (RDF) Schema Specification", Candidate Recommendation, Available at: http://w3.org./TR/1999/REC-rdf-schema. [3] T. Gruber “Towards Principles for the Design of Ontologies Used for Knowledge Sharing”, International Workshop on Formal Ontology, Padova, Italy, 1993. [4] D. Smith, M. Lopez, “Information finding and filtering for collections of semi-structured documents”, Proc. INFORSID XV, Toulouse, 353-367, 1997. [5] G. Soldar, D. Smith “Retrieving Semantics from Scientific Data: A Domain Specific Approach Using RDF”, Proceedings of the IASTED International Symposia on Applied Informatics, Innsbruck, Austria, February 2001

93

Data Fusion and Semantic Web Mining: Meta-Models of Distributed Data and Decision Fusion

Vladimir Gorodetski, Oleg Karsaev, Vladimir Samoilov St. Petersburg Institute for Informatics and Automation

gor, ok, [email protected]

Abstract:

According to the Project funded by European Office of Aerospace Research and Development (AFRL/IF) we are developing mathematical model, multi-agent architecture and technology realized as a software tool supporting design and implementation of Data Fusion (DF) applications of broad spectrum. A multitude of tasks to be solved with regard to the development of DF software tool can practically be divided into two groups. The tasks whose solutions make use of methods, models and technologies of other adjoining scientific fields, for instance, data mining and knowledge discovery, multi-agent systems, object-oriented design, etc fall into the first group. The second group includes the tasks specific for DF systems and require development of specific methods, models and technologies. In fact, the most part of tasks of the last group fall into the interests of Semantic Web Mining. Although the tasks of both above groups are the subjects of the Project, below the DF specific tasks are only highlighted.

In its essence, DF task is one of making decisions (as a rule, classifications) on the basis of distributed data sources presented by distributed databases with access through Intra- or Internet. These sources contain data that can be represented by different data structures (temporal, sequential, transactional, relational) , they can be of different physical nature (images, signals, truth values, etc) and measured in different scales (Boolean, categorical, real), be of different accuracy and reliability, they can be uncertain, contain missing values, etc. The objective of a DF system is to combine useful information from all of these sources to make decision, for instance, classification of an object, object state, situation, etc.

Within DF specific tasks two classes of them are of most significance. The first is development of meta-model of distributed data sources and the second is development of meta-model of combining decisions produced on the basis of particular sources. Note that the former task is purely Semantic Web related.

Three main issues of DF specific R&D of the Project are the subjects of presentation. 1. The ontology-centric approach which is considered as a basis of the development of meta-model of

distributed data sources. In particular, ontology-based approach aims to answer the following questions associated with the development of meta-model of distributed data sources: How to resolve the data non-congruency problem caused by heterogeneity and distribution of data sources? The particular questions are: How to provide for monosematic understanding of the terminology used in formal specification of distributed entities which, as a rule , are developed by distributed analysts? How to solve the entity identification problem, which arises due to the fact that the same entity specification is represented by its fragments in distributed databases? How to cope with the diversity of data physical natures, scales of attribute measurement, variety of data accuracy, duplication of the same attributes in different data sources? How to provide compatibility of ontology-based specifications of DF system notions and their interpretations represented in terms of a database language?

2. The second important issue of DF system design is distributed learning and meta-model of combining decisions. The questions to be answered here are as follows: What structures are used to combine particular decisions made on the basis of particular data sources to generate global decision? How to manage distributed data in order to provide correctness with regard to allocation of training and testing data used for learning particular classifiers and how to form and manage meta -data used for learning components responsible for combining decisions? What formal techniques are appropriate to combine decisions in accordance with the hierarchy of classifiers?

3. The third issue is of architectural kind. We use multi-agent architecture which to our opinion is the most appropriate for the implementation of DF systems and also for many other Semantic Web-based applications. This architecture includes specific components (intelligent agents) and protocols aiming at solving the questions formulated within both above issues.

In presentation we intend to highlight our results concerning to the above issues and their correlation with Semantic Web Mining-related tasks and problems. In addition, we intend to present an outline of a technology supported by a software tool used for design and implementation of DF software tool including design and implementation of its ontology component. In conclusion a brief outline of DF applications developed and being developed can be outlined.

94