ࡱ> (1    ! " # $ % & ' ( ) * + , - . / 0 r M T V e eeeeet5@ 0Nbjbj22 XX/fm^pspspss"zb4'''();D  I.wpxpx(, Hٲ  Ñ ő ő ő ő ő ő $h R | Q$X{ ~   TM , XXx,: ۲ ۲ ۲ y SXt)x ,Ñ ۲ { Ñ ۲ ۲ *{ ̳g , I W`'m $7 R 4uP 0  6 6 g ,JNsXX6 ?  t۲ M . ) L  ( D%V$ "VParaMor: from Paradigm Structure to Natural Language Morphology Induction Christian Monson Language Technologies Institute School of Computer Science Carnegie Mellon University Pittsburgh, PA, USA Thesis Committee Jaime Carbonell (CoChair) Alon Lavie (CoChair) Lori Levin Ron Kaplan (CSO at PowerSet) Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy  For Melinda   Abstract. Most of the worlds natural languages have complex morphology. But the expense of building morphological analyzers by hand has prevented the development of morphological analysis systems for the large majority of languages. Unsupervised induction techniques, that learn from unannotated text data, can facilitate the development of computational morphology systems for new languages. Such unsupervised morphological analysis systems have been shown to help natural language processing tasks including speech recognition (Creutz, 2006) and information retrieval (Kurimo and Turunen, 2008). This thesis describes ParaMor, an unsupervised induction algorithm for learning morphological paradigms from large collections of words in any natural language. Paradigms are sets of mutually substitutable morphological operations that organize the inflectional morphology of natural languages. ParaMor focuses on the most common morphological process, suffixation. ParaMor learns paradigms in a three-step algorithm. First, a recall-centric search scours a space of candidate partial paradigms for those which possibly model suffixes of true paradigms. Second, ParaMor merges selected candidates that appear to model portions of the same paradigm. And third, ParaMor discards those clusters which most likely do not model true paradigms. Based on the acquired paradigms, ParaMor then segments words into morphemes. ParaMor, by design, is particularly effective for inflectional morphology, while other systems, such as Morfessor (Creutz, 2006), better identify derivational morphology. This thesis leverages the complementary strengths of ParaMor and Morfessor by adjoining the analyses from the two systems. ParaMor and its combination with Morfessor were evaluated by participating in Morpho Challenge, a peer operated competition for morphology analysis systems (Kurimo, Turunen, and Varjokallio, 2008). The Morpho Challenge competitions held in 2007 and 2008 evaluated each systems morphological analyses in five languages, English, German, Finnish, Turkish, and Arabic. When ParaMors morphological analyses are merged with those of Morfessor, the resulting morpheme recall in all five languages is higher than that of any system which competed in either years Challenge; in Turkish, for example, ParaMors recall, at 52.1%, is twice that of the next highest system. This strong recall leads to an F1 score for morpheme identification above that of all systems in all languages but English. Table of Contents  TOC \o "1-3" \h \z \u  HYPERLINK \l "_Toc217315416" Abstract......  PAGEREF _Toc217315416 \h 5  HYPERLINK \l "_Toc217315417" Table of Contents  PAGEREF _Toc217315417 \h 7  HYPERLINK \l "_Toc217315418" Acknowledgements  PAGEREF _Toc217315418 \h 9  HYPERLINK \l "_Toc217315419" List of Figures  PAGEREF _Toc217315419 \h 11  HYPERLINK \l "_Toc217315420" Chapter 1: Introduction  PAGEREF _Toc217315420 \h 13  HYPERLINK \l "_Toc217315421" 1.1 The Structure of Morphology  PAGEREF _Toc217315421 \h 14  HYPERLINK \l "_Toc217315422" 1.2 Thesis Claims  PAGEREF _Toc217315422 \h 19  HYPERLINK \l "_Toc217315423" 1.3 ParaMor: Paradigms across Morphology  PAGEREF _Toc217315423 \h 20  HYPERLINK \l "_Toc217315424" 1.4 A Brief Readers Guide  PAGEREF _Toc217315424 \h 23  HYPERLINK \l "_Toc217315425" Chapter 2: A Literature Review of ParaMors Predecessors  PAGEREF _Toc217315425 \h 25  HYPERLINK \l "_Toc217315426" 2.1 Finite State Approaches  PAGEREF _Toc217315426 \h 26  HYPERLINK \l "_Toc217315427" 2.1.1 The Character Trie  PAGEREF _Toc217315427 \h 27  HYPERLINK \l "_Toc217315428" 2.1.2 Unrestricted Finite State Automata  PAGEREF _Toc217315428 \h 29  HYPERLINK \l "_Toc217315429" 2.2 MDL and Bayes Rule: Balancing Data Length against Model Complexity  PAGEREF _Toc217315429 \h 32  HYPERLINK \l "_Toc217315430" 2.2.1 Measuring Morphology Models with Efficient Encodings  PAGEREF _Toc217315430 \h 34  HYPERLINK \l "_Toc217315431" 2.2.2 Measuring Morphology Models with Probability Distributions  PAGEREF _Toc217315431 \h 37  HYPERLINK \l "_Toc217315432" 2.3 Other Approaches to Unsupervised Morphology Induction  PAGEREF _Toc217315432 \h 43  HYPERLINK \l "_Toc217315433" 2.4 Discussion of Related Work  PAGEREF _Toc217315433 \h 45  HYPERLINK \l "_Toc217315434" Chapter 3: Paradigm Identification with ParaMor  PAGEREF _Toc217315434 \h 47  HYPERLINK \l "_Toc217315435" 3.1 A Search Space of Morphological Schemes  PAGEREF _Toc217315435 \h 48  HYPERLINK \l "_Toc217315436" 3.1.1 Schemes  PAGEREF _Toc217315436 \h 48  HYPERLINK \l "_Toc217315437" 3.1.2 Scheme Networks  PAGEREF _Toc217315437 \h 54  HYPERLINK \l "_Toc217315438" 3.2 Searching the Scheme Lattice  PAGEREF _Toc217315438 \h 60  HYPERLINK \l "_Toc217315439" 3.2.1 A Birds Eye View of ParaMors Search Algorithm  PAGEREF _Toc217315439 \h 60  HYPERLINK \l "_Toc217315440" 3.2.2 ParaMors Initial Paradigm Search: The Algorithm  PAGEREF _Toc217315440 \h 64  HYPERLINK \l "_Toc217315441" 3.2.3 ParaMors Bottom-Up Search in Action  PAGEREF _Toc217315441 \h 67  HYPERLINK \l "_Toc217315442" 3.2.4 The Construction of Scheme Networks  PAGEREF _Toc217315442 \h 68  HYPERLINK \l "_Toc217315443" 3.2.5 Upward Search Metrics  PAGEREF _Toc217315443 \h 75  HYPERLINK \l "_Toc217315444" 3.3 Summarizing the Search for Candidate Paradigms  PAGEREF _Toc217315444 \h 92  HYPERLINK \l "_Toc217315445" Chapter 4: Clustering and Filtering of Initial Paradigms  PAGEREF _Toc217315445 \h 93  HYPERLINK \l "_Toc217315446" 4.1 From Schemes to Comprehensive Models of Paradigms  PAGEREF _Toc217315446 \h 93  HYPERLINK \l "_Toc217315447" 4.1.1 A Sample of Initially-Selected Scheme Models of Paradigms  PAGEREF _Toc217315447 \h 94  HYPERLINK \l "_Toc217315448" 4.1.2 ParaMors Paradigm-Processing Pipeline  PAGEREF _Toc217315448 \h 99  HYPERLINK \l "_Toc217315449" 4.2 Training Corpus Clean-Up  PAGEREF _Toc217315449 \h 101  HYPERLINK \l "_Toc217315450" 4.3 Clustering of Partial Paradigms  PAGEREF _Toc217315450 \h 104  HYPERLINK \l "_Toc217315451" 4.3.1 Three Challenges Face any Scheme-Clustering Algorithm  PAGEREF _Toc217315451 \h 104  HYPERLINK \l "_Toc217315452" 4.3.2 Clustering Large and Clustering Small  PAGEREF _Toc217315452 \h 112  HYPERLINK \l "_Toc217315453" 4.3.3 An Examination of ParaMors Scheme-Clusters  PAGEREF _Toc217315453 \h 112  HYPERLINK \l "_Toc217315454" 4.4 Filtering of Merged Clusters  PAGEREF _Toc217315454 \h 121  HYPERLINK \l "_Toc217315455" 4.4.1 Filtering of Small Scheme-clusters  PAGEREF _Toc217315455 \h 122  HYPERLINK \l "_Toc217315456" 4.4.2 Morpheme Boundary Filtering  PAGEREF _Toc217315456 \h 126  HYPERLINK \l "_Toc217315457" 4.5 ParaMors Paradigm Models  PAGEREF _Toc217315457 \h 137  HYPERLINK \l "_Toc217315458" 4.5.1 ParaMors Final Scheme-Clusters as Viable Models of Paradigms  PAGEREF _Toc217315458 \h 140  HYPERLINK \l "_Toc217315459" 4.5.2 Paradigm Learning and Vocabulary Size  PAGEREF _Toc217315459 \h 144  HYPERLINK \l "_Toc217315460" 4.6 Scheme-Clusters in Languages Beyond Spanish  PAGEREF _Toc217315460 \h 148  HYPERLINK \l "_Toc217315461" Chapter 5: Morphological Segmentation  PAGEREF _Toc217315461 \h 153  HYPERLINK \l "_Toc217315462" 5.1 The Segmentation Algorithm  PAGEREF _Toc217315462 \h 155  HYPERLINK \l "_Toc217315463" 5.2 A Sample of ParaMors Word Segmentations  PAGEREF _Toc217315463 \h 156  HYPERLINK \l "_Toc217315464" 5.3 Morpheme Segmentation Enables Evaluation  PAGEREF _Toc217315464 \h 162  HYPERLINK \l "_Toc217315465" Chapter 6: ParaMor and Morpho Challenge  PAGEREF _Toc217315465 \h 163  HYPERLINK \l "_Toc217315466" 6.1 Evaluation Methodology at Morpho Challenge 2007/2008  PAGEREF _Toc217315466 \h 164  HYPERLINK \l "_Toc217315467" 6.1.1 The Linguistic Evaluation of Morpho Challenge 2007/2008  PAGEREF _Toc217315467 \h 165  HYPERLINK \l "_Toc217315468" 6.1.2 The Task-Based Evaluation of Morpho Challenge 2007/2008  PAGEREF _Toc217315468 \h 169  HYPERLINK \l "_Toc217315469" 6.2 An Ablation Study  PAGEREF _Toc217315469 \h 169  HYPERLINK \l "_Toc217315470" 6.3 Inflectional vs. Derivational Morphology  PAGEREF _Toc217315470 \h 179  HYPERLINK \l "_Toc217315471" 6.4 Morpho Challenge 2007/2008  PAGEREF _Toc217315471 \h 181  HYPERLINK \l "_Toc217315472" 6.4.1 Linguistic Evaluation Results from Morpho Challenge 2007/2008  PAGEREF _Toc217315472 \h 182  HYPERLINK \l "_Toc217315473" 6.4.2 The Task-Based Evaluation of Morpho Challenge 2007/2008  PAGEREF _Toc217315473 \h 188  HYPERLINK \l "_Toc217315474" Chapter 7: Conclusions and Future Work  PAGEREF _Toc217315474 \h 193  HYPERLINK \l "_Toc217315475" 7.1 Improving the Core ParaMor Algorithms  PAGEREF _Toc217315475 \h 194  HYPERLINK \l "_Toc217315476" 7.2 The Future of Unsupervised Morphology Induction  PAGEREF _Toc217315476 \h 198  HYPERLINK \l "_Toc217315477" 7.2.1 Beyond Suffixation  PAGEREF _Toc217315477 \h 198  HYPERLINK \l "_Toc217315478" 7.2.2 Morphophonological Change  PAGEREF _Toc217315478 \h 199  HYPERLINK \l "_Toc217315479" 7.2.3 Mapping Morphemes to Features  PAGEREF _Toc217315479 \h 199  HYPERLINK \l "_Toc217315480" 7.3 ParaMor: A Successful Morphology Induction Algorithm  PAGEREF _Toc217315480 \h 200  HYPERLINK \l "_Toc217315481" Bibliography  PAGEREF _Toc217315481 \h 203  HYPERLINK \l "_Toc217315482" Appendix A: A Summary of Common Spanish Suffixes  PAGEREF _Toc217315482 \h 213  HYPERLINK \l "_Toc217315483" Appendix B: Scheme Clustering, a Pseudo-Code Implementation  PAGEREF _Toc217315483 \h 219  Acknowledgements I am as lucky as a person can be to obtain a Ph.D. from the Language Technologies Institute (LTI) at Carnegie Mellon University. While I started this program with little more than an interest in languages and computation, the people I met at the LTI have brought natural language processing to life. From the unparalleled teaching and mentoring of professors like Pascual Masullo, Larry Wasserman, and Roni Rosenfeld I learned to appreciate, and begin to quantify natural language. And from fellow students including Katharina Probst, Ariadna Font Llitjs, and Erik Peterson I found that natural language processing is a worthwhile passion. Still, whatever my desire for a degree in computational linguistics, this thesis never would have happened without the support, encouragement, and work of my three faculty advisors: Jaime Carbonell, Alon Lavie, and Lori Levin. Many people informed me that having three advisors was unworkableone person telling you what to do is enough, let alone three! But in my particular situation, each advisor brought unique and indispensable strengths: Jaimes broad experience in cultivating and pruning a thesis project, Alons hands-on detailed work on algorithmic specifics, and Loris linguistic expertise were all invaluable. Without input from each of my advisors, ParaMor would not have been born. But beyond the technical aspects, I would never have been able to complete this thesis were it not for the support and patience of my family. Thank you Mom and Dad for being my final line of support; Thank you Kurt and Mitzi for not kicking me out of the house. And finally, without the encouragement of my wife, without her kind words and companionship, I would have thrown in the towel long ago. Thank you, Melinda, for standing behind me when the will to continue was beyond my grasp. And thank you to James, Liesl, and Edith whose needs and smiles have kept me trying.  Christian List of Figures  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 A fragment of the Spanish verbal paradigms.............................................1 7  REF _Ref193208912 \h  \* MERGEFORMAT Figure  1.2 A finite state automaton for administrar.................................................. 19  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3 A portion of a morphology scheme network............................................. 23 Figure 2.1 A hub and a stretched hub in a finite state automaton............................... 32 Figure 3.1 Schemes from a small vocabulary............................................................. 53 Figure 3.2 A morphology scheme network over a small vocabulary.......................... 57 Figure 3.3 A morphology scheme network over the Brown Corpus........................... 59 Figure 3.4 A morphology scheme network over a Spanish corpus............................. 61 Figure 3.5 A birds-eye conceptualization of ParaMors initial search algorithm....... 63 Figure 3.6 Pseudo-code implementing ParaMors initial search algorithm................ 68 Figure 3.7 Eight search paths followed by ParaMors initial search algorithm.......... 71 Figure 3.8 Pseudo-code computing all most-specifics schemes from a corpus.......... 74 Figure 3.9 Pseudo-code computing the most-specific ancestors of each csuffix....... 76 Figure 3.10 Pseudo-code computing the cstems of a scheme.................................... 77 Figure 3.11 Seven parents of the a.o.os scheme........................................................ 79 Figure 3.12 Four expansions of the a.o.os scheme.................................................... 82 Figure 3.13 Six parent-evaluation metrics.................................................................. 86 Figure 3.14 An oracle evaluation of six parent-evaluation metrics............................ 92 Figure 4.1 A sample of initially-selected schemes...................................................... 98 Figure 4.2 Six initial schemes that model the Spanish er and/or ir paradigms......... 108 Figure 4.3 Five initial schemes licensed by the word apoyadas............................. 112 Figure 4.4 Typical Spanish scheme-clusters............................................................. 116 Figure 4.5 A portion of a scheme-cluster tree........................................................... 118 Figure 4.6 Cluster count and suffix recall vary with a cluster size threshold........... 125 Figure 4.7 Six scheme-clusters discarded by ParaMors small-cluster filter............ 127 Figure 4.8 Three scheme-clusters hypothesize morpheme boundaries..................... 129 Figure 4.9 A character trie and corresponding schemes........................................... 131 Figure 4.10 Pseudo-code for a suffix-internal morpheme boundary error filter....... 134 Figure 4.11 Three schemes hypothesize morpheme boundaries............................... 135 Figure 4.12 An illustration of ParaMors morpheme boundary error filters............. 138 Figure 4.13 Pseudo-code for a stem-internal morpheme boundary error filter......... 140 Figure 4.14 ParaMors paradigm-induction pipeline................................................ 141 Figure 4.15 Typical Spanish scheme-clusters (Figure 4.4 revisited)........................ 144 Figure 4.16 Cluster count and suffix recall vary with vocabulary size..................... 148 Figure 4.17 Scheme and scheme-cluster counts for six languages........................... 152 Figure 5.1 Pseudo-code implementing ParaMors word segmentation algorithm.... 158 Figure 5.2 A sample of ParaMors morphological segmentations of Spanish.......... 160 Figure 6.1 Answer key in the style of a Morpho Challenge Linguistic Evaluation.. 168 Figure 6.2 Ablation study: Search, filtering, and clustering impact segmentation... 174 Figure 6.3 Ablation study: A paradigm induction corpus of 20,000 unique types... 179 Figure 6.4 ParaMor and inflectional vs. derivational morphology........................... 182 Figure 6.5 Results from the Linguistic Evaluation of Morpho Challenge................ 186 Figure 6.6 Results from the Task-Based IR Evaluation of Morpho Challenge........ 192 Figure 6.7 Four reference algorithms for the Task-Based Evaluation...................... 192 Figure 7.1 Feature detection from word-aligned translated sentences...................... 203 Figure A.1 The suffixes of the ar inflection class of Spanish verbs......................... 217 Figure A.2 The suffixes of the er inflection class of Spanish verbs......................... 217 Figure A.3 The suffixes of the ir inflection class of Spanish verbs.......................... 218 Figure A.4 Suffixes of a paradigm for Number on nouns and adjectives................ 218 Figure A.5 Suffixes of a second paradigm for Number on nouns and adjectives.... 218 Figure A.6 Suffixes of the cross-product paradigm of Gender and Number.......... 218 Figure A.7 Spanish pronominal clitics...................................................................... 219 Figure A.8 Frequent derivational suffixes of Spanish............................................... 219 Introduction Most natural languages exhibit inflectional morphology, that is, the surface forms of words change to express syntactic featuresI run vs. She runs. Handling the inflectional morphology of English in a natural language processing (NLP) system is fairly straightforward. The vast majority of lexical items in English have fewer than five surface forms. But English has a particularly sparse inflectional system. It is not at all unusual for a language to construct tens of unique inflected forms from a single lexeme. And many languages routinely inflect lexemes into hundreds, thousands, or even tens of thousands of unique forms! In these inflectional languages, computational systems as different as speech recognition (Creutz, 2006), machine translation (Goldwater and McClosky, 2005; Oflazer and El-Kahlout, 2007), and information retrieval (Kurimo and Turunen, 2008) improve with careful morphological analysis. Computational approaches for analyzing inflectional morphology categorize into three groups. Morphology systems are either: Hand-built, Trained from examples of word forms correctly analyzed for morphology, or Induced from morphologically unannotated text in an unsupervised fashion. Presently, most computational applications take the first option, hand-encoding morphological facts. Unfortunately, manual description of morphology demands human expertise in a combination of linguistics and computation that is in short supply for many of the worlds languages. The second option, training a morphological analyzer in a supervised fashion, suffers from a similar knowledge acquisition bottleneck: Morphologically analyzed training data must be specially prepared, i.e. segmented and labeled, by human experts. This thesis seeks to overcome these problems of knowledge acquisition through language independent automatic induction of morphological structure from readily available machine-readable natural language text. The Structure of Morphology Natural language morphology supplies many language independent structural regularities which unsupervised induction algorithms can exploit to discover the morphology of a language. This thesis intentionally leverages three such regularities. The first regularity is the paradigmatic opposition found in inflectional morphology. Paradigmatically opposed inflections are mutually substitutable and mutually exclusive. Spanish, for example, marks verbs in the ar paradigm sub-class, including hablar (to speak), for the feature 2nd Person Present Indicative with the suffix as, hablas; But marks 1st Person Present Indicative with a mutually exclusive suffix o, hablo. The o suffix substitutes in for as, and no verb form can occur with both the as and the o suffixes simultaneously, *hablaso. Every set of paradigmatically opposed inflectional suffixes is said to fill a paradigm. In Spanish, the as and the o suffixes fill a portion of the verbal paradigm. Because of its direct appeal to paradigmatic opposition, the unsupervised morphology induction algorithm described in this thesis is dubbed ParaMor. The second morphological regularity leveraged by ParaMor to uncover morphological structure is the syntagmatic relationship of lexemes. Natural languages with inflectional morphology invariably possess classes of lexemes that can each be inflected with the same set of paradigmatically opposed morphemes. These lexeme classes are in a syntagmatic relationship. Returning to Spanish, all regular ar verbs (hablar, andar, cantar, saltar, ...) use the as and o suffixes to mark 2nd Person Present Indicative and 1st Person Present Indicative respectively. Together, a particular set of paradigmatically opposed morphemes and the class of syntagmatically related stems adhering to that paradigmatic morpheme set forms an inflection class of a language, in this case the ar inflection class of Spanish verbs. The third morphological regularity exploited by ParaMor follows from the paradigmatic-syntagmatic structure of natural language morphology. The repertoire of morphemes and stems in an inflection class constrains phoneme sequences. Specifically, while the phoneme sequence within a morpheme is restricted, a range of possible phonemes is likely at a morpheme boundary: A number of morphemes, each with possibly distinct initial phonemes, might follow a particular morpheme. Spanish non-finite verbs illustrate paradigmatic opposition of morphemes, the syntagmatic relationship between stems, inflection classes, paradigms, and phoneme sequence constraints. In the schema of Spanish non-finite forms there are three paradigms, depicted as the three columns of  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1. The first paradigm marks the Type of a particular surface form. A Spanish verb can appear in exactly one of three Non-Finite Types: as a Past Participle, as a Present Participle, or in the Infinitive. The three rows of the Type columns in  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 represent the suffixes of these three paradigmatically opposed forms. If a Spanish verb occurs as a Past Participle, then the verb takes additional suffixes. First, an obligatory suffix marks Gender: an a marks Feminine, an o Masculine. Following the Gender suffix, either a Plural suffix, s, appears or else there is no suffix at all. The lack of an explicit plural suffix marks Singular. The Gender and Number columns of  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 represent these additional two paradigms. In the left-hand table the feature values for the Type, Gender, and Number features are given. The right-hand table presents surface forms of suffixes realizing the corresponding feature values in the left-hand table. Spanish verbs which take the exact suffixes appearing in the right-hand table belong to the syntagmatic ar inflection class of Spanish verbs. Appendix A gives a more complete summary of the paradigms and inflection classes of Spanish morphology. To see the morphological structure of  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 in action, we need a particular Spanish lexeme: a lexeme such as administrar, which translates as to administer or manage. The form administrar fills the Infinitive cell of the Type paradigm in  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1. Other forms of this lexeme fill other cells of  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1. The form filling the Past Participle cell of the Type paradigm, the Feminine cell of the Gender paradigm, and the Plural cell of the Number paradigm is administradas, a word which would refer to a group of feminine gender nouns under administration. Each column of  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 truly constitutes a paradigm in that the cells of each column are mutually exclusive there is no way for administrar (or any other Spanish lexeme) to appear simultaneously in the infinitive and in a past participle form: *admistrardas, *admistradasar. The phoneme sequence constraints within these Spanish paradigms emerge when considering the full set of surface forms for the lexeme administrar, which include: Past Participles in all four combinations of Gender and Number: administrada, administradas, administrado, and administrados; the Present Participle and Infinitive non-finite forms described in  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1: administrando, administrar; and the many finite forms such as the 1st Person Singular Indicative Present Tense form administro.  REF _Ref193208912 \h  \* MERGEFORMAT Figure  1.2 shows these forms (as in Johnson and Martin, 2003) laid out graphically as a finite state automaton (FSA). Each state in this FSA represents a character boundary, while the arcs are labeled with characters from the surface forms of the lexeme administrar. Morpheme-internal states are open circles in  REF _Ref193208912 \h  \* MERGEFORMAT Figure  1.2, while states at word-internal morpheme boundaries are solid circles. Most morpheme-internal states have exactly one arc entering and one arc exiting. In contrast, states at morpheme boundaries tend to have multiple arcs entering or leaving, or both the character (and phoneme) sequence is constrained within morpheme, but more free at morpheme boundaries. This discussion of the paradigmatic, syntagmatic, and phoneme sequence structure of natural language morphology has intentionally simplified the true range of morphological phenomena. Three sources of complexity deserve particular mention. First, languages employ a wide variety of morphological processes. Among others, the processes of suffixation, prefixation, infixation, reduplication, and template filling all produce surface forms in some languages. Second, the application of word forming processes often triggers phonological (or orthographic) change. These phonological changes can obscure a straightforward FSA treatment of morphology. And third, the morphological structure of a word can be inherently ambiguousthat is, a single surface form of a lexeme may have more than one legitimate morphological analysis. Despite the complexity of morphology, this thesis holds that a large caste of morphological structures can be represented as paradigms of mutually substitutable substrings. In particular, sequences of affixes can be modeled by paradigm-like structures. Returning to the example of Spanish verbal paradigms in  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1, the Number paradigm on past participles can be captured by the alternating pair of strings s and . Similarly, the Gender paradigm alternates between the strings a and o. Additionally, and crucially for this thesis, the Number and Gender paradigms combine to form an emergent cross-product paradigm of four alternating strings: a, as, o, and os. Carrying the cross-product further, the past participle endings alternate with the other verbal endings, both non-finite and finite, yielding a large cross-product paradigm-like structure of alternating strings which include: ada, adas, ado, ados, ando, ar, o, etc. These emergent cross-product paradigms each identify a single morpheme boundary within the larger paradigm structure of a language. And with this brief introduction to morphology and paradigm structure we come to the formal claims of this thesis. Thesis Claims The algorithms and discoveries contained in this thesis automate the morphological analysis of natural language by inducing structures, in an unsupervised fashion, which closely correlate with inflectional paradigms. Additionally, The discovered paradigmatic structures improve the word-to-morpheme segmentation performance of a state-of-the-art unsupervised morphology analysis system. 2. The unsupervised paradigm discovery and word segmentation algorithms improve this state-of-the-art performance for a diverse set of natural languages, including German, Turkish, Finnish, and Arabic. 3. The paradigm discovery and improved word segmentation algorithms are computationally tractable. Augmenting a morphologically nave information retrieval (IR) system with induced morphological segmentations improves performance on an IR task. The IR improvements hold across a range of morphologically concatenative languages. And Enhanced performance on other natural language processing tasks is likely. ParaMor: Paradigms across Morphology The paradigmatic, syntagmatic, and phoneme sequence constraints of natural language allow ParaMor, the unsupervised morphology induction algorithm described in this thesis, to first reconstruct the morphological structure of a language, and to then deconstruct word forms of that language into constituent morphemes. The structures that ParaMor captures are sets of mutually replaceable word-final strings which closely model emergent paradigm cross-productseach paradigm cross-product identifying a single morpheme boundary in a set of words. This dissertation focuses on identifying suffix morphology. Two facts support this choice. First, suffixation is a concatenative process and 86% of the worlds languages use concatenative morphology to inflect lexemes (Dryer, 2008). Second, 64% of these concatenative languages are predominantly suffixing, while another 17% employ prefixation and suffixation about equally, and only 19% are predominantly prefixing. In any event, concentrating on suffixes is not a binding choice: the methods for suffix discovery detailed in this thesis can be straightforwardly adapted to prefixes, and generalizations could likely capture infixes and other non-concatenative morphological processes. To reconstruct the cross-products of the paradigms of a language, ParaMor defines and searches a network of paradigmatically and syntagmatically organized schemes of candidate suffixes and candidate stems. ParaMor s search algorithms are motivated by paradigmatic, syntagmatic, and phoneme sequence constraints.  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3 depicts a portion of a morphology scheme network automatically derived from 100,000 words of the Brown Corpus of English (Francis, 1964). Each box in  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3 is a scheme, which lists in bold a set of candidate suffixes, or csuffixes, together with an abbreviated list, in italics, of candidate stems, or cstems. Each of the csuffixes in a scheme concatenates onto each of the cstems in that scheme to form a word found in the input text. For instance, the scheme containing the csuffix set .ed.es.ing, where signifies a null suffix, is derived from the words address, addressed, addresses, addressing, reach, reached, etc. In  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3, the two highlighted schemes, .ed.es.ing and e.ed.es.ing, represent valid paradigmatically opposed sets of suffixes that constitute (orthographic) inflection classes of the English verbal paradigm. The other candidate schemes in  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3 are wrong or incomplete. Crucially note, however, that ParaMor is not informed which schemes represent true paradigms and which do not separating the good scheme models from the bad is exactly the task of ParaMor s paradigm induction algorithms. Chapter 3 details the construction of morphology scheme networks over suffixes and describes a network search procedure that identifies schemes which contain in aggregate 91% of all Spanish inflectional suffixes when training over a corpus of 50,000 types. However, many of the initially selected schemes do not represent true paradigms; And of those that do represent paradigms, most capture only a portion of a complete paradigm. Hence, Chapter 4 describes algorithms to first merge candidate paradigm pieces into larger groups covering more of the affixes in a paradigm, and then to filter out those candidates which likely do not model true paradigms. With a handle on the paradigm structures of a language, ParaMor uses the induced morphological knowledge to segment word forms into likely morphemes. Recall that, as models of paradigm cross-products, each scheme models a single morpheme boundary in each surface form that contributes to that scheme. To segment a word form then, ParaMor simply matches the csuffixes of each discovered scheme against that word and proposes a single morpheme boundary at any match point. Examples will help clarify word segmentation. Assume ParaMor correctly identifies the English scheme .ed.es.ing from  REF _Ref211051873 \h  \* MERGEFORMAT Figure  1.3. When requested to segment the word reaches, ParaMor finds that the es csuffix in the discovered scheme matches the word-final string es in reaches. Hence, ParaMor segments reaches as reach +es. Since more than one paradigm cross-product may match a particular word, a word may be segmented at more than one position. The Spanish word administradas from Section  REF _Ref192388754 \r \h  1.1 contains three suffixes, ad, a, and s. Presuming ParaMor correctly identifies three separate schemes, one containing the cross-product csuffix adas, one containing as, and one containing s, ParaMor will match in turn each of these csuffixes against administradas, and will ultimately produce the correct segmentation: administr +ad +a +s. To evaluate morphological segmentation performance, ParaMor competed in two years of the Morpho Challenge competition series (Monson et al., 2008a; 2008b). The Morpho Challenge competitions are peer operated, pitting against one another algorithms designed to discover the morphological structure of natural languages from nothing more than raw text (Kurimo, Turunen, and Varjokallio, 2008). Unsupervised morphology induction systems were evaluated in two ways during the 2007 and 2008 Challenges. First, a linguistically motivated metric measured each system at the task of morpheme identification (Kurimo, Creutz, and Varjokallio, 2008; Kurimo and Varjokallio, 2008). Second, the organizers of Morpho Challenge augmented an information retrieval (IR) system with the morphological segmentations that each system proposed and measured mean average precision of the relevance of returned documents (Kurimo, Creutz, and Turunen, 2007; Kurimo and Turunen, 2008). Morpho Challenge 2007 evaluated morphological segmentations over four languages: English, German, Turkish, and Finnish; while the 2008 Challenge added Arabic. As a stand-alone system, ParaMor performed on par with state-of-the-art unsupervised morphology induction systems at the Morpho Challenge competitions. Evaluated for F1 at morpheme identification, in English ParaMor outperformed an already sophisticated reference induction algorithm, Morfessor-MAP (Creutz, 2006); placing third overall out of the eight participating algorithm families from the 2007 and 2008 competitions. In Turkish, ParaMor identified a significantly higher proportion of true Turkish morphemes than any other participating algorithm. This strong recall placed the solo ParaMor algorithm first in F1 at morpheme identification for this language. But ParaMor particularly shines when ParaMors morphological analyses are adjoined to those of Morfessor-MAP. Where ParaMor focuses on discovering the paradigmatic structure of inflectional suffixes, the Morfessor algorithm identifies linear sequences of inflectional and derivational affixesboth prefixes and suffixes. With such complementary algorithms, it is not surprising that the combined segmentations of ParaMor and Morfessor improve performance over either algorithm alone. The joint ParaMor-Morfessor system placed first at morpheme identification in all language tracks of the Challenge but English, where it moved to second. In Turkish, morpheme identification of ParaMor-Morfessor is 13.5% higher absolute than the next best submitted system, excluding ParaMor alone. In the IR competition, which only covered English, German, and Finnish, the combined ParaMor-Morfessor system placed first in English and German. And the joint system consistently outperformed, in all three languages, a baseline algorithm of no morphological analysis. A Brief Readers Guide The remainder of this thesis is organized as follows: Chapter 2 situates the ParaMor algorithm in the field of prior work on unsupervised morphology induction. Chapters 3 and 4 present ParaMors core paradigm discovery algorithms. Chapter 5 describes ParaMors word segmentation models. And Chapter 6 details ParaMors performance in the Morpho Challenge 2007 competition. Finally, Chapter 7 summarizes the contributions of ParaMor and outlines future directions both specifically for ParaMor and more generally for the broader field of unsupervised morphology induction. A Literature Review of ParaMors Predecessors The challenging task of unsupervised morphology induction has inspired a significant body of work. This chapter highlights unsupervised morphology systems that influenced the design of or that contrast with ParaMor, the morphology induction system described in this thesis. Two induction techniques have particularly impacted the development of ParaMor: Finite State (FS) techniques, and 2. Minimum Description Length (MDL) techniques. Sections  REF _Ref212623558 \r \h  2.1 and  REF _Ref212623577 \r \h  2.2 present, respectively, FS and MDL approaches to morphology induction, emphasizing their influence on ParaMor. Section  REF _Ref212623601 \r \h  2.3 then describes several morphology induction systems which do not neatly fall in the FS or MDL camps but which are nevertheless relevant to the design of ParaMor. Finally, Section  REF _Ref212623620 \r \h  2.4 synthesizes the findings of the earlier discussion. Finite State Approaches In 1955, Zellig Harris proposed to induce morphology in an unsupervised fashion by modeling morphology as a finite state automaton (FSA). In this FSA, the characters of each word label the transition arcs and, consequently, states in the automaton occur at character boundaries. Coming early in the procession of modern methods for morphology induction, Harris-style finite state techniques have been incorporated into a number of unsupervised morphology induction systems, ParaMor included. ParaMor draws on finite state techniques at two points within its algorithms. First, the finite state structure of morphology impacts ParaMors initial organization of candidate partial paradigms into a search space (Section  REF _Ref193859274 \r \h  \* MERGEFORMAT  3.1.2). And second, ParaMor identifies and removes the most unlikely initially selected candidate paradigms using finite state techniques (Section  REF _Ref190771388 \r \h  4.4.2). Three facts motivate finite state automata as appropriate models for unsupervised morphology induction. First, the topology of a morphological FSA captures phoneme sequence constraints in words. As was presented in Section  REF _Ref192388754 \r \h  1.1, phoneme choice is usually constrained at character boundaries internal to a morpheme but is often more free at morpheme boundaries. In a morphological FSA, a state with a single incoming character arc and from which there is a single outgoing arc is likely internal to a morpheme, while a state with multiple incoming arcs and several competing outgoing branches likely occurs at a morpheme boundary. As described further in Section  REF _Ref212696175 \r \h  2.1.1, it was this topological motivation that Harris exploited in his 1955 system, and that ParaMor draws on as well. A second motivation for modeling morphological structure with finite state automata is that FSA succinctly capture the recurring nature of morphemesa single sequence of states in an FSA can represent many individual instances, in many separate words, of a single morpheme. As described in Section  REF _Ref212713404 \r \h  \* MERGEFORMAT  2.1.2 below, the morphology system of Altun and Johnson (2001) particularly builds on this succinctness property of finite state automata. The third motivation for morphological FSA is theoretical: Most, if not all, morphological operations are finite state in computational complexity (Roark and Sproat, 2007). Indeed, state-of-the-art solutions for building morphological systems involve hand-writing rules which are then automatically compiled into finite state networks (Beesley and Karttunen, 2003; Sproat 1997). The next two subsections ( REF _Ref212696175 \r \h  \* MERGEFORMAT  2.1.1 and  REF _Ref212713404 \r \h  \* MERGEFORMAT  2.1.2) describe specific unsupervised morphology induction systems which use finite state approaches. Section  REF _Ref212696175 \r \h  \* MERGEFORMAT  2.1.1 begins with the simple finite state structures proposed by Harris, while Section  REF _Ref212713404 \r \h  \* MERGEFORMAT  2.1.2 presents systems which allow more complex arbitrary finite state automata. The Character Trie Harris (1955; 1967) and later Hafer and Weiss (1974) were the first to propose and then implement finite state unsupervised morphology induction systemsalthough they may not have thought in finite state terms themselves. Harris (1955) outlines a morphology analysis algorithm which he motivated by appeal to the phoneme succession constraint properties of finite state structures. Harris algorithm first builds character trees, or tries, over corpus utterances. Tries are deterministic, acyclic, but un-minimized FSA. In tries, Harris identifies those states for which the finite state transition function is defined for an unusually large number of characters in the alphabet. These branching states represent likely word and morpheme boundaries. Although Harris only ever implemented his algorithm to segment words into morphemes, he originally intended his algorithms to segment sentences into words, as Harris (1967) notes, word-internal morpheme boundaries are much more difficult to detect with the trie algorithm. The comparative challenge of word-internal morpheme detection stems from the fact that phoneme variation at morpheme boundaries largely results from the interplay of a limited repertoire of paradigmatically opposed inflectional morphemes. In fact, as described in Section  REF _Ref192388754 \r \h  1.1, word-internal phoneme sequence constraints can be viewed as the phonetic manifestation of the morphological phenomenon of paradigmatic and syntagmatic variation. Harris (1967), in a small scale mock-up, and Hafer and Weiss (1974), in more extensive quantitative experiments, report results at segmenting words into morphemes with the trie-based algorithm. Word segmentation is an obvious task-based measure of the correctness of an induced model of morphology. A number of natural language processing tasks, including machine translation, speech recognition, and information retrieval, could potentially benefit from an initial simplifying step of segmenting complex surface words into smaller recurring morphemes. Hafer and Weiss detail word segmentation performance when augmenting Harris basic algorithm with a variety of heuristics for determining when the number of outgoing arcs is sufficient to postulate a morpheme boundary. Hafer and Weiss measure recall and precision performance of each heuristic when supplied with a corpus of 6,200 word types. The variant which achieves the highest F1 measure of 0.754, from a precision of 0.818 and recall of 0.700, combines results from both forward and backward tries and uses entropy to measure the branching factor of each node. Entropy captures not only the number of outgoing arcs but also the fraction of words that follow each arc. A number of systems, many of which are discussed in depth later in this chapter, embed a Harris style trie algorithm as one step in a more complex process. Demberg (2007), Goldsmith (2001; 2006), Schone and Jurafsky (2000; 2001), and Djean (1998) all use tries to construct initial lists of likely morphemes which they then process further. Bordag(2008) extracts morphemes from tries built over sets of words that occur in similar contexts. And Bernhard (2008) captures something akin to trie branching by calculating word-internal letter transition probabilities. Both the Bordag (2008) and the Bernhard (2008) systems competed strongly in Morpho Challenge competition, alongside the unsupervised morphology induction system described in this thesis, ParaMor. Finally, the ParaMor system itself examines trie structures to identify likely morpheme boundaries. ParaMor builds local tries from the last characters of candidate stems which all occur in a corpus with the same set of candidate suffixes attached. Following Hafer and Weiss (1974), ParaMor measures the strength of candidate morpheme boundaries as the entropy of the relevant trie structures. Unrestricted Finite State Automata From tries it is not a far step to modeling morphology with more general finite state automata. A variety of methods have been proposed to induce FSA that closely model morphology. The ParaMor algorithm of this thesis, for example, models the morphology of a language with a non-deterministic finite state automaton containing a separate state to represent every set of word final strings which ends some set of word initial strings in a particular corpus (see Section  REF _Ref193859274 \r \h  3.1.2). In contrast, Johnson and Martin  ADDIN EN.CITE <EndNote><Cite ExcludeAuth="1"><Author>Johnson</Author><Year>2003</Year><RecNum>29</RecNum><record><database name='EndNoteLibrary.enl' path='F:\Research\EndNoteLibrary\EndNoteLibrary.enl'>EndNoteLibrary.enlEndNote2910<style face='normal' font='default' size='100%'>Unsupervised Learning of Morphology for English and Inuktitut</style>(2003) suggest identifying morpheme boundaries by examining properties of the minimal finite state automaton that exactly accepts the word types of a corpus. The minimal FSA can be generated straightforwardly from a Harris-style trie by collapsing trie states from which precisely the same set of strings is accepted. Like a trie, the minimal FSA is deterministic and acyclic, and the branching properties of its arcs encode phoneme succession constraints. In the minimal FSA, however, incoming arcs also provide morphological information. Where every state in a trie has exactly one incoming arc, each state, , in the minimal FSA has a potentially separate incoming arc for each trie state which collapsed to form . A state with two incoming arcs, for example, indicates that there are at least two strings for which exactly the same set of final strings completes word forms found in the corpus. Incoming arcs thus encode a rough guide to syntagmatic variation. Johnson and Martin combine the syntagmatic information captured by incoming arcs with the phoneme sequence constraint information from outgoing arcs to segment the words of a corpus into morphemes at exactly: 1. Hub states states which possess both more than one incoming arc and more than one outgoing arc,  REF _Ref95885852 \h  \* MERGEFORMAT Figure  2.1, left. 2. The last state of stretched hubs sequences of states where the first state has multiple incoming arcs and the last state has multiple outgoing arcs and the only available path leads from the first to the last state of the sequence,  REF _Ref95885852 \h  \* MERGEFORMAT Figure  2.1, right. Stretched hubs likely model the boundaries of a paradigmatically related set of morphemes, where each related morpheme begins (or ends) with the same character or character sequence. Johnson and Martin (2003) report that this simple Hub-Searching algorithm segments words into morphemes with an F1 measure of 0.600, from a precision of 0.919 and a recall of 0.445, over the text of Tom Sawyer; which, according to Manning and Schtze (1999, p. 21), has 71,370 tokens and 8,018 types. To improve segmentation recall, Johnson and Martin extend the Hub-Searching algorithm by introducing a morphologically motivated state merge operation. Merging states in a minimized FSA generalizes or increases the set of strings the FSA will accept. In this case, Johnson and Martin merge all states that are either accepting word final states, or that are likely morpheme boundary states by virtue of possessing at least two incoming arcs. This technique increases F1 measure over the same Tom Sawyer corpus to 0.720, by bumping precision up slightly to 0.922 and significantly increasing recall to 0.590. State merger is a broad technique for generalizing the language accepted by a FSA, used not only in finite state learning algorithms designed for natural language morphology, but also in techniques for inducing arbitrary FSA. Much research on FSA induction focuses on learning the grammars of artificial languages. Lang, Pearlmutter, and Price (1998) present a state-merging algorithm designed to learn large randomly generated deterministic FSA from positive and negative data. Lang, Pearlmutter, and Price (1998) also provides a brief overview of other work in FSA induction for artificial languages. Since natural language morphology is considerably more constrained than random FSA, and since natural languages typically only provide positive examples, work on inducing formally defined subsets of general finite state automata from positive data may be a bit more relevant here. Work in constrained FSA induction includes Miclet (1980), who extends finite state ktail induction, first introduced by Biermann and Feldman (1972), with a state merge operation. Similarly, Angluin (1982) presents an algorithm, also based on state merger, for the induction of kreversible languages. Altun and Johnson (2001) present a technique for FSA induction, again built on state merger, which is specifically motivated by natural language morphological structure. Altun and Johnson induce finite state grammars for the English auxiliary system and for Turkish Morphology. Their algorithm begins from the forward trie over a set of training examples. At each step the algorithm applies one of two merge operations. Either any two states,  and , are merged, which then forces their children to be recursively merged as well; or an Ttransition is introduced from  to . To keep the resulting FSA deterministic following an Ttransition insertion, for all characters  for which the FSA transition function is defined from both  and , the states to which a leads are merged, together with their children recursively. Each arc  in the FSA induced by Altun and Johnson (2001) is associated with a probability, initialized to the fraction of words which follow the  arc. These arc probabilities define the probability of the set of training example strings. The training set probability is combined with the prior probability of the FSA to give a Bayesian description length for any training set-FSA pair. Altun and Johnsons greedy FSA search algorithm follows the minimum description length principle (MDL)at each step of the algorithm, that state merge operation or Ttransition insertion operation is performed which most decreases the weighted sum of the log probability of the induced FSA and the log probability of the observed data given the FSA. If no operation results in a reduction in the description length, grammar induction ends. Being primarily interested in inducing FSA, Altun and Johnson do not actively segment words into morphemes. Hence, quantitative comparison with other morphology induction work is difficult. Altun and Johnson do report the behavior of the negative log probability of Turkish test set data, and the number of learning steps taken by their algorithm, each as the training set size increases. Using these measures, they compare a version of their algorithm without Ttransition insertion to the version that includes this operation. They find that their algorithm for FSA induction with Ttransitions achieves a lower negative log probability in less learning steps from fewer training examples. MDL and Bayes Rule: Balancing Data Length against Model Complexity The minimum description length (MDL) principle employed by Altun and Johnson (2001) in a finite-state framework, as discussed in the previous section, has been used extensively in non-finite-state approaches to unsupervised morphology induction. The MDL principle is a model selection strategy which suggests to choose that model which minimizes the sum of: The size of an efficient encoding of the model, and 2. The length of the data encoded using the model. In morphology induction, the MDL principle measures the efficiency with which a model captures the recurring structures of morphology. Suppose an MDL morphology induction system identifies a candidate morphological structure, such as an inflectional morpheme or a paradigmatic set of morphemes. The MDL system will place the discovered morphological structure into the model exactly when the structure occurs sufficiently often in the data that it saves space overall to keep just one copy of the structure in the model and to then store pointers into the model each time the structure occurs in the data. Although ParaMor, the unsupervised morphology induction system described in this thesis, directly measures neither the complexity of its models nor the length of the induction data given a model, ParaMors design was, nevertheless, influenced by the MDL morphology induction systems described in this section. In particular, ParaMor implicitly aims to build compact models: The candidate paradigm schemes defined in Section  REF _Ref192582713 \r \h  3.1.1 and the partial paradigm clusters of Section  REF _Ref190686481 \r \h  4.3 both densely describe large swaths of the morphology of a language. Closely related to the MDL principle is a particular application of Bayes Rule from statistics. If d is a fixed set of data and m a morphology model ranging over a set of possible models, M, then the most probable model given the data is: . Applying Bayes Rule to this expression yields: , And taking the negative of the logarithm of both sides gives: . Reinterpreting this equation, the  term is a reasonable measure of the length of a model, while  expresses the length of the induction data given the model. Despite the underlying close relationship between MDL and Bayes Rule approaches to unsupervised morphology induction, a major division occurs in the published literature between systems that employ one or the other methodology. Sections  REF _Ref212891806 \r \h  2.2.1 and  REF _Ref212891829 \r \h  2.2.2 reflect this division: Section  REF _Ref212891806 \r \h  2.2.1 describes unsupervised morphology systems that apply the MDL principle directly by devising an efficient encoding for a class of morphology models, while Section  REF _Ref212891829 \r \h  2.2.2 presents systems that indirectly apply the MDL principle in defining a probability distribution over a set of models, and then invoking Bayes Rule. In addition to differing in their method for determining model and data length, the systems described in Sections  REF _Ref212891806 \r \h  2.2.1 and  REF _Ref212891829 \r \h  2.2.2 differ in the specifics of their search strategies. While the MDL principle can evaluate the strength of a model, it does not suggest how to find a good model. The specific search strategy a system uses is highly dependent on the details of the model family being explored. Section  REF _Ref212623558 \r \h  2.1 presented search strategies used by morphology induction systems that model morphology with finite state automata. And now Sections  REF _Ref212891806 \r \h  2.2.1 and  REF _Ref212891829 \r \h  2.2.2 describe search strategies employed by non-finite state morphology systems. The details of system search strategies are relevant to this thesis work as Chapters 3 and 4 of this dissertation are largely devoted to the specifics of ParaMors search algorithms. Similarities and contrasts with ParaMors search procedures are highlighted both as individual systems are presented and also in summary in Section  REF _Ref212891998 \r \h  2.4. Measuring Morphology Models with Efficient Encodings This survey of MDL-based unsupervised morphology induction systems begins with those that measure model length by explicitly defining an efficient model encoding. First to propose the MDL principle for morphology induction was Brent, Murthy, and Lundberg (1995; see also Brent, 1993). These authors use MDL to evaluate models of natural language morphology of a simple, but elegant form. Each morphology model describes a vocabulary as a set of three lists: 1. A list of stems 2. A list of suffixes 3. A list of the valid stem-suffix pairs Each of these three lists is efficiently encoded. The sum of the lengths of the first two encoded lists constitutes the model length, while the length of the third list yields the size of the data given the model. Consequently the sum of the lengths of all three encoded lists is the full description length to be minimized. As the morphology model in Brent, Murthy, and Lundberg (1995) only allows for pairs of stems and suffixes, each model can propose at most one morpheme boundary per word. Using this list-model of morphology to describe a vocabulary of words, V, there are  possible modelsfar too many to exhaustively explore. Hence, Brent, Murthy, and Lundberg (1995) describe a heuristic search procedure to greedily explore the model space. First, each word final string, f, in the corpus is ranked according to the ratio of the relative frequency of f divided by the relative frequencies of each character in f. Each word final string is then considered in turn, according to its heuristic rank, and added to the suffix list whenever doing so decreases the description length of the corpus. When no suffix can be added that reduces the description length further, the search considers removing a suffix from the suffix list. Suffixes are iteratively added and removed until description length can no longer be lowered. To evaluate their method, Brent, Murthy, and Lundberg (1995) examine the list of suffixes found by the algorithm when supplied with English word form lexicons of various sizes. Any correctly identified inflectional or derivational suffix counts toward accuracy. Their highest accuracy results are obtained when the algorithm induces morphology from a lexicon of 2000 types: the algorithm hypothesizes twenty suffixes with an accuracy of 85%. Baroni (2000; see also 2003) describes DDPL, an MDL inspired model of morphology induction similar to the Brent, Murthy, and Lundberg (1995) model. The DDPL model identifies prefixes instead of suffixes, uses a heuristic search strategy different from Brent, Murthy, and Lundberg (1995), and treats the MDL principle more as a guide than an inviolable tenet. But most importantly, Baroni conducts a rigorous empirical study showing that automatic morphological analyses found by DDPL correlate well with human judgments. He reports a Spearman correlation coefficient of the average human morphological complexity rating to the DDPL analysis on a set of 300 potentially prefixed words of 0.62 . Goldsmith (2001; 2006), in a system called Linguistica, extends the promising results of MDL morphology induction by augmenting the basic model of Brent, Murthy, and Lundberg (1995) to incorporate the paradigmatic and syntagmatic structure of natural language morphology. As discussed in Chapter 1, natural language inflectional morphemes belong to paradigmatic sets where all the morphemes in a paradigmatic set are mutually exclusive. Similarly, natural language lexemes belong to syntagmatic classes where all lexemes in the same syntagmatic class can be inflected with the same set of paradigmatically opposed morphemes. While previous approaches to unsupervised morphology induction, including Djean (1998), indirectly drew on the paradigmatic-syntagmatic structure of morphology, Goldsmiths Linguistica system was the first to intentionally model this important aspect of natural language morphological structure. The paradigm based algorithms of the ParaMor algorithm, as described in this thesis, were directly inspired by Goldsmiths success at unsupervised morphology induction when modeling the paradigm. The Linguistica system models the paradigmatic and syntagmatic nature of natural language morphology by defining the signature. A Goldsmith signature is a pair of sets , T a set of stems and F a set of suffixes, where T and F satisfy the following three conditions: For any stem t in T and for any suffix f in F, must be a word in the vocabulary, Each word in the vocabulary must be generated by exactly one signature, and Each stem  occurs in the stem set of at most one signature As in Brent, Murthy, and Lundberg (1995), a morphology model in Goldsmiths work consists of three lists. The first two are, as for Brent, a list of stems and a list of suffixes. But, instead of a list containing each valid stem-suffix pair, the third list in a Linguistica morphology consists of signatures. Replacing the list of all valid stem-suffix pairs with a list of signatures allows a signature model to potentially represent natural language morphology with a reduced description length. A description length decrease can occur because it takes less space to store a set of syntagmatically opposed stems with a set of paradigmatically opposed suffixes than it does to store the cross-product of the two sets. Following the MDL principle, Goldsmith efficiently encodes each of the three lists that form a signature model; and the sum of the encoded lists is the models description length. Notice that, just as for Brent, Murthy, and Lundberg (1995), Goldsmiths implemented morphology model can propose at most one morpheme boundary per word typealthough Goldsmith (2001) does discuss an extension to handle multiple morpheme boundaries. To find signature models, Goldsmith (2001; see also 2006) proposes several different search strategies. The most successful strategy seeds model selection with signatures derived from a Harris (1955) style trie algorithm. Then, a variety of heuristics suggest small changes to the seed model. Whenever a change results in a lower description length the change is accepted. Goldsmith (2001) reports precision and recall results on segmenting 1,000 alphabetically consecutive words from: The more than 30,000 unique word forms in the first 500,000 tokens of the Brown Corpus (Francis, 1964) of English: Precision: 0.860, Recall: 0.904, F1: 0.881. A corpus of 350,000 French tokens: Precision: 0.870, Recall: 0.890, F1: 0.880. Goldsmith (2001) also gives qualitative results for Italian, Spanish, and Latin suggesting that the best signatures in the discovered morphology models generally contain coherent sets of paradigmatically opposed suffixes and syntagmatically opposed stems. Measuring Morphology Models with Probability Distributions This section describes in depth morphology induction systems which exemplify the Bayes Rule approach to unsupervised morphology induction: Both Snover (2002) and Creutz (2006) build morphology induction systems by first defining a probability distribution over a family of morphology models and then searching for the most probable model. Both the Snover system and that built by Creutz directly influenced the development of ParaMor. The Morphology Induction System of Matthew Snover Snover (2002; c.f.: Snover and Brent, 2002; Snover, Jarosz, and Brent, 2002; Snover and Brent, 2001) discusses a family of morphological induction systems which, like Goldsmiths Linguistica, directly model the paradigmatic and syntagmatic structure of natural language morphology. But, where Goldsmith measures the quality of a morphological model as its encoded length, Snover invokes Bayes Rule (see this sections introduction)defining a probability function over a space of morphology models and then searching for the highest probability model. Snover (2002) leverages paradigmatic and syntagmatic morphological structure both in his probabilistic morphology models and in the search strategies he employs. To define the probability of a model, Snover (2002) defines functions that assign probabilities to: The stems in the model 2. The suffixes in the model 3. The assignment of stems to sets of suffixes called paradigms Assuming independence, Snover defines the probability of a morphology model as the product of the probabilities of the stems, suffixes, and paradigms. Like Goldsmith (2001; 2006), Snover only considers models of morphology where each word and each stem belong to exactly one paradigm. Hence, the third item in the above list is identical to Goldsmiths definition of a signature. Since Snover defines probabilities for exactly the same three items that Goldsmith computes description lengths for, the relationship of Snovers models to Goldsmiths is quite tight. To find strong models of morphology Snover proposes two search procedures: Hill Climbing Search and Directed Search. Both strategies leverage the paradigmatic structure of language in defining data structures similar to the morphology networks proposed for this thesis in Chapter 3. The Hill Climbing Search follows the same philosophy as the MDL based algorithms of Brent, Murthy, and Lundberg (1995) and Goldsmith (2001; 2006): At each step, Snovers Hill Climbing Search algorithm proposes a new morphology model, which is only accepted if it improves the model scoreBut in Snovers case, model score is probability. The Hill Climbing search uses an abstract suffix network defined by inclusion relations on sets of suffixes. Initially, the only network node to possess any stems is that node containing just the null suffix, . All vocabulary items are placed in this node. Each step of the Hill Climbing Search proposes adjusting the current morphological analysis by moving stems in batches to adjacent nodes that contain exactly one more or one fewer suffixes. At each step, that batch move is accepted which most improves the probability score. Snovers probability model can only score morphology models where each word contains at most a single morpheme boundary. The Hill Climbing Search ensures this single boundary constraint is met by forcing each individual vocabulary word to only ever contribute to a single network node. Whenever a stem is moved to a new network node in violation of the single boundary constraint, the Hill Climbing Search simultaneously moves a compensating stem to a new node elsewhere in the network. Snovers second search strategy, Directed Search, defines an instantiated suffix network where each node, or, in the terminology of Chapter 3, each scheme, in the network inherently contains all the stems which form vocabulary words with each suffix in that node. In this fully instantiated network, a single word might contribute to two or more scheme-nodes which advocate different morpheme boundariesa situation which Snovers probability model cannot evaluate. To build a global morphology model that the probability model can evaluate, the Directed Search algorithm visits every node in the suffix network and assigns a local probability score just for the segmentations suggested by that node. The Directed Search algorithm then constructs a globally consistent morphology model by first discarding all but the top n scoring nodes; and then, whenever two remaining nodes disagree on the segmentation of a word, accepting the segmentation from the better scoring node. To evaluate the performance of his morphology induction algorithm, while avoiding the problems that ambiguous morpheme boundaries present to word segmentation, Snover (2002) defines a pair of evaluation metrics to separately: 1. Identify pairs of related words, and 2. Identify suffixes (where any suffix allomorph is accepted as correct). Helpfully, Snover (2002) supplies not only the results of his own algorithms using these metrics but also the results of Goldsmiths (2001) Linguistica. Snover (2002) achieves his best overall performance when using the Directed Search strategy to seed the Hill Climbing Search. This combination outperforms Linguistica on both the suffix identification metric as well as on the metric designed to identify pairs of related words, and does so for both English and Polish lexicons of up to 16,000 vocabulary items. Hammarstrm (2006b; see also 2006a, 2007) presents a non-Bayes stemming algorithm that involves a paradigm search strategy that is closely related to Snovers. As in Snovers Directed Search, Hammarstrm defines a score for any set of candidate suffixes. But where Snover scores a suffix set according to a probability model that considers both the suffix set itself and the set of stems associated with that suffix set, Hammarstrm assigns a non-probabilistic score that is based on counts of stems that form words with pairs of suffixes from the set. Having defined an objective function, Hammarstrms algorithm searches for a set of suffixes that scores highly. Hammarstrms search algorithm moves from one set of suffixes to another in a fashion similar to Snovers Hill Climbing Searchby adding or subtracting a single suffix in a greedy fashion. Hammarstrms search algorithm is crucially different from Snovers in that Hammarstrm does not instantiate the full search space of suffix sets, but only builds those portions of the power set network that his algorithm directly visits. Hammarstrms paradigm search algorithm is one step in a larger system designed to stem words for information retrieval. To evaluate his system, Hammarstrm constructed sets of words that share a stem and sets of words that do not share a stem in four languages: Maori (an isolating language), English, Swedish, and Kuku Yalanji (a highly suffixing language). Hammarstrm finds that his algorithm is able to identify words that share a stem with accuracy above 90%. ParaMors search strategies, described in Chapter 3 of this thesis, were directly inspired by Snovers work and have much in common with Hammarstrms. The most significant similarity between Snovers system and ParaMor concerns the search space they examine for paradigm models: the fully instantiated network that Snover constructs for his Directed Search is exactly the search space that ParaMors initial paradigm search explores (see Chapter 3). The primary difference between ParaMors search strategies and those of Snover is that, where Snover must ensure his final morphology model assigns at most a single morpheme boundary to each word, ParaMor intentionally permits individual words to participate in scheme-nodes that propose competing morpheme boundaries. By allowing more than one morpheme boundary per word, ParaMor can analyze surface forms that contain more than two morphemes. Other contrasts between ParaMors search strategies and those of Snover and of Hammarstrm include: The scheme nodes in the network that is defined for Snovers Directed Search algorithm, and that is implicit in Hammarstrms work, are organized only by suffix set inclusion relations and so Snovers network is a subset of what Section  REF _Ref193859274 \r \h  3.1.2 proposes for a general search space. Furthermore, the specifics of the strategies that Snover, Hammarstrm, and the ParaMor algorithm use to search the networks of candidate paradigms are radically different. Snovers Directed Search algorithm is an exhaustive search that evaluates each network node in isolation; Hammarstrms search algorithm also assigns an intrinsic score to each node but searches the network in a greedy fashion; and ParaMors search algorithm, Section  REF _Ref214085008 \r \h  3.2, is a greedy algorithm like Hammarstrm s, but explores candidate paradigms by comparing each candidate to network neighbors. Finally, unlike Snover s Directed Search algorithm, neither Hammarstrm nor ParaMor actually instantiate the full suffix network. Instead, these algorithms dynamically construct only the needed portions of the full network. Dynamic network construction allows ParaMor to induce paradigms over a vocabulary nearly three times larger than the largest vocabulary Snovers system handles. Snover, Jarosz, and Brent (2002) discusses the possibility of using a beam or best-first search strategy to only search a subset of the full suffix network when identifying the initial best scoring nodes, but does not report results. Mathias Creutzs Morfessor In a series of papers culminating in a Ph.D. thesis (Creutz and Lagus, 2002; Creutz, 2003; Creutz and Lagus, 2004; Creutz, 2006) Mathias Creutz builds a probability-based morphology induction system he calls Morfessor that is tailored to agglutinative languages. In languages like Finnish, Creutzs native language, long sequences of suffixes agglutinate to form individual words. Morfessors ability to analyze agglutinative structures inspired the ParaMor algorithm of this thesis to also account for suffix sequencesalthough the ParaMor and Morfessor algorithms use vastly different mechanisms to address agglutination. To begin, Creutz and Lagus (2002) extend a basic MDL morphology model (Brent, Murthy, and Lundberg, 1995) to account for the morpheme sequences that are typical of agglutinative languages. Creutz and Lagus (2002) define an MDL model that consists of just two parts: 1. A list of morphs, character strings that likely represent morphemes, where a morpheme could be a stem, prefix, or suffix. 2. A list of morph sequences that result in valid word forms By allowing each word to contain many morphs, Creutz and Lagus Morfessor system neatly defies the single-suffix-per-word restriction found in so much work on unsupervised morphology induction. The search space of agglutinative morphological models is large. Each word type can potentially contain as many morphemes as there are characters in that word. To rein in the number of models actually considered, Creutz and Lagus (2002) use a greedy search strategy where each word is recursively segmented into two strings they call morphs as long as segmentation lowers the global description length. Creutz (2003) improves Morfessors morphology induction by moving from a traditional MDL framework, where models are evaluated according to their efficiently encoded size, to a probabilistic one, where model scores are computed according to a generative probability model (and implicitly relying on Bayes theorem). Where Snover (2002) defines a paradigm-based probability model, Creutz (2003) probability model is tailored for agglutinative morphology models, and does not consider paradigm structure. Creutz (2003) does not modify the greedy recursive search strategy Morfessor uses to search for a strong morphology model. Finally, Creutz and Lagus (2004) refine the agglutinative morphology models that Morfessor selects by introducing three categories: prefix, stem, and suffix. The Morfessor system assigns every identified morph to each of these three categories with a certain probability. Creutz and Lagus then define a simple Hidden Markov Model (HMM) that describes the probability of outputting any possible sequence of morphs that conforms to the regular expression: (prefix* stem suffix*)+. The morphology models described in this series of three papers each quantitatively improves upon the previous. Creutz and Lagus (2004) compare word segmentation precision and recall scores from the full Morfessor system that uses morph categories to scores from Goldsmiths Linguistica. They report results over both English and Finnish with a variety of corpus sizes. When the input is a Finnish corpus of 250,000 tokens or 65,000 types, the category model achieves an F1 of 0.64 from a precision of 0.81 and a recall of 0.53, while Linguistica only attains an F1 of 0.56 from a precision of 0.76 and a recall of 0.44. On the other hand, Linguistica does not fare so poorly on a similarly sized corpus of English (250,000 tokens, 20,000 types): Creutz and Lagus Morfessor Category model: F1: 0.73, precision: 0.70, recall: 0.77; Linguistica: F1: 0.74, precision: 0.68, recall: 0.80. Other Approaches to Unsupervised Morphology Induction Sections  REF _Ref212623558 \r \h  2.1 and  REF _Ref212954957 \r \h  2.2 presented unsupervised morphology induction systems which directly influenced the design of the ParaMor algorithm in this thesis. This section steps back for a moment, examining systems that either take a radically different approach to unsupervised morphology induction or that solve independent but closely related problems. Lets begin with Schone and Jurafsky (2000), who pursue a very different approach to unsupervised morphology induction from ParaMor. Schone and Jurafsky notice that in addition to being orthographically similar, morphologically related words are similar semantically. Their algorithm first acquires a list of pairs of potential morphological variants (PPMVs) by identifying, in a trie, pairs of vocabulary words that share an initial string. This string similarity technique was earlier used in the context of unsupervised morphology induction by Jacquemin (1997) and Gaussier (1999). Schone and Jurafsky apply latent semantic analysis (LSA) to score each PPMV with a semantic distance. Pairs measuring a small distance, those pairs whose potential variants tend to occur where a neighborhood of the nearest hundred words contains similar counts of individual high-frequency forms, are then proposed as true morphological variants of one another. In later work, Schone and Jurafsky (2001) extend their technique to identify not only suffixes but also prefixes and circumfixes. Schone and Jurafsky (2001) report that their full algorithm significantly outperforms Goldsmiths Linguistica at identifying sets of morphologically related words. Following a logic similar to Schone and Jurafsky (2000; 2001), Baroni, Matiasek, and Trost (2002) marry a mutual information derived semantic based similarity measure with an orthographic similarity measure to induce the citation forms of inflected forms. And in the information retrieval literature, where stemming algorithms share much in common with morphological analysis systems, Xu and Croft (1998) describe an unsupervised stemmer induction algorithm that also has a flavor similar to Schone and Jurafskys. Xu and Croft start from sets of word forms that, because they share the same initial three characters, likely share a stem. They then measure the significance of word form co-occurrence in windows of text. Word forms from the same initial string set that co-occur unusually often are placed in the same stem class. Finally, this discussion concludes with a look at some work which begins to move beyond simple word-to-morpheme segmentation. All of the unsupervised morphology induction systems presented thus far, including the ParaMor algorithm of this thesis, cannot generalize beyond the word forms found in the induction corpus to hypothesize unseen inflected words. Consequently, the induction algorithms described in this chapter are suitable for morphological analysis but not for generation. Chan (2006) seeks to close this generation gap. Using Latent Dirichlet Allocation, a dimensionality reduction technique, Chan groups suffixes into paradigms and probabilistically assigns stems to those paradigms. Over preprocessed English and Spanish texts, where each individual word has been morphologically analyzed with the correct segmentation and suffix label, Chans algorithm can perfectly reconstruct the suffix groupings of morphological paradigms. In other work that looks beyond word segmentation, Wicentowski and Yarowsky (Wicentowski, 2002; Yarowsky and Wicentowski, 2000; Yarowsky, Ngai, and Wicentowski, 2001) iteratively train a probabilistic model that identifies the citation form of an inflected word from several individually unreliable measures including: relative frequency ratios of stems and inflected word forms, contextual similarity of the candidate forms, the orthographic similarity of the forms as measured by a weighted Levenstein distance, and in Yarowsky, Ngai, and Wicentowski (2001) a translingual bridge similarity induced from a clever application of statistical machine translation style word alignment probabilities. Wicentowskis work stands out in the unsupervised morphology induction literature for explicitly modeling two important but rarely addressed morphological phenomena: non-concatenative morphology and morphophonology. Wicentowski and Yarowskys probabilistic model explicitly allows for non-concatenative stem-internal vowel changes as well as phonologic stem-boundary alterations that may occur when either a prefix or a suffix attaches to a stem. More recent work has taken up the work on both non-concatenative morphological processes and morphophonology. Xanthos (2007) builds an MDL-based morphology induction system specifically designed to identify the interstitial root-and-pattern morphology of Semitic languages such as Arabic and Hebrew. While in other work, Demberg (2007) takes a close look at morphophonology. Dembergs system extracts, from forward and backward tries, suffix clusters similar to Goldsmiths signatures. Demberg then measures the edit distance between suffixes in a cluster and notes that suffixes with a small distance are often due to morphophonologic changes. Probst (2003) pursues another less-studied aspect of unsupervised morphology induction: Unlike any other recent proposal, Probsts unsupervised morphology induction system can assign morphosyntactic features (i.e. Number, Person, Tense, etc.) to induced morphemes. Like Yarowsky, Ngai, and Wicentowski (2001), Probst uses machine translation word alignment probabilities to develop a morphological induction system for the second language in the translation pair. Probst draws information on morphosyntactic features from a lexicon that includes morphosyntactic feature information for the first language in the pair and projects this feature information onto the second language. Probsts work, as well as Chans and that of Yarowsky and Wicentowskis, take small steps outside the unsupervised morphology induction framework by assuming access to limited linguistic information. Discussion of Related Work ParaMor, the unsupervised morphology induction system described in this thesis, fuses ideas from the unsupervised morphology induction approaches presented in the previous three sections and then builds on them. Like Goldsmiths (2001; 2006) Linguistica, ParaMor intentionally models paradigm structure; and like Creutzs (2006) Morfessor, ParaMor addresses agglutinative sequences of suffixes; but unlike either, ParaMor tackles agglutinative sequences of paradigmatic morphemes. Similarly, ParaMors morpheme search space is a synthesis of the paradigmatic/syntagmatic morphology structure modeled by Snover (2002) on the one hand, and the finite state phoneme sequence description of morphology (Harris, 1955; Johnson and Martin, 2003) on the other. In a related vein, while ParaMor does not model morphology in a minimum description length (MDL) framework as Brent, Murthy, and Lundberg (1995), Baroni (2000), and Goldsmith (2001; 2006) do, and while ParaMor does not define a probabilistic model as Snover (2002) and Creutz (2003) do, ParaMor does acknowledge the premise of these systems that a compact morphology model is desirable. Indeed, the basic building blocks of the network search space defined in Chapter 3, schemes, are a compact representation of morphological structure. The work proposed for this thesis does not directly extend every promising approach to unsupervised morphology described in the first three sections of this chapter. For example, ParaMor only models suffixing morphology, and does not follow Baroni (2000) or Schone and Jurafsky (2001) in addressing prefixes, or Xanthos (2007) in modeling nonconcatenative processes. ParaMor also postpones for future work the modeling of morphophonologic change, al a Wicentowski (2002). Finally, the work by Schone and Jurafsky (2000), Wicentowski (2002), and others on identifying morphologically related word forms by analyzing their semantic and syntactic relatedness is both interesting and promising. While this thesis does not pursue this direction, integrating semantic and syntactic information into ParaMors existing algorithms is an exciting path for future work on unsupervised morphology induction. Clearly, ParaMor owes a great debt to previously proposed ideas for unsupervised morphology induction. Without the example of previously built systems, ParaMor would not have been able to spearhead new work on topics including a comprehensive search space of candidate paradigms, innovative measures to search that space, or how to transform candidate paradigms that individually model a single morpheme boundary into agglutinative analyses consisting of multiple morphemes. As the next chapters will show, biasing the morphology induction problem with the paradigmatic, syntagmatic, and phoneme sequence structure inherent in natural language morphology is the powerful leg-up needed for an unsupervised solution to the morphology induction problem. Paradigm Identification with ParaMor This thesis describes and motivates ParaMor, an unsupervised morphology induction algorithm. To uncover the organization of morphology within a specific language, ParaMor leverages paradigms as the language independent structure of natural language morphology. In particular ParaMor exploits paradigmatic and syntagmatic relationships which hold cross-linguistically among affixes and lexical stems respectively. The paradigmatic and syntagmatic properties of natural language morphology were presented in some detail in Section  REF _Ref192388754 \r \h  \* MERGEFORMAT  1.1. Briefly, an inflectional paradigm in morphology consists of: 1. A set of mutually substitutable, or paradigmatically related, affixes, and 2. A set of syntagmatically related stems which all inflect with the affixes in 1. This chapter describes and motivates ParaMors unsupervised strategies to initially isolate likely partial models of paradigmatic structures. These initial paradigm models are partial in two ways. First, most of ParaMors initial models will only describe a subset of the affixes in any particular paradigm. The clustering algorithm described in Chapter 4 is specifically designed to join initial models that describe subsets of the same underlying paradigm. Second, while many natural languages combine several morphemes to form a single word, each of ParaMors paradigm models can only describe a single morpheme boundary lying between a pair of morphemes. The word-to-morpheme segmentation algorithm described in Chapter 5 will recombine ParaMors individual paradigm models to segment a single word into more than two morphemes. ParaMors paradigm discovery algorithms begin with the definition of a search space over natural groupings, or schemes, of paradigmatically and syntagmatically related candidate suffixes and candidate stems, Section  REF _Ref97712489 \r \h  3.1. As Chapter 1 indicated, this thesis focuses on identifying suffix morphology. Then, with a clear view of the search space, ParaMor searches for those schemes which most likely model the paradigm structure of suffixes within the language, Section  REF _Ref192388888 \r \h  3.2. A Search Space of Morphological Schemes Schemes The constraints implied by the paradigmatic and syntagmatic structure of natural language can organize candidate suffixes and stems into the building blocks of a search space in which to identify language specific models of paradigms. This thesis names these building blocks schemes, as each is an orderly combination of related parts (The American Heritage Dictionary, 2000). Scheme organization begins by proposing candidate morpheme boundaries at every character boundary in every word form in a corpus vocabulary. The scheme-based approach to unsupervised morphology induction is designed to work on orthographies which at least loosely code each phoneme in each word with a separate character; and, as such, ParaMors induction approach does not extend to the standard writing systems of many East Asian languages, including Chinese and Japanese. Languages often mark a specific feature value combination by explicitly not changing the form of a stem. One way to describe these empty changes is with an attached null affix, . To account for null suffixes, the set of candidate morpheme boundaries the ParaMor algorithm proposes include the boundary after the final character in each word form. Since this thesis focuses on identifying suffixes, it is assumed that each word form contains a stem of at least one character. Hence, the boundary before the first character of each word form is not considered a candidate morpheme boundary. Call each string that occurs before a candidate morpheme boundary a candidate stem or cstem, and each string after a proposed boundary a csuffix. Let V be a set of stringsa vocabulary of word types. Let  be the set of all cstems generated from the vocabulary and be the corresponding set of all csuffixes. With these preliminaries, define a scheme  to be a pair of sets of strings (, ) satisfying the following four conditions: , called the adherents of C , called the exponents of C   EMBED Equation.3   The first three conditions require each of the syntagmatically related cstems in a scheme to combine with each of the paradigmatic csuffixes of that scheme to form valid vocabulary words. The fourth condition forces a scheme to contain all of the syntagmatic cstems that form valid words with each of the paradigmatic csuffixes in that scheme. The number of cstems in  is the adherent size of , and the number of csuffixes in  is the paradigmatic level of . At this point, it is worth noting two facts about the definition of a scheme. First, the definition of a scheme allows a single word to contribute to two or more schemes which divide the word differently into cstem and csuffix. Although seemingly inoccuous, the fact that different schemes, derived from the same vocabulary, can model different morpheme boundaries in the same word will be the key to ParaMors ability to segment words into sequences of agglutinative morphemes, see Chapter 5. The second fact of note is that the fourth condition in the definition of a scheme is intentionally asymmetric. Condition four only requires a scheme to contain all the cstems that combine with all the csuffixes in the scheme, but not necessarily all the csuffixes that combine with all the cstems of a scheme. This asymmetry in the scheme definition has two far-reaching consequences. First, the asymmetry creates a unique scheme for any and every set of csuffixes derived from a vocabulary. This one-to-one correspondence between each scheme, C, and the set of csuffixes in C allows this thesis to unambiguously refer to a scheme by its set of csuffixes. But more importantly, it is the one-to-one correspondence between schemes and sets of csuffixes that permits, as discussed in Section  REF _Ref193859274 \r \h  3.1.2, the organization of schemes into a search space in which ParaMor identifies initial candidate paradigms. The second consequence of the asymmetry in the definition of a scheme will impact the algorithms, described in Section  REF _Ref192388888 \r \h  3.2, that ParaMor uses to do the searching for likely paradigms. Specifically, the consequence is as follows: given two schemes C = (, ) and = (, ) where , if  is a cstem in , then  must also be in . In other words, any stem that is part of a scheme, C, is also in every scheme that is built from a subset of the suffixes in C. Schemes: An Example To better understand how, in practice, schemes succinctly capture both the paradigmatic and syntagmatic regularities of natural language morphology, let us look at a few illustrative schemes from a small example. Each box in  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1 contains a scheme derived from one or more of the word forms listed in the top portion of the figure. The vocabulary of  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1 mimics the vocabulary of a text corpus from a highly inflected language where few, if any, lexemes will occur in the complete set of possible surface forms. Specifically, the vocabulary of  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1 lacks the surface form blaming of the lexeme blame, solved of the lexeme solve, and the root form roam of the lexeme roam. Proposing, as ParaMors procedure does, morpheme boundaries at every character boundary in every word form necessarily produces many ridiculous schemes such as the paradigmatic level three scheme ame.ames.amed, from the word forms blame, blames, and blamed and the cstem bl. Dispersed among the incorrect schemes, however, are also schemes that seem very reasonable, such as .s, from the cstems blame and solve. Schemes are intended to capture both the paradigmatic and syntagmatic structure of morphology. For example, the fact that the paradigmatically related csuffixes , and s each concatenate onto both of the syntagmatically related cstems solve and blame suggests that the cstem blame should be an adherent of the scheme .sjust as the given definition of a scheme requires. But note that if schemes were limited to containing cstems that concatenate only the csuffixes in that scheme, then the cstem blame could not be part of the .s scheme, as the cstem blame also occurs in the form blamed. Before moving on, observe two additional intricacies of scheme generation. First, while the scheme .s arises from the pairs of surface forms (blame, blames) and (solve, solves), there is no way for the form roams to contribute to the .s scheme because the surface form roam is not in this vocabulary. Second, as a result of English spelling rules, the scheme s.d, generated from the pair of surface forms (blames, blamed), is separate from the scheme s.ed, generated from the pair of surface forms (roams, roamed). One Scheme, One Morpheme Boundary Behind each scheme, C, is a set of licensing word forms, W, which contribute cstems and csuffixes to C. Each csuffix in C which matches the tail of a licensing word, , segments w in exactly one position. Although it is theoretically possible for more than one csuffix of C to match a particular licensing word form, in empirical schemes that arise from natural language text, each word w typically matches just one csuffix in C. Hence, a naturally occurring scheme, C, models only a single morpheme boundary in each word w that licenses C. But words in natural language may possess more than one morpheme boundary. In Spanish, as discussed in Section  REF _Ref192489643 \r \h  1.1 of the thesis introduction, Past Participles of verbs contain either two or three morpheme boundaries: one boundary after the verb stem and before the Past Participle marker, ad on ar verbs; one boundary between the Past Participle marker and the Gender suffix, a for Feminine, o for Masculine; and, if the Past Participle is plural, a final morpheme boundary between the Gender suffix and the Plural marker, s; see  REF _Ref192490406 \h  \* MERGEFORMAT Figure  1.1 on p.15. Although a single scheme models just a single morpheme boundary in a particular word, together separate schemes can model all the morpheme boundaries of a class of words. In Spanish Past Participles, the .s scheme can model the paradigm for the optional Number suffix, while another scheme, a.as.o.os, models the cross-product of the Gender and Number paradigms, and yet another scheme, which includes the csuffixes ada, adas, ado, and ados, models the cross-product of three paradigms: Verbal Form, Gender, and Number. In one particular corpus of 50,000 types of newswire Spanish the .s scheme contains 5501 cstems, the a.as.o.os scheme 892 cstems, and the scheme ada.adas.ado.ados contains 302 cstems. Notice that it is only when a scheme models the final morpheme boundary of the schemes licensing words, that a scheme can model a full traditional paradigm. When a scheme captures morpheme boundaries that are not word final, then the schemes csuffixes encapsulate two or more traditional morphemes. Schemes which model more than one morpheme in a single csuffix no longer correspond to a single traditional paradigm, but instead capture a cross-product of several paradigms. Although the only traditional paradigms that schemes can directly model are word-final, schemes still provide this thesis with a strong model of natural language morphology for two reasons. First, as noted in the previous paragraph, while any particular scheme cannot by itself model a single word-internal paradigm, in concert, schemes can identify agglutinative sequences of morphemes. Second, the cross-product structure captured by a scheme retains the paradigmatic and syntagmatic properties of traditional inflectional paradigms. Just as suffixes in a traditional paradigm can be interchanged on adherent stems to form surface forms, the csuffixes of a cross-product scheme can be swapped in and out to form valid surface forms with the adherent cstems in the scheme: Just as a traditional paradigm might replace the final s in the Spanish word form administradas with to yield the valid form administrada, a scheme might suggest replacing the final cross-product suffix as with o to form the grammatical Spanish word form administrado. And it is the paradigmatic and syntagmatic properties of paradigms (and schemes) which ParaMor exploits in its morphology induction algorithms. Ultimately, restricting each scheme to model a single morpheme boundary is computationally much simpler than a model which allows more than one morpheme boundary per modeling unit. And, as Chapters 5 and 6 show, algorithms built on the simple scheme allow ParaMor to effectively analyze the morphology of even highly agglutinative languages such as Finnish and Turkish. Scheme Networks Looking at  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1, it is clear there is structure among the various schemes. In particular, at least two types of relations hold between schemes. First, hierarchically, the csuffixes of one scheme may be a superset of the csuffixes of another scheme. For example the csuffixes in the scheme e.es.ed are a superset of the csuffixes in the scheme e.ed. Second, cutting across this hierarchical structure are schemes which propose different morpheme boundaries within a set of word forms. Compare the schemes me.mes.med and e.es.ed; each is derived from exactly the triple of word forms blame, blames, and blamed, but differ in the placement of the hypothesized morpheme boundary. Taken together the hierarchical csuffix set inclusion relations and the morpheme boundary relations impose a lattice structure on the space of schemes.  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2 diagrams a scheme lattice over an interesting subset of the columns of  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1. Each box in  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2 is a scheme, where, just as in  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1, the csuffix exponents are in bold and the cstem adherents are in italics. Hierarchical csuffix set inclusion links, represented by solid lines ( SHAPE \* MERGEFORMAT ), connect a scheme to often more than one parent and more than one child. The empty scheme (not pictured in  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2), containing no csuffixes and no cstems, can be considered the child of all schemes of paradigmatic level 1 (including the scheme). Horizontal morpheme boundary links, dashed lines ( SHAPE \* MERGEFORMAT ), connect schemes which hypothesize morpheme boundaries which differ by a single character. In most schemes of  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2, the csuffixes all begin with the same character. When all csuffixes begin with the same character, there can be just a single morpheme boundary link leading to the right. Similarly, a morphology scheme network contains a separate leftward link from a particular scheme for each character which ends some cstem in that scheme. The only scheme with explicit multiple left links in  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2 is , which has depicted left links to the schemes e, s, and d. A number of left links emanating from the schemes in  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2 are not shown; among others absent from the figure is the left link from the scheme e.es leading to the scheme ve.ves with the adherent sol. Section  REF _Ref190771388 \r \h  4.4.2 defines morpheme boundary links more explicitly. The scheme network of  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2 is a portion of the search space in which the ParaMor unsupervised morphology induction algorithm would operate given the small vocabulary of  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1. But it is important to note that the ParaMor algorithm would not necessarily instantiate every scheme node in  REF _Ref69705450 \h  \* MERGEFORMAT Figure  3.2. The number of schemes in a full scheme network is the size of the powerset of the set of all word-final strings in the vocabulary! Much too large to exhaustively search or build for even moderately sized corpora. As Sections  REF _Ref214372965 \r \h  3.2.2 and  REF _Ref192653564 \r \h  3.2.4 discuss, ParaMor dynamically creates just those portions of the search space most likely to harbor schemes that correctly model true paradigms of a language. Two additional graphical examples, these generated from naturally occurring text, will help visualize scheme-based search spaces.  REF _Ref87327235 \h  \* MERGEFORMAT Figure  3.3 contains a portion of a search space of schemes automatically generated from 100,000 tokens of the Brown Corpus of English (Francis, 1964); while  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4 diagrams a portion of a hierarchical lattice over a Spanish newswire corpus of 1.23 million tokens (50,000 types). As before, each box in these networks is a scheme and the csuffix exponents appear in bold. Since schemes in each search space contain more cstem adherents than can be listed in a single scheme box, abbreviated lists of adherents appear in italics. The number immediately below the list of csuffixes is the total number of cstem adherents that fall in that scheme. The scheme network in  REF _Ref87327235 \h  \* MERGEFORMAT Figure  3.3 contains the paradigmatic level four scheme covering the English csuffixes .ed.ing.s. These four suffixes, which mark combinations of Tense, Person, Number, and Aspect, are the exponents of a true sub-class of the English verbal paradigm. This true paradigmatic scheme is embedded in a lattice of less satisfactory schemes. The right-hand column of schemes posits, in addition to true inflectional suffixes of English, the derivational suffix ly. Immediately below .ed.ing.s, a scheme comprising a subset of the suffixes of the true verbal sub-class appears, namely .ed.ing. To the left, .ed.ing.s is connected to d.ded.ding.ds, a scheme which proposes an alternative morpheme boundary for 19 of the 106 cstems in the scheme .ed.ing.s. Notice that since left links effectively slice a scheme on each character in the orthography, adherent count monotonically decreases as left links are followed. Similarly, adherent count  monotonically decreases as csuffix set inclusion links are followed upward. Consider again the hierarchically related schemes .ed.ing.s and .ed.ing, which have 106 and 201 adherents respectively. Since the .ed.ing.s scheme adds the csuffix s to the three csuffixes already in the .ed.ing scheme, only a subset of the cstems which can concatenate the csuffixes , ed, and ing can also concatenate s to produce a word form in the corpus. And so, only a subset of the cstems in the .ed.ing scheme belong in .ed.ing.s. Now turning to  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4, this figure covers the Gender and Number paradigms on Spanish adjectival forms. As with Spanish Past Participles, adjectives in Spanish mark Number with the pair of paradigmatically opposed suffixes s and . Similarly, the Gender paradigm on adjectives consists of the pair of strings a and o. Together the Gender and Number paradigms combine to form an emergent cross-product paradigm of four alternating strings: a, as, o, and os.  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4 contains: 1. The scheme containing the true Spanish exponents of the emergent cross-product paradigm for Gender and Number: a.as.o.os. The a.as.o.os scheme is outlined in bold. 2. All possible schemes whose csuffix exponents are subsets of a.as.o.os, e.g. a.as.o, a.as.os, a.os, etc. 3. The scheme a.as.o.os.ualidad, together with its descendents, o.os.ualidad and ualidad. The Spanish string ualidad is arguably a valid Spanish derivational suffix, forming nouns from adjectival stems. But the repertoire of stems to which ualidad can attach is severely limited. The suffix ualidad does not form an inflectional paradigm with the adjectival endings a, as, o, and os.  Searching the Scheme Lattice Given the framework of morphology scheme networks outlined in Section  REF _Ref97712489 \r \h  \* MERGEFORMAT  3.1, an unsupervised search strategy can automatically identify schemes which plausibly model true paradigms and their cross-products. Many search strategies are likely capable of identifying reasonable paradigmatic suffix sets in scheme networks. Snover (2002), for example, describes a successful search strategy over a morphology network in which each network node is assigned a global probability score (see Chapter 2). In contrast, the search strategy presented in this section gauges a schemes value by computing a local score, non-probabilistic, over the schemes network neighbors. A Birds Eye View of ParaMors Search Algorithm  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 abstractly visualizes ParaMor s initial search for partial paradigms. The two dimensional area of  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 represents ParaMor s lattice-structured search space. Embedded in ParaMor s search space are schemes whose csuffixes match the suffixes of true paradigms in  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5, the large circles, atop the shaded triangles, symbolize these correct schemes. ParaMor s paradigm search seeks to find these correct schemes. However, ParaMor is often unable to extract sufficient evidence from the input corpus during this initial search phase to fully reconstruct the paradigms of a language. And so, the primary goal of ParaMors initial search is to identify schemes which individually model subsets of the suffixes in true paradigms, and which jointly cover as many correct suffixes as possible. It will then be up to ParaMors clustering algorithm, described in Section  REF _Ref190686481 \r \h  4.3, to merge the initially selected scheme-models of partial paradigms. In a scheme lattice, the descendents of a scheme, C, are those schemes whose csuffix sets are subsets of the csuffix set in C.  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 conceptualizes the descendents of a scheme as a cone projecting downward. In particular, the shaded triangles in  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5, headed by the large paradigm circles, represent schemes whose csuffix sets are subsets of true paradigms. At a high level, then, ParaMors search algorithm should seek to identify as many schemes as possible within the shaded cones, preferably high in the lattice and preferably as widely spaced as possibleto maximize csuffix coverage. To identify candidate partial paradigms, ParaMor adopts a bottom-up search strategy. (ParaMor will harness the horizontal morpheme boundary links, which also connect networked schemes in a later stage of the algorithm, see Section  REF _Ref190771388 \r \h  4.4.2). At the bottom of a scheme network, and, conceptually, at the bottom of  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5, are the paradigmatic level one schemes which contain a single csuffix each. A priori, ParaMor does not know which sets of csuffixes model paradigms, and so, the initial search algorithm attempts to grow strong models of paradigmatic suffix sets by starting with these single csuffix schemes. To ensure a wide spacing in the selection of schemes, ParaMor pursues a separate upward search path from each level one scheme. By starting separate paths from all individual csuffixes, ParaMor awards each csuffix the chance for consideration as a true suffix. This magnanimous policy is recall-centric: ParaMors initial search hopes to find scheme models of as many true suffixes as possible. However, in starting a separate search path from each csuffix, ParaMor also increases the number of incorrectly selected csuffixesmany search paths begin from poor csuffix models. ParaMor relies on a series of filters, described in Chapter 4, to weed out the poorly chosen initial schemes. From each level one scheme, ParaMor follows a single upward path. In a greedy fashion, ParaMor selects a single csuffix to add to the current scheme. ParaMor limits each upward search to a single greedy path for computational reasonsthe full search space of candidate paradigm schemes is immense and following all reasonable paths quickly leads to an explosion in paths to follow (see Section  REF _Ref192653564 \r \h  3.2.4). Returning to  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5: Each jagged line in  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 is a search path ParaMor pursues. Search paths originate at the bottom of the figure and move upward through the lattice. The terminal schemes of each search path, marked by small circles in  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5, are the final output of ParaMor s initial search. Some of ParaMor s search paths originate within a shaded trianglethese are paths that begin from a csuffix that models a true suffix in some paradigm. But other search paths inevitably begin from csuffixes that model no suffix in the language. Furthermore, some paths which begin at schemes which do model suffixes eventually take a wrong turn, introducing an incorrect csuffix and thereby leaving the shadow of a true paradigm. The final point to note from  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 concerns the left most paradigm in the figure. For several of the terminal schemes of search paths which originate beneath this paradigm,  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5 indicates csuffix subset cones with dashed lines. The union of the bases of these dashed cones closely coincides with the base of the shaded cone of the left-most paradigm. The base of each subset cone contains all schemes which contain just a single csuffix from the scheme at the cones apex. Thus if the csuffixes from the schemes at the bases of the dashed cones were reunited, the resultant paradigm model would closely match a true paradigm. This is exactly ParaMors strategy. Section  REF _Ref190686481 \r \h  4.3 will describe a clustering algorithm that unites schemes which each model subsets of the true suffixes of individual paradigms. Why a Bottom-Up Search? The choice to design ParaMor s initial paradigm identification algorithm as a repeated series of bottom-up search paths through a scheme network, a la  REF _Ref214525226 \h  \* MERGEFORMAT Figure  3.5, was explicitly taken so as to leverage the paradigmatic and syntagmatic structure that is captured by the vertical csuffix set inclusion links of the scheme network search space. At the bottom of a network of schemes, syntagmatic stem alternations are evident but each scheme contains only a single csuffix. At successively higher levels, the networked schemes contain not only successively more paradigmatically opposed csuffixes, but also successively fewer syntagmatic cstems. ParaMors search strategy moves upward through the network, trading off syntagmatic cstem alternations for paradigmatic alternations of csuffixes. Consider the paradigmatic and syntagmatic structure captured by and between the schemes of the Spanish network in  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4 from p.59. The schemes at the bottom of this network each contain exactly one of the csuffixes a, as, o, os, or ualidad. The syntagmatic cstem evidence for the level 1 schemes which model productive inflectional suffixes of Spanish, namely a, as, o, and os, is significantly greater than the syntagmatic evidence for the unproductive derivational csuffix ualidad: The a, as, o, and os schemes contain 9020, 3182, 7520, and 3847 cstems respectively, while the ualidad scheme contains just 10 cstems. Moving up the network, paradigmatic-syntagmatic tradeoffs strongly resonate. Among the 3847 cstems which allow the csuffix os to attach, more than half, 2390, also allow the csuffix o to attach. In contrast, only 4 cstems belonging to the os scheme form a corpus word with the csuffix ualidad: namely, the cstems act, cas, d, and event. Adding the suffix a to the scheme o.os again reduces the cstem count, but only from 2390 to 1418; and further adding as, just lowers the cstem count to 899. There is little syntagmatic evidence for adding csuffixes beyond the four in the scheme a.as.o.os. Adding the non-paradigmatic csuffix ualidad, for example, drastically reduces the count of syntagmatic cstems to a meager 3. It is insightful to consider why morphology scheme networks capture tradeoffs between paradigmatic and syntagmatic structures so succinctly. If a particular csuffix, f, models a true inflectional suffix (or cross-product of suffixes), then, disregarding morphophonologic change, the paradigmatic property of inflectional morphology implies f will be mutually substitutable for some distinct csuffix . Consequently, both f and  will occur in a text corpus attached to many of the same syntagmatically related cstems. In our example, when f is the csuffix os and  the paradigmatically related o, many cstems to which os can attach also allow o as a word-final string. Conversely, if the f and  suffixes lack a paradigmatic relationship in the morphological structure of some language, then there is no a priori reason to expect f and  to share cstems: when f is os and  ualidad, a csuffix which is not paradigmatically opposed to os, few of the cstems which permit an os csuffix admit ualidad. ParaMors Initial Paradigm Search: The Algorithm With the motivation and background presented in the previous subsection, the specifics of ParaMors search algorithm follow: ParaMors bottom-up search treats each individual non-null csuffix as a potential gateway to a model of a true paradigm cross-product. ParaMor considers each one-suffix scheme in turn beginning with that scheme containing the most cstems, and working toward one-suffix schemes containing fewer cstems. From each bottom scheme, ParaMor follows a single greedy upward path from child to parent. As long as an upward path takes at least one step, making it to a scheme containing two or more alternating csuffixes, ParaMors search strategy accepts the terminal scheme of the path as likely modeling a portion of a true inflection class. To take each greedy upward search step, ParaMor first identifies the best scoring parent of the current scheme according to a particular scoring function. Section  REF _Ref193274112 \r \h  3.2.5 will propose and evaluate one reasonable class of parent scoring function. ParaMor then applies two criteria to the highest scoring parent. The first criterion is a threshold on the parents score. ParaMors upward search will only move to a parent whose score passes the set threshold. The second criterion governing each search step helps to halt upward search paths before judging parents worth becomes impossible. As noted in the second half of Section  REF _Ref214372949 \r \h  3.2.1 above, cstem counts monotonically decrease with upward network moves. But small adherent cstem counts render statistics that assess parents strength unreliable. ParaMor s policy is to not move to any scheme that does not contain more cstems than it has csuffixes. Moreover, this second criterion serves ParaMor well for two additional reasons. First, requiring each path scheme to contain more cstems than csuffixes attains high suffix recall by setting a low bar for upward movement at the bottom of the network. Hence, search paths which begin from schemes whose single csuffix models a rare but valid suffix, can often take at least one upward search step and manage to be selected. Second, this halting criterion requires the top scheme of search paths that climb high in the network to contain a comparatively large number of cstems. Reigning in high-reaching search paths, before the cstem count falls too far, captures path-terminal schemes which cover a large number of word types. In a later stage of ParaMors paradigm identification algorithm, presented in Sections  REF _Ref216074389 \r \h  4.3.1 and  REF _Ref215319787 \r \h  4.3.2, these larger terminal schemes effectively vacuum up the useful smaller paths that result from the more rare suffixes. At times, ParaMor s search algorithm can reach the same scheme along more than one path. But, since ParaMors upward search from any particular scheme is deterministic, there is no need to retrace the same best path more than once from any particular scheme. While it might be reasonable to follow a next-best path each time a scheme is re-encountered, ParaMor instead simply abandons the redundant path. In practice, the large number of individual paths that ParaMor follows ensures that ParaMor discovers all the prominent paradigms, even if a few redundant paths are discarded. Pseudo code for ParaMor s bottom-up initial search is given in  REF _Ref213917055 \h  \* MERGEFORMAT Figure  3.6. The core search() method calls searchOneGreedyPath() once for each scheme which contains a single csuffix. Most of the function names in the pseudo code are self-explanatory, but note that the initializeDynamicNetwork() method does not build the full scheme lattice. This initialization method merely prepares the necessary data structures to enable ParaMor to quickly build schemes as needed. Details on ParaMor s dynamic network generation appear in Section  REF _Ref192653564 \r \h  3.2.4. Experiments with a range of languages, including Spanish, English, German, Finnish, and Turkish, show that ParaMor s bottom-up search algorithm takes less than 10 minutes to complete, when given a paradigm induction corpus containing 50,000 unique types. ParaMor is currently implemented in Java and runs on a standard Linux server.  ParaMor s Bottom-Up Search in Action  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7 contains a number of search paths that ParaMor followed when analyzing a Spanish newswire corpus of 50,000 types and when using one particular metric for parent evaluation (See Section  REF _Ref214336453 \r \h  3.2.5). Most of the paths in  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7 are directly relevant to the analysis of the Spanish word administradas. As mentioned in Chapter 1, the word administradas is the Feminine, Plural, Past Participle form of the verb administrar,  to administer or manage. The word administradas gives rise to many csuffixes including: stradas, tradas, radas, adas, das, as, s, and . Of these candidate suffixes, s marks Spanish plurals and is a word final string of 10,662 word forms in this same corpusmore than one fifth of the unique word forms. Additionally, in the word administradas, the left edges of the word-final strings as and adas occur at Spanish morpheme boundaries. All other derived csuffixes incorrectly segment administradas: The csuffixes radas, tradas, stradas, etc. erroneously include part of the stem; while das, in the analysis of Spanish morphology adopted by this thesis, places a morpheme boundary internal to the Past Participle morpheme ad; and does not mark any morphosyntactic feature on Spanish adjectives when it occurs after a Plural suffix. Of course, while we can discuss which csuffixes are reasonable and which are not, an unsupervised morphology induction system has no a priori knowledge of Spanish morphology. ParaMor does not know what strings are valid Spanish morphemes, nor is ParaMor aware of the feature value meanings associated with morphemes. Each search path of  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7 begins at the bottom of the figure and proceeds upwards from scheme to scheme. In Spanish, the non-null csuffix that can attach to the most stems is s; and so, the first search path ParaMor explores begins from s. This search path is the right-most search path shown in  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7. At 5513 cstems, the null csuffix, , can attach to the largest number of cstems to which s can attach. The parent-evaluation function gave the .s scheme the highest score of any parent of the s scheme, and the score of .s passed the parent score threshold. Consequently, the first search step moves to the scheme which adds to the csuffix s. ParaMors parent-evaluation function then identifies the parent scheme containing the csuffix r as the new parent with the highest score. Although no other csuffix can attach to more cstems to which s and can both attach, r can only form corpus words in combination with 281 or 5.1% of the 5513 cstems which take s and . Accordingly, the score assigned by this particular parent-evaluation function to the .s.r scheme falls below the stipulated threshold; and ParaMor does not add r, or any other suffix, to the now closed partial paradigm s.. Continuing leftward from the s-anchored search path in  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7, ParaMor follows search paths from the csuffixes a, n, es, and an in turn. The 71st csuffix from which ParaMor grows a partial paradigm is rado. The search path from rado is the first path to build a partial paradigm that includes the csuffix radas, a csuffix potentially relevant for an analysis of the word administradas. Similarly, search paths from trado and strado lead to partial paradigms which include the csuffixes tradas and stradas respectively. The search path from strado illustrates the second criterion restricting upward search. From strado, ParaMors search adds four csuffixes, one at a time: strada, str, strar, and stradas. Only seven cstems form words when combined singly with all five of these csuffixes. Adding any additional csuffix to these five brings the cstem count in this corpus down at least to six. Since six cstems is not more than the six csuffixes which would be in the resulting parent scheme, ParaMor does not add a sixth csuffix. The Construction of Scheme Networks It is computationally impractical to build full morphology scheme networks, both in terms of space and time. The space and time complexities of full network creation are directly related to the number of schemes in a network. Returning to the definition of a scheme in Section  REF _Ref192582713 \r \h  3.1.1, each scheme contains a set of csuffixes, , where  is the set of all possible csuffixes generated by a vocabulary. Thus, the set of potential schemes from some particular corpus is the powerset of , with  members. In practice, the vast majority of the potential schemes have no adherent cstemsthat is, for most there is no cstem, t, such that  is a word form in the vocabulary. If a scheme has no adherent cstems, then there is no evidence for that scheme, and network generation algorithms would not need to actually create that scheme.  Unfortunately, even the number of schemes which do possess adherent cstems grows exponentially. The dominant term in the number of schemes with a non-zero cstem count comes from the size of the power set of the scheme with the largest set of csuffixesthe highest level scheme in the network. In one corpus of 50,000 Spanish types, the highest level scheme contains 5816 csuffixes. The number of schemes in this network is thus grater than , a truly astronomical number, larger than . Or more schemes, by far, than the number of hydrogen atoms in the observable universe. Because of the difficulty in pre-computing full scheme networks, during the initial upward scheme search (described in Sections  REF _Ref214372949 \r \h  \* MERGEFORMAT  3.2.1 and  REF _Ref214372965 \r \h  3.2.2) any needed schemes are calculated dynamically. The rest of this section spells out ParaMor s scheme-building procedure in detail. Most-Specific Schemes ParaMor builds individual schemes dynamically from most-specific schemes. Like full-blooded schemes, each most-specific scheme, M, is a pair of sets of strings, (,):  a set of cstems, and  a set of csuffixes. Furthermore, as in the definition of a scheme found in Section  REF _Ref192582713 \r \h  3.1.1, four conditions stipulate which cstems and csuffixes occur in each most-specific scheme. The first three conditions for most-specific schemes are identical to the first three conditions on full schemes:   where is the set of all cstems generated by a vocabulary   the set of all csuffixes generated by a vocabulary 3.   V being the corpus vocabulary But the fourth condition changes in the definition of most-specific scheme. The new condition is: 42 .  This new fourth condition requires each  to form vocabulary words with exactly and only the csuffixes in  (exactly the situation imagined in the case of the cstem blame and the scheme .s in Section  REF _Ref192582713 \r \h  3.1.1 on p.52). A consequence of this new fourth restriction that differentiates most-specific schemes from basic schemes is that each corpus cstem occurs in precisely one most-specific scheme: Suppose a cstem t belonged to two most-specific schemes, M = (,) and = (,), . Since , without loss of generality, there must exist  such that . And because , it follows that . But now we have a contradiction as there exists a cstem in , namely t, and a csuffix that is not in , namely , that concatenate to form a vocabulary item. Hence, t cannot be an element of both  and . The idea of the most-specific scheme has been proposed previously: translating into the terminology of this thesis, Demberg (2007) stores cstems that are not surface types themselves in most specific schemes. Additionally, as the csuffixes of a most-specific scheme, C, exactly specify the yield of the cstems in C, the states in the minimal (forward) character-based finite state automaton that exactly accepts a vocabulary are in a one-to-one correspondence with the set of most-specific schemes generated by that vocabulary. And so, finite state approaches to morphology induction, such as Johnson and Martin (2003), also implicitly compute most-specific schemes. But to my knowledge, no one has suggested dynamically generating full-fledged schemes from their most-specific cousins. Computing Most-Specific Schemes Since the number of most-specific schemes is bounded above by the number of cstems in a corpus, the number of most-specific schemes grows much more slowly with vocabulary size than does the total number of schemes in a network. In the 50,000 type Spanish corpus from above, a mere 28,800 (exact) most-specific schemes occurred. A two-part algorithm can quickly compute the set of most-specific schemes from a corpus. The first half of the algorithm associates each cstem, t, with the set of csuffixes, , where each csuffix in  forms a vocabulary word when combined with t. To find  for each t, ParaMor loops through each (cstem t, csuffix f) division of each vocabulary word. Whenever ParaMor encounters a specific cstem, t, ParaMor adds the corresponding f csuffix of the current pair to . In the second half of the algorithm that computes most specific schemes, ParaMor associates each unique set of csuffixes, F, that was computed in the algorithms first half with the set of cstems that individually pointed to identical copies of F. Pseudo-code for the creation of most-specific schemes is given in  REF _Ref214350828 \h  \* MERGEFORMAT Figure  3.8. Overloading the | | operator to denote both the size of a set and the length of a string, the time complexity of this two-part algorithm to compute most-specific schemes is , where, as above, V is the vocabulary size and is the set of all cstems generated from a vocabulary. In practice, computing the most-specific schemes from corpora of 50,000 types takes the current Java implementation of ParaMor less than five minutes on a standard Linux server, with ParaMor spending about equal time in the first and second halves of the algorithm. To Schemes from Most-Specific Schemes From the full set of most specific schemes, the cstems, , of any particular scheme, C = (, ), can be directly computed from the given . Since a single cstem can belong to multiple schemes but to only one most-specific scheme, the cstems of the C scheme are scattered among various most-specific schemes. Recall from the definition of a scheme (Section  REF _Ref192582713 \r \h  3.1.1) that the cstems in  are all those cstems which form words with all csuffixes in . Let M = (, ) be a most-specific scheme where , and let . Since all cstems in  form words with all csuffixes in , and since ,  must form a word with all csuffixes in . Hence,  belongs in . The converse, that if  then  does not belong in , follows from similar logic. Consequently, to dynamically compute , ParaMor must place in  all cstems from all most-specific schemes whose csuffix sets are (proper or improper) supersets of . Let  denote the set of all most-specific schemes whose csuffix sets are supersets of . And in general, given a set of csuffixes, F, define  as the set of most specific schemes whose csuffixes are a (proper or improper) superset of F. To quickly compute the cstems of any particular scheme, C = (, ), ParaMor must efficiently compute , i.e. without exhaustively comparing  to the csuffix set of every most specific scheme. Consider , the set of all most-specific schemes whose csuffix sets contain , for each . These  are easy to compute: for each most-specific scheme, M=(, ), and for each csuffix , simply add M to the set of all most-specific schemes that  occurs in. Furthermore,  for any particular  need only be computed once.  REF _Ref214443964 \h  \* MERGEFORMAT Figure  3.9 give pseudo-code, that ParaMor calls just once, to compute  for all individual  csuffixes that are generated by a corpus. But to dynamically compute the cstems of a scheme, C, ParaMor s must find , not individual . As the following Lemma shows, for any scheme, C,  can be computed from the set of all  where  ranges over all csuffixes in : Lemma:  To show : If M = (, ) , that is, if , then , , and . And so , . And thus . To show : If M = (, ) , then , . And By the def of , , ; and of course . But suppose , then there must exist  such that a contradiction. So,  and . The lemma has been proved. In summary then, for any scheme C = (, ), with a specified csuffix set, , the cstems  are the union of the cstems from all most-specific schemes in , that is, , while the members of each are found by intersecting the pre-computed sets of most-specific schemes in the , or . The key is that both the intersection step as well as the unification step are quick to compute. Pseudo code for the full algorithm to dynamically compute a scheme from a set of most-specific schemes is given in  REF _Ref214461298 \h  \* MERGEFORMAT Figure  3.10. Upward Search Metrics As described in detail in Sections  REF _Ref214372949 \r \h  3.2.1 and  REF _Ref214372965 \r \h  3.2.2 of this chapter, at each step of ParaMor s bottom-up search, the system selects, or declines to select, a parent of the current scheme as most likely to build on the paradigm modeled by the current scheme. A key element of ParaMors bottom-up search is the exact form of the parent-evaluation metric used to select the most promising parent. The parent metric must reliably determine where and when to expand the current candidate paradigm using only the limited information available: As an unsupervised induction system, ParaMor has no knowledge of, for example, the part of speech or morphosyntactic features that individual words mark. Hence, ParaMors parent-evaluation function must work from co-occurrence statistics of words and of word substrings, e.g. csuffix and cstem counts in schemes. Clearly, ParaMors parent-evaluation procedure directly impacts performance. A variety of parent-evaluation measures are conceivable, from simple local measures to statistical measures which take into consideration the schemes larger contexts. This section investigates one class of localized metric and concludes that, at least within this metric class, a conceptually and computationally simple metric gives as accurate an indication of the paradigmatic worth of a parent scheme as more complex and expensive metrics. The Parent Selection Dilemma To motivate the metrics under investigation, consider the plight of an upward search algorithm that has arrived at the a.o.os scheme when searching through the Spanish morphology scheme network of  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4 from p.59. All three of the csuffixes in the a.o.os scheme model inflectional suffixes from the cross-product paradigm of Gender and Number on Spanish adjectives. In  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7 on p.69, the second upward search path that ParaMor pursues brings ParaMor to the a.o.os scheme (the second search path in  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7 is the second path from the right). In both  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4 and in  REF _Ref192516178 \h  \* MERGEFORMAT Figure  3.7, just a single parent of the a.o.os scheme is shown, namely the a.as.o.os scheme. But these two figures cover only a portion of the full scheme network derived from the 50,000 types in this Spanish corpus. In the full scheme network there are actually 20,949 parents of the a.o.os scheme! The vast majority of the parents of the a.o.os scheme occur with just a single cstem. However, 1,283 parents contain two cstems, 522 contain three cstems, and 330 contain four. Seven of the more likely parents of the a.o.os scheme are shown in  REF _Ref192774275 \h  \* MERGEFORMAT Figure  3.11. Out of nearly 21,000 parents, only one correctly builds on this adjectival inflectional cross-product paradigm of Gender and Number: The a.as.o.os parent adds the csuffix as, which marks Feminine Plural. The parent scheme of a.o.os that has the second most cstem adherents, also shown in  REF _Ref192774275 \h  \* MERGEFORMAT Figure  3.11, adds the csuffix amente. Like the English suffix ly, the Spanish suffix (a)mente derives adverbs from adjectives quite productively. The other parents of the a.o.os scheme given in  REF _Ref192774275 \h  \* MERGEFORMAT Figure  3.11 arise from csuffixes that model verbal suffixes or that model derivational morphemes. The verbal csuffixes are ar, e, and es, while the derivational csuffixes are and ualidad. The primary reason for the fairly high cstem counts among the verbal parents of the a.o.os scheme is that Spanish syncretically employs the strings a and o not only as adjectival suffixes marking Feminine and Masculine, respectively, but also as verbal suffixes marking 3rd Person and 1st Person Present Indicative. However, for a cstem to occur in a verbal parent of a.o.os, such as a.ar.o.os, the cstem must somehow combine with os into a non-verbal Spanish word formas the csuffix os does not model any verbal inflection. In the a.ar.o.os scheme in  REF _Ref192774275 \h  \* MERGEFORMAT Figure  3.11, the four listed cstems cambi, estudi, marc, and pes model verb stems when they combine with the csuffixes a and ar, but they model, often related, noun stems when they combine with os, and the Spanish word forms cambio, estudio, marco, and peso ambiguously can be both verbs and nouns. Motivation for Paradigmatic-Syntagmatic Parent-Evaluation Metrics The task laid before the parent-evaluation function in ParaMors bottom-up scheme search algorithm is, then, to select, out of potentially thousands of candidates, at most a single parent which best extends the paradigm modeled by the current scheme. As detailed in the second half of Section  REF _Ref214372949 \r \h  3.2.1, ParaMor relies on the paradigmatic-syntagmatic frequency structure of scheme networks to make these parent selection decisions. To briefly recapitulate the paradigmatic-syntagmatic tradeoffs captured in scheme networks: suppose  is a set of suffixes which form a paradigm or paradigm cross-product. And let  be the set of csuffixes in some scheme. Because the suffixes of  are mutually substitutable, it is reasonable to expect that, in any given corpus, many of the stems which occur with  will also occur with other suffixes, , . Hence, when moving upward through a scheme network among schemes all of whose csuffixes belong to a single paradigm, adherent cstem counts should fall only slowly. Conversely, there is no reason to expect that some csuffix  will form valid words with the same cstems that form words with csuffixes in P. The C-Stem Ratio: A Simple and Effective Parent-Evaluation Metric Taking another look at  REF _Ref192774275 \h  \* MERGEFORMAT Figure  3.11, the parents of the a.o.os scheme clearly display paradigmatic and syntagmatic structure. More than 63% of the cstems in the a.o.os scheme form a word with the csuffix as, a csuffix which is paradigmatically tied to a, o and os through the Spanish adjectival paradigm. Only 10% of the cstems in a.o.os form corpus words with the csuffix ar, part of a productive inflectional verbal paradigm with a and o. And only 0.2% of the cstems in the a.o.os scheme form words with the derivational csuffix ualidad, which is not part of a productive inflectional paradigm with any of the three csuffixes a, o or os. Among all parents of the a.o.os scheme, the correct paradigmatic parent, a.as.o.os, retains by far the highest fraction of the childs cstems. Parent-child cstem ratios are surprisingly reliable predictors of when a parent scheme builds on the paradigmatic csuffix interaction of that scheme, and when a parent scheme breaks the paradigm. Indeed, while the remainder of this chapter presents and motivates a variety of parent metrics, an empirical evaluation discussed below suggests that no examined metric significantly outperforms the parent-child cstem ratio at paradigm identification. Parent-Evaluation Metrics beyond C-Stem Ratios Parent-child cstem ratios are a simple measure of a parent schemes paradigmatic strength, but it seems reasonable that a more sophisticated measure might more accurately predict when a parent extends a childs paradigm. In particular, the cstem ratio metric does not at all account for the fact that some suffixes occur less frequently than others. In Spanish, for example, while the derivational suffix (a)mente is so productive that its paradigmatic behavior is nearly that of an inflectional suffix, in naturally occurring text relatively few unique types end in amente. In comparison, the verbal infinitive suffix ar is much more common. In one corpus of 50,000 unique types 1448 word forms ended in ar but only 332 ended in amente. But the parent-child cstem ratio has difficulty differentiating between the parent which introduces the productive csuffix amente and the parent which introduces the non-paradigmatic verbal csuffix ar: Both the schemes a.amente.o.os and a.ar.o.os have very nearly the same number of cstems, and so have very similar parent-child cstem ratios of 0.122 and 0.102 respectively. The sheer number of ar forms ensures that a reasonable number of cstems adhere to the a.ar.o.os scheme. Five additional parent-evaluation metrics are discussed below. All five incorporate the frequency of the expansion csuffix, that is, they incorporate the frequency of the csuffix introduced by the parent scheme under consideration. In a scheme network, the frequency of the expansion scheme is found in the level 1 scheme which contains the single csuffix which expands the current scheme.  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12 depicts expansion schemes, with their cstem frequencies, for four parents of the a.o.os scheme. These four parents were chosen to aid exposition. The expan-  sion schemes are: as, amente, ar, and ualidad, respectively, the fourth member of the adjectival cross-product paradigm, a productive derivational suffix, a verbal suffix that is not paradigmatically related to os, and an unproductive derivational suffix. The Dice Metric There are many ways the cstem information from expansion schemes might be combined with predictions from parent-child cstem ratios. One combination method is to average parent-expansion cstem ratios with parent-child cstem ratios. In the a.amente.o.os example, the ratio of cstem counts from the parent scheme a.amente.o.os to the expansion scheme amente, at 173/332, can be averaged with the ratio of cstems from the parent scheme to the child scheme a.o.os, namely 173/1418. The bottom-up search of scheme networks particularly seeks to avoid moving to schemes that do not model paradigmatically related csuffixes. Taking the harmonic mean of the parent-expansion and parent-child cstem ratios, as opposed to the arithmetic mean, captures this conservative approach to upward movement. Compared with the arithmetic mean, the harmonic mean comes out closer to the lower of a pair of numbers, effectively dragging down a parents score if either cstem ratio is low. Interestingly, after a bit of algebra, it emerges that the harmonic mean of the parent-expansion and parent-child cstem ratios is equivalent to the dice similarity metric on the sets of cstems in the child and expansion schemes. The dice metric, a standard measure of set similarity (Manning and Schtze, 1999, p. 299), is defined as , for any two arbitrary sets X and Y. In the context of schemes, the intersection of X and Y is the intersection of the cstem sets of the child and expansion schemesor exactly the cstem set of the parent scheme. As hoped, the relative difference between the dice scores for the amente and ar parents is larger than the relative difference between the parent-child cstem ratios of these parents. The dice scores are 0.198 and 0.101 for the amente and ar parents respectively, a difference of nearly a factor of two; as compared with the relative difference factor of 1.2 for the parent-child cstem ratios of the amente and ar parents. Note that it is meaningless to directly compare the numeric value of a parent-child cstem ratio to the numeric value of the dice measure of the same parentthe two scores operate on different scales. The parent-child cstem ratio metric and the dice metric are the first two of six metrics that ParaMor investigated as potential guiding metrics for the vertical network search described in Sections  REF _Ref214372949 \r \h  3.2.1 and  REF _Ref214372965 \r \h  3.2.2. All six investigated metrics are summarized in  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13. Each row of this figure details a single metric. After the metric s name which appears in the first column, the second column gives a brief description of the metric, the third column contains the mathematical formula for calculating that metric, and the final four columns apply each rows metric to the four illustrative parent schemes of the a.o.os scheme from  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12: For example, the parent-child cstem ratio to the a.o.os.ualidad parent is given in the upper-right cell of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13 as 0.002. The variables in the mathematical notation of the third column of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13 refer to cstem counts of schemes involved in the evaluation of parent schemes. For example, the formula, in the first row of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13, to calculate a parent-child cstem ratio is P/C, where P is the count of the cstems in the parent scheme, and C is the count of cstems in the current scheme. Comparing the metric formulas in the third column of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13, the parent-child cstem ratio is by far the simplest of any metric ParaMor examined. In addition to P and C, the formulas in  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13 refer to two other variables: E represents the total number of cstems that can attach to the csuffix that expands the current scheme into the parent (i.e. the cstem adherent size of the expansion scheme), and V is the size of the vocabulary over which the network was built. The formula for the dice metric uses E in addition to P and C: 2P/(C+E). Pointwise Mutual Information as a Parent-Evaluation Metric Of the six parent-evaluation metrics that ParaMor examined, the four which remain to be described all look at the occurrence of cstems in schemes from a probabilistic perspective. To convert cstem counts into probabilities, ParaMor must estimate the total number of cstems which could conceivably occur in a single scheme. ParaMor takes this upper limit to be the corpus vocabulary size. With this estimate in hand, the maximum likelihood estimate of the cstem probability of any given scheme, S, is the count of the adherent cstems in S over the size of the corpus vocabulary, or, in the notation of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13, C/V. Note that the joint probability of finding a cstem in the current scheme and in the expansion scheme is just the probability of a cstem appearing in the parent scheme. The first parent-evaluation metric that makes use of the probabilistic view of cstem occurrence in schemes is pointwise mutual information. The pointwise mutual information between values of two random variables measures the amount by which uncertainty in the first variable changes when a value for the second has been observed. In the context of morphological schemes, the pointwise mutual information registers the change in the uncertainty of observing the expansion csuffix when the csuffixes in the current scheme have been observed. The formula for pointwise mutual information between the current and expansion schemes is given on the third row of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13. Like the dice measure, the pointwise mutual information identifies a large difference between the amente parent and the ar parent. As Manning and Schtze (1999, p. 181) observe, however, pointwise mutual information increases as the number of observations of a random variable decrease. And since the expansion schemes amente and ualidad have comparatively low cstem counts, the pointwise mutual information score is higher for the amente and ualidad parents than for the truly paradigmatic asundesirable behavior for a metric guiding a search that must identify productive, and therefore likely frequent, paradigmatic suffixes. Three Statistical Tests for Parent-Evaluation Metrics While the three heuristic parent-evaluation metrics: parent-child cstem ratios, dice similarity, and pointwise mutual information scores seem intuitively reasonable, it would be theoretically appealing if ParaMor could base an upward search decision on a statistical test of a parents worth. Just such statistical tests can be defined by viewing each cstem in a scheme as a successful trial of a Boolean random variable. Taking the view of schemes as Boolean random variables, the joint distribution of pairs of schemes can be tabulated in 2x2 grids. The grids beneath the four expansion schemes of  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12 hold the joint distribution of the a.o.os scheme and the respective  expansion scheme. The first column of each table contains counts of adherent cstems that occur with all the csuffixes in the current scheme. While the second column contains an estimate of the number of stems which do not form corpus words with each csuffix of the child scheme. Similarly, the tables first row contains adherent counts of cstems that occur with the expansion csuffix. Consequently, the cell at the intersection of the first row and first column contains the adherent stem count of the parent scheme. The bottom row and the rightmost column of each 2x2 table contain marginal adherent counts. In particular, the bottom cell of the first column contains the count of all the cstems that occur with all the csuffixes in the current scheme. In mirror image, the rightmost cell of the first row contains the adherent count of all cstems which occur with the expansion csuffix. The corpus vocabulary size is the marginal of the marginal cstem counts, and estimates the total number of cstems. Treating sets of csuffixes as Bernoulli random variables, we must ask what measurable property of random variables might indicate that the csuffixes of the current scheme and the csuffix of the expansion scheme belong to the same paradigm. One answer is correlation. As described both earlier in this section as well as in Section  REF _Ref214372949 \r \h  3.2.1, suffixes which belong to the same paradigm are likely to have occurred attached to the same stems this co-occurrence is statistical correlation. Think of a big bag containing all possible cstems. We reach our hand in, draw out a cstem, and ask: Did the csuffixes of the current scheme all occur attached to this cstem? Did the expansion csuffixes all occur with this cstem? If both sets of csuffixes belong to the same paradigm, then the answer to both of these questions will often be the same, implying the random variables are correlated. A number of standard statistical tests are designed to detect if two random variables are correlated. In designing ParaMors parent-evaluation metric, three statistical tests were examined: Pearson s 2 test, The Wald test for the mean of a Bernoulli population, and A likelihood ratio test for independence of Binomial random variables Pearson s 2 test is a nonparametric test designed for categorical data, that is, data in which each observed trial point can be categorized as belonging to one of a finite number of types. Pearsons test compares the expected number of occurrences of each category with the observed number of occurrences of that category. In a 2x2 table, such as the tables of cstem counts in  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12, the four central cells in the table are the categories. Pearson s 2 test relies on the observation that if two random variables are independent, then the expected number of observations in each cell is the product of the marginal probabilities along that cell s row and column. Pearson s test statistic for a 2x2 table, given in the fourth row of  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12, converges to the 2 distribution as the size of the data increases. (DeGroot, 1986 p. 536). The second statistical test investigated for ParaMors vertical scheme search is a Wald test of the mean of a Bernoulli population (Casella and Berger, 2002 p. 493). This Wald test compares the observed number of cstems in the parent scheme to the number which would be expected if the child csuffixes and the expansion csuffixes were independent. When the current and expansion schemes are independent, the central limit theorem implies that the statistic given in the fifth row of  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13 converges to a standard normal distribution. Since the sum of Bernoulli random variables is a Binomial distribution, we can view the random variable which corresponds to any particular scheme as a Binomial. This is the view taken by the final statistical test investigated as a parent-evaluation metric for ParaMors initial bottom-up scheme search. In this final test, the random variables corresponding to the current and expansion schemes are tested for independence using a likelihood ratio statistic from Manning and Schtze (1999, p. 172). When the current and expansion schemes are not independent, then the occurrence of a cstem, t, in the current scheme will affect the probability that t appears in the expansion scheme. On the other hand, if the current and expansion schemes are independent, then the occurrence of a cstem, t, in the current scheme will not affect the likelihood that t occurs in the expansion scheme. The denominator of the formula for the likelihood ratio test statistic given in  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13 describes current and expansion schemes which are not independent; while the numerator gives the independent case. Taking two times the negative log of the ratio produces a statistic that is 2 distributed. One caveat, both the likelihood ratio test and Pearson s 2 test only assess the independence of the current and expansion schemes, they cannot disambiguate between random variables which are positively correlated and variables which are negatively correlated. When csuffixes are negatively correlated it is extremely likely that they do not belong to the same paradigm. Clearly, ParaMors search strategy should not move to parent schemes whose expansion csuffix is negatively correlated with the csuffixes of the current scheme. Negative correlation occurs when the observed frequency of cstems in a parent scheme is less than the predicted frequency assuming that the current and expansion csuffixes are independent. When evaluating parent schemes with either the likelihood ratio test or Pearson s 2 test, ParaMor explicitly checks for negative correlation, preventing ParaMor s search from moving to a parent scheme whose expansion csuffix is negatively correlated with the current scheme. Looking in  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13, the values of the three statistical tests for the four parents of the a.o.os scheme suggest that the tests are generally well behaved. For each of the tests, a larger score indicates that an expansion scheme is more likely to be correlated with the current schemealthough, again, comparing the absolute scores of one test to the numeric values from another test is meaningless. All three statistical tests score the unproductive derivational ualidad scheme as the least likely of the four expansion schemes to be correlated with the current scheme. And each test gives a large margin of difference between the amente and the ar parents. The only obvious misbehavior of any of these statistical tests is that Pearson s 2 test ranks the amente parent as more likely correlated with the current scheme than the as parent. An Empirical Comparison of Parent-Evaluation Metrics To quantitatively compare parent-evaluation metrics at their ability to identify paradigmatically coherent schemes, ParaMor performed an oracle experiment. This oracle experiment places on an even playing field all six parent-evaluation metrics summarized in  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13. Looking at Spanish data, the oracle evaluation assessed each metric at its ability to identify schemes in which every csuffix is string identical to a suffix of some single inflectional paradigm. The inflectional paradigms of Spanish used as the oracle answer key in this experiment were hand complied from a standard Spanish textbook (Gordon and Stillman, 1999), and are detailed in Appendix A. The methodology of the oracle experiment is as follows: ParaMor visited every scheme that contained only csuffixes which model suffixes from a single inflectional paradigmcall such schemes sub-paradigm schemes. Each parent of a sub-paradigm scheme is either a sub-paradigm scheme itself, or else the parents csuffixes no longer form a subset of the suffixes of a true paradigm. The oracle experiment evaluated each metric at its ability to classify each parent of each sub-paradigm scheme as either being a sub-paradigm scheme itself or as introducing a non-paradigmatic csuffix. Notice that the oracle experiment did not directly evaluate parent metrics in the context of ParaMor s greedy upward search procedure described in Sections  REF _Ref214372949 \r \h  3.2.1 and  REF _Ref214372965 \r \h  3.2.2. Following the methodology of this oracle experiment allows a direct comparison between parent-evaluation metrics: where ParaMors greedy search is not guaranteed to visit identical sets of schemes when searching with different upward metrics or with different halting thresholds, the oracle experiment described here evaluates all metrics over the same set of upward decisions. Also, the methodology of this oracle experiment necessitated using a considerably smaller corpus than other experiments that are reported in this thesis.  REF _Ref193104570 \h  \* MERGEFORMAT Figure  3.14 gives results of an oracle evaluation over a corpus of free-running Spanish newswire text containing 6,975 unique types. A small corpus is necessary because the oracle experiment visits all parents of all sub-paradigm schemes, and a larger corpus would create a search space that is too large to instantiate. To compare parent-evaluation metrics ParaMor must factor out threshold effects. Each metrics performance at identifying sub-paradigm schemes varies with the cutoff threshold below which a parent is believed to not be a sub-paradigm scheme. For example, when considering the cstem ratio metric at a threshold of 0.5, ParaMor would take as a sub-paradigm scheme any parent that contains at least half as many cstems as the current scheme. But if this threshold were raised to 0.75, then a parent must have at least three-quarters the number of cstems as the child to pass for a sub-paradigm scheme. Moreover, while each of the six metrics described in the previous section score each parent scheme with a real value, the scores are not normalized. The ratio and dice metrics produce scores between zero and one, Pearson s 2 test and the Likelihood Ratio test produce non-negative scores, while the scores of the other metrics can fall anywhere on the real line. But even metrics which produce scores in the same range are not comparable. Referencing  REF _Ref192997341 \h  \* MERGEFORMAT Figure  3.13, the ratio and dice metrics, for example, can produce very different scores for the same parent scheme. Furthermore, while statistical theory can give a confidence level to the absolute scores of the metrics that are based on statistical tests, the theory does not suggest what confidence level is appropriate for the task of paradigm detection in scheme networks. The ratio, dice, and pointwise mutual information metrics lack even an interpretation of confidence. To remove threshold effects and fairly compare parent-evaluation metrics, the oracle experiment performs a peak-to-peak comparison. The oracle evaluation measures the precision, recall, and F1 of each metric over a range of threshold values relevant to that metric. And the maximum F1 value each metric achieves is its final score.  REF _Ref193104570 \h  \* MERGEFORMAT Figure  3.14 reports the peak F1 score for each of the six metrics presented in this section. Two results are immediately clear. First, all six metrics consistently outperform the baseline algorithm of considering every parent of a sub-paradigm scheme to be a sub-paradigm scheme. Although a seemingly weak baseline, the always-move-up rule correctly classifies the paradigmatic integrity of scheme parents more than 70% of the time. The second result evident from  REF _Ref193104570 \h  \* MERGEFORMAT Figure  3.14 is that the most simple metric, the parent-child cstem ratio, does surprisingly well, identifying parent schemes which contain no extra-paradigmatic suffixes just as consistently as more sophisticated tests, and outperforming all but one of the considered metrics. The primary reason the parent-child cstem ratio performs so well appears to be that the ratio metric is comparatively robust when data is sparse. In 79% of the oracle decisions that each metric faced, the parent scheme had fewer than 5 cstems! Interestingly, Hammarstrm (2006b) has also found that simple cstem ratios are surprisingly effectively at identifying suffixes that paradigmatically contrast. In the context of a stemming algorithm designed for information retrieval, Hammarstrm reports that, in assessing the cstem coherence of sets of csuffixes, he has not had much success with standard vector similarity measures. Hammarstrm then turns to parent-child cstem ratios to define a novel measure of the extent to which csuffixes share cstems. Hammarstrms metric, which, unfortunately, this oracle evaluation does not investigate, looks only at parent-child cstem ratios that occur at the bottom of scheme networks, namely, between paradigmatic level 2 and level 1 schemes. In addition to strong performance in this oracle evaluation, the parent-child cstem ratio has a second, unrelated, advantage over other metrics: simplicity of computation. In taking each upward step, ParaMors bottom-up search procedure, described in Sections  REF _Ref214372949 \r \h  3.2.1 and  REF _Ref214372965 \r \h  3.2.2, greedily moves to that parent with the highest score. Hence, to find the best greedy parent, ParaMor need not necessarily compute the scores of parent schemes, all ParaMor must do is identify which parent would have the highest score. As the cstem ratio metric does not incorporate the frequency of the expansion csuffix, and as the number of cstems in the current scheme does not change, the parent scheme with the largest cstem ratio is always that parent with the most cstemsan easily computable statistic. On the basis of both the sturdy performance in the oracle evaluation, as well as the simplicity and speed of identifying the parent with the largest cstem ratio, all further experiments in this thesis use the parent-child cstem metric to guide ParaMors vertical search. Setting the Halting Threshold in ParaMors Bottom-Up Search Having settled on the parent-child cstem ratio as the parent-evaluation metric of choice, ParaMor must next select a threshold value at which to halt upward search paths. The oracle experiment described in this section was designed to ascertain which parent-evaluation metric most accurately identifies paradigmatic parents. Somewhat in contrast, ParaMors greedy search algorithm, described in Sections  REF _Ref214372949 \r \h  3.2.1 and  REF _Ref214372965 \r \h  3.2.2, prizes caution ahead of accuracy. As discussed in Section  REF _Ref214372949 \r \h  3.2.1, ParaMor s greedy search takes a separate upward path from all individual single-csuffix schemes. Pursuing many paths in parallel, it is likely that most csuffixes that model a true suffix will be selected as part of some path. ParaMor then relies on a clustering algorithm, described in Chapter 4, to bring together paradigmatically related csuffixes that only appear in separate search paths. A cautious initial search helps ParaMor to keep initially selected models of partial paradigms largely free of non-paradigmatic csuffixes. The threshold value at which the parent-child cstem ratio achieved its peak F1 in the oracle experiment is 0.05. At this threshold, for ParaMors greedy search to accept a parent scheme, the parent need only contain one-twentieth the number of cstems as the current scheme. A qualitative examination of ParaMors selected schemes suggests a threshold of 0.05 is not cautious enough; Over a corpus of 50,000 Spanish types, at the 0.05 threshold, ParaMors greedy search selects many schemes that include both inflectional paradigmatic csuffixes and csuffixes that model only marginally productive derivational suffixes. Hence, the remainder of this thesis sets the parent-child cstem ratio threshold at the more cautious value of 0.25. It is possible that a threshold value of 0.25 is sub-optimal for paradigm identification. And future work should more carefully examine the impact that varying this threshold has on morphological segmentation of words (Chapter 5). However, the quantitative evaluations of Chapter 6 will show that the current setting of 0.25 leads to morphological segmentations of a diverse set of natural languages that out-perform the segmentations of state-of-the-art unsupervised morphology induction systems. Summarizing the Search for Candidate Paradigms This chapter has presented the strategy that ParaMor employs to identify initial candidate paradigms. This strategy has three main steps. First, ParaMor organizes candidate stems and candidate suffixes into natural groupings, or schemes, that model potential inflectional paradigms. Second, ParaMor relates the scheme groupings to one another, in the process forming a network of candidates. And third, ParaMor searches the candidate paradigm network in a recall-centric fashion for schemes which individually cover diverse portions of the paradigms of a language. Although, as the next chapter will show quantitatively, ParaMors initial paradigm search successfully uncovers most suffixes of languages like Spanish, the output of the search is still far from a tight model of paradigm structure. With ParaMors recall-centric search, many of the initially selected candidate paradigms are erroneously selected. And of those which are correct, many redundantly cover the same portions of the same paradigms. Overcoming these two shortcomings of the initial search procedure is the topic of the next chapter. Clustering and Filtering of Initial Paradigms This chapter builds on the initial models of paradigms that are selected by the greedy bottom-up search strategy described in Chapter 3. Taking in the initial set of scheme models, this chapter consolidates and purifies a focused set of paradigm models. And, in turn, the algorithms described in Chapter 5 wield these focused models to segment words into morphemes. From Schemes to Comprehensive Models of Paradigms The bottom-up search strategy presented in Chapter 3 is a solid first step toward identifying useful models of productive inflectional paradigms. However, as Section  REF _Ref214372949 \r \h  3.2.1 explains, ParaMor s greedy bottom-up search strategy narrowly focuses on finding partial models of paradigms that, in aggregate, focus on recall. ParaMor s recall-centric strategy of starting upward paths from all individual csuffixes inevitably seeds some paths with csuffixes which do not model suffixes. And while ParaMors initial scheme-models jointly recover many inflectional suffixes, individual schemes will likely cover only portions of full linguistic paradigms. Section  REF _Ref215812367 \r \h  4.1.1 illustrates, by way of an extended example, the successes and shortcomings of ParaMor s initially selected schemes, while Section  REF _Ref215812423 \r \h  4.1.2 outlines the strategies that ParaMor takes to mitigate the shortcomings. Sections  REF _Ref190686436 \r \h  4.2 REF _Ref190686481 \r \h  4.3, and  REF _Ref190686503 \r \h  4.4 then present each mitigation strategy in detail. And finally, Section  REF _Ref215812860 \r \h  4.5 analyzes the paradigm models that ParaMor ultimately produces. A Sample of Initially-Selected Scheme Models of Paradigms To better see the strengths of the paradigm models that ParaMor initially selects, as well as to motivate the steps ParaMor takes to overcome gaps in the initial models,  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 presents a range of schemes selected during a typical run of ParaMor s bottom-up search procedure, as described in Chapter 3. Each row of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 lists a scheme selected while searching over a Spanish newswire corpus of 50,000 types using the parent-child cstem ratio metric at a halting threshold of 0.25 (see Chapter 3). On the far left of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, the Rank column states the ordinal rank at which that row s scheme was selected during the search procedure: the .s scheme was the terminal scheme of ParaMors 1st upward search path, a.as.o.os the terminal scheme of the 2nd path, ido.idos.ir.ir the 1592nd, etc. The right four columns of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 present raw data on the selected schemes, giving the number of csuffixes in that scheme, the csuffixes themselves, the number of adherent cstems of the scheme, and a sample of those cstems. Between the rank on the left, and the scheme details on the right, are columns which categorize the scheme on its success, or failure, to model a true paradigm of Spanish. Appendix A lists the inflectional paradigms of Spanish morphology. A dot appears in the columns marked N, Adj, or Verb if multiple csuffixes in a rows scheme clearly represent suffixes in a paradigm of that part of speech. The verbal paradigm is further broken down by inflection class, ar, er, and ir. A dot appears in the Derivation column if at least one csuffix of a scheme models a derivational suffix. The remaining six columns of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 classify the correctness of each row s scheme. The Good column of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 is marked if the csuffixes in a scheme take the surface form of true suffixes. Initially selected schemes in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 that correctly capture real paradigm suffixes are the 1st, 2nd, 4th, 5th, 12th, 30th, 40th, 127th, 135th, 400th, 1592nd, and 2000th selected schemes. Most true inflectional suffixes of Spanish are modeled by some scheme that is selected during the run of ParaMor s initial search that is presented in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1: The bottom-up search identifies partial paradigms which, between them, contain 91% of all string-unique suffixes of the Spanish verbal inflectional paradigms. (Appendix A summarizes Spanish inflectional morphology). If we ignore as undiscoverable all suffix strings which occurred at most once in the Spanish newswire corpus, ParaMors coverage jumps to 97% of unique verbal suffixes. In addition to verbal suffixes, ParaMor identifies schemes which model: 1. Each of the two phonologic inflection classes that express number on nouns: the 1st selected scheme models .s, while the 4th selected scheme models .es; and 2. The full adjectival cross-product paradigm of gender and number, a.as.o.os, in the 2nd selected scheme. But, while most true inflectional suffixes are modeled by some scheme selected in the initial search, ParaMors initially selected schemes also display two broad shortcomings. A First Shortcoming of Initial Schemes: Fragmentation No single initially selected scheme comprehensively models all the suffixes of the larger Spanish paradigms. Instead, csuffix models of paradigm suffixes are fragmented across large numbers of schemes. The largest schemes that ParaMor identifies from the newswire corpus are the 5th and 12th selected schemes: Shown in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, both of these schemes contain 15 csuffixes which model suffixes from the ar inflection class of the Spanish verbal paradigm. But the full ar inflection class has 36 unique surface suffixes. In an agglutinative language like Turkish, the cross-product of several word-final para-  digms may have an effective size of hundreds or thousands of suffixes, and ParaMor will only identify a minute fraction of these in any one scheme. In  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, the Complete column is marked when a scheme contains separate csuffixes that correspond to each suffix of a paradigm or paradigm cross-product. On the other hand, if a schemes csuffixes model suffixes of some paradigm of Spanish, but manage to model only a portion of the full paradigm, then  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 has a dot in the Partial column. Among the many schemes in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 which faithfully describe significant partial fractions of legitimate paradigms are the 5th, 12th, and 400th selected schemes. These three schemes each contain csuffixes which clearly model suffixes from the verbal ar paradigmbut each contains csuffixes that correspond to only a subset of the suffixes in the full ar inflection class. Notice that while some legitimate inflectional suffixes occur in only one ar scheme, e.g. aban and arse in the 5th selected scheme, other csuffixes appear in two or more schemes that model the ar paradigm, e.g. a, ados, . Indeed, the ados csuffix occurs in 31 schemes that were selected during this run of ParaMors initial search. The search paths that identified these 31 schemes each originate from a distinct initial csuffix, including such csuffixes as: an, en, acin, amos, etc. Separate scheme patchworks cover the other inflection classes of Spanish verbs as well. For example, schemes modeling portions of the ir inflection class include the 30th, 135th, 1592nd, and 2000th selected schemes. Consider the 1592nd scheme, which contains four csuffixes. Three of these csuffixes, ido, idos, and ir, occur in other schemes selected during the initial search, while the uncommon 1st Person Singular Future Tense suffix ir is unique to the 1592nd selected schemein all the schemes selected during this run of ParaMors initial bottom-up search, the csuffix ir occurred in only one, the 1592nd selected scheme. ParaMors recall-centric search has correctly identified the ir suffix, but as yet ir is isolated from most other csuffixes of the ir paradigm. A Second Shortcoming of Initial Schemes: Poor Precision The second broad shortcoming of ParaMor s initial search, apparent from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, is simply that many schemes do not satisfactorily model suffixes. The vast majority of schemes with this second shortcoming belong to one of two sub-types. The first sub-type of unsatisfactory paradigm model comprises schemes which systematically misanalyze word forms. These systematically incorrect schemes come in two flavors: a scheme will either hypothesize morpheme boundaries in its licensing words that fall internal to true stems or a scheme will propose boundaries that are consistently internal to underlying suffixes. Schemes which misanalyze morpheme boundaries in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 are marked in the Error: Stem-Internal or Error: Suffix-internal columns, and comprise the 3rd, 10th, 11th, 20th, 200th, 1000th, and 5000th selected schemes. Of these, the 3rd and 11th selected schemes place a morpheme boundary at suffix-internal positions, truncating the full suffix forms: Compare the 3rd and 11th selected schemes with the 5th and 12th. In symmetric fashion, a significant fraction of the csuffixes in the 10th, 20th, 200th, 1000th, and 5000th selected schemes hypothesize morpheme boundaries for their licensing word forms internal to real Spanish stems. Placing morpheme boundaries internal to stems, these schemes inadvertently include the final characters of verb stems as leading characters on their csuffixes. In a random sample of 100 schemes from the 8339 schemes which the initial search strategy selected, 48 schemes incorrectly placed morpheme boundaries stem-internally, while one scheme hypothesized morpheme boundaries at locations inside the suffixes of the schemes licensing forms. The second sub-type of unsatisfactory paradigm model, exemplified in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, occurs when a scheme s csuffixes are related not by belonging to the same paradigm, but rather by a chance string similarity of surface type. Schemes which arise from chance string collisions are marked in the Error: Chance column of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, and include the 20th, 100th, 3000th, and 4000th selected schemes. In the random sample of 100 schemes selected by ParaMor s initial bottom-up search, 40 were schemes produced from a chance similarity between word types. These chance schemes are typically small along two distinct dimensions. First, the string lengths of the cstems and csuffixes of these chance schemes are often quite short. The longest cstem of the 100th selected scheme is two characters long; while both the 100th and the 3000th selected schemes contain the null csuffix, , which has length zero. Short cstem and csuffix lengths in selected schemes are easily explained combinatorially: While the inventory of possible strings grows exponentially with the length of the string, there just arent very many length-one or length-two strings. With few strings to choose from, it should come as no surprise when a variety of csuffixes happen to occur attached to the same set of very short cstems. Schemes arising through a chance string similarity of word types are small on a second dimension as well. Chance schemes typically contain few cstems, and, by virtue of the details of ParaMors search procedure (see Chapter 3), even fewer csuffixes. For example, the 3000th selected scheme contains just three cstems and two csuffixes. The evidence for this 3000th scheme arises, then, from a scant six (short) types, namely: li, a Chinese name; lo, a Spanish determiner and pronoun; man, part of an abbreviation for Manchester United in a listing of soccer statistics; lizano, a Spanish name; lozano, a Spanish word meaning leafy; and manzano, Spanish for apple tree. Schemes formed from a chance string similarity of a few types, such as the 3000th selected scheme, are particularly prevalent among schemes chosen later in the search procedure, where search paths originate from level 1 schemes whose single csuffix is less frequent. Although there are less frequent csuffixes that do correctly model portions of true paradigms (including the csuffix ir, which led to the paradigmatically coherent 1592nd selected scheme, see above) the vast majority of less frequent csuffixes do not model true suffixes. And because the inventory of word final strings in a moderately sized corpus is enormous, some few of the many available csuffixes happen to be interchangeable with some other csuffix on some few (likely short) cstems of the corpus. ParaMors Paradigm-Processing Pipeline This chapter describes the algorithms ParaMor adopts to remedy the two shortcomings of ParaMors initially selected schemes that were identified in Section  REF _Ref215204016 \r \h  4.1.1. To consolidate the patchwork modeling of paradigms and to corral free csuffixes into structures which more fully model complete paradigms, ParaMor adapts an unsupervised clustering algorithm to automatically group related schemes. Meanwhile, to remove schemes which fail to model true suffixes, ParaMor takes a two pronged approach: First, a clean-up of the training data reduces the incidence of chance similarity between strings, and second, targeted filtering algorithms identify and discard schemes which likely fail to model paradigms. To organize these additional steps of paradigm identification, ParaMor adopts a pipeline architecture. ParaMors initial morphology network search algorithm, described in Chapter 3, becomes one step in this pipeline. Now ParaMor must decide where to add the pipeline step that will cluster schemes which model portions of the same paradigm, and where to add steps that will reduce the incidence of incorrectly selected schemes. At first blush, it might seem most sound to place steps that remove incorrectly selected schemes ahead of any scheme-clustering stepafter all, why cluster schemes which do not model correct suffixes? At best, clustering incorrect schemes seems a waste of effort; at worst, bogus schemes might confound the clustering of legitimate schemes. But removing schemes before they are clustered has its own dangers. Most notably, a discarded correct scheme can never be recovered: If the distraction of incorrect schemes could be overcome, corralling schemes into monolithic paradigm models might safeguard individual useful schemes from imperfect scheme-filtering algorithms. And by the same token, scheme filters can also mistake incorrect schemes for legitimate models of paradigms. Hence, if a clustering algorithm could place together such misanalyses as the 3rd and 11th selected schemes from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, which both model the same incorrect morpheme boundaries in their licensing types, then clustering incorrect schemes might actually facilitate identification and removal of misanalyzed schemes. As Section  REF _Ref190686481 \r \h  4.3 explains, ParaMor s clustering algorithm can, in fact, accommodate schemes which hypothesize incorrect morpheme boundaries, but has more difficulty with non-paradigmatic schemes which are the result of chance string similarity. To retain a high recall of true suffixes within the framework of a pipeline architecture, ParaMor takes steps which reduce the inventory of selected schemes only when necessary. Section  REF _Ref190686436 \r \h  4.2 describes a technique that vastly reduces the number of selected schemes which result from chance string similarity, while insignificantly impacting correctly selected schemes. Section  REF _Ref190686481 \r \h  \* MERGEFORMAT  4.3 then describes ParaMor s scheme-clustering algorithm. And Section  REF _Ref190686503 \r \h  \* MERGEFORMAT  4.4 presents two classes of filtering algorithm which remove remaining incorrectly selected, but now clustered, schemes. Training Corpus Clean-Up ParaMor s clustering algorithm (Section  REF _Ref190686481 \r \h  4.3) is specifically tailored to leverage the paradigmatic structure of schemes, and, as such, is ill-suited to schemes which do not exhibit a regular paradigmatic alternation. Schemes which result from chance similarities in word forms pointedly lack such paradigmatic structure. Thus ParaMor seeks to remove chance schemes before the scheme-clustering step. As mentioned in Section  REF _Ref215820051 \r \h  4.1.1, the string lengths of the csuffixes and cstems of chance schemes are typically quite short. And if the csuffixes and cstems of a scheme are short, then the underlying types which license the scheme are also short. These facts suggest the simple data clean up step of excluding short types from the vocabulary that ParaMor uses to induce paradigms. As described below, this simple word-length requirement in the paradigm-induction vocabulary virtually eliminates the entire category of chance scheme. For all of the languages considered in this thesis ParaMor has raw text corpora available that are much larger than the 50,000 types used for paradigm induction. Consequently, for the experiments reported in this thesis, ParaMor does not merely remove short types from the induction vocabulary, but replaces each short word with a new longer word. ParaMors word-to-morpheme segmentation algorithm, presented in Chapter 5, is independent of the set of types from which schemes and scheme-clusters are built. Consequently, removing short types from training does not preclude these same short types from being analyzed as containing multiple morphemes during segmentation. The string length below which words are removed from the paradigm-induction vocabulary is a free parameter. ParaMor is designed to identify the productive inflectional paradigms of a language. Unless a productive paradigm is restricted to occur only with short stems, a possible but unusual scenario (as with the English adjectival comparative, c.f. faster but *exquisiter), we can expect a productive paradigm to occur with a reasonable number of longer stems in a corpus. Hence, ParaMor neednt be overly concerned about discarding short types. A qualitative examination of Spanish data suggested excluding types 5 characters or less in length. All experiments reported in this thesis which exclude short types only permit types longer than this 5 character cutoff. An Evaluation of Schemes from a Vocabulary that Excludes Short Types When restricted to a corpus containing types longer than 5 characters in length, the schemes that ParaMor selects during the initial search phase are remarkably similar to the schemes that ParaMors search algorithm selects over a corpus containing types of all lengthsexcept for the notable absence of incorrect chance schemes. In a random sample of 100 schemes selected by ParaMor over a type-length restricted corpus of Spanish newswire containing 50,000 unique word forms, only 2 schemes resulted from a chance similarity of word formsthis is down from 40 chance schemes in a random sample of 100 schemes selected from a corpus unrestricted for type length. The 3000th selected scheme, .zano, shown in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 and discussed in the final subsection of Section  REF _Ref215240305 \r \h  4.1.1, is an example of the kind of scheme ParaMor s search algorithm no longer selects once a type-length restriction is in place. The .zano scheme contains just three cstems: li, lo, and man. Because the .zano scheme contains the null csuffix, , the three surface word types, li, lo, and man, are among those that license this scheme. And because all three of these word types are not longer than 5 characters in length, all three are excluded from the Spanish paradigm induction corpus. Removing these three types strips the .zano scheme of all evidence for the csuffix, and the 3000th scheme cannot be selected by the search algorithm. In all, ParaMors search algorithm selects 1430 fewer schemes, 6909 vs. 8339 when training over a type-length restricted corpus. However, including additional long types in the paradigm-induction vocabulary can actually increase the fragmentation of true paradigms across schemes. For example, the number of schemes that contain the csuffix ados, which models a suffix from the ar verbal paradigm of Spanish, increases from 31 to 40 when restricting the paradigm-induction corpus to contain no short word types. Because the training corpus has changed, schemes selected from a corpus restricted for type length are often not identical to schemes selected from an unrestricted corpus. But, in Spanish, ParaMor continues to identify schemes which model the major inflection classes of nouns, adjectives, and verbs. Moreover, with one notable exception, the schemes which model these major paradigms of Spanish contain numbers of csuffixes that are similar to the numbers of csuffixes of corresponding schemes induced over an unrestricted corpus: From the type-length unrestricted corpus, ParaMor directly models the .es inflection class of Number on Spanish nouns with the scheme .es (see  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1); But from the corpus that is restricted for type-length, ParaMor only models the two suffixes and es, of this less common nominal paradigm, in combination with derivational suffixes, in schemes like .es.idad.idades.mente. In Spanish, although nouns, such as fin end, pluralize by adding an es suffix, fines, many words which end in es cannot strip off that es to form a new surface form. For example, to form the Singular of the Plural adjective grandes big, only the s is removed, yielding the word grande. It so happens that in the type-length restricted Spanish corpus only 751 of the 3045 word strings which end in es can remove that es to form a new word that occurred in the Spanish corpus. And 751 out of 3045 is 24.7%just short of the 25.0% cstem ratio threshold used in the upward search algorithm. Additionally, since the ParaMor search strategy does not begin any search path from the scheme, the only search paths which include the csuffixes and es necessarily begin from some third csuffixlike the fairly productive derivation suffix idad. The idad suffix is the Spanish analogue of the English derivational suffix ity. As Chapter 5 discusses, ParaMor can still analyze the morphology of inflected forms despite the distracting presence of derivational suffixes in a scheme. Thus, although ParaMor is no longer able to identify the .es scheme in isolation, the type-length restriction does not cause ParaMor to lose the ability to morphologically analyze Spanish nouns that mark Plural with es. Clustering of Partial Paradigms With the careful weeding of the paradigm-induction corpus described in the previous section, ParaMor largely avoids selecting schemes which arise through chance string similarity. And with the non-paradigmatic chance schemes out of the way, ParaMor is free to apply a clustering algorithm to group paradigmatically related schemes that hypothesize compatible morpheme boundaries. As an unsupervised algorithm, ParaMor must use an unsupervised clustering algorithm to merge schemes. A variety of unsupervised clustering algorithms exist, from kmeans to self-organizing kohonen maps. ParaMors current clustering algorithm is an adaptation of bottom-up agglomerative clustering. Two reasons underlie the choice of bottom-up agglomerative clustering. First, agglomerative clustering is simple, there are just two steps: 1. Clustering begins with each item, i.e. scheme in this application, as its own separate cluster; and 2. At each time step, unless a halting criterion is met, that pair of clusters which is most similar is merged to form a new cluster. The second reason ParaMor adopts bottom-up agglomerative clustering is that the algorithm produces a tree structure that can be examined by hand. And as the remainder of this section describes, careful examination of scheme-cluster trees directly led to adaptations of vanilla agglomerative clustering which accommodate the unique structure of paradigmatic schemes. Three Challenges Face any Scheme-Clustering Algorithm ParaMors scheme-clustering algorithm must address three challenges that arise when schemes are the items being clustered. Two of the three challenges concern intrinsic properties of linguistic paradigms; and two of the three involve schemes as computational models of paradigms. First, the purely linguistic challenge: morphological syncretism. Common cross-linguistically, syncretism occurs when distinct paradigms in a single language contain surface-identical suffixes. Syncretism implies there will be schemes which should not be merged even though they contain some identical csuffixes. Second, ParaMor faces the wholly computational challenge of deciding when to halt clustering. Whenever possible, an unsupervised algorithm, such as ParaMor, should avoid introducing free parameters that must somehow be tuned. To decide when to stop merging clusters, bottom-up agglomerative clustering typically uses a free parameter: Clustering ends when the similarity score between the pair of most similar clusters falls below a threshold. To avoid introducing this tunable halting parameter, it would be desirable to devise a threshold-free method to end the clustering of schemes. The third challenge ParaMor must overcome has both linguistic and computational roots: competing schemes may hypothesize rival morpheme boundaries in a common set of surface types, but such competing schemes should not be merged into the same cluster as they are distinct models of morphological structure. Different individual schemes hypothesize different morpheme boundaries in a single word type both for the computational reason that ParaMor does not know, a priori, where the correct morpheme boundaries lie, but also because, in natural language, morphological agglutination allows a single word type to legitimately contain more than one morpheme boundary. The following sub-sections motivate and present adaptations to the basic bottom-up agglomerative clustering algorithm that address these three challenges that face scheme clustering. The First Challenge: Syncretism Some examples will help elucidate the first challenge facing ParaMors clustering algorithm: surface identical csuffixes in distinct paradigms, or syncretism.  REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2 lists six syncretic schemes taken from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1: As a refresher, the schemes in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 were selected by ParaMors initial search over a Spanish corpus of 50,000 types unrestricted for type length. In Spanish, verbs that belong to the er paradigm systematically share many inflectional suffixes with verbs of the ir paradigm. In  REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2 the 30th, 135th, and 2000th selected schemes all contain only csuffixes which model suffixes found in both the er and the ir inflection classes. But grammars of Spanish still distinguish between an er and an ir inflection class because er and ir verbs do not share all inflectional suffixes. In  REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2, the 127th selected scheme contains the csuffixes er, er, and era which all only occur on er verbs, while the 1592nd selected scheme contains the csuffixes ir and ir which only occur on ir verbs. A scheme-clustering algorithm must not place the 127th and 1592nd selected schemes into the same cluster, but should instead produce distinct clusters to model the er and ir inflection classes. More insidious than the suffix overlap between the er and ir inflection classes of Spanish verbs, is the overlap between the ar inflection class on the one hand and the er and ir inflection classes on the other. While shared surface suffixes in the er and ir inflection classes consistently mark identical sets of morphosyntactic features, most suffixes whose forms are identical across the ar and the er/ir inflection classes mark different morphosyntactic features. The Present Tense suffixes as, a, amos, and an mark various Person-Number features in the Indicative Mood on ar verbs but Subjunctive Mood on er and ir verbs. Conversely, es, e, emos (imos on ir verbs), and en mark Indicative Mood on er verbs, but these eforms mark Subjunctive on ar verbs. Of course, ParaMor is unaware of morphosyntactic features, and so an appeal to syntactic features is irrelevant to ParaMor s current algorithms. But clearly, ParaMor must carefully consider the 12th and 30th selected schemes ( REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2), which ultimately model the ar and er/ir classes respectively, but which contain the csuffixes e and en in common. Syncretic paradigm overlap is widespread in languages beyond Spanish. In English, many nouns form their plural with the suffix s, but verbs can also take an s suffix, marking 3rd Person Present Tense. Important for the evaluation of the work in this thesis, paradigm overlap also occurs in German, Finnish, Turkish, and Arabic which, together with English, comprise evaluation languages of the Morpho Challenge competitions (see Chapter 6). When modeling the inflectional paradigms of English, Spanish, German, Finnish, Turkish, Arabic or any other language, ParaMor must retain distinct models of each paradigm, though some of their suffixes might be string identical. Suffix string identity overlap between paradigms has direct implications for the scheme similarity measure ParaMor employs during clustering. Bottom-up agglomerative clustering, like most unsupervised clustering algorithms, decides which items belong in the same cluster by measuring similarity between and among the items being clustered. To cluster schemes, ParaMor must define a similarity between pairs of scheme-clusters. Since natural language paradigms are essentially sets of suffixes, perhaps the most intuitive measure of scheme similarity would compare schemes csuffix sets. But because a suffix surface form may appear in more than one paradigm, comparing csuffix sets alone could spuriously suggest that schemes which model distinct but syncretic paradigms be merged. Two facts about paradigm structure can rescue csuffix-based scheme similarity measures from the complication of paradigm overlap. First, as a robust rule of thumb, while two paradigms may share the surface forms of one or more suffix, each paradigm will also contain suffixes the other does not. For example, although the ar, er, and ir inflection classes of the Spanish verbal paradigm have suffixes in common, each contains suffixes which the others do notthe infinitive suffixes ar, er, and ir themselves uniquely distinguish each inflection class. Similarly, in English, although verbal and nominal paradigms share the string s as a suffix, the verbal paradigm contains the suffixes ing and ed which the nominal paradigm does not; and the nominal paradigm contains suffix surface forms which the verbal paradigm does not: the English possessive ending, which appears after the optional number suffix, yields the orthographic paradigm cross-product forms: s, ss, and s, which are unique to nouns. The second useful fact about paradigm structure is that a single lexical item can only belong to a single paradigm. In Spanish, any particular verb will belong to exactly one of the three inflection classes: ar, er, or ir. In the Spanish newswire corpus of 50,000 types that  REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2 is built from, for example, no cstem forms a word type by attaching the csuffix adas while also forming a separate word by attaching idas: The csuffixes adas and idas mark the Past Participle Feminine Plural in the ar and in the ir/er inflection classes respectively. Since adas and idas share no cstem, it is reasonable to propose that adas and idas belong to distinct paradigms and that the 12th selected scheme, which contains adas, and the 30th selected scheme, which contains idas (see  REF _Ref215290439 \h  \* MERGEFORMAT Figure  4.2), should not be merged Adopting exactly this strategy, ParaMor permits no scheme-cluster to be formed which would contain a pair of csuffixes which share no cstems. The general technique of preventing induced clusters from containing items that are significantly dissimilar is known as discriminative clustering. ParaMor easily calculates the set of cstems that two csuffixes, f1 and f2, have in common by revisiting the morphology scheme network used in the initial search for candidate schemes, see Chapter 3. The level 2 network scheme subsuming f1 and f2 exactly holds the cstems which form words in combination with both f1 and f2. Typically, the f1.f2 scheme will be empty of cstems exactly when, without loss of generality, f1 is a csuffix that is unique to a paradigm to which f2 does not belong. Note that since the initial search will only select schemes containing a positive number of cstems, if f1 and f2 are not mutually substitutable on at least one cstem, then no single selected scheme can contain both f1 and f2. ParaMors discriminative clustering requirement that every pair of csuffixes in a cluster share some cstem is not foolproof. First of all, the general rule that distinct paradigms contain some distinct surface suffix is not a language universal. Even in Spanish, the two suffixes in the paradigm of Number on adjectives are identical to the suffixes of the largest inflection class of the Number paradigm on nounsboth paradigms contain exactly the two suffixes and s. And, indeed, both adjectives and nouns appear in ParaMors final cluster which contains the .s scheme. At the same time, it is also possible for two lexical items that participate in two different paradigms to have surface-identical stems. In English, many verbal lexical items have a nominal counterpart: to run vs. a run, a table vs. to table, etc. And in Spanish, where stems do not readily convert part-of-speech without a change in the form of the stem, there are still lexical collisions. Among the most egregious collisions are pairs of unrelated Spanish verbs whose stems are surface identical but which belong to distinct inflectional paradigms: parar to stop vs. parir to give birth, crear to create vs. creer to believe. Thus, pairs of csuffixes that uniquely distinguish paradigms, such as the infinitive endings ar and ir, or ar and er, can still share cstems. In practice, pairs of stem-identical lexical items are rare enough, and paradigm unique suffixes common enough, that with the discriminative restriction on clustering in place ParaMor is able to identify paradigms. Nevertheless, when a paradigm has no distinguishing suffix, or when two distinct paradigms contain many lexemes with surface-identical stems, ParaMor struggles to prevent distinct paradigms from mergingthis is a clear area for future research. The Second Challenge: Halting ParaMors Clustering In addition to helping to keep distinct paradigms separate, ParaMors discriminative restriction on the csuffixes which can belong to a cluster also provides a principled halting criterion that avoids the introduction of an arbitrary similarity cutoff parameter. ParaMor allows agglomerative clustering to proceed until merging any pair of clusters would place into the same cluster two csuffixes that share no cstem. Thus, discriminatively restricting clustering by csuffixes solves two of the three challenges facing a scheme-clustering algorithm. The Third Challenge: Isolating Distinct Morpheme Boundaries Now ParaMor must solve the third challenge: To not coalesce schemes that model competing morpheme boundary hypotheses. Consider the 1st, 2nd, 3rd, 5th, and 12th selected schemes, as first presented in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 and repeated in  REF _Ref215296727 \h  \* MERGEFORMAT Figure  4.3 for convenience. Each of these five schemes is licensed by the Spanish word apoyadas  supported (Adjectival Feminine Plural) : The 1st selected scheme contains the cstem apoyada and the csuffix s, the 2nd scheme has the cstem apoyad and the csuffix as, the 3rd contains apoya and das, while the 5th and 12th selected schemes each contain the cstem apoy and the csuffix adas. Between them, these five schemes model four distinct morpheme boundaries in the same word: apoyada+s, apoyad+as, apoya+das, and apoy+adas. As the 5th and 12th selected schemes hypothesize the same boundary, apoy+adas, it is reasonable to consider placing this pair of schemes in a single cluster. The other schemes licensed by apoyadas each hypothesize distinct morpheme boundaries and should remain in separate clusters. Now consider the implications for a scheme similarity metric that arise from competition between schemes over their morpheme boundary hypotheses. Since schemes possess syntagmatic information in their sets of adherent cstems, it seems reasonable to add information from cstems to information from csuffixes when evaluating scheme similarity. It also seems reasonable to give cstem similarity and csuffix similarity approximately equal weight. Fair representation of cstems and csuffixes is of concern as the number of cstems can far outstrip the number of csuffixes in a scheme: the 1st scheme selected in the ParaMor run of  REF _Ref215296727 \h  \* MERGEFORMAT Figure  4.3 contains just two csuffixes but more than five thousand cstems. To fairly weight syntagmatic and paradigmatic information in a scheme similarity metric, ParaMor compares schemes based on the sets of word types which license each scheme. Comparing schemes by their sets of licensing types weights the aggregate evidence from cstems equally with the aggregate evidence from csuffixes because each licensing type consists of one cstem and one csuffix. But, because a single word, such as apoyadas, may license two or more schemes which hypothesize distinct morpheme boundaries, directly measuring scheme similarity by comparing sets of licensing word types could erroneously merge schemes such as .s and a.as.o.os which model distinct morpheme boundaries in many of the same words! Addressing the problem of separating distinct models of morpheme boundaries, ParaMor annotates each licensing type of each scheme with the morpheme boundary proposed by that scheme. For example, ParaMor annotates the licensing type apoyadas as apoyada+s for the .s scheme, but as apoyad+as for the a.as.o.os scheme. Then to compute the similarity between any pair of schemes, ParaMor measures the similarity between the pairs sets of morpheme boundary annotated types. Morpheme boundary annotated types retain the naturalness of measuring the similarity of schemes through their licensing word types, while distinguishing between schemes that model distinct morpheme boundaries, To measure the similarity between pairs of scheme-clusters, as opposed to pairs of lone schemes, ParaMor must specify for each cluster the set of morpheme boundary annotated word types the cluster covers. There are at least two reasonable ways to define the set of boundary annotated types covered by a cluster. First, ParaMor could define this set as the union of the sets of morpheme boundary annotated types which belong to any scheme in the cluster. Alternatively, ParaMor could define a clusters set of morpheme boundary annotated types as the cross-product of all cstems and all csuffixes contained in any scheme in the cluster. ParaMor opts for the more conservative first option. By allotting to each cluster just those morpheme boundary annotated types which explicitly license some selected scheme, ParaMor guarantees that reconcatenating the cstem and csuffix portions of any boundary annotated type in any cluster will produce a word that directly occurred in the paradigm induction corpus. As long as scheme-cluster similarity is measured over sets of morpheme boundary annotated licensing types, ParaMors performance at clustering schemes is not significantly affected by the particular set similarity metric used. For the experiments reported in this thesis, ParaMor uses the cosine metric of set similarity to measure the relatedness of sets of licensing morpheme boundary annotated word types. The formula for the cosine similarity of two arbitrary sets  and  is: . Clustering Large and Clustering Small To finish out the description of ParaMors clustering algorithm, this section presents an adaptation ParaMor makes to the basic agglomerative clustering algorithm that prevents schemes which are all individually licensed by very few types from joining forces to form a single larger cluster. The adaptation described in this section has a relatively mild impact on the final set of scheme-clusters that ParaMor induces. However, this sections adaptation is described for completeness, as all experiments reported in this thesis take this step. Because individual schemes that receive support from few licensing types may still introduce valid csuffixes, ParaMor does not want to simply discard all small selected schemes. Instead, ParaMor leverages larger selected schemes to rope in the valid small schemes. Specifically, ParaMor requires at least one large scheme for each small scheme a cluster contains, where the size of a scheme is measured as the number of unique morpheme boundary annotated word forms that license it. The threshold size above which schemes are considered large is a free parameter. As described in Section  REF _Ref190686503 \r \h  4.4, the scheme size threshold is reused during ParaMor s filtering stage. Section  REF _Ref190768256 \r \h  4.4.1 details the setting of this scheme size threshold. An Examination of ParaMor s Scheme-Clusters By tailoring bottom-up agglomerative clustering of schemes with first, a discriminative restriction over the csuffixes a cluster may contain, and second, a similarity measure over sets of morpheme boundary annotated types, ParaMor solves the three challenges that face a scheme-clustering algorithm. ParaMor: 1. Recognizes syncretic csuffix overlap in paradigms, 2. Halts the agglomerative clustering algorithm without introducing a free parameter, and 3. Preserves the differing morpheme boundary hypotheses proposed by competing schemes. As a technical summary of ParaMor s scheme-clustering algorithm, Appendix B contains a pseudo-code implementation of bottom-up agglomerative clustering adapted for schemes as described in Sections  REF _Ref215319774 \r \h  4.3.1 and  REF _Ref215319787 \r \h  4.3.2. This section focuses instead on illustrating the output of ParaMor s agglomerative clustering algorithm by looking at clusters built over Spanish selected schemes.  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 contains typical scheme-clusters that ParaMor builds after the three pipelined steps of: 1. Data clean-up (Section  REF _Ref190686436 \r \h  \* MERGEFORMAT  4.2), 2. Initial scheme selection from a morphology network (Chapter 3), and 3. Cluster aggregation (Sections  REF _Ref215377056 \r \h  \* MERGEFORMAT  4.3.1 and  REF _Ref215377063 \r \h  \* MERGEFORMAT  4.3.2). Like  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 found on p. 96,  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 was built from a Spanish newswire corpus of 50,000 types, but all word types in the corpus from which the clusters in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 were built are longer than five characters. Because the corpus underlying  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 is not identical to the corpus backing  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, the schemes from which the clusters of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 were built are not identical to the schemes in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1. But most schemes from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 have a close counterpart among the schemes which contribute to the clusters of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. For example,  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 contains a .s scheme, modeling the most frequent inflection class of Number on Spanish nouns and adjectives. A .s scheme also contributes to the first cluster given in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, but where the .s scheme of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 contains 5501 cstems, the .s scheme contributing to the 1st cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 has a cstem count of 5399. Note that only full clusters are shown in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, not the .s scheme, or any other scheme, in isolation. As another example of similarity between the schemes of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 and those of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, turn to the third cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. This  third cluster contains a scheme model of Gender and Number on Spanish adjectives consisting of the same csuffixes as the 2nd selected scheme in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, namely a.as.o.os. Further correspondences between the clusters of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 and the schemes of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 are given in the second column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, labeled Corresponds to  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1. If the cluster of a row of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 contains a scheme whose set of csuffixes is identical, or nearly identical, to that of a scheme in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, then the rank of the corresponding scheme of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 is given outright in the Corresponds column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4; if the majority of the csuffixes of a scheme of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 appear in a cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, but no particular scheme in that cluster exactly corresponds to the scheme of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, then the Corresponds column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 gives the rank of the  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 scheme in parentheses. The clusters in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 are sorted by the number of unique morpheme boundary annotated surface types which license schemes of the cluster this number of unique licensing types appears in the third column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. Because most cstems do not occur in all of a cluster s schemes, the number of unique licensing types of a cluster is not simply the number of csuffixes multiplied by the number of cstems in the cluster. The fourth column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 gives the number of schemes which merged to form that row s cluster. The only other column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 which does not also appear in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 is the column labeled Allo. The Allo. column is marked with a dot when a row s cluster models a morphophonological alternation. Clusters marked in the Allo. column are discussed further in Section  REF _Ref190771388 \r \h  4.4.2. For additional explanation of the other columns of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, reference their description Section  REF _Ref215326291 \r \h  4.1.1. Zooming in close on one scheme-cluster,  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 contains a portion of the clustering tree for the scheme-cluster from  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 with the 4th most licensing types a cluster covering suffixes which attach to ar verbs. The cluster tree in  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 is of course binary, as it was formed through bottom-up agglomerative clustering. Schemes in  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 appear in solid boxes, while intermediate clusters consisting of more than one scheme are in dotted boxes. Each scheme or cluster lists the full set of csuffixes it contains,  where a cluster contains all csuffixes that appear in any scheme it subsumes. Leaf schemes also report their full sets of cstems; and clusters state the cosine similarity of the sets of boundary annotated licensing types of the clusters two children. It is interesting to note that similarity scores do not monotonically decrease moving up the tree structure of a particular cluster. Non-decreasing similarities are a consequence of computing similarities over sets of objects, in this case sets of morpheme boundary annotated types, which are unioned up the tree. The bottom-most cluster of  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 is built directly from two schemes. Two additional schemes then merge in turn with the bottom-most cluster. Finally, the top-most cluster of  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 is built from the merger of two clusters which already have internal structure. The full cluster tree continues upward beyond the small branch shown in  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 until the cluster contains 23 schemes. Although ParaMor can form clusters from children which do not both introduce novel csuffixes, each child of each cluster in  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5 brings to its parent some csuffix not found in the parent s other child. In each intermediate cluster, any csuffix which does not occur in both children is underlined in  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5. Keeping in mind the scheme-clusters presented in Figures 4.4 and 4.5, examine ParaMor s scheme-clusters in the light of the two broad shortcomings of initially selected schemes discussed in Section  REF _Ref215326291 \r \h  4.1.1: 1. The fragmentation of language paradigms across many scheme models, and 2. The poor precision of selected models against underlying paradigms. Clustering Improves Fragmentation Scheme clustering was specifically designed to address the patchwork fragmentation of paradigms across schemes. One of the most striking features of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 are the clusters which merge schemes that jointly and correctly model significant fractions of a single large Spanish paradigm. One such significant model is the cluster with the 4th largest number of licensing types. A portion of this 4th largest cluster appears in the cluster-tree of  REF _Ref190840539 \h  \* MERGEFORMAT Figure  4.5. All told, the 4th cluster contains 41 csuffixes, more than any other scheme-cluster. These 41 csuffixes model Spanish suffixes which attach to ar verb stems: 7 csuffixes model agglutinative sequences of a non-finite inflectional suffix followed by a pronominal clitic, namely: arla, arlas, arlo, arlos, arme, arse, and ndose; 9 of the csuffixes are various inflectional forms of the relatively productive derivational suffixes acin, ador, and ante; And more than half of the csuffixes in this cluster are the surface forms of inflectional suffixes in the ar inflection class: This 4th cluster contains 24 csuffixes which model inflectional ar suffixes, as presented in Appendix A; while one additional csuffix, ase, is a less common alternate form of the 3rd Person Singular Past Subjunctive. Counting just the 24 csuffixes, this scheme-cluster contains 64.9% of the 37 unique suffix surface forms of the ar inflection class of Spanish verbs listed in Appendix A. Among the 24 inflectional suffixes are all of the 3rd Person endings for both Singular and Plural Number for all seven Tense-Mood combinations that are marked in Spanish. These 3rd Person endings are: a, an, , aron, aba, aban, ar, arn, ara, aran, e, en, ara, and aran. Since newswire is largely written in the 3rd Person, it is to be expected that the 3rd Person morphology is most readily identified from a newswire corpus. Focusing in on one suffix of the ar verbal paradigm, ados, an example suffix followed throughout this chapter, clustering reduces the fragmentation of this one suffix across partial paradigms. Before clustering, the csuffix ados occurs in 40 schemes, after clustering it is present in just 13 distinct clusters. Clearly, ParaMors clustering algorithm has considerably consolidated the fragmented models of the Spanish ar verbal paradigm that were output by the initial bottom-up scheme search. The ar inflection class is the most frequent of the three regular Spanish verbal inflection classes, and so is most completely identified. But the clusters with the 11th and 17th most licensing types cover, respectively, the er and ir inflection classes nearly as completely as the 4th cluster covers the ar inflection class: The 11th cluster models 19 of the 37 unique inflectional suffixes in the er inflection class, 4 inflection+clitic sequences, and 6 derivational suffixes; And the 17th cluster contains 14 of the 36 unique surface forms of inflectional suffixes in the ir inflection class, 4 inflection+clitic sequences, and 2 derivational suffixes. Cluster Precision Now consider how the clusters of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 stack up against the second broad shortcoming of ParaMor s initially selected schemes: the many original schemes that did not model paradigms. First, the data clean-up step, described in Section  REF _Ref190686436 \r \h  4.2, which excludes short types from ParaMors paradigm-induction corpus, virtually eliminated the first subclass of unsatisfactory schemes, those schemes which resulted from chance string similarities between word types. None of the scheme-clusters in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, for example, are built from schemes that arise from chance lexical collisions. Although ParaMor now avoids constructing non-paradigmatic schemes that result from accidental string similarities, because ParaMor s bottom-up scheme search begins an upward path from each individual csuffix, frequent or infrequent (Section  REF _Ref214372965 \r \h  3.2.2), the initial search algorithm constructs many schemes that are licensed by very few word types. Some of these small schemes are absorbed into larger clusters, but ParaMor s csuffix discriminative restriction on scheme clustering (Section  REF _Ref215840852 \r \h  4.3.1), in combination with ParaMor s heuristic restriction on the number of small schemes which a cluster may contain (Section  REF _Ref215319787 \r \h  4.3.2), prevents the majority of these small schemes from joining any cluster. From the 6909 original schemes that ParaMor selects when training on a corpus of types longer than 5 characters, clustering only reduces the total number of separate paradigm models to 6087 scheme-clusters. The last six rows of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 all contain  clusters consisting of just a single scheme that was prevented from merging with any other scheme. All six singleton clusters are licensed by no more than 35 word types. And none of the clusters correctly models inflectional affixes of Spanish. Five of the six singleton clusters misanalyze the morpheme boundaries in their few types, while the cluster with the 300th most licensing types treats a set of unproductive verbal prefixes as cstems, placing valid verb stems into the csuffix set. Section  REF _Ref190768256 \r \h  4.4.1, directly addresses ParaMor s strategy for removing the many small incorrect singleton clusters that the clustering procedure produces. In addition to the large number of incorrect singleton clusters, many initially created clusters misanalyze morpheme boundaries. But these incorrect morpheme boundary models are expected: As described in Section  REF _Ref215464457 \r \h  4.1.2, ParaMor intentionally postpones discarding schemes which hypothesize unlikely morpheme boundaries until after the schemes have been clustered, in the hope that clustering might aggregate schemes which all model the same incorrect boundaries. Half of the clusters in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 hypothesize inferior morpheme boundaries in their licensing types. The largest such cluster is the cluster with the 2nd most licensing types. Like the 2nd selected scheme of  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1, which it subsumes, the 2nd cluster places morpheme boundaries after the a vowel which begins most suffixes in the ar inflection class. And exactly as hoped, the 2nd cluster has nicely unified schemes which all hypothesize the same morpheme boundaries in a large set of typesonly this time, the hypothesized boundaries happen to be incorrect. Section  REF _Ref190771388 \r \h  4.4.2 describes steps of ParaMor s pipeline which specifically remove clusters which hypothesize incorrect morpheme boundaries. A New Shortcoming: Overgeneralization Before moving on to a discussion of the algorithms ParaMor employs to improve the csuffix precision of scheme-clusters, note that clustering introduces a new shortcoming into ParaMors models of paradigms: overgeneralization. Each scheme, C, is a computational model that the specific set of cstems and csuffixes of C are paradigmatically related. When ParaMor merges C to a second scheme, C2 , the paradigmatic relationship of the cstems and csuffixes of C is generalized to include the cstems and csuffixes of C2 as well. Sometimes a merger s generalization is well founded, and sometimes it is misplaced. When both C and C2 model inflectional affixes of the same paradigm on syntactically similar stems, then the cstems of C usually do combine to form valid words with the csuffixes of C2 . For example, the suffixes ir and imos are regular inflectional suffixes of the ir inflection class of Spanish verbs. Although the csuffix ir never occurs in any scheme with the csuffix imos, and although the Spanish word cumplimos we carry out never occurs in the Spanish corpus from which ParaMor learns paradigms, the cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 with the 21st most licensing types places the csuffixes ir and imos in the same cluster together with the cstem cumpl correctly predicting that cumplimos is a valid Spanish word form. On the other hand, when a csuffix, f, of some scheme, C, models a syntactically or idiosyncratically restricted suffix, it is unlikely that f forms valid words with all the cstems of a merged cluster C2 . Consider the 1st scheme-cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 which joins the scheme .s with the schemes .mente.s and menente.mente. The csuffixes and s mark Singular and Plural Number, respectively, on nouns and adjectives; The suffix (a)mente productively converts an adjective into an adverb, something like the suffix ly in English; and the string menente is simply a misspelling. Where the .s scheme contains 5399 cstems, the scheme .mente.s contains 253, and the scheme menente.mente contains just 3 candidate stems: inevitable, unni, and nica. The 1st scheme-cluster contains many Spanish cstems that represent only nouns, including hombro  shoulder listed in the CStem column of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. These nominal cstems originate in the .s scheme and do not form legitimate Spanish adverbs by attaching mente: *hombromente *shoulderly. Furthermore, productively assuming that the csuffix menente can attach to any candidate stem is wrong. Thus this 1st cluster has overgeneralized in merging these three schemes. Overgeneralization is endemic to all clustering algorithms, not just unsupervised bottom-up agglomerative clustering of schemes. And in the particular case of scheme clustering, it would be difficult for any unsupervised method to reliably distinguish between infrequent inflectional affixes on the one hand and reasonably frequent derivational affixes, such as mente, on the other. Chapters 5 and 6 of this thesis describe applying ParaMors induced scheme-clusters to a morphological analysis task: Specifically ParaMor segments word forms into constituent morphemes. But before ParaMor could be applied to a generation task that would propose novel full form words, the problem of overgeneralization in scheme-clusters would need to be seriously addressed. Filtering of Merged Clusters With most valid schemes having found a safe haven in a cluster with other schemes that model the same paradigm, ParaMor focuses on improving precision by removing erroneous scheme-clusters. ParaMor applies two classes of filters to cull out unwanted clusters. These two filter classes address the two remaining types of cluster-precision error described in Section  REF _Ref215396639 \r \h  4.3.3, p. 119. The first filter class, detailed in Section  REF _Ref190768256 \r \h  4.4.1, targets the many scheme-clusters with support from only few licensing types. The second class of filter, presented in Section  REF _Ref190771388 \r \h  4.4.2 identifies and removes remaining scheme-clusters which hypothesize incorrect morpheme boundaries. Filtering of Small Scheme-clusters ParaMors first class of filtering algorithm consists of just a single procedure which straightforwardly removes large numbers of erroneous small clusters: the filter discards all clusters that are licensed by less than a threshold number of morpheme boundary annotated word types. To minimize the number of free parameters in ParaMor, the threshold below which this filter discards small clusters is tied to the clustering threshold described in Section  REF _Ref215396791 \r \h  4.3.2, which restricted the number of small schemes that may join a cluster to be no more than the number of large schemes in the cluster. These two thresholds can be reasonably tied together for two reasons. First, both thresholds limit the influence of small erroneous schemes. Second, both thresholds measure the size of a cluster as its number of licensing morpheme-boundary annotated types.  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6 graphs the number of clusters that ParaMor identifies over a Spanish corpus after first clustering schemes with a particular setting, k, of the cluster-size threshold and then filtering out those clusters which are not licensed by at least k word types.  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6 also contains a plot of suffix recall as a function of these tied thresholds. ParaMor calculates suffix recall by counting the number of string-unique surface forms of Spanish inflectional suffixes, as given in Appendix A, that appear in any identified cluster. The technique described in this section for removing small clusters was developed before ParaMor adopted the practice of only training on longer word types; and  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6 presents the cluster count and suffix recall curves over a corpus that includes types of all lengths.  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6 presents results over corpora that are not restricted by type length because it is from the unrestricted corpus data that the threshold for small cluster filtering was set to the value which ParaMor uses in experiments throughout this thesis. Over a Spanish corpus that only includes types longer than 5 characters, the effect of filtering by licensing type count is similar. Looking at  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6, as the size threshold increases the number of clusters that ParaMor retains quickly drops. But suffix recall only slowly falls during the steep decline in cluster count: ParaMor is therefore discarding mostly bogus schemes containing illicit suffixes. Because recall degrades gracefully, the exact size threshold below which clusters are discarded should have a relatively minor effect on the paradigms that ParaMor induces. At a threshold size of 37 morpheme-boundary annotated licensing word types, ParaMor retains more than 80% of the string-unique inflectional suffixes of Spanish while significantly reducing the number of clusters that ParaMor selects. Except where otherwise noted, all experiments reported in this thesis use a cluster-size threshold of 37. At this threshold value, all but 137 of the 7511 clusters that formed from the 8339 originally selected schemes are removed, a 98.2% reduction in the number of clusters. Note that the vertical scale on  REF _Ref191194358 \h  \* MERGEFORMAT Figure  4.6 goes only to 1000 clusters. Counting in another way, before filtering, the scheme-clusters contained 9896 unique csuffixes, and after filtering just 1399, an 85.9% reduction. The recall of unique inflectional suffixes at a threshold value of 37 licensing types is 81.6%, or 71 out of 87. Before filtering schemes for the number of licensing types they contain, 92.0% of the unique suffixes of Spanish morphology appeared as a csuffix in some scheme-cluster. But this automatically derived value of 92.0%, or 80 of 87, is somewhat misleading. At a threshold value of 37, nine unique csuffixes which are string identical to true Spanish suffixes are lost. But six of the nine lost unique csuffixes appear in schemes that clearly do not model Spanish inflectional suffixes. For example, one csuffix that is removed during cluster size filtering is eras, which is string identical to a 2nd Person Singular Present Conditional Spanish verbal suffix. But the csuffix eras appears in only one clustera cluster which clearly does not model Spanish verbs. The cluster in which eras occurs consists of the single scheme era.eras.o.os with its cstems escud.ganad.grad.libr.mercad. Although the csuffixes era, eras, and o are all string identical to suffixes in the Spanish er verbal paradigm, in this scheme these csuffixes do not arise from verbal licensing forms: Reconcatenating cstems and csuffixes, most word forms which both license this cluster and end in era or eras, like librera/libreras library/libraries and ganadera/ganaderas ranch/ranches, are not verbs but nouns; And forms with the o/os endings, like libro/libros book/books and ganado/ganados cattle (sg/pl), are derivationally related nouns. After ignoring csuffixes which appear in schemes where the csuffix does not model a Spanish inflectional suffix given in the summary of Spanish inflectional morphology of Appendix A, only three true csuffixes are lost during this first filtering step that removes small clusters. When training from a Spanish corpus consisting only of types longer than 5 characters in length, the numbers follow a similar pattern to those given above. Before clustering, ParaMor identifies 6909 schemes; this is reduced to 6087 after clustering; and after filtering at a threshold of 37, only 150 clusters remain. The recall of unique suffixes, when training over a Spanish corpus restricted for type length, are identical to the values over the unrestricted corpus: 92.0% before filtering and 81.6% after, although, oddly, ParaMor does not either identify or filter exactly the same set of inflectional suffixes. Of the clusters presented in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, the six clusters in the last six rows each contain fewer than 37 licensing types, and are therefore discarded. For convenience these six small clusters are repeated in  REF _Ref215456615 \h  \* MERGEFORMAT Figure  4.7 And finally, paradigm fragmentation improves for the particular suffix of the ar inflection class that this chapter follows: After filtering, 7 clusters remain out of the original 13 clusters which contained the csuffix ados. Two of the six discarded clusters that contain ados are clearly correctly discarded as they are among the few surviving clusters which arise from chance string similarities between word types. Of the additional four clusters containing ados that ParaMor removes, each models at least one relatively unproductive derivational suffix. The discarded cluster ado.ados.amento.amentos.ando, for example, contains csuffixes which model the inflectional suffixes ado, ados, and ando, but also c-suffixes modeling forms of the derivational suffix amento/amentos, which forms nouns from some ar verbs. As ParaMor is designed to identify productive inflectional morphology, it is not unreasonable or unexpected that unproductive derivational suffixes are lost. Morpheme Boundary Filtering In Spanish, as described in Section  REF _Ref190768256 \r \h  4.4.1, filtering scheme-clusters by thresholding the number of types that must license each cluster drastically reduces the number of clusters that ParaMor proposes as models for inflectional paradigms. From the thousands of initially created scheme-clusters, type-license filtering leaves less than two hundred. This is progress in the right direction. But as Spanish has fewer than ten productive inflectional paradigms, see Appendix A, ParaMor still vastly over estimates the number of Spanish paradigms. A hand analysis of scheme-clusters reveals that the major type of erroneous cluster which persists after size filtering are those clusters that incorrectly model morpheme boundaries. After training and filtering over a Spanish corpus of 50,000 types that are each longer than 5 characters in length, exactly 150 scheme-clusters remain. And of these 150 clusters, more than two-thirds, 108, hypothesize an incorrect morpheme boundary in their licensing types. That misplaced morpheme boundaries are the major source of error in the remaining scheme-clusters is not surprising. Morpheme boundary errors comprised one of the major error classes of the initially selected schemes identified in Section  REF _Ref215464223 \r \h  4.1.1. And thus far in ParaMor s processing pipeline no algorithm has specifically dealt with boundary errors. Indeed, as described in Section  REF _Ref215464223 \r \h  4.1.1, ParaMor has intentionally postponed the removal of schemes and scheme-clusters that hypothesize incorrect morpheme boundaries until after ParaMors scheme-clustering step. The Problem: Mutual Substitutability ParaMors initial bottom-up morphology network search strategy, described in Chapter 3, is designed to detect the hallmark of inflectional paradigms in natural language: mutual substitutability between sets of affixes. But, when the csuffixes of a scheme break not at a morpheme boundary, but rather at some character boundary internal to a true morpheme, the incorrect csuffixes are sometimes still mutually substitutable.  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8 contains three scheme-clusters, built over a Spanish corpus, that illustrate non-paradigmatic mutual substitutability. The schemes of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8 were first shown in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. The scheme-cluster on the 1st row of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8 incorrectly hypothesizes a morpheme boundary that is after the a vowel which begins many inflectional and derivational ar verb suffixes. In placing the morpheme boundary after the a, this scheme-cluster cannot capture the full paradigm of ar verbs. Compare, for example, the cluster on the 2nd row of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8, which includes inflectional suffixes such as o 1st Person Singular Present Indicative and 1st Person Singular Past Indicative which do not begin with a. Nevertheless, the (incorrect) csuffixes which appear in the 1st row s cluster are mutually substitutable: in the Spanish word administrados administered (Adjectival Masculine Plural), the erroneously segmented csuffix dos can be replaced by to form the Spanish word administra administer (3rd Person Singular Present Indicative), or dos can be replaced by the csuffix da to form administrada (Adjectival Feminine Singular), etc. Similarly, the scheme-cluster on the 3rd row of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8 models some of the many Spanish adjective stems which end in t, abierto  open , cierto  certain , pronto  quick etc. but this cluster incorrectly prepends the final t of these adjective stems to the adjectival suffixes, forming csuffixes such as: ta, tas, and to. And these prepended csuffixes are mutually substitutable on adjectives whose stems end in t: abierto becomes abiertas when to is replaced by tas. Encountering schemes like those that contribute to the clusters on the 1st and 3rd rows of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8, ParaMor s initial search strategy discovers mutual substitutability and erroneously selects these incorrectly segmented csuffix sets. Since ParaMor cannot rely on mutual substitutability of suffixes to identify correct morpheme boundaries, ParaMor turns to a secondary characteristic of paradigms. The Solution: Letter Successor Variety The secondary characteristic that ParaMor adapts in order to identify those scheme-clusters which hypothesize incorrect morpheme boundaries is letter successor variety. Letter successor variety is an idea that was originally proposed by Harris (1955). Take any string t. Let F be the set of strings such that for each  EMBED Equation.3 , t.f is a word form of a particular natural language. Harris noted that when the right edge of t falls at a morpheme boundary, the strings in F typically begin in a wide variety of characters; but when t divides a word form at a character boundary internal to a morpheme, any legitimate word final string must first complete the erroneously split morpheme, and so will begin with the same character. This argument similarly holds when the roles of t and f are reversed. Harris harnesses this idea of letter successor variety by first placing a corpus vocabulary into a character tree, or trie, and then proposing morpheme boundaries after trie nodes that allow many different characters to immediately follow. Consider Harris algorithm over a small English vocabulary consisting of the twelve word forms: rest, rests, resting, retreat, retreats, retreating, retry, retries, retrying, roam, roams, and roaming. The upper portion of  REF _Ref191198647 \h  \* MERGEFORMAT Figure  4.9 places these twelve English words in a trie. The bottom branch of the trie begins r-o-a-m. Three branches follow the m in roam, one branch to each of the trie nodes , i, and s. Harris suggests that such a high branching factor indicates there may be a morpheme boundary after r-o-a-m. The trie in  REF _Ref191198647 \h  \* MERGEFORMAT Figure  4.9 is a forward trie in which all vocabulary items share a root node on the left. A vocabulary also defines a backward trie that begins with the final character of each vocabulary item.  Adapting Letter Successor Variety to Schemes Interestingly there is a close correspondence between trie nodes and ParaMor schemes. Each circled sub-trie of the trie in the top portion of  REF _Ref191198647 \h  \* MERGEFORMAT Figure  4.9 corresponds to one of the four schemes in the bottom-right portion of the figure. For example, the right-branching children of the y node in retry form a sub-trie consisting of and i-n-g, but this same sub-trie is also found following the t node in rest, the t node in retreat, and the m node in roam. ParaMor conflates all these sub-tries into the single scheme .ing with the four adherent cstems rest, retreat, retry, and roam. Notice that the number of leaves in a sub-trie corresponds to the paradigmatic csuffix level of a scheme, e.g. the level 3 scheme .ing.s corresponds to a sub-trie with three leaves ending the trie paths , i-n-g, and s. Similarly, the number of sub-tries which conflate to form a single scheme corresponds to the number of adherent cstems belonging to the scheme. By reversing the role of csuffixes and cstems, a backward trie similarly collapses onto ParaMor schemes. Just as schemes are analogues of trie nodes, ParaMor can link schemes in a fashion analogous to transition links between nodes in forward and backward tries. Transition links emanating to the right from a particular scheme, C, will be analogues of the transition links in a forward trie, and links to the left, analogues of transition links in a backward trie. To define forward scheme links from a scheme, C, let the set  EMBED Equation.3  consist of all csuffixes of C which begin with the same character, h. Strip h from each csuffix in  EMBED Equation.3 , forming a new set of csuffixes,  EMBED Equation.3 . Link C to the scheme containing exactly the set of csuffixes  EMBED Equation.3 . Schemes whose csuffixes all begin with the same character, such as t.ting.ts and t.ting, have exactly one rightward path that links to the scheme where that leading character has been stripped from all csuffixes. For example, in  REF _Ref191198647 \h  \* MERGEFORMAT Figure  4.9 the t.ting.ts scheme is linked to the .ing.s scheme. Leftward links among schemes are defined by reversing the roles of cstems and csuffixes as follows: for each character, h, which ends a cstem in a particular scheme, C, a separate link takes C to a new scheme where h starts all csuffixes. For example, the .ing.s scheme contains the cstems rest and retreat, which both end in the character t, hence there is a link from the .ing.s scheme to the t.ting.ts scheme. Note that when all the csuffixes of a scheme, C, begin with the same character, the rightward link from C to some scheme, C2 , exactly corresponds to a leftward link from C2 to C. Drawing on the correlation between character tries and scheme networks, ParaMor ports Harris trie based morpheme boundary identification algorithm quite directly into the space of schemes and scheme-clusters. Just as Harris identifies morpheme boundaries by examining the variety of the branches emanating from a trie node, ParaMor identifies morpheme boundaries by examining the variety in the trie-style scheme links. ParaMor employs two filters which examine trie-style scheme links: the first filter seeks to identify scheme-clusters, like the cluster on the 1st row of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8, whose morpheme boundary hypotheses are placed inside true suffixes; while the second filter removes scheme-clusters that hypothesize stem-internal morpheme boundaries, as the cluster on the 3rd row of  REF _Ref215471559 \h  \* MERGEFORMAT Figure  4.8 does. Filtering Suffix-Internal Morpheme Boundary Errors To identify scheme-clusters whose morpheme boundary hypotheses are incorrectly suffix-internal, ParaMor s first boundary filter examines the Harris-style letter successor variety of the leftward trie links of the schemes in a cluster. When a scheme places its morpheme boundaries at locations that are internal to true suffixes, leftward trie links lead back toward the stem and back toward the correct morpheme boundary of each licensing type. Following Hafer and Weiss (1974) in a trie algorithm and Goldsmith (2001; 2006) in a paradigm-based system, ParaMor measures letter successor variety using entropy. Each leftward scheme link, l, can be weighted by the number of cstems in that scheme whose final character advocates l. In  REF _Ref191198647 \h  \* MERGEFORMAT Figure  4.9 two cstems in the .ing.s scheme end in the character t, and thus the leftward link from .ing.s to t.ting.ts receives a weight of two. ParaMor then measures the entropy of the distribution of the cstem weighted links. A leftward link entropy close to zero indicates that the cstems of a scheme have little variety in their final character; And minimal character variety is the sign of a boundary hypothesis placed at a morpheme-internal position. Goldsmith (2001; 2006) To judge whether a cluster of schemes hypothesizes an incorrect morpheme boundary, ParaMors suffix-internal boundary filter examines the leftward link entropy of each scheme in each cluster. Each scheme with a leftward link entropy below a threshold is flagged as an error. And if at least half of the schemes in a cluster are flagged, then ParaMors suffix-internal error filter discards that cluster. ParaMors suffix-internal filter is conservative in discarding schemes. ParaMor uses a threshold value, 0.5, over the leftward link entropy that only flags a scheme as containing a boundary error when virtually all of a schemes cstems end in the same character.  REF _Ref215484430 \h Figure  4.10 is a pseudo-code implementation of ParaMor s algorithm to filter out suffix-internal morpheme boundary errors. Filtering Stem-Internal Morpheme Boundary Errors ParaMor employs a second morpheme boundary filter to identify scheme-clusters which incorrectly hypothesize boundaries that are internal to the stems of the words which license the cluster. This filter over stem-internal boundary errors is designed to remove schemes like that found on the 2nd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11, where the final t of true Spanish stems has been stripped from the stem and erroneously added to the front of all csuffixes in the scheme.  But consider ParaMor s quandary when examining the morpheme boundaries proposed by the schemes on the 1st and 3rd rows of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11. The scheme on the 1st row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 is taken from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 on p. 96, which lists schemes selected in one particular Spanish run of ParaMor s initial scheme-search procedure. The scheme on the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 is a valid scheme which ParaMor could have selected, but did not. This 3rd row scheme arises from the same word types which license the 1st row s scheme. The schemes on the 1st and 3rd rows of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 both attempt to model inflectional suffixes of the verbal ar paradigm. And indeed, it is somewhat ambiguous where to place the morpheme boundary in Spanish verbs. All of the cstems of the 1st rows scheme end with the character a, but all of the csuffixes of the 3rd rows scheme begin with the character a. While traditional grammarians would likely prefer the boundaries hypothesized by the scheme on the 3rd row, many linguists would argue that the inflectional suffix of a Spanish verb begins after the characteristic vowel, as the 1st rows scheme suggests. ParaMor adopts a pragmatic view, when there is ambiguity about a morpheme boundary, ParaMor retains the left-most reasonable boundary in  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11, the scheme on the 3rd row proposes the left-most boundary that is consistent with the word forms which license the 1st and the 3rd rows schemes. Biasing ParaMors morpheme boundary filters toward preferring the left-most reasonable boundary requires that a filter identifying misplaced stem-internal boundaries cannot merely be the mirror image of the suffix-internal error filter. As described in the previous subsection, the suffix-internal boundary error filter will note the low entropy of the cstem-final character distribution among the cstems of the scheme on the 1st row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11; consequently the suffix-internal error filter will flag the 1st row s scheme as hypothesizing an incorrect morpheme boundary. If the stem-internal error filter were the mirror image of the suffix-internal filter, then, because all of the csuffixes of the 3rd row s scheme begin with the same character, the scheme on the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 would also be flagged as not modeling a correct morpheme boundary! ParaMor s stem-internal error filter achieves asymmetry from the suffix-internal filter by examining both rightward and leftward scheme links. Indeed, the core subroutine of the suffix-internal error filter that measures the entropy of leftward scheme links is called from ParaMors stem-internal filter. The first requirement that a scheme, C, must meet to be flagged as a stem-internal boundary error is that all the csuffixes in C must begin with the same character, or equivalently, that the entropy in the distribution of the initial character of C s csuffixes must be zero. If all the csuffixes of C begin with the same character, then ParaMor strips that character from each csuffix in C and examines the scheme, C2 , that contains exactly the stripped set of csuffixes. For example, since all the csuffixes in the scheme on the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 begin with the character a: a.aba.aban.ada.adas.ado.ados.an.ando.ar.ara.aron.arse.ar.arn, ParaMor s stem-internal boundary error filter follows the single rightward path along the character a to the scheme found on the 1st row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11, namely .ba.ban.da.das.do.dos.n.ndo.r.ra.ron.rse.r.rn. Similarly, because all of the csuffixes that belong to the scheme on the 2nd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 begin with the character t: ta.tamente.tas.to.tos, ParaMor examines the scheme a.amente.as.o.os. Once ParaMor has identified a C2 scheme that lies along an unambiguous rightward scheme link, ParaMor considers the likelihood that C2 is a (left edge of a) morpheme boundary. If C2 is a likely morpheme boundary then ParaMor discards the original C scheme ParaMor assumes that the original C misplaced the morpheme boundary a little bit to the left of the correct location, accidentally moving inside a true stem. To assess the likelihood that C2 correctly models a morpheme boundary, ParaMor simply asks the suffix-internal boundary error filter to measure the entropy of the leftward distribution of cstem-final characters of C2 . For example, when C is the scheme on the 2nd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 and C2 is the scheme a.amente.as.o.os, which correctly models morpheme boundaries that divide adjectival stems and suffixes, the suffix-internal boundary error filter finds that C2 does indeed mark a morpheme boundary and so ta.tamente.tas.to.tos is rightly flagged as erroneously hypothesizing a stem-internal morpheme boundary. Conversely, when C is the scheme on the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 and C2 is the 1st row s scheme, the 3rd row s scheme cannot be discarded, because the suffix-internal error filter will not flag the 1st row s scheme as lying at a morpheme boundary the cstems of the 1st row scheme end in a wide variety of characters. All the csuffixes of the scheme on the 2nd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 and all the csuffixes of the scheme on the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 begin with the same single character these two schemes have misplaced their morpheme boundaries exactly one character inside the stems of their licensing types. But it is not uncommon for ParaMor to select schemes that clip two or more characters off the tail of each stem. To catch stem-internal boundary errors that are placed deep inside stems, ParaMor follows chains of rightward scheme links. Specifically, if ParaMor can reach any scheme along a non-branching rightward path that scores as a likely morpheme boundary, then the original scheme is discarded. Two examples will help illustrate rightward scheme chains in the stem-internal filter. First, as mentioned, when ParaMors stem-internal error filter evaluates the scheme from the 3rd row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11, the filter first visits the scheme on the 1st row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 and decides that the 1st row s scheme does not model a morpheme boundary. At this point, the stem-internal error filter attempts to strip off yet another character from the front of the csuffixes that belong to the 1st rows scheme. But the csuffixes of the scheme on the 1st row of  REF _Ref215541334 \h  \* MERGEFORMAT Figure  4.11 do not all begin with the same character: the null csuffix, , begins with the null character, the csuffix ba begins with b, the csuffix da with d, etc. Since multiple rightward paths emanate from the 1st row s scheme, the stem-internal error filter cannot examine any more schemes. And, as the stem-internal filter encountered no rightward scheme with high leftward entropy, ParaMor accepts the scheme a.aba.aban.ada.adas.ado.ados.an.ando.ar.ara.aron.arse.ar.arn as modeling a valid morpheme boundary. As a second example of rightward scheme chains, consider the diagram in  REF _Ref215571062 \h  \* MERGEFORMAT Figure  4.12. Each circle in this diagram abstractly represents a scheme. With its high leftward scheme-link entropy, the double-circled scheme in the center of the figure is correctly identified as a morpheme boundary. The far-left scheme-circle in  REF _Ref215571062 \h  \* MERGEFORMAT Figure  4.12 hypothesizes a stem-internal boundary error, but to discover the far-left scheme s error, ParaMor s stem-internal filter must follow two rightward links to reach the double-circled scheme. In summary, ParaMor will flag a scheme, C, as likely hypothesizing a stem-internal boundary if there exists a non-branching rightward path from C leading to a scheme, C2 , such that ParaMor believes C2 falls at the left edge of a correct morpheme. Like the suffix-internal morpheme boundary error filter, to discard a cluster of schemes, the stem-internal error filter must flag half of the schemes in a cluster as hypothesizing an incorrect morpheme boundary.  REF _Ref215574539 \h  \* MERGEFORMAT Figure  4.13 contains a pseudo-code implementation of ParaMor s stem-internal morpheme-boundary error filter. ParaMor s Paradigm Models Section  REF _Ref190686503 \r \h  4.4 completes the description of all steps in ParaMor s paradigm identification pipeline. The description of ParaMor s pipeline began in Chapter 3 with the initial scheme search and continued in this chapter with scheme clustering and filtering.  REF _Ref191364304 \h  \* MERGEFORMAT Figure  4.14 is a graphical representation of ParaMor s pipelined paradigm discovery algorithms. Beginning at the top of the figure: A monolingual natural language corpus is screened of all words 5 characters or less in length (Section  REF _Ref190686436 \r \h  \* MERGEFORMAT  4.2). From the new corpus of longer types, ParaMor searches a network of candidate paradigms, or schemes, for those which most likely model true inflectional paradigms (Chapter 3). The many overlapping and fragmented scheme models of partial paradigms are then clustered into unified models of individual inflectional paradigms (Section  REF _Ref190686481 \r \h  \* MERGEFORMAT  4.3). And finally, three filtering algorithms remove those clusters which, upon closer inspection, no longer appear to model inflectional paradigms: one filter removes small singleton clusters; while two others examine for errors the morpheme boundaries that the proposed scheme-clusters hypothesize in their licensing types. After these six steps, ParaMor outputs a relatively small and coherent set of scheme-clusters which it believes model the inflectional paradigms of a language. The remainder of this section assesses the quality of ParaMors Spanish paradigm induction performance after the full set of pipelined induction steps has been taken. Section  REF _Ref217137097 \r \h  4.6 in this chapter, together with Chapters 5 and 6, will broaden the evaluation of ParaMor to include English, German, Finnish, Turkish, and Arabic. Section  REF _Ref191979081 \r \h  4.5.1 concludes the in-depth examination of ParaMor s Spanish paradigm models that began in Section  REF _Ref215204016 \r \h  4.1.1. Where Section  REF _Ref215204016 \r \h  4.1.1 discussed the initial schemes that ParaMor s bottom-up search procedure produces, Section  REF _Ref191979081 \r \h  4.5.1 takes a detailed look at ParaMor s final set of filtered Spanish scheme-clusters. Section  REF _Ref215980624 \r \h  4.5.2 then presents an experiment that measures the robustness of ParaMor s paradigm building pipeline to a reduction in the size of the induction corpus.  ParaMor s Final Scheme-Clusters as Viable Models of Paradigms This section evaluates the final scheme-clusters that ParaMor produces over a Spanish newswire corpus of 50,000. As described in Section  REF _Ref190768256 \r \h  4.4.1, after scheme search, scheme clustering, and the filtering out of small scheme-clusters, ParaMor retains 150 scheme-clusters. Of these 150, the stem-internal and suffix-internal morpheme boundary error filters remove all but 42; And of the 108 scheme-clusters which hypothesized an incorrect morpheme boundary only 12 are not discarded by the morpheme boundary filters. Unfortunately, ParaMors filtering of schemes which incorrectly model morpheme boundaries does have collateral damage: Recall of string-unique Spanish suffixes drops from 81.6% to 69.0%. All together, 11 unique csuffixes that were string identical to Spanish inflectional suffixes given in Appendix A no longer appear in any cluster that ParaMor retains. Four of these unique csuffixes were only found in clusters which did not model a Spanish paradigm. Consider the fate of the csuffix iste which is string identical to the er/ir suffix that marks 2nd Person Singular Past Indicative. The only cluster in which the csuffix iste occurs is iste.isten.istencia.istente.istentes.istiendo.istir.isti.istaa cluster which ParaMors stem-internal morpheme boundary error filter correctly deletes. The iste-containing cluster is an incorrect segmentation of verb forms whose stems end in the string ist, such as consist + e, exist + e, persist + e, etc. But 7 of the 11 lost unique csuffixes model true Spanish suffixes. All 7 of these lost csuffixes model Spanish pronominal clitics. And all 7 were lost when the only cluster which modeled these Spanish clitics was removed by the suffix-internal morpheme boundary error filter. The specific cluster that was removed is: .a.emos.la.las.le.lo.los.me.on.se..n.a.an. In this cluster, the csuffixes la, las, le, lo, los, me, and se are all pronominal clitics, the csuffix correctly captures the fact that not all Spanish verbs occur with a clitic pronoun, and the remaining csuffixes are incorrect segmentations of verbal inflectional suffixes. While it is worrisome that an entire category of Spanish suffix can be discarded with a single mistake, Spanish clitics had two counts against them. First, ParaMor was not designed to retrieve rare paradigms, but pronominal clitics are very rare in formal Spanish newswire text. And second, the pronominal clitics which do occur in newswire text almost exclusively occur after an infinitive morpheme, usually ar. Exacerbating the problem, the csuffixes which appear alongside the clitics in this cluster are incorrect segmentations whose cstems also end in r. As so many cstems in the .a.emos.la.las.le.lo.los.me.on.se..n.a.an cluster end in r, the suffix-internal boundary-error filter believes this cluster to hypothesize an erroneous morpheme boundary. ParaMors bias toward preferring the left-most plausible morpheme boundary will fail whenever the csuffixes of a cluster consistently follow the same suffix, or even when they consistently follow the same set of suffixes which all happen to end with the same character. This is a weakness of ParaMors current algorithm. Note however, that ParaMor retains clusters that contain csuffixes that model cross-product sequences of verbal inflectional suffix + clitic. For example, turning to  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15, one cluster that ParaMor retains is the scheme-cluster on the 4th row of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15. This 4th row cluster contains such inflectional suffix + clitic cross-product csuffixes as: arla, arse, and ndose.  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 is a reprint of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 which contained a sampling of 17 clusters that ParaMor constructs from initially selected schemes before any filtering. Keeping in mind ParaMor s full paradigm induction pipeline, here are the fates of the scheme-clusters first introduced in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4: The cluster on the top row of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 models the most prevalent inflection class of Number on Spanish nouns and adjectives, containing the scheme .s. This 1st cluster is correctly retained after all filtering steps. The 2nd scheme-cluster in  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 incorrectly places a morpheme boundary after the epenthetic vowel a which leads off most suffixes in the ar inflection class. ParaMors suffix-internal morpheme boundary error filter correctly and successfully removes this 2nd cluster. ParaMor correctly retains the scheme-clusters on the 3rd, 4th, 7th, and 8th rows of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15. These clusters have respectively the 3rd, 4th, 11th, and 17th most licensing types. The 3rd scheme-cluster covers the scheme, a.as.o.os, which models the cross-product of Gender and Number on Spanish adjectives, and which was the 2nd selected scheme during ParaMors initial search. The other candidate suffixes in this cluster include a version of the adverbial suffix (a)mente, and a number of derivational suffixes that convert adjectives to  nouns. The 4th, 11th, and 17th scheme-clusters in  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 are correct collections of inflectional and derivational suffixes that correspond to, respectively, the verbal ar, er, and ir paradigms. The 5th scheme-cluster in  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 segments a Spanish nominalization internally. But ParaMors morpheme boundary filters are unsuccessful at removing this scheme-cluster because this Spanish nominalization suffix has four allomorphs: sion, cion, sin, and cin. The 5th scheme-cluster places a morpheme boundary immediately before the i in these allomorphs. The suffix-internal morpheme boundary error filter is unable to remove the cluster because some cstems end in s while others end in c, increasing the leftward link entropy. But the stem-internal boundary error filter is also unable to remove the cluster, because, from a majority of the schemes of this cluster, after following a rightward link through the initial i of these csuffixes, ParaMors stem-internal error filter reaches a scheme that, although still morpheme-internal, has two rightward trie-style paths, one following the character o and one following the character . In fact, the largest class of errors in ParaMors remaining 42 scheme-clusters consists of clusters which, like the 5th cluster, involve allomorphic variation in the cstems or csuffixes: Thirteen final clusters contain allomorphic errors. In  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15, in addition to the 5th cluster, the cluster on the 9th row, with the 21st most licensing types, is also led astray by an allomorphic alternation in the form of a morphemethis time it is the stem that changes form. The 21st cluster hypothesizes a morpheme boundary to the left of the true stem boundary. Although the cluster with the 21st most licensing types includes the final characters of verb stems within csuffixes, the 21st cluster is modeling the allomorphic change that in Spanish orthography stem final c becomes zc in some Spanish verbs. The only way ParaMor can model an allomorphic stem change is by expanding the csuffixes of a scheme or cluster to include the variable portion of the verb stems. Both the 5th cluster and the 21st cluster are marked in the Allo. column of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15. Nearing the end of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15, the scheme-cluster on the 6th row of  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15, with the 10th most licensing types and the scheme-cluster on the 10th row, with the 100th most licensing types hypothesize morpheme boundaries in adjectives at stem-internal positions. Both the 6th and the 10th clusters are rightly removed by ParaMors stem-internal morpheme boundary error filter. Correctly, neither the suffix-internal nor the stem-internal morpheme boundary filter removes the scheme-cluster on the 11th row, with the 122nd most licensing types, which models Plural Number on nouns. Finally, as mentioned in Section  REF _Ref190768256 \r \h  4.4.1, the last six scheme-clusters in  REF _Ref215989076 \h  \* MERGEFORMAT Figure  4.15 are removed by ParaMor s cluster size filter that looks at the number of licensing types in a cluster. In summary, the scheme-clusters that ParaMor produces as models of Spanish paradigms are generally quite reasonable. In the course of ParaMors paradigm induction pipeline, ParaMor builds models of all major inflectional paradigms of Spanish including: both major sub-paradigms marking Number on nouns, the paradigm cross-product of Gender and Number on adjectives, and all thre sub-paradigms of Spanish verbs. In Chapter 5, ParaMor will take the paradigms that are built for any particular natural language and morphologically segment word-forms of that language. Paradigm Learning and Vocabulary Size As a final examination of paradigms that ParaMor builds from a Spanish corpus, this section considers the lower limit of the vocabulary size from which ParaMor can confidently learn morphological paradigm structure. Where the majority of this and the previous chapter held the vocabulary size constant and focused on devising strong algorithms for identifying the paradigms of a language in an unsupervised fashion, this section takes ParaMors full paradigm identification pipeline as given and investigates the effects of limiting the size of the vocabulary. While ParaMors algorithms are accepted as given, this vocabulary experiment still requires careful consideration of how to set ParaMors free parameters. ParaMor has five free parameters: one parameter thresholds the scheme cstem ratios in ParaMors bottom-up initial search phase; two parameters set the entropy values at which ParaMor discards scheme-clusters that seem to model incorrect morpheme boundaries; one parameter sets the minimum word length for the vocabulary; and finally, one parameter thresholds the minimum number of word types which must license a scheme-cluster. Of these five parameters, only the word-licensing cutoff for clusters need vary with vocabulary size. In thresholding a ratio of cstems, the parameter controlling ParaMor bottom-up search is inherently agnostic to absolute counts; as described in Section  REF _Ref190771388 \r \h  4.4.2 the thresholds over morpheme boundary errors are already set conservatively; and this experiment will leave unchanged the requirement that all input words be at least five characters long. The cluster size threshold, on the other hand is sensitive to vocabulary size with fewer types in the induction vocabulary, a scheme built from the same csuffix set will have fewer adherent cstems. And, indeed, when the cluster size threshold is left at the value set in Section  REF _Ref190768256 \r \h  4.4.1, namely 37, ParaMor removes an unreasonable number of scheme-clusters at lower vocabulary sizes. Hence, for this experiment, the cluster size threshold is linearly scaled with vocabulary size such that at a vocabulary of zero types the cluster size threshold would also be zero. This section uses two metrics to evaluate ParaMors performance at identifying the suffixes of true paradigms. First, to gauge success at uncovering suffixes, ParaMors final scheme-clusters are measured for global morpheme recall, that is, the recall in any scheme-cluster of any string-unique suffix from a true paradigm. Second, to evaluate ParaMors ability to succinctly group suffixes into paradigms, this section reports the final number of clusters ParaMor identifies. It is reasonable to expect morpheme recall to drop as the size of the induction vocabulary falls simply because in a smaller vocabulary the more rare suffixes may just not occur. The more serious question for ParaMor is whether there is a vocabulary size at which ParaMors morphology induction algorithms break down; a lower limit on vocabulary size where ParaMor is unable to identify even those suffixes which do occur in the data. Similarly, in measuring the number of clusters ParaMor produces, the objective is to discover a vocabulary size at which ParaMor begins to fail at reliably grouping suffixes into paradigms. ParaMors final cluster count would reveal a failure to form paradigms if, as vocabulary size decreased, the number of clusters grew unreasonably.  REF _Ref213230715 \h  \* MERGEFORMAT Figure  4.16 plots ParaMor s paradigm identification performance over a corpus of Spanish as the vocabulary size varies down from 50,000 types to 5,000. Three separate series are plotted in  REF _Ref213230715 \h  \* MERGEFORMAT Figure  4.16. The bottom, dashed, line reports ParaMor s final cluster count at each vocabulary size, while the two upper lines report recall scores. The central, solid, line is ParaMor s global recall of all string-unique Spanish suffixes. The upper, dashed, line gives ParaMors recall of all morphemes that occurred as a word-final string on at least two types in that vocabularyit seems reasonable to allow any morphology induction system to discount as noise those morphemes that only occur once in a corpus. Note, however, that when the induction vocabulary is varied, as in this experiment, the set of morphemes that occur at least twice in the vocabulary will change. In particular, because the smaller vocabularies in this experiment are subsets of the larger vocabularies, the set of achievable morphemes can only decrease as the vocabulary size shrinks. Also,  since, the number of achievable morphemes will never be larger than the full set of morphemes in a language, achievable recall will never fall below the strict global recall. Look first at the central solid global recall curve in  REF _Ref213230715 \h  \* MERGEFORMAT Figure  4.16. While morpheme recall does fall as vocabulary size decreases, the fall is initially gradual. Reducing the vocabulary size from 50,000 types to 20,000 results in only an 8.0% absolute drop in global recall: from 67.8% to 59.8%. Reducing the vocabulary size further begins a more precipitous decline in recall. Moving from 20,000 to 10,000 vocabulary types reduces recall by 10.4% absolute, down to 49.4%; and then moving to 5,000 reduces recall by an additional 13.8% absolute, falling all the way to 35.6%. ParaMors performance at achievable recall supports the conclusion that ParaMors morphology induction algorithms are less robust at finding Spanish suffixes when the vocabulary size is below 20,000 types. From 50,000 types down to 20,000 achievable recall declines only 4.5% absolute. But then, like global recall, when provided with 10,000 unique types, achievable recall falls significantly more steeply, by 8.6%, and again by an additional 11.2% when moving to 5,000 types. In all, ParaMor identifies less than half, just 47.7%, of all true suffix strings that occurred two or more times in the corpus of 5,000 unique types. At these smaller vocabulary sizes, ParaMors algorithms are no longer able to find even those suffixes that are present in the corpus. For ParaMor to find a suffix there must be paradigmatic and syntagmatic evidence for that suffix. At a minimum, ParaMors initial search strategy cannot identify any csuffix as likely modeling a true suffix without 1) the syntagmatic evidence of the csuffix appearing attached to at least three cstems, and without 2) the paradigmatic evidence of those same three cstems all accepting some other csuffix. At vocabulary sizes below 20,000 types in Spanish, these paradigmatic and syntagmatic requirements are just not met. Turn now to ParaMors performance at grouping suffixes into paradigms. Empirically, as the vocabulary size decreases from 50,000 types, the number of clusters that ParaMor groups the selected csuffixes into first moves up from 41 to 50 and then back down to 31. But examining this empirical behavior in two halves reveals an interesting behavior: When the vocabulary size is above 20,000 types ParaMors cluster count is relatively stable between 41 and 50; Then, as suffix recall falls off at vocabulary sizes below 20,000, the number of clusters that ParaMor identifies also drops. At the smaller vocabulary sizes, the number of clusters roughly follows the number of identified suffixes. It may be that ParaMor is reasonably grouping into paradigms those suffixes that ParaMor is able to identify. Indeed, although individual clusters contain fewer csuffixes and cstems when induced from a small vocabulary, a qualitative analysis of ParaMors Spanish scheme-clusters could find no significant difference in the paradigmatic coherence of individual clusters formed from 50,000 vocabulary types and those formed from 5,000. Although paradigmatic coherence of scheme-clusters does not significantly change at smaller vocabulary sizes, there is one crucial difference between the clusters formed from large and those formed at small vocabulariesbut this difference is a failure of suffix identification not a failure of paradigm grouping. At smaller vocabulary sizes, clusters for some paradigms are simply missing. At 10,000 types there is no paradigm that corresponds to the .es paradigm that marks Number on Spanish nouns, and at 5,000 types ParaMor also finds no scheme-cluster to model those suffixes of the Spanish ir verbal paradigm that are not shared with the er paradigm. In summary then, the preceding analysis finds that, at least for Spanish, to identify a significant fraction of the suffixes of a language requires a corpus of at least 20,000 unique types. However, with some confidence, ParaMor will organize the suffixes that are identified into reasonable paradigm-like structures, even for very small vocabulary sizes. All of the analysis in this section has been over Spanish. For languages with a morphological system different than that of Spanish, it may be that a vocabulary larger than 20,000 is necessary. Section  REF _Ref217137149 \r \h  6.2 investigates the question of the necessary vocabulary size for English and German. Scheme-Clusters in Languages Beyond Spanish Thus far, this thesis has described ParaMor s performance at paradigm identification by closely examining the paradigm models that ParaMor constructs when analyzing Spanish text. But ParaMors unsupervised induction methods are intended to enable paradigm discovery in any language. Directly and quantitatively assessing the quality of ParaMors induced paradigms for a language, as has been done for Spanish throughout Chapters 3 and 4, requires compiling by hand a definitive set of the paradigms of a language. Deciding on a single set of productive inflectional paradigms can be difficult even for a language with relatively straightforward morphology such as Spanish. Appendix A describes the challenge of deciding whether Spanish pronominal clitics are inflectional paradigms. Moreover, for an agglutinative language like Turkish, the combinatorial number of potential suffix sequences makes a single list of paradigm cross-products extremely unwieldy. Hence, rather than separately define paradigm sets for each language that ParaMor analyzes in this thesis, Chapter 5 will apply ParaMors induced paradigm models to the task of word-to-morpheme segmentation. And Chapter 6 will empirically evaluate ParaMors morphological segmentations in English, German, Finnish, Turkish, and Arabic. But it is still enlightening to look at a few global statistics describing the kinds of scheme-clusters that ParaMor builds over a range of languages.  REF _Ref216410304 \h  \* MERGEFORMAT Figure  4.17 tabulates, for all six natural languages examined in this thesis, the counts of schemes (or scheme-clusters) that ParaMor constructs after each step in ParaMor s paradigm induction pipeline. The six languages in  REF _Ref216410304 \h  \* MERGEFORMAT Figure  4.17, Spanish, English, German, Finnish, Turkish, and Arabic, fall into three general categories. First, Spanish, English, and German have relatively simple morphological structure that is primarily suffixing. Of these first three languages, English has the least inflectional morphology, and German has the added complication of productive, written, compounding. The second group of languages examined for this thesis encompasses Finnish and Turkish: These two languages have rich morphological structures of agglutinative suffixes. In agglutinative languages multiple affixes (suffixes in the case of Finnish and Turkish) can occur on a single surface word form one after the next. Finally, among this set of six languages Arabic is in a class by itself. Like Finnish and Turkish, Arabic has a rich morphological structure. But Arabic words contain not only suffixes, but also prefixes and templatic morphology that interleaves sequences of vowels and consonants to form new words. The scheme and cluster counts of  REF _Ref216410304 \h  \* MERGEFORMAT Figure  4.17 reflect this three-way classification of these six languages. The counts for Spanish, English, and German are roughly similar. In particular, the final sets of scheme-clusters that ParaMor produces for these three languages are in a similar range, at 40, for English, and 48 for German. Many of schemes and scheme-clusters which contribute to the higher counts for German that occur before ParaMor s  Small Cluster Filtering and  Morpheme Boundary Filtering steps (Sections  REF _Ref190768256 \r \h  4.4.1 and  REF _Ref190771388 \r \h  4.4.2 respectively) are scheme-models of morpheme boundaries that occur between compound nouns. But noun compounding does not exhibit the same paradigmatic regularity as inflectional morphology and many of the compounding scheme-clusters are discarded during filtering. The scheme and cluster counts for the agglutinative languages Finnish and Turkish are also quite similar to one another. ParaMors processing of both languages finishes with cluster counts about double those for Spanish, English, or German, at or just above 100. Considering that Finnish and Turkish have more complex morphological structures than Spanish, English, or German, it is to be expected that ParaMor requires additional scheme-clusters to adequately describe the morphology of these two languages. Finally, ParaMors scheme-cluster counts for Arabic follow a pattern unlike that for any of the other languages. Initially, ParaMors bottom-up scheme-search procedure selects more schemes for Arabic than for any other language. But most of these schemes are discarded by ParaMors Small Cluster Filter of Section  REF _Ref190768256 \r \h  4.4.1. And ParaMor ultimately retains the fewest scheme-clusters for Arabic out of any language, just 27. The many clusters that ParaMor discards for Arabic primarily attempt to model morpheme boundaries between prefixes and stems. These erroneous prefix models place prefixes into schemes cstem sets and stems into schemes csuffix sets. While ParaMors initial search finds enough paradigmatic evidence to select these prefixational schemes, ParaMors paradigmatic clustering algorithm is unable to meaningfully group them, and consequently the vast majority of these schemes are removed by ParaMors filtering steps. This preliminary look at ParaMors performance at paradigm identification for languages beyond Spanish is preparation for their more extensive evaluation in the next two chapters. Chapter 5 will lay the groundwork, defining a methodology to segment words into constituent morphemes using the scheme-cluster models of paradigms that ParaMor produces when following the paradigm induction pipeline described in Chapters 3 and 4. Chapter 6 then reports on two evaluations of ParaMors morphological segmentations. First, a linguistically motivated metric directly measures ParaMors precision and recall at morpheme identification; and second, ParaMors morphological segmentations augment an information retrieval (IR) system. Morphological Segmentation Chapters 3 and 4 presented the steps of ParaMors paradigm discovery pipeline. This chapter and the next will apply ParaMors induced paradigms to the natural language processing task of morphological word segmentation. The job of a morphological segmentation algorithm is to break full form words at morpheme boundaries. For example, the correct segmentation of the Spanish word apoyar would be apoy+ar. This apoy+ar segmentation is correct because in the word apoyar to support, the stem apoy carries the lexical meaning of support, while the ar marks Infinitive. Crucially, however, a morphological segmentation algorithm is not required to associate morphosyntactic features with the morpheme pieces that it proposes: When segmenting the word apoyar a word-to-morpheme algorithm is not asked to notate the meaning of the stem apoy, nor must a segmentation algorithm state that ar signifies the Infinitive. While not a full morphological analysis from a linguistic perspective, morphological segmentation can nonetheless advance performance in a variety of natural language processing tasks. Oflazer and El-Kahlout (2007) improve a Turkish-English statistical machine translation system by morphologically segmenting Turkish. The morphological segmentation system that Oflazer and El-Kahlout use is hand-built. In contrast, Creutz (2006) significantly improves a Finnish speech recognition system using an unsupervised morphology induction system called Morfessor to morphologically segment the training data for the systems language model. Linguistically nave word stemming, a crude form of morphological segmentation, is also standard in information retrieval. And Section  REF _Ref217137370 \r \h  6.4.2 of this thesis discusses a simple embedding of ParaMor s unsupervised morphological segmentations into an information retrieval system, giving promising results. For two reasons, the paradigm models that ParaMor produces are well-suited to the task of word-to-morpheme segmentation. First, while ParaMor can identify morpheme segments within word forms, ParaMor is unable to provide a more complete morphosyntactic analysis of words: ParaMor does not associate morphosyntactic features with the csuffixes in each discovered cluster. ParaMor might know that the csuffix ar attaches to a class of cstems including apoy, but ParaMor does not know that apoy carries the lexical meaning of support, or even that the apoy-class of cstems models Spanish verbs; nor does ParaMor know that the ar suffix forms the Infinitive. Second, ParaMors paradigm models are best suited to morphological tasks, like segmentation, that analyze word forms, as opposed to generation tasks that propose novel surface forms. As noted in Section  REF _Ref215396639 \r \h  \* MERGEFORMAT  4.3.3, most of ParaMor s clusters contain csuffixes which do not form valid surface words with all the cstems in the cluster. These invalid (cstem, csuffix) pairs are incorrect generalizations that would hurt performance in a morphological generation task. Despite the suitability of ParaMors scheme-cluster models of paradigms to the task of morphological segmentation, ParaMors segmentation algorithm must be carefully devised so as to segment a wide range of words. Even though clustering introduces sometimes erroneous model generalization, the scheme-clusters that ParaMor produces remain highly specific. By associating a set of csuffixes with a particular set of cstems, a strict interpretation of ParaMors paradigm models would constrain ParaMors analyses only to the word types covered by each (cstem, csuffix) pair in the scheme-cluster. Section  REF _Ref216322123 \r \h  5.1 proposes a segmentation algorithm that successfully applies the specific scheme-clusters to morphologically segment the wide range of word forms that occur in a large corpus. The Segmentation Algorithm Two principles guide ParaMors approach to morphological segmentation. First, ParaMor only segments word forms when the discovered scheme-clusters hold paradigmatic evidence of a morpheme boundary. Second, ParaMors segmentation algorithm must generalize beyond the specific set of types which license individual scheme-clusters. In particular, ParaMor will be able to segment word types which did not occur in the data from which ParaMor induced its scheme-clusters. ParaMors segmentation algorithm is perhaps the most simple paradigm inspired segmentation algorithm possible that can generalize beyond the specific set of licensing types in each scheme-cluster. To segment any word, w, ParaMor examines all segmentations of w into an initial non-null stem, t, and a final non-null suffix, f, such that t.f = w. For each such stem + suffix segmentation of w, ParaMor identifies all scheme-clusters, C, that contain f as a csuffix. If there is some second csuffix, f2 , in C such that t.f2 is a word form found either in the corpus from which ParaMor induced its scheme-clusters or else found in the corpus of text that ParaMor is to segment, then ParaMor segments w between t and f. The rationale behind ParaMor s segmentation algorithm is that since f and f2 are mutually substitutable suffixes from the same paradigm, ParaMor has found evidence of a morpheme boundary. Note that the csuffix f2 need not arise from the same original scheme as f, but merely from the same scheme-cluster C. If ParaMor finds no complex analysis, then ParaMor proposes w itself as the analysis of the word. On the other hand, if ParaMor discovers more than one potential morpheme boundary in w, ParaMor accepts them allproducing a single analyzed form of w containing multiple morpheme boundaries.  REF _Ref216323373 \h Figure  5.1 gives a pseudo-code implementation of ParaMor s word-to-morpheme segmentation algorithm. A Sample of ParaMor s Word Segmentations  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 contains examples of Spanish word-to-morpheme segmentations that ParaMor produces. ParaMor segmented the word forms of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 using scheme-cluster models of paradigms that were induced over the same newswire corpus of 50,000 types used throughout Chapters 3 and 4. But the segmented words in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 come from a larger newswire corpus of 100,000 types which includes the 50,000 type corpus as a subset. Each row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 contains segmentation information about a single word. The word forms of the first thirteen rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 were hand selected to illustrate the range of analyses that ParaMor s segmentation algorithm is capable of. And then, to provide a flavor for the typical segmentations that ParaMor outputs, the forms in the last six rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 were randomly selected from the words in ParaMor s segmentation corpus of 100,000 unique word forms. Starting with the leftmost column, each row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 specifies: 1. The row number; 2. A particular Spanish word which ParaMor segmented; 3. An English gloss for that word form; 4. The word s correct morphological segmentation; 5. A full morphosyntactic analysis of the Spanish word form; 6. Each separate morpheme boundary that ParaMor matches; 7. ParaMors full segmentation of the word; and 8. The rank of one or more of ParaMors scheme-clusters which provide paradigmatic support for ParaMors segmentation of the rows word form. Elaborating on the 8th column of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2: Whenever ParaMor proposes a morpheme boundary in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 that is backed by a scheme-cluster in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4, p.114, then the rank of the  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 cluster is given in the final column of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2. In the few cases where no  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 cluster supports a morpheme boundary that is proposed in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2, then the rank of a supporting cluster appears in parentheses. Many morpheme boundaries that ParaMor proposes gain paradigmatic support from two or more scheme-clusters. For any particular morpheme boundary,  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 only lists the rank of more than one supporting cluster when each supporting cluster appears in  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. The 1st row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 contains ParaMor s segmentation of the monomorphemic word form sacerdote  priest . ParaMor s segmentation algorithm correctly analyzes sacerdote as containing no morpheme boundaries. The 2nd row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 segments sacerdotes  priests . Since both the word forms sacerdote and sacerdotes occurred in the Spanish corpora, and because ParaMor contains a scheme-cluster which contains both  of the csuffixes s and , namely the cluster from  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 with rank 1, ParaMor detects sufficient paradigmatic evidence to suggest a morpheme boundary before the final s in sacerdotes, giving sacerdote +s. ParaMor similarly correctly segments the form regulares ordinary (Plural) before the final es, using the rank 122 scheme-cluster, as regular +es; and the form chancho filthy (Masculine Singular) before the final o, drawing on the rank 3 cluster, giving chanch +o. The particular forms sacerdote, sacerdotes, regulares, and chancho of the first four rows in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 illustrate the ability of ParaMor s segmentation algorithms to correctly generalize. The forms sacerdote and sacerdotes directly contribute to the .s scheme that participates in the rank1 scheme-cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. And so to segment sacerdotes required no generalization whatsoever. On the other hand, the cstem regular does not occur in the rank 122 scheme-cluster, and the cstem chanch does not occur in the rank 3 cluster, but ParaMor was yet able to generalize from the rank 122 cluster and the rank 3 cluster to properly segment the word forms regulares and chancho. ParaMor segmented regulares because: 1.The rank 122 scheme-cluster contains the and the es c-suffixes; and 2. The word forms regular and regulares both occur in the corpus from which ParaMor learned its scheme-clusters. The occurrence of regular and regulares provides the paradigmatic evidence that ParaMor requires to suggest segmentation. ParaMors justification for segmenting chancho is similar to the reasoning behind regulares but takes generalization one step furtherthe form chancho did not occur in the corpus from which ParaMor induced paradigms, but only occurred in the larger corpus of 100,000 word types that ParaMor segments. The 5th row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 illustrates ParaMor s ability to segment a single word into more than two morphemes. The fifth row contains the Plural Feminine form of the Spanish adjective incgnito  unknown : incgnitas. The gender is marked by the a in this form, while Plural Number is marked in the final s. And so, the correct segmentation contains three morphemes and two morpheme boundaries, incgnit +a +s. ParaMors scheme-cluster models of paradigms successfully identify both morpheme boundaries in this word. But ParaMors segmentation algorithm is not always so perfect when it proposes multiple morpheme boundaries. In the segmentation of the word form agradecimos we thank, as agrade+c+imos, on row ten of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2, while the morpheme boundary before the final imos is reasonable, the character c is rightfully part of this verb s stem and should not be segmented off as a suffix. The 6th row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 gives an example of the rank 4 scheme-cluster from  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 correctly segmenting a Spanish verb; the rank 4 scheme-cluster, capturing suffixes from the Spanish verbal ar paradigm, contains more csuffixes than any other cluster that ParaMor induces over this Spanish corpus. The 7th row in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 is an example of ParaMor s failure to analyze Spanish pronominal clitics. As discussed in Section  REF _Ref191979081 \r \h  \* MERGEFORMAT  4.5.1, ParaMor s suffix-internal morpheme boundary error filter discarded the scheme-cluster which contained the majority of Spanish clitics. Where there should be a morpheme boundary before the reflexive clitic se, in accidentarse, ParaMor places none. ParaMor s correct segmentation of the adjectival verb, errados  wrong (Masculine Plural) as err +ad +o +s, on row 8 of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 contrasts with the incorrect oversegmentation of another adjectival verb, investido invested (Masculine Singular) as invest+i+d+o, on the tables 17th row. In the segmentation of investido, there should be no morpheme boundary between the characters i and d. The 9th, 10th, and 11th rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 illustrate some of the incorrect segmentations that ParaMor produces from scheme-clusters which involve morphophonemic change. The 9th and 10th rows contain the 1st Person Singular Present Indicative and the 1st Person Plural Past Indicative forms, respectively, of the Spanish verb agradecer to thank. The stem of the verb agradecer has two written surface forms: before suffixes which begin with the back vowels a and o, the stem agradezc occurs; but, before the front vowels e and i, the stem is agradec. In fact, this c/zc alternation is a feature of a reasonably sized minority of Spanish verbs, and hence, the rank 21 cluster of  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 attempts to model this stem alternation by fusing the variable stem material to the front of csuffixes. Thus, where the 1st Person Singular Present Indicative morpheme is o, ParaMor removes zco; and where the morpheme marking 1st Person Plural Past Indicative in er verbs (like agradecer) is imos, ParaMor identifies both a csuffix that matches the correct imos morpheme but also a csuffix that matches cimos, ultimately producing the oversegmented form agrade+c+imos. In ParaMors defense, there is no simple segmentation of the word form agradezco, on the 9th row of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 which contains the verb stem agradec. In contrast to the varying stems of the word forms in the 9th and 10th rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2, it is the suffix of antelacin, on the 11th row, that appears in different surface forms in different words. As discussed in Section  REF _Ref216448569 \r \h  4.5.1, the four related suffixes cin, sin, cion, and sion confound ParaMor s morpheme boundary error filters such that ParaMor is unable to remove the 5th ranked scheme-cluster which contains the csuffix in. The consequence of this failure to detect a morpheme boundary error in ParaMors scheme-cluster paradigm models is that the word antelacin is incorrectly oversegmented before the word-final string in. Whereas most of ParaMors segmentations in  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 split off inflectional suffixes, the segmentations that ParaMor gives for the word forms of the 11th and 12th rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 segment the derivational morphemes acin and ador respectively. The suffix acin forms abstract nouns from verbs, while ador is the agentive. The short word form vete, veto (3rd Person Singular Present Subjunctive) on the tables 13th row, is correctly segmented by ParaMor even though its four characters excluded it from the corpus from which ParaMor induced scheme-clusters. The final six rows of  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2 place ParaMor in the wild, giving segmentations of a small random sample of Spanish words. ParaMor successfully leaves unsegmented the Indonesian proper name bambamg and the monomorphemic Spanish word sabidura  wisdom . ParaMor correctly segments the verbal form clausurar conclude (3rd Person Singular Future Indicative) as clausur +ar; but oversegments the three forms hospital hospital, investido, invested (Adjectival Masculine Singular) and pacificamente peaceably. The oversegmentations of both investido and pacificamente occur because of errors that ParaMor makes while attempting to model legitimate morphemes: the Adjectival Masculine Singular suffix ido on er and ir verbs, and the productive adjectival suffix amente, respectively. But the double mis-segmentation of the monomorphemic hospital is conspicuous. Both of the mis-hypothesized morpheme boundaries in the word hospital, that before al and that before the final character l (el), occur because of an incorrect overgeneralization in ParaMors segmentation algorithm. The boundary before the word-final string al occurs when ParaMor incorrectly applies a cluster that contains, among other csuffixes, the adjectival csuffixes al Adjectival Singular and ales Adjectival Plural. This adjectival cluster is derived from Spanish words that include accidental accidental (Adjectival Singular) and accidentales accidental (Adjectival Plural)). Because the Spanish noun hospital happens to end in the character sequence al, and because the plural form of hospital is hospitales, ParaMors segmentation algorithm overgeneralizes to apply this adjectival cluster to segment a noun. In a similar fashion, ParaMor places a boundary before the final character l in hospital because of a cluster that contains incorrect stem-internal segmentations of these same adjectival al and ales suffixesthe incorrect cluster contains the csuffixes l and les. Morpheme Segmentation Enables Evaluation ParaMors word-to-morpheme segmentation algorithm, defined in this chapter, provides a practical way to evaluate the quality of ParaMors scheme-cluster models of natural language morphological paradigms. Cannonical morphological analyses are available for the words of a range of languages; And ParaMors morphological segmentations can be directly compared to these cannonical analyses. Moreover, in a range of natural language processing applications, it is possible to replace raw word-forms with morphological segmentations and then measure the (hopefully positive) impact of using the segmented forms. Chapter 6 will evaluate ParaMors induced scheme-cluster models of paradigms using both the cannonical analysis method as well as the task-embedding method. ParaMor and Morpho Challenge With the development, in Chapter 5, of an unsupervised morphological segmentation algorithm, the ParaMor morphology induction system can now analyze the morphology of individual words. To evaluate ParaMors word-to-morpheme segmentations, ParaMor competed in two years of the Morpho Challenge competition series. In May 2007 and again in June 2008, the Adaptive Informatics Research Center at the Helsinki University of Technology sponsored a Morpho Challenge. This series of Challenges pits against one another algorithms that, like ParaMor, are designed to discover the morphological structure of natural languages from nothing more than raw text (Kurimo, Turunen, and Varjokallio, 2008). Evaluating ParaMor through participation in the Morpho Challenge competitions permits direct comparison of ParaMors morphological analyses to the analyses produced by other state-of-the-art unsupervised morphology induction systems. The remainder of this chapter is structured as follows: Section  REF _Ref216925536 \r \h  6.1 will describe the evaluation methodology used in the 2007 and 2008 Morpho Challenge competitions. Then, using the evaluation methodology of Morpho Challenge, Sections  REF _Ref216925572 \r \h  6.2 REF _Ref216858117 \r \h  6.3 examine ParaMor s performance at morphological segmentation from two perspectives: Section  REF _Ref216925572 \r \h  6.2, an ablation study, weighs the contributions that each of ParaMor s major sub-algorithms separately make toward the final morpheme segmentations that ParaMor produces; And Section  REF _Ref216858117 \r \h  6.3 considers ParaMor s ability to identify inflectional, as opposed to derivational, morphology. Finally, Section  REF _Ref216925677 \r \h  6.4 presents ParaMor s performance in the Morpho Challenge competitions proper. Evaluation Methodology at Morpho Challenge 2007/2008 The Morpho Challenge competitions of 2007/2008 appraised participating algorithms on their morphological analyses of five languages: English, German, Finnish, Turkish, and Arabic. As the ParaMor algorithm was developed while analyzing Spanish data, the Morpho Challenge competitions will allow ParaMor to show its language independence. ParaMors induction algorithms were specifically designed around the natural organizing structure of inflectional morphology: the paradigm (see Chapters 1 and 3). Hence, ParaMor will likely be able to learn the morphological structure of languages other than Spanish. However, ParaMor has several free parameters, and it is conceivable that parameter settings that produce strong paradigm models for Spanish will yield imperfect paradigms for other languages. Nevertheless, setting ParaMors parameters anew for each separate language would void ParaMor status as an unsupervised algorithm. Hence, for all languages, ParaMor holds parameters at the settings which produce reasonable Spanish suffix sets, as determined in Chapters 3 and 4. Section  REF _Ref216925677 \r \h  6.4 will demonstrate that these parameter settings do, in fact, reasonably transfer to the languages of the Morpho Challenge competitions. For each of the five language tracks in the competition, the Morpho Challenge organizing committee provided text corpora with vocabulary sizes much larger than the 50,000 Spanish types that the ParaMor algorithms were developed over. The English corpus contains nearly 385,000 unique types; the German corpus, 1.26 million types; Finnish, 2.21 million; Turkish, 617,000; while the Arabic corpus is the smallest, with just under 143,000 unique types. To increase the likelihood that ParaMors parameter settings will transfer from Spanish to other language corpora, ParaMor keeps the size of the paradigm induction vocabulary constant at 50,000 types. As ParaMor has a choice over which types to place in the paradigm induction vocabulary, ParaMor selects, from each larger language corpus, the 50,000 most frequent word types that pass ParaMors string-length criterion (Section  REF _Ref190686436 \r \h  4.2). Once ParaMor has learned paradigm models of a language from the 50,000 most frequent types, ParaMor segments all the word types in the Morpho Challenge corpus for that language, following the methodology of Chapter 5. The Morpho Challenge competitions held in 2007 and 2008 scored each contending algorithms morphological analyses in two ways: First, a Linguistic Evaluation measured a systems morpheme identification against an answer key of morphologically analyzed word forms (Kurimo and Varjokallio, 2008; Kurimo, Creutz, and Varjokallio, 2008). And second, a Task-Based Evaluation embedded each algorithms morphological analyses in an information retrieval (IR) system (Kurimo and Turunen, 2008; Kurimo, Creutz, and Turunen, 2007). Sections  REF _Ref216926683 \r \h  6.1.1 REF _Ref216926690 \r \h  6.1.2 describe the Linguistic and Task-Based Evaluation procedures in turn. The Linguistic Evaluation of Morpho Challenge 2007/2008 While the majority of the unsupervised morphology induction systems, including ParaMor, which have participated in Morpho Challenge competitions, perform simple morphological segmentation, the Linguistic Evaluation of Morpho Challenge was not purely a word segmentation task. The Linguistic Evaluation compared each systems automatic morphological analyses against an answer key containing the full morphological analysis of each word form. Although a more challenging standard than word-to-morpheme segmentation, evaluating an unsupervised morphology induction system against full morphosyntactic analyses is less arbitrary than evaluating against an artificial segmentation standardOnly in an idealized world are morphemes consistently strung together in a purely concatenative fashion. In actual natural languages, morphophonological processes often change the surface form of both stem and affix morphemes at the time of affixation; and moreover, in a language like Arabic, morphemes are often non-concatenative to begin with, rendering the idea of a gold-standard morphological segmentation meaningless. The morphosyntactic answer keys of Morpho Challenge are in the same format as the analyses in the Morphosyntactic Analysis column of  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1. Extracted from  REF _Ref216326495 \h  \* MERGEFORMAT Figure  5.2, the rows and columns of  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1 contain morphological analyses of five Spanish words. The analyses in the Morphosyntactic Analysis column of  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1, and the analyses of words in a Morpho Challenge answer key, contain one or more lexical stems and zero or more inflectional or derivational morpheme feature markers. Both stems and feature markers are strings: feature markers have a leading  + . In addition to the Morpho Challenge style morphosyntactic analyses for each of five Spanish words,  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1 gives an English gloss of the Spanish word and the morphological segmentation that ParaMor produces for the word. As a morphosyntactic answer key, distinct surface forms of the same morpheme are marked with the same lexical stem or feature marker. For example, Spanish builds the Plural of sacerdote priest by appending an s, while Plural is marked on the Spanish form regular with es. But in both cases, a Morpho Challenge style morphosyntactic answer key would mark Plural with the same feature marker+pl in  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1. The organizing committee of Morpho Challenge has designed the linguistic answer keys for each language to contain feature markers for all and only morphosyntactic features that are overtly marked in a word form. Since Singular forms are unmarked on Spanish nouns, the Morpho Challenge style analysis of the word sacerdote in  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1 does not contain a feature marker indicating that sacerdote is Singular. Also important for the ParaMor algorithm is that the morphosyntactic answer keys for the Linguistic Evaluation of Morpho Challenge analyze not only inflectional but also derivational morphology (See Section  REF _Ref216858117 \r \h  6.3). Against each morphosyntactic answer key, the Linguistic Evaluation of Morpho Challenge assessed systems precision and recall at identifying the stems and feature markers of each word form. But to calculate these precision and recall scores, the Morpho Challenge Linguistic Evaluation must account for the fact that label names assigned to stems and to feature markers are arbitrary. In  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1, morphosyntactic analyses mark Plural Number with the space-saving feature marker +pl, but another human annotator might have preferred the more verbose +plural in fact, any unique string would suffice. Since the names of feature markers, and stems, are arbitrary, the Linguistic Evaluation of Morpho Challenge does not require a morphology analysis system to guess the particular names used in the answer key. Instead, to measure recall, the automatic Linguistic Evaluation selects a large number of word pairs such that each word pair shares a morpheme in the answer key. The fraction of these word pairs which also share a morpheme in a systems automatic analyses is the Morpho Challenge recall score for that system. Precision is measured analogously: a large number of word pairs are selected where each pair shares a morpheme in the automatically analyzed words. Out of these pairs, the number of pairs that share a morpheme in the answer key is the precision. To illustrate the scoring methodology of the Linguistic Evaluation of Morpho Challenge, consider a recall evaluation of the Spanish words in  REF _Ref216707798 \h  \* MERGEFORMAT Figure  6.1. To calculate recall, the linguistic analysis might select the pairs of words: (agradezco, agradecimos) and (sacerdotes, regulares) for sharing in the answer key the stem agradecer and the feature marker +pl, respectively. ParaMor would get recall credit for its agrade +zco and agrade +c +imos segmentations as these segmentations share the morpheme string agrade. Note here that the stem in the answer key, agradecer, is different from the stem ParaMor suggests, agrade, but ParaMor still receives recall credit. On the other hand, ParaMor would not get recall credit for the (sacerdotes, regulares) pair, as ParaMors segmentations sacerdote +s and regular +es do not contain any common pieces. It is possible for a word pair to share more than one morpheme. In this case the Linguistic Evaluation of Morpho Challenge credits a systems precision or recall with a fractional count of correctly identified morphemes. For example, suppose the morphological answer key for a language states that a word pair (w1, w2) shares two morphemes in common, but a particular morphological analysis system analyzes w1 and w2 to share just one morpheme. In this situation, the denominator of the recall calculation increments by a full word pair, while the numerator is incremented by one-half a pairfor finding one of the two morphemes. The Linguistic Evaluation of Morpho Challenge also normalizes precision and recall scores when a surface word has multiple analyses. It is not uncommon for a single word type to be ambiguous between two or more morphological analyses: consider he dances where dances is a 3rd Person Singular verb and the dances where dances is a Plural noun. Because morphology can be inherently ambiguous, the Morpho Challenge Linguistic Evaluation permits for each word: 1. Multiple morphological analyses in the answer key for a language, and 2. A morphological analysis system to propose more than one analysis. The probability with which the Morpho Challenge Linguistic Evaluation selects any particular word, w, for participation in a morpheme-sharing word pair during the calculations of precision and recall is proportional to the number of morphological analyses that w hasthe more ambiguous morphological analyses there are of w, the more pairs that w will appear in. To reduce the influence that ambiguous word forms have on precision and recall, the Linguistic Evaluation down-weights word-pairs that contain the ambiguous word w by the number of analyses that exist for w. Once a precision and a recall score has been calculated, the Morpho Challenge Linguistic Evaluation uses F1, the harmonic mean of precision and recall, as a single overall performance measure for each algorithm and each language. The official specification of the Linguistic Evaluation in Morpho Challenge appears in Kurimo, Creutz, and Varjokallio(2008). The Task-Based Evaluation of Morpho Challenge 2007/2008 The 2007/2008 Morpho Challenge competitions balanced the Linguistic Evaluation described in the previous section against a Task-Based Evaluation in which the morphological analyses of each competing morphology induction system are embedded in an information retrieval (IR) system. The IR Evaluation consists of queries over a language specific collection of newswire articles. To measure the effect that a particular morphological analysis algorithm has on newswire IR, the Task-Based Evaluation replaces all word forms in all queries and all documents with their morphological decompositions, according to that analysis algorithm. Separate IR tasks were run for English, German, and Finnish, but not Turkish or Arabic. For each language, the IR task made at least 50 queries over collections ranging in size from 55K (Finnish) to 300K (German) articles. The evaluation data included 20K or more binary relevance assessments for each language. The IR Evaluation employed the LEMUR toolkit (Ogilvie and Callan, 2002), a state-of-the-art retrieval suite; and used okapi term weighting (Robertson, 1994). To account for stopwords, terms in each run with a frequency above a threshold, 75K for Finnish, 150K for English and German, were discarded. The performance of each IR run was measured with Uninterpolated Average Precision. For additional details on the IR Evaluation of Morpho Challenge please reference Kurimo and Turunen (2008) and Kurimo, Creutz, and Turunen (2007). An Ablation Study Before delving into the official results of the Morpho Challenge competitions from 2007 and 2008, this section uses the Morpho Challenge evaluation methodology to explore the contributions toward final word segmentations that are made by each major sub-piece of ParaMors paradigm induction Algorithm. ParaMor algorithm for unsupervised paradigm induction consists of the pipelined sub-algorithms described in Chapters 3 and 4. At a high level, ParaMors paradigm induction breaks down into three major steps: The initial search for scheme-models of paradigms, as described in Chapter 3; The agglomerative scheme-clustering algorithm, detailed in Section  REF _Ref190686481 \r \h  \* MERGEFORMAT  4.3; and The filtering procedures, Sections  REF _Ref190686436 \r \h  \* MERGEFORMAT  4.2 and  REF _Ref190686503 \r \h  \* MERGEFORMAT  4.4, that were designed to both limit the creation of and to then discard unlikely paradigm models. To examine the separate impacts that the search, clustering, and filtering algorithms each have on morphological word segmentations, this section evaluates the English and German word segmentations that four configurations of these three sub-algorithms produce. All four configurations begin with ParaMors initial search: ParaMors scheme search procedure is prerequisite to both clustering and filtering. The first of the four configurations consists solely of the initial search, skipping entirely both the clustering and all filtering procedures. The second configuration clusters schemes, but does not filter out the poorer candidates. The third configuration filters out unlikely candidate paradigms, but does not cluster the initially selected schemes into coherent paradigmatic groups. And the fourth configuration applies the full suite of algorithms: first searching for candidate schemes, then clustering the candidates into consolidated paradigms, and finally filtering out the least promising clusters. A few specifics on the experimental setup are in order. To begin, although Chapter 4 presented four distinct procedures designed to filter out less desirable paradigm models, this section only measures their aggregate effect. In addition to the cluster-size filter that removes clusters with support from few licensing types (Section  REF _Ref190768256 \r \h  4.4.1), and the two filters designed to detect and discard schemes which hypothesize incorrect morpheme boundaries (Section  REF _Ref190771388 \r \h  4.4.2), this experiment counts as a filtering algorithm the technique described in Section  REF _Ref190686436 \r \h  4.2 that requires all word types in the paradigm induction corpus to consist of more than five characters. The corpus word length requirement is categorized as a filtering algorithm here because it was specifically designed to eliminate (i.e. filter) the production of schemes that result from accidental string similarities between corpus word types, see Section  REF _Ref190686436 \r \h  4.2. Also concerning experimental setup: The order in which ParaMor invokes the scheme-search, filtering, and clustering algorithms is the fixed order presented in  REF _Ref191364304 \h  \* MERGEFORMAT Figure  4.14 of Section  REF _Ref216169257 \r \h  \* MERGEFORMAT  4.5. Specifically, the filtering step of excluding short types from the paradigm induction vocabulary always occurs immediately before ParaMors scheme search; ParaMors clustering algorithm is run just following scheme search; and after clustering, ParaMor removes scheme-clusters that have support from few licensing word types, discards scheme-clusters that propose morpheme boundaries suffix-internally, and finally discards clusters whose boundaries fall stem-internally. When a specific experimental configuration omits the clustering or filtering steps, the clustering algorithm or all filtering procedures are simply skipped in the pipeline.  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 tabulates ParaMor s Morpho Challenge-style precision, recall, and F1 scores for word segmentation of English and German under each of the four search-cluster-filter configurations. After each precision, recall, or F1 value,  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 gives (in parentheses) the standard deviation for that value. The standard deviation values were obtained by calculating precision, recall, and F1 scores, after the manner of the Morpho Challenge Linguistic Evaluation, on multiple non-overlapping sets of 1000 feature-sharing word pairs. For each experimental-configuration of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 and for each language, ParaMor applied the paradigm induction pipeline to a corpus of 50,000 unique word types. With the resultant models of morphological paradigms, ParaMor then segmented the same full data used in the Morpho Challenge 2007/2008 competitions: an English corpus containing almost 385,000 unique words, and a German corpus of 1.26 million types. The top two lines of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 are the experimental configurations that exclude ParaMor s filtering steps, while the lower two rows include filtering. And the top-most row of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 gives the word-to-morpheme segmentation performance that ParaMor attains in the absence of both the clustering and all filtering procedures. In this search-only configuration ParaMors scheme search algorithm identifies paradigm models that have high recall and low precision: recall near 80% with precision in the mid-teens. That ParaMors initially selected schemes trade high morpheme boundary recall for low precision is not unexpected. In beginning separate search paths from all individual csuffixes, a bias toward high recall, at the expense of precision, was intentionally built into ParaMors search procedure (See Section  REF _Ref214085008 \r \h  3.2). Now consider the effect that scheme clustering has on ParaMor s performance at morphological segmentation. ParaMor s clustering algorithm was designed to group together all csuffixes that belong to the same paradigm. Particularly relevant for this word segmentation experiment is the fact that clustering can join pairs of csuffixes into a single paradigm that did not co-occur in any individual scheme. Recall that ParaMor segments words exactly when a pair of csuffixes that co-occur in a scheme-cluster are mutually interchangeable on some cstem from a corpus. Therefore, increasing the number of csuffixes which co-occur in paradigm models can increase the number of morpheme boundaries that ParaMor will be able to place. More frequent segmentation may translate into higher recall but may also lower precision. With ParaMors segmentation recall already quite high, it is questionable how much clustering could improve recall. Empirically, in the absence of filtering, ParaMors clustering algorithm has a negligible effect on both precision and recall. Apparently, most common pairs of csuffixes already co-occur in some scheme. In contrast to ParaMors clustering step, scheme filtering has a significant effect on ParaMors morpheme segmentation performance. The paradigm filtration steps were designed to increase the initially low precision of ParaMor s selected schemes (See Section  REF _Ref215396639 \r \h  4.3.3). ParaMor s filtering algorithms reduce the number of morpheme boundaries that ParaMor hypothesizes; And with such low initial precision, as long as a reasonable majority of the morpheme boundaries that ParaMor drops are the incorrect ones, precision will increase. ParaMors filtering algorithms will decrease the number of morpheme boundaries that ParaMor proposes in two ways. First, individual erroneous csuffixes can be entirely eliminated from ParaMors paradigm models when all schemes or scheme-clusters that contain a particular csuffix are removedif a csuffix doesnt exit, then it cant match against the tail of a word. Second, just as clustering can join pairs of csuffixes, deleting a scheme or a scheme-cluster can remove the hypothesis that a particular pair of csuffixes is in a paradigmatic relationship. And lacking paradigmatic evidence, ParaMor will be unable to propose certain morpheme boundaries. Of course, ParaMors filtration algorithms are imperfect, so it is likely that some correct schemes and some correct csuffixes will be discardeddecreasing recall. The 1st and 3rd rows of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 contain the configurations of ParaMor s paradigm induction steps that contrast in paradigm filtering in the absence of scheme clustering. In both English and German, precision rises significantly when scheme filtering is added: from 14.2% to 63.0% in English, and from 13.1% to 48.4% in German: These are improvements of 48.8% and 35.3% absolute for English and German respectively. Unfortunately, among the paradigm models that are discarded during the scheme filtering steps are a significant number of models that propose correct morpheme boundaries: recall declines from 82.8% to 47.8% in English and from 78.6% to 41.6% in German, absolute percentage falls of 35.0% and 37.0% respectively. On balance, however, after filtering, the harmonic mean of precision and recall immensely improves for both languages: F1 approximately doubles in each case to 54.3% for English and to 44.7% for German. Now compare the final two rows of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2. Both of these experimental configurations invoke both the search and the filtering routines, but the next-to-last row has no clustering while the final row does merge schemes. In English, as was the case when moving from the non-clustering to the clustering configuration without filtering (top two rows), precision falls and recall risesalthough the magnitudes of the precision and recall changes are significantly greater in the presence of ParaMors filtering routines. However, while in English precision falls and recall rises when introducing ParaMors scheme-clustering algorithm in the presence of the filtering routines, interactions between ParaMors clustering algorithm and filtering algorithm make it possible for recall to fall and precision to rise, as happens in this corpus of German. During clustering, a scheme, C, which alone is not removed by ParaMors morpheme boundary error filters, can form a cluster with schemes that ParaMor does believe mark morpheme boundaries. When ParaMors morpheme boundary error filters flag 50% or more of the schemes in a cluster as hypothesizing a morpheme boundary, the cluster is removedconsequently, in the presence of ParaMors scheme clustering, the C scheme will now be filtered from ParaMors set of paradigm models. If C was the basis of hypothesized morpheme boundaries, then these boundaries will not be proposed when ParaMors filtering algorithms are run after clustering. Hence, recall can drop and precision rise. In German, the scheme .n fills the role of the C scheme in the previous paragraph. A word-final n can mark a variety of morphosyntactic features in German including Plural Number and/or Dative Case on nouns. The stems of many, but not all, German words which inflect with a final n end in e, c.f. Auge eye (Singular) becomes Augen eye (Plural), but Fenster window (Singular Nominative) can inflect to Fenstern window (Plural Dative). Because the cstems of the .n scheme end in a sufficient variety of characters, the .n scheme is not flagged as modeling an incorrect morpheme boundary by ParaMors suffix-internal morpheme boundary error filter. However, during ParaMors scheme-clustering phase, the .n scheme is merged with several more-specialized schemes whose sets of cstems end exclusively in the character e. ParaMor flags these now-merged eschemes as incorrectly hypothesizing morpheme boundaries that lie internal to true suffixes, and the .n scheme is consequently removed. Significantly for German, the .n scheme introduces some 133,891 morpheme boundaries into the 1.26 million unique German words that ParaMor segments! More than 10% of German words end in n and alternate with a surface form that lacks the n. Thus, the choice to include or discard this single .n scheme from among ParaMors paradigm models has a significant impact on the precision and recall scores that ParaMor receives for morphological segmentation. Comparing the two experimental configurations on the bottom two rows of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2: when ParaMor filters but does not cluster schemes (thus retaining the .n scheme) recall lies at 41.6% and precision at 48.4%; but when ParaMor both clusters schemes and filters the resulting clusters (in the process losing the .n scheme), recall of course falls to 38.9% (as ParaMor is unable to correctly analyze words like Augen) but precision also rises considerably, to 54.1%. ParaMors precision significantly increases when the .n scheme is removed because in many German words a final n does not constitute a morpheme. One common error caused by the .n scheme occurs in German verbs. Consider three inflected forms of one particular verb: spielen play (Infinitive), spiele play (1st Person Present Indicative), and spielt play (3rd Person Past Indicative). The correct morphological analysis of the verb spielen treats the string spiel as the verb stem and segments off the infinitive morpheme en, i.e. spiel+en. However, the existence of the form spiele incorrectly allows the .n scheme to segment spilen as spiele +n. Overall, the significant increase in ParaMors precision leads to a slight increase in F1 for German when schemes are clustered in addition to being filtered. While scheme clustering increases F1 in German, in English, ParaMors clustering algorithm actually lowers F1 in the presence of scheme filtering, from 54.3% down to 52.7%. The scheme-clustering procedure manages to raise English recall only slightly, from 47.8% to 48.8%, while significantly lowering precision through the introduction of many incorrect overgeneralized morpheme boundaries. Precision falls from 63.0% to 57.2% As the ultimate effect of ParaMors scheme-clustering algorithm on F1 can differ from language to language, it is difficult to discern whether scheme clustering is fully appropriate for the word segmentation task. However, Section  REF _Ref215396639 \r \h  4.3.3 demonstrated that scheme clustering significantly reduces the number of separate partial models that ParaMor produces. Because of this clear reduction in paradigm fragmentation, ParaMor s word segmentation submissions to the Morpho Challenge competition (Section  REF _Ref216925677 \r \h  6.4) are produced when including the scheme-clustering algorithm in ParaMor s paradigm induction pipeline. Limiting the Vocabulary As a second look at the individual contributions that ParaMor s major sub-algorithms make toward morphological segmentation, consider ParaMors performance when inducing paradigm models from a much smaller corpus. Where all other word-to-morpheme segmentation experiments in this thesis run ParaMors paradigm induction algorithms over corpora of 50,000 unique word types, the results in  REF _Ref216591034 \h  \* MERGEFORMAT Figure  6.3 were obtained from a corpus of 20,000 types. Like  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2,  REF _Ref216591034 \h  \* MERGEFORMAT Figure  6.3 reports precision, recall, and F1 scores (and the value of one standard deviation in parentheses) for morphological segmentations of English and German for four configurations of ParaMors scheme-search, clustering, and filtering algorithms. The smaller corpus size of 20,000 unique types was chosen because the experiments in Section  REF _Ref215980624 \r \h  4.5.2 demonstrated that, at least in Spanish, the quality of ParaMor s induced paradigm models begins to quickly degrade below a vocabulary size of 20,000. Before turning to the experimental results, it is important to keep in mind that both this experiment over corpora of 20,000 types and that of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 over 50,000 types, as well as all other word segmentation experiments reported in this thesis, segment words over corpora that are much larger than the corpora from which paradigms are initially learned. The large size of the segmentation corpus is important because the more unique word types that are present in the segmentation corpus the more likely that a particular lexeme will occur in more than one inflected form, providing the paradigmatic evidence that ParaMor requires to propose a morpheme boundary. By holding the segmentation corpus fixed, this experiment evaluates the quality of ParaMors induced paradigms, as applied to the word-segmentation task, when the size of the paradigm induction vocabulary is reduced. Overall, ParaMors word segmentation algorithm is remarkably resilient to a reduction in the size of the paradigm induction corpus. In particular, the F1 scores for the full ParaMor algorithm of search, clustering, and filtering at a vocabulary of 20,000 types ( REF _Ref216591034 \h  \* MERGEFORMAT Figure  6.3) are within two standard deviations of the F1 scores for paradigms trained over 50,000 types ( REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2). In German, F1 is slightly lower at 44.7% vs. 45.2%; And in English, F1 is actually higher when paradigms are learned from the smaller vocabulary of 20,000 types, 53.8% vs. 52.7%. This increase in English F1 at word segmentation reflects small increases in both precision (from 57.2% to 57.6%) and in recall (from 48.8% to 50.6%) when learning paradigms from a smaller corpus. At first blush, increases in recall, let alone precision and F1, are counterintuitive for paradigm induction over a smaller data set. It would seem that ParaMor should find fewer total csuffixes from a smaller corpus, and that this smaller set of csuffixes should propose fewer morpheme boundaries. If this smaller set of discovered csuffixes is more focused, then precision might increase, but surely recall should go down. Indeed, a pattern of lower recall scores at smaller vocabulary sizes does occur among the experimental configurations that do not include ParaMors filtering algorithms: Recall is consistently at least 5% absolute higher for both English and German in the top two rows of  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2 than in the top two rows of  REF _Ref216591034 \h  \* MERGEFORMAT Figure  6.3. However, in three of the four the experiments that include ParaMor s filtering algorithms (the bottom two rows of Figures 6.2 and 6.3), the morpheme recall is higher when vocabulary size is smaller. These recall statistics suggest that ParaMor s filtering algorithms are behaving in a less aggressive fashion at lower vocabulary sizes. As discussed in Section  REF _Ref215980624 \r \h  4.5.2, ParaMor has one parameter that is not invariant to vocabulary size. The filtering algorithm that discards scheme-clusters which are not licensed by a sufficient number of word forms (Section  REF _Ref190768256 \r \h  4.4.1) employs a threshold that must be adjusted with the size of the vocabulary. To compensate for the change in vocabulary size, the experiments in this section set the cluster-size threshold using the same procedure as the experiments in Section  REF _Ref215980624 \r \h  4.5.2: the cluster-size threshold is linearly scaled with the vocabulary. In the Spanish experiments of Section  REF _Ref215980624 \r \h  4.5.2, the linear adjustment of the cluster-size threshold was sufficient to ensure that the number of unique correct csuffixes discovered by ParaMor should decrease with vocabulary size. But empirically, a linear scaling of the cluster-size filtering threshold does not prevent the inventory of discovered csuffixes in English and German from increasing at lower vocabulary sizes. And hence, recall increases. Now look beyond the morphology segmentation scores that the full ParaMor algorithm achieves in the small-vocabulary scenario of  REF _Ref216591034 \h  \* MERGEFORMAT Figure  6.3 to the experimental configurations that omit the scheme clustering or filtering steps, or both. The segmentation scores of these other experimental configurations reveal a pattern broadly similar to that of the larger vocabulary experiments from  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2: When filtering is omitted, ParaMor s segmentation recall is high, precision low, and clustering has little effect. And again, as when learning from the larger vocabulary, invoking ParaMor s filtering algorithms brings recall and precision into closer balance. At the smaller vocabulary size of 20,000 words types, ParaMor attains a maximum F1 score, for both English and German, when ParaMors full suite of search, clustering, and filtering algorithms is applied. Inflectional vs. Derivational Morphology As mentioned in Section  REF _Ref216926683 \r \h  6.1.1 the Linguistic Evaluation of Morpho Challenge explicitly requires analyzing both inflectional and derivational morphology. But ParaMor is designed to discover paradigms the organizational structure of inflectional morphology. The experiment of  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 makes concrete ParaMor s relative strength at identifying inflectional morphology and relative weakness at analyzing derivational morphology.  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 contains Morpho Challenge style linguistic evaluations of English and German. For English and German, the official answer keys used in the 2007 and 2008 Morpho Challenge competitions were created from the widely available Celex morphological database (Burnage, 1990). To create the official Morpho Challenge answer keys, the Morpho Challenge organization extracted from Celex both the inflectional and the derivational structure of word forms. For the experiment in  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4, I constructed from Celex two Morpho Challenge style answer keys for English and two for German. First, because the Morpho Challenge organization did not release their official answer key, I constructed, for each language, an answer key very similar to the official Morpho Challenge answer keys where each word form is analyzed for both inflectional and derivational morphology. Second, I constructed, from Celex, answer keys for both English and Germen which contain analyses of only inflectional morphology. From the 50,000 most frequent types in the Morpho Challenge English and German data, ParaMor built scheme-cluster models of paradigms. And then  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 evaluates ParaMor s morphological segmentations against both the answer key which contains only inflectional morphology and against the answer key which contains inflectional and derivational morphology. As described in Section  REF _Ref216955967 \r \h  6.2, a minor modification to the Morpho Challenge scoring script allowed the calculation of the standard deviation of F1, reported in the  column of  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4.  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 reveals that ParaMor attains remarkably high recall of inflectional morphemes for both English, at 70.1%, and for German, at 66.5%. When evaluated against analyses which include both inflectional and derivational morphemes, ParaMor s morpheme recall is about 30 percentage points lower absolute, English: 34.0% and German: 37.1%. In addition to the evaluations of ParaMor s segmentations,  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 evaluates segmentations produced by a morphology analysis system called Morfessor Categories-MAP v0.9.2 (Creutz, 2006), a state-of-the-art minimally supervised morphology induction algorithm that has no bias toward identifying inflectional morphology. To obtain Morfessors segmentations of the English and German Morpho Challenge data used in this experiment, I downloaded the freely available Morfessor program and ran Morfessor over the data myself. Morfessor has a single free parameter. To make for stiff competition,  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4 reports results for Morfessor at that parameter setting which maximized F1 in each separate evaluation scenario. Morfessor s unsupervised morphology induction algorithms, described briefly in Chapter 2, are quite different from ParaMor s. While ParaMor focuses on identifying productive paradigms of usually inflectional suffixes, Morfessor is designed to identify agglutinative sequences of morphemes. Looking at  REF _Ref216952908 \h  \* MERGEFORMAT Figure  6.4, Morfessor s strength emerges as the accurate identification of morphemes: both inflectional and derivational. In English and in German, Morfessors precision against the answer key that contains both inflectional and derivational morphology is significantly higher than ParaMors. And, as compared with ParaMor, a significant portion of the morphemes that Morfessor identifies are derivational. Morfessors relative strength at identifying derivational morphemes is particularly clear in German. Against the German answer key of inflectional and derivational morphology, Morfessors precision is higher than ParaMors; but ParaMor has a precision comparable to Morfessors when identifying just inflectional morphemesindicating that many of the morphemes Morfessor correctly identifies are derivational. Similarly, while both ParaMor and Morfessor score lower at recall when required to identify derivational morphology in addition to inflectional; Morfessors recall falls much less than ParaMorsindicating that many of Morfessors suggested segmentations which were dragging down precision against the inflection-only answer key were actually modeling valid derivational morphemes. Clearly, the ParaMor and Morfessor morphology induction systems focus on very different areas of morphology. These two systems complementary nature suggests pooling their morphological segmentations. Hence, in addition to the ParaMor algorithm in isolation, a joint ParaMor-Morfessor system was submitted to the Morpho Challenge competitions described in the next section, Section  REF _Ref216925677 \r \h  6.4. Morpho Challenge 2007/2008 Where the ablation study of Section  REF _Ref217017701 \r \h  6.2 clarified the contributions of the sub-components of the ParaMor algorithm toward morphological segmentation, and Section  REF _Ref216858117 \r \h  6.3 examined ParaMor's relative performance at learning inflectional and derivational morphological structure, this section focuses on the full ParaMor algorithm under the strict requirements of the Morpho Challenge comp $(<?KLQRd= > º¶xlZ#jh8L56U\mHnHuh<hr&6CJaJ+jh<hr&6CJUaJmHnHuh<hd6CJaJh<hd6 hd6jh6UmHnHuhdhdCJ$aJ$hdCJaJhd:>*CJ4NHaJ4hd:>*CJ4aJ4hd:CJaJhd:CJ4NHaJ4hd:CJ4aJ4 $:QRd  9 : ; !H!H!H!H!!!!!!!!!!!!!$ H d^ `a$ $`a$ $dx`a$ $8`a$d`$ Pdx`a$$dxh^`a$u z 8{ M; < = t x `uvGHlF!!!!!! ! ! !!3 & Fgd< $`a$gd<$`a$$ H d^ `a$ b c        c d k l   q r w vw{|+,#ƻƻhh hdNHhdh8LB*KHphh<hdKHNHh<hdKHh<h<6h<hd56\)jh<hr&56U\mHnHu#jh<56U\mHnHu#jhl56U\mHnHu3# `a¹Ÿ”tй¹ j}h8LUmHnHujh8LUmHnHuh8LmHnHu2jhhSh8L>*B*UmHnHphuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHujhdU hdH* hd0J hdNHhd(%&'ABCDEFGHIefghxyzźӑőwźfӑő jqh8LUmHnHu2jhhSh8L>*B*UmHnHphuh8LmHnHu jwh8LUmHnHujh8LUmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu    %&'ABCEFGHIJfgźӑőwőźfӑő jeh8LUmHnHu2jhhSh8L>*B*UmHnHphuh8LmHnHu jkh8LUmHnHujh8LUmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu"ghilmżűӼżwżűfӼż jYh8LUmHnHu2jhhSh8L>*B*UmHnHphu j_h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$ !$%IJKefgijklmnżűӼżwżűfӼż jMh8LUmHnHu2jhhSh8L>*B*UmHnHphu jSh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$#$%?@ACDEFGHdefgjkżűӼżwżűfӼż jA h8LUmHnHu2j hhSh8L>*B*UmHnHphu jG h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$FmvwL eB*la T!!"" ### !"'(JKLfghjklmnożűӼżwżűfӼż j5 h8LUmHnHu2j hhSh8L>*B*UmHnHphu j; h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j hhSh8L>*B*UmHnHphu$STUopqstuvwxżűӼżwżűfӼż j)h8LUmHnHu2j hhSh8L>*B*UmHnHphu j/ h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j hhSh8L>*B*UmHnHphu$TUVpqrtuvwxyżűӼżwżűfӼż jh8LUmHnHu2jhhSh8L>*B*UmHnHphu j#h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$)*+EFGIJKLMNjkżűӼżwżűfӼż jh8LUmHnHu2jhhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$klmpq     *+żűӼżwżűfӼż jh8LUmHnHu2jhhSh8L>*B*UmHnHphu j h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$+,-23BCD^_`bcdefgżűӼżwżűfӼż jh8LUmHnHu2j|hhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jhhSh8L>*B*UmHnHphu$ !;<=?@ABCD`abchiżűӼżwżűfӼż jh8LUmHnHu2jphhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jvhhSh8L>*B*UmHnHphu$ #$%'()*+,HIJKPQtuvżűӼżwżűfӼż jh8LUmHnHu2jdhhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jjhhSh8L>*B*UmHnHphu$IJKefgijklmnżűӼżwżűfӼż jh8LUmHnHu2jXhhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j^hhSh8L>*B*UmHnHphu$   > ? @ Z [ \ ^ _ ` a b c  żűӼżwżűfӼż jh8LUmHnHu2jLhhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jRhhSh8L>*B*UmHnHphu$ !!!! ! !1!2!3!M!N!O!Q!R!S!T!U!V!r!s!żűӼżwżűfӼż jh8LUmHnHu2j@hhSh8L>*B*UmHnHphu jh8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jFhhSh8L>*B*UmHnHphu$s!t!u!x!y!!!!!!!!!!!!!!!!!!!!!!""""""""";"<"żűӼżwżűfӼż j!h8LUmHnHu2j4!hhSh8L>*B*UmHnHphu j h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j: hhSh8L>*B*UmHnHphu$<"=">"C"D"y"z"{"""""""""""""""""""### # # # ###+#,#żűӼżwżűfӼż j#h8LUmHnHu2j(#hhSh8L>*B*UmHnHphu j"h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j."hhSh8L>*B*UmHnHphu$,#-#.#3#4#_#`#a#{#|#}#########################$$żűӼżwżűfӼż j%h8LUmHnHu2j%hhSh8L>*B*UmHnHphu j$h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j"$hhSh8L>*B*UmHnHphu$$$ $$$1$2$3$M$N$O$R$S$T$U$V$W$s$t$u$v${$|$$$$$$$$$$$$$$$żűӼżwżűfӼż j'h8LUmHnHu2j'hhSh8L>*B*UmHnHphu j&h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j&hhSh8L>*B*UmHnHphu$#U$$%%&&&V''8((!))%***S++],,6-- .o..T//00$$$$$$$$%%%%%%%%%;%<%=%>%C%D%%%%%%%%%%%%%%%żűӼżwżűfӼż j)h8LUmHnHu2j)hhSh8L>*B*UmHnHphu j(h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j (hhSh8L>*B*UmHnHphu$%%%%%%%% &&&&&&&&&3&4&5&6&9&:&e&f&g&&&&&&&&&&&&żűӼżwżűfӼż ju+h8LUmHnHu2j*hhSh8L>*B*UmHnHphu j{*h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j)hhSh8L>*B*UmHnHphu$&&&&&&&&&&&&&&&&&''''''2'3'4'N'O'P'S'T'U'V'W'X't'u'żűӼżwżűfӼż ji-h8LUmHnHu2j,hhSh8L>*B*UmHnHphu jo,h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j+hhSh8L>*B*UmHnHphu$u'v'w'z'{'''''''''''''''''''(((0(1(2(5(6(7(8(9(żűӼżwggVӼ j]/h8LUmHnHuhhSh8L0JmHnHsHu2j.hhSh8L>*B*UmHnHphu jc.h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j-hhSh8L>*B*UmHnHphu!9(:(V(W(X(Y(c(d(((((((((((((((((((((())))) )!)")#)?)@)λ񰡰v񰡰e jQ1h8LUmHnHu2j0hhSh8L>*B*UmHnHphu jW0h8LUmHnHujh8LUmHnHuh8LmHnHu$jhhSh8L0JUmHnHu2j/hhSh8L>*B*UmHnHphuh8LmHnHuhhSh8L0JmHnHu'@)A)B)G)H)))))))))))))))))))******"*#*$*%*&*'*C*D*żűӼżwżűfӼż jE3h8LUmHnHu2j2hhSh8L>*B*UmHnHphu jK2h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j1hhSh8L>*B*UmHnHphu$D*E*F*I*J*[*\*]*w*x*y*|*}*~**********************++żűӼżwżűfӼż j95h8LUmHnHu2j4hhSh8L>*B*UmHnHphu j?4h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j3hhSh8L>*B*UmHnHphu$+++++/+0+1+K+L+M+P+Q+R+S+T+U+q+r+s+t+y+z+++++++++++++++żűӼżwżűfӼż j-7h8LUmHnHu2j6hhSh8L>*B*UmHnHphu j36h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j5hhSh8L>*B*UmHnHphu$+++,,9,:,;,U,V,W,Z,[,\,],^,_,{,|,},~,,,,,,,,,,,,,,,,,żűӼżwżűfӼż j!9h8LUmHnHu2j8hhSh8L>*B*UmHnHphu j'8h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j7hhSh8L>*B*UmHnHphu$,,,,,---.-/-0-3-4-5-6-7-8-T-U-V-W-Z-[---------------żűӼżwżűfӼż j;h8LUmHnHu2j:hhSh8L>*B*UmHnHphu j:h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j9hhSh8L>*B*UmHnHphu$--------.... . . . . .).*.+.,.1.2.K.L.M.g.h.i.l.m.n.o.p.q...żűӼżwżűfӼż j =h8LUmHnHu2j<hhSh8L>*B*UmHnHphu j<h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j;hhSh8L>*B*UmHnHphu$.......................0/1/2/L/M/N/Q/R/S/T/U/V/r/s/żűӼżwżűfӼż j>h8LUmHnHu2j>hhSh8L>*B*UmHnHphu j>h8LUmHnHujh8LUmHnHuh8LmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2j=hhSh8L>*B*UmHnHphu$s/t/u////////////////////0000000008090źӑőwźfӑő j@h8LUmHnHu2jt@hhSh8L>*B*UmHnHphuh8LmHnHu j?h8LUmHnHujh8LUmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jz?hhSh8L>*B*UmHnHphu 90:0;0v0w0x000000000000000 11-11122222233/333334b4c44źӑ{uuumuii{hh'`hmHsH h'`hNH hdNHh'`hhdjhdUh8LmHnHu jAh8LUmHnHujh8LUmHnHuh8LmHnHuhhSh8L0JmHnHu$jhhSh8L0JUmHnHu2jnAhhSh8L>*B*UmHnHphu)000/35J8L8d889:;Z<<!=|==2>>>=?!!!!!!!!!!!!!!!!!! !gdG !gdk !gd{/ !gdHj !gdu !gdIgdgd'`h & Fgd ;4455555:6;6666D7E7T7t7>8@8X8Z8888888888889F9b999999999999: :::::H:N:Z:\:hd0J>NH hd0J>jBhdUhkhuhIh{/hkchkcmHnHuhkchkcmHnHu hkchkcjhBhdUjhdUh h'`hNHh'`hhd hdNH6\:d:::::::; ;";0;2;8;:;;;;;7<V<Y<<<<<<== =,=]=x={=|============'>.>1>2>x>>>>>>>/?9?@=???`@@7AAB]BB(CCCMDD EdEEFsFF8GGGGH!!!!!!!!!!!!!!!!!!!!!!!! !gdk !gd:V !gd  !gdG???@@$@\@_@f@@@@@@@@3A5A6AxAAAAAAABBBB9B;BCBEBOBTBVBXB[B\B]BlBBBBBBBBBBB#C&C'CUCCCCCCCh:Vh7&huh#ihbh 0J>h hkhHjM_EcEEEE FFFeFnFpFrFFFFFG3G7G}GGGGGGGGG&HFH{HHHHHHHHHHHHHHH=I>IIIeIIIIIIIITJWJJJJJK KdKgKKKKKKܾܵh8Ch8Ch8C0J> hUNHhX hoNH hA6NHhohA6hUhX h7h3-Ah$h7&hHjh,hkFGHHHJIIIXJJ KhKK*LLL9MMMfNiNvN RRR!!!!!!!!!!!!!!!!!! !!! & F & F4gd< !gd8C !gdX  !gdo !gdU !gdX KKKKKK&L)LDLELILJLKLLLLLLLLLLLMM2M3MaMnMtMyMMMMfNiNmNnNNNO OzO{OOO2P3PPPPPQQQQRRRR|S}SSSSSTT/U0U^U_UUUU̽h hd6 hdNHhdhkhk0JQhkh8C0JQhkh8Ch8C0J> h8C0J> h8Ch8Ch8CJRR'SV$VwZ]y_ipkx{csZ„%ZQ!! !! ! !!! !! ! !!!!!!!!!!!!x & Fx^`/ & FUUqVrVVV&W'W2W>WWWWWXX+X,X1X4X5X;X hd6hd hdNHU{\|\\\\%]&]E]O]P]U]v]x]~]]]]@^A^t^^^^__q_r_x_y_z___8`:```.a0a\a^a~aaaaaaaaaaXbZbbbbbbbc*cZcϼ hd0J@hkchkc\mHnHuhkchkc\mHnHuhkchkc\jChdUjhdU hd0J hd0J> hd6 hdNHhd hd0J?hd0J?NHg@ggghhih̾̾jDhdU hd0J> hd0J? hdNHhkchkc\mHnHuhkchkc\mHnHuhkchkc\j\DhdUjhdUhd hd0J@?ihrhthhh&i'i=i>iNiOijjj jRjTjjjjjjjk*k@kTknkvkkkkkkkkkkk\l^llllllllll m mmƸƸƸhd0J?NHjPFhdUjEhdU hd0J@ hd0J?hd0J>NHhkchkc\mHnHuhkchkc\mHnHuhkchkc\jVEhdUjhdU hdNH hd0J>hd4m,m4mRmbm|mmmmmmmn$nlnnnnn o o,o.ooDoFovoxo ppvpwppppp&q1qBqRqoquqzqqqqqqqqqqqqqqqqr r6rĵhd0J>NH hd6hkchkc\mHnHuhkchkc\mHnHuhkchkc\jFhdUjhdU hdNH hd] hd0J> hd0J? hd0J@hd<6r8rXrZrhrjrprrrvrrrrrrrRs^srsvsxsssssssssttttuu5u6uFuGuvvv vNvPvv𽵽蝏}r蝏}jDHhdU hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjGhdU hd0J?hd0J>NH hd0J>hkchkc\mHnHuhkchkc\mHnHuhkchkc\jhdUhdjJGhdU,vv*x+xXxYxkxlxxxyyyypzqzzz/{0{{{{{&|'|||||}}~~~ ~~"~~~~~~~~~,.RT^`OPRhd0J@NH hd6 hd0J> hd0J@hkchkcmHnHuhkchkcmHnHu hkchkcjHhdUjhdUjhdUmHnHuhd hdNH=RTVW]_߀CFHLNQSWY]_acd]^OPу҃qr„ńpqDEdhux{|͆Ά}~҇Ӈ~IJBDdfj>IhdUjhdU hd6 hd>*h#i hdNHhd hd0J>Oftv|~X` 4;=FHIJQS]_dfnwx H^h~ؓړӼӼӼӼӼӼӼӼӼӼөӼӼӞjJhdUj8JhdUhd0J>NH hd0J>jIhdU hdNHhdjhdUhkchkcmHnHuhkchkcmHnHu hkchkc>(*8:@B{|;< !yz&'%& Tb&4hv~ĝƞj *hdUj2KhdU hd0J>j *hdUmHnHu hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhd?"$X\`bnpz|$(9;PQԠ'(gFo89?jݤޤ45CDťƥABEF01ƿܻܻܻܶܶ hdH*h99 hdNH\ hd\ h99\ hdNH hd0J>hdhkc hkcj *hdUjK *hdUHJTH&\ X~a '!! !!!!!!! ! !!!!! ! !!!!! !!gd(9 & F   gd3-Aب٨ԩթ LMQRBCHN <>@BHJıƱ8:lnpj&MhdUjhdUjL *hdUhkc hkcj,L *hdUj *hdUh#ih3-Ajhd0JU hdNHhd@prxz"$ XZ ιйڹܹ prֻ$& ꕢjO *hdUj *hdUh&-jN *h4haBNUh4j *h4Uj NhdUjMhdU hdNHhdjhdUhkc hkc:>~¿Ŀƿпҿ  pq#$=>N LNnpr|~PRjQhdUjQhdUjPhdUh'-jPhdUjhdU hdNHhkc hkcjO *hdUj *hdUh&-hd<Z\=> FG#$Z\RSijCD=>BC"LhYe hdH*jRhdU hdNHjRhdUhdhkc hkcjhdUNLMlm/efklgh TV^_efSTJK,-VW_`jrUhdEHUjShdEHUjhdUhkc hkcjS *hdUj *hdUhYe hdNHhdhdmHsHD2402RTbdjlfg>DEHm!"qrqr{|󬥬jhdUmHnHuhdmHsH hd6NH hd6 hdH*jWhdU hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjeWhdUhdjhdU8.& M|   B F  |x!!!!! ! !!!!!!!!!!!!!! !!!3.gd(9 & F ]gd<$gdi F^  !WX$&Rڶڮڢڮږڊj|dhdEHPJUjxbhdEHPJUj`hdEHPJUhdNHPJjz^hdEHPJUjp\hdEHPJUjeZhdEHPJU hdPJj[XhdEHPJUhi hdNHhd hd66RT|~z{@Avwz{ DE/0rsab& '    : H J | μβj *hdUh&-h#i hdNHh<$hdjhhdEHPJUjfhdEHPJUhdNHPJ hdPJ hd6PJC| ~           " $ * ,   ? @ \ ]           B C h i y z   ܸܸܸܿܿܮܿܤܚܐܿ܆yjZy *hdUjvhdEHUj{thdEHUjqhdEHUjEnhdEHUjkhdEHU hd6] hdNHjxkhaBNUjh4Uh4hdhkc hkcj *hdUjj *hdU/,.02<>BD~  >@BDNP`bRTDF,.`bdfjH| *hdU hd6]j{ *hdUjN{ *hdUjz *hdUjTz *hdU hdNHhkc hkcjy *hdUj *hdUhd8flnxz op 9: VXfh:>j` a z {   !!-!.!!!!!!!j<~hdEHU hd6]hYej} *hdUjB} *hdU hkcj| *hdU hdNHhdj *hdUhkcB/XH!$G&(]-h../B/23655F667799!!!! !! !!!!!! !!!!!!!!!! & F gdf1 & F gdf & F gd,YgdYegdYe  h^h`!!!1"2"h"i"""""""##$$`$a$$$2%3%%%%%9&:&&&&&&&/'^'@(A(((((()M)N)j))))I*J*++++++B,C,[,\,,,U-V------..jhdEHU hd] hd6NH hd6hdmHsHh,YjhdEHU hd6] hdNHhYehdI.. .....*.+.8.9.>.?.`.a.u.v.z.{.........////B/c///J0K0V0W01 1]1^1112333333333l4m44455555566<6=6E6F6666697:7 hdH*hih,YjVhdEHUhfjMhdEHU hdNHhd5>*\ hd6]hdL:777H8I888r9s9|9}999`:a:;;];^;`;a;;;<<<<====> >p>q>>>>>u?v????@@2@3@Y@Z@@@@@A)A*A/A5A;AUAsA{A|AAABBNBOBBBBBBBC'C|C}CCCbDcDEEE hd0J> h,Y] hd] hd6hGUh,Y hdNHhdV9;<<=Y==?@DEI MfQ8SV_ `{bcd?def i!!!!!!!! !! ! ! !! !!!!!!!!!!  h^h` gdNPgd,Y & FEEF,FNFOFFFGG}G~GGG HHH-HRHSHHH1I2IZI[IIIJJNKOKKK(L)LTLULLLMM)M-MBMIMkMlMNNOO>P?PPPVQWQQQdReRRRRR%S&SSSSSTT U UUUVV[W\WWWX he\he\ he\hd hd6] hd]hd hdNHTXXHXJXLXNXXXZXXXYY-Z.ZZZZZ [ [%[&[\\\ \H\J\%^&^{^|^^^__(_)_-_/_6_7_______ a aaaebfbtbubbbbc0cOcUcVcXc]ckclczc}ccc hd6]hNPh,YjljhdUjhdU hdNHhkc hkcjJ *hdUhdj *hdUGcddddZe[eeeffff5g6ggggg]h^hhhii_j`jjjjjkkxlylllmmFmGmnnlnnnnn:ohdUjhdUh3-A hd6] hdNHhdhNPNЛћ.024:<8: ,.`bdflnxzpq|ʡˡDE  st£ãuvUcgm hdEH hd0J> hd6] hd6j6 *hdUhkc hkcj *hdUj *hdUhd hdNHEžҞ¢Ǧ 2!¹A b!!!!!!!!!!!!!!! !!!!!!!!gd(9gd'q 1gd3" & Fgdfh¥ߥBCmnostuǦȦ}vldjhdUjhdEHU hfhfjĘhdEHUhf hd6]jhdEHUjhdEHUjhdEHUhdmHnHujhdEHUhd56>* hd6hd5>*CJaJjhdEHU hdNHjhdEHUhd hd5>*%|ҨӨۨIͫѫ .12ԶԱԧԡԗԱԍˁˁˁˁˁˁvo hd6]hd]mHnHuhd6]mHnHuj hdEHUj hdEHU hdNHjhdEHU hd6jhdEHUhd6mHnHuhdmHnHuhdjhdU hdEHjchdEHUjCL hdCJUVaJ+2HJLNXZ "TVXZ`b\^fhlnz|úúukajhdEHUjhdEHUjhdEHUjhdEHUjhdEHUjhdEHUhd6]mHnHu juhdUmHnHuhkcmHnHuhkcmHnHu jhdUmHnHujhdUmHnHuhdmHnHu hdNHhd#ȱʱرڱ45~xy 24(*JLZ\bdζж ʾʾ񝋃x񝋃jhdUhkchkc hkchkcj.hdUjhdU hdNH$jhP0J+6U]mHnHuhd6]mHnHuhdmHnHujhdEHUj-hdEHUj,hdEHUhdj=hdEHU- .0>@FHxܷ !"̸͸ "24NO&'.dekmºغݺӱӬӬӱӱӞӬ hd6\ hdNH] hd] hd\ hdNHjhdUmHnHuhc$^hd0J) hd0J>hkchkc hkchkcjhdUhdj$hdU>ݺ9>|}ٻܻ}¼üļʼ)-.bc߽UVdmnrsxz{ſſŸŸŸŸŸŸ hd6] hdNHhd hd5\ hd6\ hdNH\ hd\ hd6 h'q ] hd0J> hdNH] hd]HUVʿ˿68jl#  *129ƶ񬞬񬞬 hd0J@hd0J?NH hd0J> hd0J?hkc hkcjhdUjhdUjhd0J+Ujhd0J<U hd6] hdNHhdjhdEHU<9kqtu 4>JW_ejp RS[^stjhd0JU hd5 hd6 hd0J?h(9hkchkc\mHnHuhkchkc\mHnHuhkchkc\j<hdUjhdU hd0J> hdNH hd0J@hd:xybcde!'(AMXY .0>@FH*,9Apt")afhntz:<hkchkc hkchkcjhdUjhdUhd0J>NH hdNHhd hd0J>LbA R hr #!! ! ! ! ! !!!!!!!!!! !!!!! ! !gdaBNgdjgd4gd\~3gd\~  gdB6gdB6 .0Z\|~$&46<>~hj~jhdUmHnHu hdNH hd0J hd0J7jhdUj*hdUhkchkcjhdUhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdj4hdU0z|QRopx89NO_`   68X˿볮˿j6hdUjhdUj+hdUjhdUmHnHu hd\ hd0J> hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdUj hdU3XZhjprz|fhx.024>@lnҾҳҮҨҞҨhkc hkcj, *h4haBNUh4j *h4U hdNH hd5jhdUjhdUmHnHu hd0J>hdhkchkcmHnHuhkchkcmHnHu hkchkcjhdU/ .068<>VXLN}~ LNPR\蓏蓏jhdUhkc hkcjhdUjhdUhkchkcj$hdU hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdU4\^pr^`46bd\^  "XYno ln hd0J>hkchkc0J5aJj hdU hd0J hd0J7jhdUhkchkcmHnHuhkchkcmHnHu hkchkcjhdU hdNHhdjhdU4(,8<HT`;<ABJYdtUV9DHPmn+,.0 hd5he\jhB6UmHnHuhB60J75aJhd0J75NHaJhd0J7aJ hd0J>hB6 hdNH hd0J@hdhd0J75aJB069OP <HR^(4HLTVtLNRVZ\hlpǹhd0J@NH hd6 hd0J? hd0J@hkchkcmHnHuhkchkcmHnHu hkchkcjhdUjhdUhB6hdCJOJQJaJ hdNH hd0J>hd8prX^ciktz+LY]dy/6MNz{}˿ hd5 hd0J> hd0J@jhd0JU hdNHhkchkcmHnHuhkchkcmHnHu hkchkc *hdj *hdUhdj *hdU9    "z|"@A  "028:dfxzⶨⶨjyhdUhkchkcmHnHuhkchkcmHnHu hkchkcjhdU hdNHhkc hkcjhdUjhdUhdh\~jh\~UmHnHu9 D K                                6 8 F H N P   𲮣jmhdUjhdUhkc hkcjshaBNUh4jh4U hd6] hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdU4 YZ .0  .0IJ![\01hj hd6]hkchkcmHnHuhkchkcmHnHu hkchkcjghdUjhdUhkc hkcjhaBNUh4jh4Ujhs UmHnHu hdNHhd2  02^`&(.0hj$%ۜۖۖۋۖۖۖj[hdU hdNHjhdUhkchkcmHnHuhkchkcmHnHu hkchkcjahdUjhdUhdhkc hkcjhjUhjjhjU3 ^` "NPpr1 2   L!M!!!!!!!""" """##j$l$$$$$$$$$%%5&𮪟jOhdUhkc hkcjhaBNUhaBNjhaBNUjUhdU hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdU6 #<#'+`.P22}5d8PACHHKLKKTaY[`abedegjjm! !! ! !! !!!!!!!!!! !! !!!!!!gdzgdC~gd@Ggd}gdaBN5&6&&&''''((((()))))$)(),).))))))))))* **e*f*************++++++++I+J+\+c+++,,?,䱪 hd0J> h(9h(9h(9hkchkcmHnHuhkchkcmHnHu hkchkcjh(9hdU h(9hdjh(9hdUhd hdNH@?,A,i,j,y,{,,,,,,,,,,,,,,- - -S-U-------. .(.).>.?...//a/b/////////0000000000000000001111q1r11jhdEHUjxhdEHUj`hdEHUjIhdEHU hd6]hB6 hdNHhd hd0J>J11111111111111225272G2J2K2N222(3)333I4J444*5+5p5q55555@6B6t6v6x6z666:7<7888899)9*9:: :ŵjS *hdUj *hdUhkc hkcjhdUjhdUhd0J>NH hdNHjhdEHU hd0J>jhdEHUhd hd6]< ::Z;\;^<_<<<n=o=G>H>>>>??E?G?H?a?b?@@ @@@@L@N@P@R@\@^@l@@@KBLBBBCCXDZDDDDDDDDDDDDر򩗏hkchkc hkchkcjhdUjhdUjMhaBNUhkc hkcjhaBNUjhaBNUhaBNh(9hdNH h(9hd hdNHhdj *hdU7DDEEE0ELFFFFGGXHZHdHfHHHHHHHHIJII2J K"KHKJKKKKKKKKKKL:MNH hd0J? hdNH hd0J> hd6hkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdUBQQRR R hd6:`XvXwX|X}XXXXXXXKYMY]Y^YaYrYYYYY Z Z+Z-Z1Z3ZDZFZfZgZZZZZZZZZZZZ[[[[ [J[K[v[w[[[[[[[[[[[\\\ \p\r\v\ŷhkchkcmHnHuhkchkcmHnHu hkchkcj5hdUjhdUhbhd0J>h(N hd0J> hdNHh}hdCv\x\|\\\\\\\\&].]\]d]]]]^?^L^Q^R^k^p^u^{^^^^^^^^^__1_7_o_u_w_{_}______````oataxaaaa.b0bbbdbfbhbrbtbbbbbjThdEHUj/hdEHUhkc hkcj *hdUj *hdUhC~h@G hdNH hdH*hd hd0J>Eb c cdd d d1d2dzd{ddddddd*e+ebecedeeeee g!gLgMgLhNhhhhhhhhhhhhxkjx *hdUj *hdUhkc hkcjhdUjhdUjhdEHUjhdEHUh@G!jh@G6UaJmHnHujhdEHU hd6]jhdEHUjVhdEHUjUhdEHU hdNHhd*hhhiiii>jJjOjdjjjjjjjjjjjjjjkkk!k3k4kMkNkll llzl~lll|m~mmmmm*nŻߡ鎙zjKhdEHUj*hdEHUjhdUjhdU hNHjhdEHUjhdEHUjhdEHUhzjhdEHU hd6] hd6h hdNHhdhkc hkcj *hdU.m*nnnooqt*wJwxnz|+QeoƊ!!!!!!! !!!!!!! !!!!! ee]e^egd!gd!gdFw1`  h^h`2 h^h`gd}~ & Fgd}~ & F*n,n2n4nnnnnnoooo p pLpNppppqq2q8qPqRqqqqqqqqq,r?rrrrrrrrrr񩣩񐛋ujhdEHUhh(9hkc hkcjhdUjhdU hd0J> hd]jhdEHUjhdEHUj hdEHU hdNH hd6]jI hdEHUh}~jZhdEHUhdj:hdEHU,rrrrrrrrrrrrrrss s!s/s0sBsCssssssssssssst t tvlhhj.hdEHUj,hdEHU hd6]j*hdEHUj(hdEHUj&hdEHUjv$hdEHUjT"hdEHU hdNHj/ hdEHUj hdEHUjhdEHUjhdEHUjhdEHUhdjhdEHU$ t ttttJtVt^t_ttttt!u"uPuQuRueupuquvuzu{uuvv v!v%vFvGv^vvvvvvww)wwwLxMxPx[xyy%y&yDyEy^y_yyyyyyy߾j8hdEHUj6hdEHUj4hdEHU hd6]haBN hzNH hFwNH hFw6]hFw hdNHhzhdj2hEHUhj0hdEHU@Acdno񯺪~tjjShdEHUjQhdEHUjOhdEHUjMhdEHUjKhdEHUjIhdEHUhkc hkcjQIhdUjhdU hdNHjcGhEHUhjuEhdEHUjChdEHU hd6]hdjAhdEHU)01`ltuNO|yoe[j^qhdEHUj;ohdEHUjMmhdEHU hdNHj^khdEHUjoihdEHUhjghdEHUjehdEHUjschdEHUjahdEHUj_hdEHUj]hdEHUj[hdEHUjYhdEHUjWhdEHUhdjUhdEHU"|}لڄ "%&():;MNtu  *+BDGHJKders{qjhdEHUjhdEHUjhdEHUjhPEHUjmhdEHUj~hdEHU hdEHjL}hPEHUj^{hdEHUj;yhdEHUjMwhdEHUj^uhdEHU hdNHjLshdEHUhd hd6],sІц ,pvx~ƈ؈܈ވ nprϹ~tjshdEHU h!6]jhPEHUjIhdEHU hdNHhkchkcmHnHuhkchkcmHnHu hkchkcj̑h!Uh!jh!UjݏhPEHUjhPEHU hd6]jhdEHUhd+‰ :;BCbcefkmnwx}vlbjӲhdEHUjhdEHUjhdEHUj]hPEHUjohdEHUjhdEHUjhPEHUjhPEHU hd>* h-Gv>*j hdEHUjhdEHUjhPEHU hdNHjȚhdEHU hd6]hdjhPEHU#ŠÊΊϊԊ֊يڊ܊݊ߊ%&>|rh^jhPEHUjhdEHUjhdEHUjahdEHUjZhPEHUjhPEHUjhdEHUj$hPEHUj6hdEHUjHhdEHU hd6]jhPEHUjûhPEHUj_hPEHUj;hdEHUhdjhdEHU >?JKabghыҋ!"-.JKh:;xtj`ZZ hdNHj%hPEHUjhPEHUh!jhdEHUj<hPEHUjhdEHUj*hdEHUj<hdEHUjNhdEHUj_hdEHU hd6]jhwhUmHnHujhPEHUj`hPEHUjAhdEHUhdj&hdEHU#Ɗ<Ք Ѩ6ڮ^a7LPV!!! !!! ! ! !!! !!!!!!!!!! !gd-Gv gd-GvgdZ'gdNgd(91gd! hh]h^h;yz  <>ȎʎԎ֎  `bab&)dovyzȒӒXYϓhPA?h} hdNHjhdUhkc hkcjchdUjh-GvUmHnHuhkchkcmHnHuhkchkcmHnHu hkchkcjhdUjhdUhd7ϓГ89LR•Õ ht$Z\—ėЗԗ "jvŹ圍hkchkc\mHnHuhkchkc\mHnHuhkchkc\jhdU hd0JSh(9hkchkcmHnHuhkchkcmHnHu hkchkcj]hdUjhdUhbhd0J>hd hdNH- $&Йҙ(*JLZ\bd*,EK|$%17LMc豣}y}hPA? hdNHhbhd0J>jQhdUhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhkchkc\mHnHuhkchkc\mHnHuhkchkc\jhdUhdjWhdU.cdtu   "4hlnԟTV(,Vf02^`𧢧jhdU hd56jKhdU hd6 hd0JR hd5hbhd0J> hd0JS hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhdU4]c֤ͤڤ "&C|*,FZnpΦЦئڦ(,4@Fڽڽڷڷڽڽڷڷڬޥ hkchkcjEhdU hdNH hd0JRjh-GvUmHnHuhbhd0J>hdjhdUhkchkcmHnHuhkchkcmHnHuAF§Ƨ&(EKMTV[aekl˨̨QRЩѩ "$&02&'OP  0134ޭ߭02rjThdEHUj;hdEHUj!hdEHUj4hdEHUjFhdEHUj,hdEHUj?hdEHUhkc hkcjhdUjhdUh} hdNHhbhd0J>hd5rzʮЮҮԮP\ܯޯ{ذٰ۰ܰ(.SUcd./UVXY]_zf¶ hdNHhbhd0J>hkchkcmHnHuhkchkcmHnHu hkchkcjuhdUjhdU hZ'hd hd6]hdhNBfgstz"12PRŷEKbc¸XZùι78NO_`  @^~Ҽǹhe-hkchkcmHnHuhkchkcmHnHu hkchkcjhdUjhdUh}h=hbhd0J>hNhd hdNHAҼ024DFJPRTX|  `aij #)lmxvw˽ hd6]jo hdEHU hdNHh-GvmHsHh-GvmHsH h-GvNHhbh-Gv0J>jhYUmHnHuhYh-Gv hSwNHhdh=he-hSw>789>?V  LRWYvw  *,68jlnpz| .0>@Hکhkchkc hkchkcj! hdUj hdUhkc hkcj' hdUjhdUhbhd0J> hSw6]hSw hdNHhd?,r Dn/ .&& !! ! !! !! !!! ! !!!!! !!! ! !1 & F gdgdgdgdSwHJHI*+;<  t @BPRZ\r$&FHVX`bbĸҥ󡛡yhkch hkchjhUjhU hNHhhkchkcjhdUhkchkcmHnHuhkchkcmHnHu hkchkcj hdUhbhd0J> hdNHhdjhdU/bd*,.0@B|~68FHPRgh~ hNHjhUjhUh5>*\ h6 h6]hkch hkchjhUhjhUB67ab  _`  lm $&OP3hihmHsHhbh0J>j hUhkch hkchjhUjhU hNHh h6]D3:>Fnp67BDdftv~ <=FG !)RS OPiŹ󵧵󵣟j *hdU hdNHh`1hX}jhdUmHnHuhdhkchmHnHuhkchmHnHu hkchjhUjhU hNHhhbh0J>8ij |~$%<>BC%&67  (*24\^3ܾɷܒɷ܎hijhdUhkchkcmHnHuhkchkcmHnHu hkchkcjhdUjhdU hdH* hdNHhdhkc hkcj *hdUj *hdU834MN &(ln:<#$12NOhi*-12\]z}  "#34    hdH*jhdU hd6hdmHsHhn*hkchkc hkchkcjzhdUjhdUhihd hdNHE ">BDTV),./Z\h\]t z       0 <      hn*hbhd0J>hkchkc hkchkcjthdUjhdU *hd hd6 hdNHhdh`1 hdH*F B i< !"6%)+51G227:<<?:A|ABC!!!!!! !!!! ! !!! ! !!!!!!!gd*gdgd!#gd\qgdRgd tQgd4]0gd  H J v x       vwab(<gh2@"$&(24<>%JKstjhdUhkc hkcjnhdUh hd6]hkchkc hkchkcjhdUjhdUhd hdNH@  `b4545{|()>@46VXfhpr    ]!^!!!""$"P"R"""" hdH*hkchkcjhdU hdH* hd6] hd6 hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjhhdU:""##### #F#H###B$C$$$%%U%V%l%m%}%~%&& & &&&''T(U((()) )[)\)****++++,,,---.. hkcj\hdUjh4]0UmHnHuh4]0jhdU hdNH hdH*hkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdjbhdU8. ....L.N.P.R.\.^..////000000!1"1117282G2O2[2d22233330424d4f4h4j4t4v444444444>5@5r5t5v5x5555ܾܾjPhdUjhdUjVhdUhRhn* hdNH h tQNHh tQhd6]h tQ hkcjhdUhdjhdUhkc>555577,8-888888:9;9>9]9^9`99999Y:Z:::;;;;;;;;;;O<x<<<<<<<8=9===X>Y>>>>>|????>@_@@@@@9ANAgAiAzA{A|AAAAÿhbh[Dh:OhKh h!#h!#h3-Ah!# *hd h tQNHh tQh\q hdH*hd hdNHJAABB1B>BWBlBmBxBBBBBBBBBGCSCTCqCrCCCCCCDD DD DzDDDDDDDDE0E4E|EEEEEEFFF5F摍hj<h:Oh:Oh~cU6] hNHhkc hkcjh+Ujh~cUUh~cUh0,}h*NH h0,}h* hQLh*h* hbNHhf[C h[DNHhbhh[Dh55F6F>F@FPFQFXF`FgF{FFFFFFFFFFGGHH HH&H\H^HfHHHHHIIHIJILINIXIZI\IbIvIIIIIIIJJ4J6J8J:J@JDJvJ־޸޾ޭޥ¥jDh)(Uh)(jh)(UjhqU hqNHhRyhkc hkcjJhqUjhqUhqhbhf[Chhh:Ohj< hj<NH:vJxJzJ|JJJJJJJJJJJ`KbKKKKKKKLLDLELKLXL\LfL}LLLLLLLLLLLLLL M¾ߨtjhPhPU hPhP hPhbhhPh5r"NH hPh5r" hPhyhbhyh5r"hVK h*NHh* h*h*jh)(Uj>h)(Uh)(hkc hkcjh)(Ujh)(U,CLKLU$X2\_R``paaBfHfrgnqqu{ !! !!!!!!!!!! !!! !! !gdPgdgdq/gdBLgd)3gd@gd!#gdo1gdogdogdogdogdbgd*gd)( M MMMNNN N NNnNpNxNNNOOOO O"ONOPOpOrOOOOOOOOOPLPNPlPPPPPPPQQQDQʾߺʾ߲|hh5r" hh!#jhPhPUhP h#h!#h)(hb h!#NHh!#hVKhyhkchkcmHnHuhkchkcmHnHu hkchkcjhPhPU hPhPj8hPhPU.DQFQfQhQvQxQ~QQQQ R R3R6RZR[RdRfRhRiR{RRRRRRRRRRRRRRR SSTTTʾߺzrdʾjhPhPUhPmHsHh1DHh!#H*mHsHh!#mHsHh!kh)(H*h)(h!kh!#H*hbh!#0J> h!#NHhc$^h!#0J)h!#hkchkcmHnHuhkchkcmHnHu hkchkcjhPhPU hPhPj2hPhPU'T TUU7V8VvVVVVVVVVVVV'W(WeWfWWWWWWWWWWWWWWWWWWWXX$X&X`XbXXپٶ٭٭̞پّ{ hPhojhPhoUhohc$^hc$^hb0J)hc$^h!#0J)NHh"hbh!#0J>h;h!#6 h!#NHhbhbhc$^h!#0J)hh!# h!kh!#h5r"mHsHh!#mHsHjhPhPU-XXXXXXXXXX2Y:YPYRY~YYYYYYYYZZZZZZZZ[[r[v[|[[[[[[[[[[[[[ʾߺʾߺʾߺh1DHhoH*hhoH*j&!hPhoUj hPhoUhc$^ho0J) hoNHhohkchkcmHnHuhkchkcmHnHu hkchkcjhPhoU hPhoj, hPhoU.[[[[[[[[\ \\\`]b]]]]]]]]]]^-^0^1^6^8^<^^^K_P_V_W_^_a_f_g___```ʼѵ󔋔~~w h56ho hoNH hShohw/hoNH hw/hohkchkcmHnHuhkchkcmHnHu hkchkcj!hPhoU hPhojhPhoU hI6hoh1DHhoH*h0hoH*hoh5hoH*,`!`.`0`H`L`````haiaHbIbbbbbbbbbbbccddd ddddddd$e(ef@fBfDfHfĶ˯sjh!#UmHnHuh!#hjho0J> hc$^hohkchkcmHnHuhkchkcmHnHu hkchkcj "hPhoU hPhojhPhoUhKhoH* hoNHhc$^hoH* h56hohoh56ho0J>+Hf,g.gngpgxgzggggggggggg'h(hohphhhhhhhhhhiii1i2iBiCijjj j.j4jȺvȺp hj0J)j#hPhjU hPhjjhPhjUhjh=b hSh}hc$^h}0J)hkchkcmHnHuhkchkcmHnHu hkchkcj"hPh}U hPh}jhPh}U h{%h} h}NHh}*4jh)h:sh:s0J>h=bhq&h:s0J> h:sNHh1DHh:sH*hh:sH*h:s hq NH hShq h}hkchkcmHnHuhkchkcmHnHu hkchkcj#hPh#U hPh#jhPh#Uh#hq hjhq 0J)%m$m)m-m2m:m h!#NHhVh=bh!#hq h@2NHhq h)0J>hq h@20J>h@2h:sh:s0J>hq&h)0J>h:s h)NHhh)H*h:sh)0J>h)7ooooooooopp7pEpZp[peplpmpppqpppppppppppppqqq.q/q0q1qlqoqqqqĻ᳧េ}w}n}n}hq hCX0J> hCXNHhCXh=hVH*mHsH hVNHh=bhBLmHsHhVmHsHh=h!#H*mHsHh!#mHsHhq hV0J>hVhV0J>hq hV0J*NHhq hV0J*hV h!#NHhq h!#0J>h1DHh!#H*h!#+qqqqqqqdrfrrrrrrrrrrrrrr(t4t5t:tCtItftottttu0u2uCuDuuuuuuǹβᎈᎀzvlzh#h? 6]h h? NHh? hBL hAFNHhAF h!#NHhkchkcmHnHuhkchkcmHnHu hkchkcj$hPhjU hPhjjhPhjUhjh#h!#hq/hShq hCX0J>hCX hCXNH*uuuuuuuuvvv v0v>vXv`vnvvvvvvvvvvvvw ww"wfwjwxw|wwwwx7x9x@xBx^x`xgxixnxxx÷hBLhjhAF h!#NHhu?h!#H*h-h!#H*h#hAF0J)h#h!#0J)hkchkcmHnHuhkchkcmHnHu hkchkcj$hPh0U hPh0jhPh0Uh0h!#2xxxxxxxxxxxxx%y&y'y)y+y,yEyhyoyyyyyyyyyyyyyyyzz zGzNzOzWzizxzzzzzzzzzzzzz{{7{8{9{H{I{_{ý󵪣 hPhjjhPhjUhjhP hu~NHhu~ h!#NH hBLNHhBLhAFhh? hu?h!#H*h-h!#H*h!#h-h!#H*@_{`{p{q{||| |||P|\|^|||}}}}}}}~~8~:~H~J~P~R~z~~~~~~~~~~~ ʾߺʾwooghu?h!#H*h^h!#H*h'Sh!#H*j%hPh0U hPh0jhPh0Uh0hh!#0J) h!#NHhPh!#hjhkchkcmHnHuhkchkcmHnHu hkchkcjhPhjU hPhjj%hPhjU( $%6PUԀր "67OP΁=CLQf~˂̂MN UVik$<BTZabƱƱ hh hh!#hf;h!#H* hNHhhjh%?QhPhh!#0J>hch!#H* hh!#h h!#NHh!#h? CbtyzMNmnۆ܆&'*+<V[]_`mxyYZ؈&'4]tu 񮲨hkc hkcj&hURUjhURU h#ENHhURh#Eh*hjh!1h!#H*h/gZ\mHsH h/gZh/gZ h!#NHhh!#H*h/gZhh!#0J>h!# hh!#h7؈(ՎDסxyYo ! !! !! ! !!!!! !!! !! !!!!gd@EgdwgddngdCgd`Mgdb'sgd'eBgdx_tgduCgd gd=Hgd*gd/gZgd PR^`ƌ +egv{|֍׍-.ÎĎԎՎ&'*J=]sxّؑ*,6^bptВҒ»j&hPhURU hPhURjhPhURUhnph!#H*h]1h h/gZh=Hh#E h!#NHh!#hURjhURUD  \^ޓ  hޔ>HJZhjr/0 HſŻŻŷůŤŻŻſŻſſſůŐŻſj'h!#Uhkc hkcj'h!#Ujh!#Uh/gZh h!#NHh!# hURNHhURhkchkcmHnHuhkchkcmHnHu hkchkcjhPhURU8LN~ԙؙ֙ޙ "fhƚȚΚК,jlx"$&68jlnpvxzҜ0P~Ǜ hb'sNHhb'sj(h!#Ujh!#Uh'eBhG h!#NHh!#jy(h h!#UhuChkc hkc h h]1j'h h!#Ujh h!#Uh h!#NH h h!#4 BDdtv)*Ѡ  ,=?_dlmv֡ס+,aw}~ޢ3Mdvw~ƣ,nowxŤƤΤ12ʥ˥֥ܥ h!#6 hb'sNH h`MNHh`Mhkc hkcjs)hb'sUjhb'sUhb'sh+NhuCh!# h!#NHI()Ԧզ./y 1PT^z˨רijovw*2@nvx੢ܢhb hb'shYnh!#NH hYnh!# hMhM hYnhMh>0 hleh!#h}UhlehMNH hlehM hgNHhghMh`Mh+Nh6 h!#NHh!# h!#69ª 68XZhjpr֫ث  $".0xõʮՎՆ{vri`h'hEP0J>hYnh!#NHhkc hkcjm*h'Ujh'U h'NHhkchkcmHnHuhkchkcmHnHu hkchkcj)hPh'U hPh'jhPh'Uh' hleh!#hEPh'h!#0J>hYnh!#H* hYnh!#hg%x|   "#$&(*0345Jbj~ݮ$DE\^߯ԿԶԲԶԮԪ h!#5\ hdnNHhdnh!#hghYnh!#NHhYnh!#H*h'h!#0J>hb hYnh!#hqh'hC0J>hChEP hYnhEPh'hEP0J><:;yzα;<]}ʲ ,-2óԳ$fnz| hPhCjhPhCU hC\h}Uh!#0JQ h}U0JQh}Uh!#0J>h h}U6]h}U hCNH hleh!#h%?Qh h!#NH hYnh!#hChq hdnNHhdnh!# hdn5\.ʴ̴ڴܴ $&J%&=WXYfgn}ʾߺ{wqwmdh}Uh9F0J>h9F hGlNHhGlh}Uh!#0J> h!#NHh h}Uh}Uh}Uh}U0J> h}UNH h h}Uh}Uh!# h\h!#hChkchkcmHnHuhkchkcmHnHu hkchkcjhPhCU hPhCj*hPhCU'ɶ˶"%*+01:;BSTZrsy{|з⵮﮵ﵡh}Uh!#0J>h6ghYnh!#NH hYnh!#h!#hC h}UhGlh}UhGl0J* h}U0J* hGlNHh}UhGl0J>hGlh}Uh9F0J>h9Fh}Uh}Uh}U0J><3489PTVlnu ?AEHTs Qmn}~ʺκϺ.01>TX޵ hwNHh}Uhw0J>hYnh!#NHhg h!#\h}Uh!#0J> h!#NHh h}U hYnh!#hwh!# h6gNHh6gh}Uh6g0J>AX^_ejlno޻߻ ;Ҽڼ   YZ6Z{Ǿо%&>@cstѹѹѳѽѽѽѹhV hiNH hNHh@Ehiht h!#NH h=^jNHhh=^j hx_th!#h\;yh}Uh!#0J>h}Uh!#0J* h}U0J*h!#B þ*UHh`v%<\j6!!!!!!! !!! ! !! ! !! !!gdk$gdgdD+gdgdwgd gdxugdQ7gdVgd rgd^5Ugd6Ngdi1gdgdgdпԿڿۿ BHRSfops28?GJTUem,>PQ^k?@ 澺ؾ곪h:Qh!#NH h:Qh!#h6Nhih:=h!#NH h:=h!# h!#NHhuC htNHhth:h!# hVNHh@EhVhC %,Echyz+AIJ #$*+034;CFHhh^hd*ghuCh6N h!#NHh rh r6]h!#h r h^5UNHh^5UhVhha^ h@ENHh@EJ>@`bprxz*,2468nvxۿ条撋}vkd h"th rjh"th rU hK!h!#j+h"th^U h"th^jh"th^Uh rhQ7hkchkcmHnHuhkchkcmHnHu hkchkcjg+hd*gh^U hd*gh^jhd*gh^Uh^hVh!#hd*g h"tNHh"t'DE`kʾߺxtib hd*gh^jhd*gh^Uh rhd*gh"t0J> hd*gh"t h"th"thd*gh!#NH hd*gh!#h^5Uh"t0J>h"th!# h^NHh^hQ7hkchkcmHnHuhkchkcmHnHu hkchkcjh"th rU h"th rja,h"th rU!   &*>BFH &Z\^ʾߺʾ߶|h%]h!#H*j[-hd*gh^Uh h!#6 h!#6h^5Uh!#0J> h!#NHhJ[jh!#H*h!#h^hkchkcmHnHuhkchkcmHnHu hkchkcjhd*gh^U hd*gh^j,hd*gh^U0.26<HPrt #%.5=acln$&+-GHnp樚 h aJjh rUmHnHu hEaaJ h!#NHh rhRh!#H*huC h raJh%]h!#H*aJh!#NHaJh!#aJmHsH h!#aJh^5Uh!#0J> h%]h!#h!#:fgxz/25?PSVZ]prtuw|}h+h!#0J>h+h!#0J*h+h!#0JQNHh+h!#0JQh^5Uh!#0J*hEah!#6]aJh!#NHaJh^5Uh!#0J> h!#aJD@BZ^hjlp!%'(+45SWabdhjkuvʼѵї܁xtptph h!#h+hxu0J> hxu0J> h!#6aJh+h!#0J>hxuhkchkcmHnHuhkchkcmHnHu hkchkcj-hd*ghxuU hd*ghxujhd*ghxuU hxuaJ h aJh@h!#H*aJh!#NHaJ h!#aJ) !%=wxVW  }%6R\:Lh»姟hChAhwhuCh hEah hh!# hhhh+h!#0J* h!#NHh+h!#0J>h!# h h!#h 0J*6] h h =hjlntwxy   12 !"'12`dl~  &'jhd*ghUhsphGFvhxuhA0J> hANHhAhC hwNHhhw h!#NHh!#hwh!#0J>F'=>NO &(B"$DNV`r~  $)[`di˿໷hhGFvH* hGFv6hxuhxuhGFv0J*hxuhxu0J* hGFvNHhxuhGFv0J>hsphGFvhhkchkcmHnHuhkchkcmHnHu hkchkcjhd*ghUjU.hd*ghU hd*gh2    8:<LN^r\)*Bƿƪƌ|v|v| hNHhhuChzh |\hhkchkcmHnHuhkchkcmHnHu hkchkcj.hd*ghxuU hd*ghxujhd*ghxuUhxu hxu0J> hGFvNHhhGFvH*hGFvhxuhsp0J>hsp hspNH,B_fo/012789}~ !"#./st Żh!Dhk$ h5\hThhTh5\ h6]hTh6H*] hTh6]hTh h!#NHh |\6H*] h |\6]h |\hh!#D#$>(GHMQ[%&DEOR[\nty%'.279AEM^_潹潹 h@LNHhxuh@L0J>hD+ hWNHhWh@LhxuhC0J>hxuhC6]hxuhC0JQhxuhxu0JQhxu hCNHhChhhnS?&SYmn %&)*679>@AEGLNOSYiju16MShi˨˨˼hBA hy5\ h5\ h9NHh9h hD+hD+hyhhy0J>hhD+0J>hhW0J> hD+NHhhk$hD+hW@ -.56i26?F^ersyz{~  |  FSTVY޼͵ީ柛hahaH*hahk hB%NHh5\hB%huC hyh!#h9hBAhw/h!#NH hw/h!#hA h!#NHhk$hh!#hVh hD+NHhD+h:6io   XG`~!!! ! !!!!!!!!!gdJ??gdmegdL4*gdq)gd1gdJ??gd WgdgdQ$gdk|gd['UgdB%gd0Fgdk$gdVY[^`cemoqr *,.Z\|~0@ھ󠙎ysjh5\ha0J> h5\NHj/h5\h5\U h5\h5\jh5\h5\U h5\h>h0FhkchkcmHnHuhkchkcmHnHu hkchkcjO/h"th5\U h"th5\jh"th5\Uh5\h> haNHhahahaH*(@BDVXZbdnp018;HIKRTUWX^dhkxz  "$-/8:CIRST\ľh> h>h> haNH haha h0FNHh0Fh5\ha0J>hahaH*h5\ha0J*NHh5\ha0J*h>xha6ha ha5\?\^eg)AFSThino$%Ne rtտ~hkchkcmHnHuhkchkcmHnHu hkchkcjI0h5\hB%U h5\hB%jh5\hB%UhB%hLh!#H*hd h!#NHh!# hah-h0F h>NHh5\h>0J>h>h2h>H*1 :<DFvxBCNOfij !%4>FPXY]`gٲϨϨϨϟϟϟϟ hxNHhB%h)0J> h)NHh['U h)h)hB%hx0J>h;&h5\h)0J>h) hypNHhyp h%h% h%NHh%hxjhk$UmHnHu<2:jks $*+N  B | }   3 4          : D L Z [ |       C H ^ b d ԻԴΰԬ԰h Wh  hoah% h%h% h GNH hk|NH h%NHh%hyph G hxNHhk|h['UhuC h)NHh)hxDd e m n |          R g j k o p    DLtu >Siu./HIÿùÿõõÿjh7h!#Ujh!#U h&GNHhdh!#h&G hNHhuCh h>bNHh|{h>bhQ$j4hw|h%EHUj2hnh%EHUj0hnh%EHUh Wh% h%NH3 Xz*+EMer  'FJKNOSTbϷ곫귫 hLMNHhLMh/6&hL4*hq)huCh: h!#NH h WNHh1hQ$hJ??h Wj7h!#Uh!#jh!#Uhkc hkcAblHXt"P~Xhj*,<>@^`¾Һ¾ꡫꡛ h2UNHh2U hmeNHhmej8hL4*Uh1hkc hkcjb8hL4*UjhL4*UhJ??huChdhaE@hq)hL4* hLMNHhLMh!#;`b <\^~  $&ʾ~yuo h!#NHhkc hkc *h Wj9 *hvh!#Uj *h!#Uh W hCoNHhCoh!#huChJ?? h2Uh!#hkchkcmHnHuhkchkcmHnHu hkchkcj\9h2Uh2UU h2Uh2Ujh2Uh2UU(~~+++:-:7@0F2FHPQBRRRU\!!!!!!gdI*gdI*gd*xgd] gdI*gd7 gd7gdgd/gd, gd, `gd?`gd!#gd gd W(*JLNXZdf  ,.<>DFHdhjĽung h,h!# h,h,hkchkcmHnHuhkchkcmHnHu hkchkcjP;h,h2UU h,h2Ujh,h2UU h2Uh2U h2Uh!#hCoh!#j:hJ??Uh,hkc hkch WjV:hJ??UhJ??jhJ??U#t "(*z   " ( * j x z |  ʾ߷ʾ|ugu|ʾ||uj<h2UhJ??U h2UhJ??jh2UhJ??Uh/EjJ<h2Uh2UU h2Uh2Ujh2Uh2UUh2UhJ??h!# h2Uh!#hkchkcmHnHuhkchkcmHnHu hkchkcjh,h2UU h,h2Uj;h,h2UU'        4!6!8!d!f!!!!!!!!!""."0"P"R"`"b"h"j""""""""""##,#.###ʾߺʾʾtʾj>h2Uh2UUj>>h2Uh2UU h!#NHj=h2Uh2UU h2Uh2Ujh2Uh2UUh2Uh!#hkchkcmHnHuhkchkcmHnHu hkchkcjh2UhJ??U h2UhJ??jD=h2UhJ??U-######$$ $$*$,$.$Z$\$|$~$$$$$$$$$$%0%x%%%%%%%&&<&>&L&N&T&V&b&n&v&~&&&&&ʾʾʾj2@h2Uh2UUh2Uh!#0JQ h!#NHh2Uh!#0J>j?h2Uh2UUh2Uh!#hkchkcmHnHuhkchkcmHnHu hkchkcj8?h2Uh2UU h2Uh2Ujh2Uh2UU1&&&&&&&&4'<'t'x''''''''''''''&(((((((((((())d))))))))*ʾߺʾߍʾߺ{{h|6jAh2Uh2UUhCoj,Ah2Uh2UUh2U h!#NHhh!#H*h2Uh!#0J>h!#hkchkcmHnHuhkchkcmHnHu hkchkcjh2Uh2UU h2Uh2Uj@h2Uh2UU/**$*&*4*6*<*>*F*H*Z*\***********+++0+2+R+T+b+d+j+l+x+z++++++++z,ʾߺʾߺʾߺ}tth|6h, 0JQjh!#UmHnHuh, j Ch2Uh2UU h]h!#jBh2Uh2UUh|6h2Uh!#hkchkcmHnHuhkchkcmHnHu hkchkcjh2Uh2UU h2Uh2Uj&Bh2Uh2UU*z,~,,,,,,,----"-4---------..,...Z.\.|.~....... / /,/.//D/F/Z/x/̾̾̾̾vh{Zh, 0J)jEh2Uh, UjDh2Uh, UjDh2Uh, Uh/Eh, 0J>hkchkcmHnHuhkchkcmHnHu hkchkcjCh2Uh, U h2Uh, jh2Uh, Uh, h]h, H*-x/z///////////00J0L0l0n0|0~000P1R1~111111111122F2H2¼±{±m{¼±_jGh2Uh, UjFh2Uh, UhkchkcmHnHuhkchkcmHnHu hkchkcjFh2Uh, U h2Uh, jh2Uh, U h, NHh, hkchkc0J)hkchkc0J)jEh{Zh, 0J)Uh{Zh, 0J)jh{Zh, 0J)U%H2h2j2x2z22222222332343B3D3J3L33333444444,404L4N4z4|4444444Z5\55555ɕyjHh2Uh, UjHh2Uh, U h}`h, jHh2Uh, UjGh2Uh, Uh|6h, :h{Zh, 0J)h, hkchkcmHnHuhkchkcmHnHu hkchkcjh2Uh, U h2Uh, /5555555664666V6X6f6h6n6p6666666666777,7.7Z7\777777777&8(8884969b9d9ЫЗЗЗ{jJh2Uh, UjsJh2Uh, U h, NHjIh2Uh, U h}`h, jyIh2Uh, U h2Uh, h{Zh, 0J)h, jh2Uh, UhkchkcmHnHuhkchkcmHnHu hkchkc0d999999999~::2;4;X;Z;h;j;l;;;;;;;;;;$<&<<<<<<<<<<<<(=*=8=:=ɹͮͮyuh7 h!#NHjKh2UhoUh/Eh!#hojmKh2UhoU h2Uhojh2UhoUjh, 0JU h, NHh, hkchkcmHnHuhkchkcmHnHu hkchkcjh2Uh, U h2Uh, +:=<=h=j==============J>L>>>>>0?2?d?f?h?j?t?v??????@ʾ~yujc h2Uhojh2UhoUhkc hkcjLh!#Ujh!#Uhoho:]ho h!#NHhoh!#:] h!#6 h{Z0J)h!#hkchkcmHnHuhkchkcmHnHu hkchkcjgLh2Uh7U h2Uh7jh2Uh7U$@@0@2@@@B@H@J@N@V@X@`@@@@@@@@@@A,A0A2A^A`AAAAAAAAAA BB(Bʾߺ~p~ʾ~jfh_x h!#NHj[Nh4h!#Ujh4h!#U h4h!#huChkc hkcjMhoUjhoU hoNHhoh!#hkchkcmHnHuhkchkcmHnHu hkchkcjh2UhoU h2UhojaMh2UhoU&(B*BVBXBxBzBBBBBBBBBCC0C6ClCnCCCCCCCCCDDDDDDDDDD6EPEEEʾʾuʾqmh0h7jOh4h!#UjUOh4h!#U h4h!#jh4h!#Uh_xh!#0J> h!#NHh0Th!#H*h!#hkchkcmHnHuhkchkcmHnHu hkchkcjNh2Uh_xU h2Uh_xjh2Uh_xU)EEEFF$F/F0F1F2FyFzFFFF/G0G=G>GGGGGHHHHHHHHIII^I`InIpI{mhkchkcmHnHu hkchkcjOPh4h!#U h4h!#jh4h!#Uh_xh/Eh!#6NH]h/Eh!#6] h/Eh!# h!#NH h0Th!# hNHh#j *hqXh7UmHnHuh7h.D\h!#h0h/E&pIvIxIIIIJJJJJJJJJJJJJKKK0L2L^L`LLLLLLLLLL@MBMMMNN4N6NDNFNLNNN辰䬡~辰jQh4h!#UjIQh4h_xU h4h_xjh4h_xUh_xhkchkcmHnHu hkchkcjPh4h!#U h4h!#h h!#NHh!#jh4h!#UhkchkcmHnHu.NNNNNN&OOOOOOOOOP PPPPJPZPPPPPP0Q2QfQQQQQQQQQQQRR;RhuC hO$h] hth] H*h] hkchkcmHnHuhkchkcmHnHu hkchkcjSh4h] U h4h] jh4h] U4YYYYYYYYYYYYY-Z0ZYZ}ZZZZZZZZZZZ][h[y[[[[[[[[[[[[[[[[[[[[[[[[\\\ \ \\\\\\\hh!#0J> hh!#hh0J* hhhh~Kh!#0JQ h!#NHhuCh~Kh!#0J* hN1h!#hN1h!#H*h!#h~Kh~Kh!#0J>?\ \%\'\(\*\,\.\1\7\;\=\f\p\\\\\\\\\ ]]])]*]G][]\]_]]]]]]]]]]]]]^ ^^^=^A^B^]^m^n^~^^^^^^^^^^!_"_r_켸߸ɼ߸߸߼ߴߴߥߴ߼ɼhBOh0J> hNHhhc hs]h!# h!#NHhBOh!#0J>h~Kh*x0J>h*xhhBOh!#0J*h!#hh!#0J> hh!#A\^\anafpn s}R}ގ֑f '!!!! !!!!gd_mgdgd6gdNngd,gdx_tgdBAgdx_tgdgdcgdmDugdV`gd7gd9gd*xgdgdcr_t_{_}___________``'`(`Z`\````` a aJaKa\anaaaaaaaabbb bXbZbbbbbbc*c:ccԇЇh hkchkcmHnHuhkchkcmHnHu hkchkcj7Th2UhU h2Uhjh2UhUhh*xhdEhhBh!#H* h!#NHhBOh!#0J>h!#hh!#H*3cccccccc d/dRdSdcdddddddddddddddd e efff f&ffffNhhhhÿô~zvrhJh!&+hmDuhkchkcmHnHuhkchkcmHnHu hkchkcj1Uh2Uh*xU h2Uh*xjh2Uh*xUhuCh*x h9NHh9h h!#NHhs$hkc hkcjTh!#Uh!#jh!#U)hhhhhhhhhhhhiiiHiZi\iii"jjjjjjjjjjjjkkkk@kkkkkk*l,l.l0l:luVuXuuu,vvvxvvv$w&wǧꘈh hcNH h!#NHh!#jXhV`UjhV`UhV`hchmn hn,NH huuNHhkc hkcjW *hWhuuUj *huuUhmDuhuuhn,hx_hn,6]1&w(wTwVwvwxwwwww^x`x|xxxxxxxyy>y@yNyPyVyXyrytyyyyy z#z%zIzPzRzhziz{{ {{ʾʾ}yqjh!#Uhuuh<1Ch!#H*hchh!#0J>hF9h!#H*jYh2UhUhhl&h!#H* h!#NHh!#hkchkcmHnHuhkchkcmHnHu hkchkcjXh2UhU h2Uhjh2UhU,{'{({|| ||||||}}}P}x}~~ ~&~A~P~v~~~~~~~8:΁ЁځށTV(,dlt~Ѓփ$'9ϼ϶ϼ϶ϼϼ϶ϼϼ϶ϼϼ϶ϼϼϭϭ϶ϭ϶ϭhBOh0J> hNH h6]huChchmDuh hh h!#NHhkc hkcjh!#UjYh!#Uh!#C9;UYy{|؄ "&~268Dbḋ·žũyrk h6] h}?hh5mHsH h5hZ 2hH*hkchkcmHnHuhkchkcmHnHu hkchkcjZh2UhU h2Uhjh2UhU hNHhV`hh6]hBOh0J>NHhBOh0J>hh[8*· ,jl<@B^fhƉȉΉЉ#0ACFGINW[bhɻдzhh0JQhh0J* h5hh0J>hkchkcmHnHuhkchkcmHnHu hkchkcjZh2UhU h2Uhjh2UhUhuChefhH* hNHh^hV` h6]h./;>AHI]hrȋʋϋՋ؋ڋ܋"Tlx|ΌЌ h2Uh^jh2Uh^Uh^h^: h^6]h^h^0J>huCh^h^H* h:0J>hh0J>NH h^NHh[8h^ hNHhhh0J>hh]5 .068<z (+`gÎĎ܎ݎ23_ʾߺwoibi h ?h hNHh/lh6h ?hH*h^h^6] h^0J>h^h^0J>hh0J>hhj[h[8hh^0J>hh^hkchkcmHnHuhkchkcmHnHu hkchkcjh2Uh^U h2Uh^j [h2Uh^U&_89HM]Ðѐ56Ji͑ՑIaJKW`az{  $bzѼƼƼјhNnhkc hkcj[h^Ujh^UhBAh^ h!#NHhZ hx_th!#h!#huC hNHhhh[80J> h[8NHh[8 hNHhNO Lz6`p%<@OPboq|0O ~ǻ hh9NHhh9hhvhP?Eh!#0J>h5hP?Ehh90J*hl4h!#0J>hh9h6huC hL.h!#hkchkcmHnHuhkchkcmHnHu hkchkcj_hith!#U hith!#jhith!#Uh!# h!#NH0°ưȰͰӰ԰   $QRfgʱϱѱұױݱޱ $%,G[\]`nsty躱觛躱觺 h 0J> h 6]hP?Eh!#6NH]hP?Eh!#6]hP?Eh 0J>h h!#6h6hIrh56]hIr h5NHhP?Eh50J>h5h!#hP?Eh!#0J>NHhP?Eh!#0J>8y{|²Ѳݲ"$=@I\]{̳:;o2?@Լ h2Uhjh2UhUhhIr h$ h!# h NHh_mh6 h!#NHhP?Eh h:hP?Eh 0J> h 6]hP?Eh!#6]h!# h 0J>724BDJL*bntvθи  (p78ʾߺʾyuquhubuqub h!#NHhP?Eh!#0J>hW h!#jhxLUmHnHuhP?Ej`hhUjhhUhhNH hh hNHhhkchkcmHnHuhkchkcmHnHu hkchkcjh2UhU h2Uhjl`h2UhU&^zc[2YLar~.!!! !!!!!gd6>gdgdwgdqgd`D`gdSH3gdqgd=$gdP,egd gd]>Lgd]>LgdH|gdx_tgdW gdIr8?HUWeikp@AW[\ahipqrλ!0167:<@FJCDS] hkc hkcjfah!#Ujh!#UhW h.NH h.h.hP?Eh!#6]hxL hNHhP?Eh0J>h h!#NHhP?Eh!#0J>h.h!#= &B ahvw*267EFMOTikrs~*Xabms ݽݲhkc hkcjah 3Ujh 3U hc6QNHhvh.hc6Q h9NHh 3h9hH| h!#NHhuCh!#jh!#UB"$VXZ\fhz|56?@JX`cnKLPafbdf hnvhVjhnvhVUhV hnvh!#h. h!#NH hk'NHhk'h)_huC hc6QNHh]>Lhkc hkcj`bh 3Ujh 3Uh: h 3NHh 3h!#hc6Q6  HR  ,.<>DFRprʾߺʾ|n|ʾ|ggg hnvh!#jchVhVUjhVhVU hVhVjZchnvhP,eU hnvhP,ejhnvhP,eUhP,eh)_huChVhkchkcmHnHuhkchkcmHnHu hkchkcjhnvhVU hnvhVjbhnvhVU'&(*VXxz  9:ADYZ~Ƹۥ桘wpbjdhnvhVU hnvhVjhnvhVUhVhVH*hVhuC h!#NHh.h!#0J>h!# hnvh.hkchkcmHnHuhkchkcmHnHu hkchkcjTdhnvhnvUjhnvhnvU hnvh!#h h h H* hnvhnv&  dfhjl >@^`ln !/=>YZ[awŸůŦŦɠɘŦŇŦɦɠɀxhw=h!#6 hw=h!#h)_hSHh!#] h!#6hVhVH* hVNHhSHh!#0J>hSHh!#0J*h.h!#0J>h/"h!#hVhkchkcmHnHuhkchkcmHnHu hkchkcjhnvhVU hnvhV/%&35MNQRSTZ[\ky۾۠۾ەykhkchkcmHnHu hkchkcjNehnvhVU hnvhVjhnvhVUhVhVH*huCjhvUmHnHuh/"h!#hSHh!#0J> hw=h!#hSHhV0J>hV hV]hSHh!#]hSHh!#0J*hSHh!#0J*NH% (X\ NOklnstu~׺׺ĥĝĝhP,ehP,eH* hP,eNH h2!0J>hSHh2!0J>h2! h!#NHhP,e h!#6hSHh!#0J>>*hSHh!#0J>h]>Lh!#jhnvhVUhkchkcmHnHu;&'  ^fhj 2Yv{ʾuh2!h2!6NH]h2!h2!6]huCh- heRh!#h]>L h!#NHh!#hh h NHh h2!hP,ehkchkcmHnHuhkchkcmHnHu hkchkcjehnvhP,eU hnvhP,ejhnvhP,eU.%.Jdefklm 29<=>xyz|}noptuv|}ϲ򡊆h!# h|h!#heRh!#NH hSH6]hSHjHfh=$h=$EHU#j;L h=$h=$CJUVaJjh=$h=$U h=$h=$h=$hSH6] h=$hSH h=$h!#h- heRh!# heRh2!4P^v  #%)+06=TUkl|} fǫhkchkcmHnHuhkchkcmHnHu hkchkcjPhhw=h!#U hw=h!#jhw=h!#UhSHh!#0J>NHhSHh!#0J> h!#NHh|h!#NHh!# h|h!#9frtxVX\^jl68:<>@BDFH`b쿸{hqho h!#NHhkchkcmHnHuhkchkcmHnHu hkchkcjhhw=h!#U hw=h!#jhw=h!#U hAh!#h|h!#]h|h!#6]hSHh!#0J>h!#h|h!#NH h|h!#0xzL &2VXnpz|~  &'04õʮՋՋՂՋՋxxՋՋՋՋՋՋh|h!#6]h|h!#NHhSHh!#0J>hkchkcmHnHuhkchkcmHnHu hkchkcjJihw=h!#U hw=h!#jhw=h!#U h|h!#h!#h]>LjhJ^UmHnHuhJ^ hoNHho/46=@Asy;Bqr &rst)*+89 h=$h=$h=$hSH6] h=$hSH h=$h!# hSH6]hSH h!#NHjh!#0JUh|h!#6h`D`h|h!#6]hSHh!#0J>h|h!#NHh!# h|h!#79:MNOPmno FGH[\]^`aǫ۞rm *hy joh=$h=$EHUjmh=$h=$EHU#jD;L h=$h=$CJUVaJjkh=$h=$EHUh=$h!#5>*\h=$hSH6] h=$hSH h=$h!#jih=$h=$EHU#j-;L h=$h=$CJUVaJ h=$h=$jh=$h=$U*anz|HX-./0LM\]^_uvw̾ӷ昑hY6\] hY\ h!#NH\ h!#\hkchkcmHnHuhkchkcmHnHu hkchkcjqhw=h!#U hw=h!#jhw=h!#UhhqhYh!#0J> h!#NHh!# hy h!#408FP>@\^bhjlnp"$ABDRhnpxyľĺľľijīĠ hkchkcj0rhnvhwU hnvhwjhnvhwUh^?shwH* h^?shwhuC hwNHhw hth!# h!#NH\ hY6]hYh!#hY6\] hY\ h!#\hYh!#0J>3 &|  8|~<\n׾׺׺ׯơƯƖӒגגג{u hNHhh3b hy\hy\h]>LhGh!#h hkchkcjrhwhwUjhwhwUhyThwhwH* hwhw hwNHhuChwjhnvhwUhkchkcmHnHuhkchkcmHnHu.(vy1Kfhp,.0:<hjwhkchkcmHnHu hkchkcj*shw=h!#U hw=h!#jhw=h!#Uh hY6]hYhQ-h6> h3bNH h3bh3bh3bh3b6NH]h3bh3b6]hhw h!#NHh!#h3bh+Zjp)GV_d-./8IT|}4¶¶ͪͦͺͪͺ互h/5 h9NHjh?AUmHnHuhw hNH hNHhh9h[h h_>h!#hhQ- h!#NHhYh!#0J>h!#jhw=h!#UhkchkcmHnHu64:@IQ$)/9<=FNOSYk  8:hLNpȽ̹ȟțܛȗhYh]>LhGhn hgrNHhkcmHnHuhkcmHnHuhkcjshgrUhgrjhgrUhh+h9 h/5NHh/5 h!#NHhh!#hh6>8.N((*v/7;@@JRRRV-Y[!!!!!gd (gdhhgd*gdzgdgd:\gd*-gd Fgd`HgdnRgdygd+ZgdS`)gdngd gd]>Lgd6>@IJWdjv  02(*,6bȺϳ~zzvrnrh[hlVhh!#jh?AUmHnHuh hLGh^0J>hkchkcmHnHuhkchkcmHnHu hkchkcj$thh^U hh^jhh^Uh^h^H* h^NHh^hLGh,h+ZhuChG)  "NPpr*,46TӺȳ{tft{{juhh^U hh^jhh^UhlVh^H*h^ hgrh^hkchkcmHnHuhkchkcmHnHu hkchkcjthhUjhhU hhhlVhlVH*hnhYAPhhH*hhlV hlVNH(TV(*02Bhjr  0 2 @ B J L   ʾ{tft{ʾ{jvhh^U hh^jhh^UhYAPh^H*hYAPh^NHhn h^NH hYAPh^h^h  hnh  hnhnhkchkcmHnHuhkchkcmHnHu hkchkcjuhnh^U hnh^jhnh^U( V \ ` d r ~             " 0 2 : < > F h       ! A B U [ ] _ l   ˽Ҷ变|shgrhg J0J>hhg JH* hgrhg Jhg JhYh!#0J>hhkchkcmHnHuhkchkcmHnHu hkchkcjvhh^U hh^jhh^Uh^h^H*h!#hk[hYAPh^H* hgrh^h^*         , - 9 ;     $LN  $&(PT`окко̯yumihS`)hnhnH*h~hkchkcmHnHuhkchkcmHnHu hkchkcjwhhnU hhnjhhnUh+Z hk[NHhk[hk[H*hnhk[ hLGhk[h hLGhLG0J>hLGhg JNHhLGhg JH* hLGhg J)  %'(),/X[eisvw~ 245<=ABH`hʽΰ蝖jwhhLGU hhLGjhhLGUhLGhLGH* hg JNH h~h^ hgrh!# h[NHhchg Jh~h+Zh  hS`)NHh[hLGh^hnhS`)hS`)H*hS`)4 (*H 8@^04BP(*Șȱtj xhhU hhjhhUh hg Jhg JH* h~h h!#NHh!#hchhH*h hgrh!#hg JhnhLGjhhLGUhkchkcmHnHuhkchkcmHnHu hkchkc-*246BJ 12BCHRUk$45 z|̾ĺĺȴȴȰȺě h}i6] h NHh h}i hV6] h[NHh[hS`) hkjNHhV hLGNHh+ZhkjhLG h!#NHh\Ch!#6h!# h0J>jhhUhkchkcmHnHu6|.0@BdhPTdf dfrvؼ|jhhpUhphpH*h+Zhy\h}i0J> hpNHhkchkcmHnHuhkchkcmHnHu hkchkcjxhh}iU hh}ijhh}iUh}ih}iH*hp h}iNH h}i6]h}i0"$,.>@VZjl  :L\^v n    ˿໷˿໒~zvzo hh/%6]hS`)hh/% hp6]h+Zhy\hp0J> hpNHjyhhpUhphpH*hy\h}i0J>h}ihphkchkcmHnHuhkchkcmHnHu hkchkcjhhpUjyhhpU hhp+ ! !!!b!t!!!!!!!!!!2"4"""## #####@$B$z$|$$$$$$$$$$*%,%L%N%\%^%f%h%yhkchkcmHnHuhkchkcmHnHu hkchkcjzhhh/%U hhh/%jhhh/%Uhh/%hh/%H*hh/%6NH] hh/%hh/% hh/%NHh+Z h 6]hS`)h hh/%hh/%6] hh/%6]hh/%/h%r%v%x%%%%%%%%%%%% & &&&&&&&&&$'4'`'f'v''''( ( (((*(X(\(l(n(((((((ۼ۴۩hkchkcmHnHu hkchkcj}zhheeU hheejhheeUheeheeH* hee6]h h~ hee0J>hy heeNHhee hh/%0J>hy\hh/%0J> hh/%hh/% hh/%6]hh/%/((((((($)()D)P)V)j)l)))))*<*>*D*n*r******+&+*+:+<+h+j++++++տDZ䭩}vhhkchkcmHnHu hkchkcjzhh*;U hh*;jhh*;Uh*;h*;H*h*;h hphyhyH* heeNHhS`)h~hyhyhy6]heeheeH* hee6]heejhheeUhkchkcmHnHu(++,,,,D,F,f,h,v,x,,,,B------...".#.*.,...8.9.S.v.w................4/m/n/u///🣟h3ghIy h_=NHh<hyh*;h_=hkchkcmHnHuhkchkcmHnHu hkchkcjw{hh U hh jhh Uh h H*h jhh*;U6///V0Z0j0l0000000000141D1F1r1t11111111122222222 333344 4 4`444444455 5ɻޡɻޓɻhy\h<0J>j|hhIyUjq|hhIyUhkchkcmHnHuhkchkcmHnHu hkchkcj{hhIyU hhIyjhhIyUhIyhIyH*h<hIy6 5$5&5,56585F55555(6)6+6-6E6f66666666666677777:7Z7[7q7r77788 8 888`9999~h.|hnRhkchkcmHnHuhkchkcmHnHu hkchkcjk}h?Ah?AUjh?Ah?AU h?Ah?Ajh-9UmHnHu h<NHhyh?Ah?AH*h[h?Ahy\h<0J>h<hIy19::8:::H:J:R:T:b:d:~::::: ;;p;r;;;;;;<< <"<F<H<T<V<Z<t<<<<=="=:=H=Z=\=`=b====˿໵~~~h`Hhm h!#NH hy\6]hy\h!#hLGhp heNHhehnRh~ h?ANHh?AhkchkcmHnHuhkchkcmHnHu hkchkcjh?Ah?AUj}h?Ah?AU h?Ah?A1=====>z>>>6?8?f?h?j???????????@&@*@,@<@`@j@t@~@@@@@@@@@@@չꛗꗑ|sh FmHnHuhZh F hnRNH h{h{ h{NHh{h1FhkchkcmHnHuhkchkcmHnHu hkchkcje~h{hnRU h{hnRjh{hnRUhe hpNHhnR h`HNHh`Hhph!#*@@AAAAAA*B8BStSxSSSSSSSSS@TNT`TeTsTTTTTTTTTTTTGUϽƴϴxϴϴtplpph^h!#h*hImHnHuhkcmHnHuhkcmHnHu jăh^UmHnHujh^UmHnHuh^mHnHuhuCmHnHuhMmHnHuh!#mHnHu hhm hzh xhuCh xhzjhzUmHnHu)GUcUdUhUyUUUUUUUUVV3VQV~VVVVVVVVVVWW-W?WKWUWZW{W}W~WWWWWWWWWWWWWWWWWW5XkXlXXXXXXXþñҫ֥֜h (hhh0J> hhhNH h^NHh (h!#0J*hM h!#6h (h!#0J> h^0J>h^hhh hMNHhMh ( h!#NHh*h!# hINHhI=XXXYYYY Y Y Y Y YYYYYYYYYY!Y"Y#Y$Y%Y,Y-YYYY"Z5Z;Z=Z>ZmZZZZZZZZZZZZZZZZZZ[[[[[[[[s\t\\\!]"]_]a]c]d]w]h (hMh* h!#NHh=jh*h*0J>>* h*0J>>*h (h!#0J>>*h (h!#0J>NHh (h!#0J>h!#Hw]x]]]]]]]^N^^^_^^^^^^^^^^^^^O_P______"`J`$a&aBa\arataaaaaa˻|u hkchkcjAh/ha%U h/ha%jh/ha%Uha%h/ h/hh/hO NH h/hO h/hhhh/h!#NH h/h!# h!#NHh^h (h!#0J>h!#h (h (hM0J>hM hMNH+aaaaa2bNb^b`bdbtbvbbbbbbbbbbb4c6cccccccccccd d.d0d8d:dFdȵȧȵӘӒӋӃӋxjx۠j;h/ha%U h/ha%h!#h/mHsH h/hO h/NHh/h/H* hkchkcjh/hUjh/hUh/hH* h/hhuCh/ha%jh/ha%UhkchkcmHnHuhkchkcmHnHu([cnnnn qmu{#Tvx'!!! !!! !! gd8gd/6&gdgd,:gd Vgdwgd]|\gdo gd':;gd!#gd':;gd/FdZd\d^ddddddddde$eee"f$f2frffffffff gggggggTgVgǻ~pd`hahkch':;mHnHuhkch':;mHnHu hkch':;j5ha%h':;U ha%h':;jha%h':;U hyNHhyh*h!#ha%hkchkcmHnHuhkchkcmHnHu hkchkcjha%ha%Ujha%ha%U ha%ha%h':;#VgXggggggggg h.hhhhhhh4i8i:iVi\i^i`iiiiiiiii(j)jOjRjnjjjjjjjjʾ|ʾxrxx h&-NHh&-j/h/haUhahuChAh!#H* h!#NHhbkh!#H*hah!#0J>hah!#0JQh!#hkchkcmHnHuhkchkcmHnHu hkchkcjh/haU h/hajh/haU,jjjjjk k kkkkkk%k&k'k=k>kNkOkll l l\l`lfljlrlvlllllllll mmfmrm|mmmmmmmn̾ӷ{hJh!#H*hah!#0JQhah!#0J>hAh!#H*hkchkcmHnHuhkchkcmHnHu hkchkcjh/haU h/hajh/haUhahefh!#H*h6h!#H*huC h!#NHh!#0mnnnynzn{n|n}nnnnnnnnnnnnnnnnnnnnnoo&o'opp p pBpDpFp򡚌wkd hyh':;hkch':;mHnHuhkch':;mHnHu hkch':;j)h/h':;U h/h':;jh/h':;UhuChefh':;H*hAh':;H*hO jhO UmHnHuhtHho hnh':;hah':;0J>hah':;]h':;h!# h!#NH%FpHpppppppqq q*q.q0qLqTqVqqqqqqqqqqqYrgrrrrrrrrrrrrrrr ssesfsss˽Ҷ|h &[ho0J>hMu1hoH*haho0J> hoNHhkchkcmHnHuhkchkcmHnHu hkchkcjh/hoU h/hojh/hoUhuCh6hoH*hohah':;0J>h':; h':;NH0sssssttttttGuIuSuTuiujuuuuuvvvvww"w$w,w.w>w@wZw^wwwwwxxnxoxsxvxxxxxyֺh?hoH* h?hohhhoH*hLhoH*hkchkcmHnHuhkchkcmHnHu hkchkcj#h/hoU h/hojh/hoUhZhoH*huC hoNHh &[ho0J>ho0yyFyHyWygyhyxyyyyyyyyyyyyyzzzzzzz{{{{{{>{F{H{^{`{{{{{{{䯪䟘uhkchkcmHnHu hkchkcjh/hoU h/hojh/hoU ho6"ho5:CJOJQJ\]aJjh]|\0J<UhLhoH*h &[ho0J> hoNHhoh?hoH* h?hoh?hoNH*{{{{{{{"|$|D|F|T|V|^|`|l||||||||||}}}}}4}8}t}}}}}}@~A~R~T~_~a~~~~ ݯݧ֙y h+NHh+h &[hoH* hoNHhefhoH*jh/hoUh?hoH*huChkchkcmHnHu hkchkcjh/hoU h/ho h?hohojh/hoUhkchkcmHnHu/  *46HJn| ,HRT\^xof]h!#mHnHuhZmHnHuh+mHnHuhJ!h+mHnHuhkchkcmHnHuhkchkcmHnHu hkchkcjh/h+U h/h+jh/h+Uhkc hkcjh+Ujh+Uh &[h+0J*hhh+H*huCh+ h+NH$́ =ÂĂ "#&q"#s˄̄ل"#ʅ˅MNX{RS߇fmwΈֈ^_ÿhhhuChU\ hdNHh,:hdhh!#h,:hw0JQNHh,:hw0JQh,:mHnHuhwmHnHuh!#mHnHuhuCmHnHu=  f8FRSȏ֏*rsܐݐGHtuJK\]PQ"ۻ۷ۻۻ۩jhdU h,:NHh,:huCjhfU hd6]hU\ hdNHhdhkc hkcjhfUhjhUB"$DFTV^`(ĕƕ$&.0PR\^lܘ\]nstuvwxΚϚ !äàà h/6&NHh8jh8UmHnHuh/6& hU\NHh,:h$ hdNHjhdUhU\hkchkcmHnHuhkchkcmHnHu hkchkcjhdUhdj hdU5!&'59AF^_uv  02$%+,͞Ξ'7RX|}ȢTUv[\efǥإĸhx! hd6] hdNHhkchkcmHnHuhkchkcmHnHu hkchkcjhdUjhdUh#ih*8hdh$h/6&BXyѦ٦ܦ9ͧۧ"12wx]iΩϩөԩ./=BKV\_cdΪЪ ѫcd¬ì: hd6] hd5\hx!hx!0JQ hd0J7hX h*8NHhuCh*8 hx!NHhx!hd hdNHFy  74uJKO}V8S(!!!! !!! !!!gdzgdeRgddKgd gd0OUb ]b gd3-Agdgd:\gd$ gdgdq=<gd%gd:\gd*8gdx!:;TU ,.:MN޲߲12DLó<Fd´ɴ*+,ÿͻ͵éåjh LLh LLU h LLh LL h LLh h*U<h7)h h7)NH h7)aJ h:\aJh LLh%h7) h:\NHh:\hzhdhkc hkcj *hhUhj *hU4,BCST   Զ~ "68,<LMVw˿˿{u{ h%NHhyi hFNHj|h LLhFU h LLhFjh LLhFUhFh8 h LLNHhq=<h%h LLhkchkcmHnHuhkchkcmHnHu hkchkcjh LLh LLUjh LLhFU h LLh LL.ϺкѺҺ!$&=GPRSX|ڻۻ"#EWɼʼ/0  ̿LǫhQhkchkcmHnHuhkchkcmHnHu hkchkcjh LLh8U h LLh8jh LLh8Ujh8UmHnHuh:\ hq=<NH h8NHh%hq=<h8hF7  JLNPZ\672g opq (02@ýùýùõ𵱫jphU hbNH h<NHh<hbh% hhGNHhhGh8 hQNHjhUhkc hkcjvhUjhUhhQ</0=YjkYZbhqrtu,-/t{DEJKLMNObkɿ磟 hx_thzhdhzh:\h=mNHaJ h$ aJ h=maJh$ h( NHh( hhGjh<0J<Uhf h=mNHh=m hb6] hbNHhbh<:kloqrs} 478=>?@BCEFJO]iz{ !о٬٬٬٬٣ٟvrih0OUh-0J>h>h0OUheR0J>h0OUh0OU0J>hyiheR0J>hN% heRNHheRh-h0OUmHnHuhemHnHuh>mHnHuh-mHnHuhWmHnHuh2mHnHuheRmHnHu hzNHhzh3-A hx_thzhx_thzNH)!$&SY[equxy:Zbctz  X]ghr|нƹнЬХССеС hhE2hE2 heRNHheR h>h-h0OUh>0J*hWhN%h0OUh>0J>h0OU h>NHh>h0OUh-0J* h-NH h-6]h-h0OUh-0J>h0OUh0OU0J>8mn2C_ 34 4Rfhn,RTVdhsomHnHuhAmHnHuh{mHnHuh-mHnHuh-h-mHnHu heRNHhsohkc hkcjh Ujh Uh hWNHhWheR hE2NHhE2h0OU5)=OV^r GJKLsw#&79AJXYeλβΎ΁{ hANHhAhoh-0J h-0J hdK6] hdKNHhdKh0OUh &[h0OU0J>h0OUh0OU0J> h-NHh &[h-0J>h{h-h-mHnHuhAmHnHuhE2mHnHuh0OUmHnHuhWmHnHu.ex~ ()9 xzXbd*OW_my.<>վÛÑhuC hNHh hdKNHh4Wh{ hohN% hN%NHhkc hkch-jjhoh-Ujhoh-Uhoh-NH hoh-hsohdKhAhN%9>Cde~"^~%678RS$'(橣機橛h6hu+ hzNHhzhzmHnHuheRmHnHu h bkheRhkc hkcjhdKUjhdKUhsohuChdKhN% h-NHh-h= "#,-.=FMN\emnyz}&:>HJ`hjĻ۵ۻ۵ۻۯۻۻᵠۻᵠhzNHaJh k6]aJ huCaJ hu+aJ h kaJhZ.6]aJhg6]aJhgNHaJ hgaJ hZ.aJ hzaJ hg6]hghsohz hzNH9 *6<F,^dnz|~.0JLhj~#*+콵˽ hsoaJhHNHaJ hHaJhzNHaJh k6]aJ hWwaJ hRaJ huCaJ hu+aJ hzaJ h kaJh kNHaJA(\YZ !!!!!!!!!gdjgdX  gdX gdzgdsJ1gd;gd6,gd0gdN*gd;gdegdBgd0+,dijk%&IJOPQghi |xjf[jhdhdUhejhBUmHnHuhz hnMNHhdhkc5\mHnHuhkc5\mHnHuh7hkc5\jdhdVUhRjhRUhnMh0NHaJh06]aJ h0aJh k6]aJ h kaJ hWwaJ hzaJhso6]aJ"@Bbdrtz|~ 8:jln@BV"$&R˿z˿t h0NHj^hdhN*U hdhN*jhdhN*UhN*h; hzNHhdVhdh0hz hdhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdhdUjhdhN*U hdhd-RTtv4BRTZtvx0LNP  ʾߺʾߺ~wiw~ʾ~jՕhdh6,U hdh6,jhdh6,Uh6,h; hzNHjXhdhN*UhN* h0NHh0 hj]hzhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdhN*U hdhN*j۔hdhN*U)JLj 08D&(68>@Tjln~xtphIrh0 hzNHjϖhdhN*U hdhN*jhdhN*UhN*hzhkchkcmHnHuhkchkcmHnHu hkchkcjRhdh6,U hdh6,jhdh6,Uhp h6,h6,h; h6,NHh6,*$&P"KL[\] %):;XYmoz{üjLhdhN*U hdhN*jhdhN*Uhi(hi(H*hphIrhN*h6,h0h; hzNHhzhi(h"n}D @JLnpr&(*VXxz "028żԵԣŜߑqjØhqhhqhUhqhjFhqhhqhUjhqhhqhU hqhhqhh;jɗhdhN*U hdhN*hqhhzNH hqhhzhzhN*jhdhN*UhkchkcmHnHuhkchkcmHnHu hkchkc,8:.0\^~  :<\^lntv|~ŷŷڈŷ~zvhehuCh4W hzNHj:hdhN*UjhqhhqhU hqhhqhhkchkcmHnHuhkchkcmHnHu hkchkcj@hdhN*U hdhN*jhdhN*UhN*hzjhqhhqhU+FHhjxz02BDFrt"$24:ʾʾ|qj\jqʾjhdh6,U hdh6,jhdh6,Uh6,hphpH*hphIrj4hqhhqhU hqhhqhjhqhhqhUhqh hzNHhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdheU hdhejhdheU$:<>|>@HJL^*,24:<H\^`npZv}hsJh#(huChkchkcmHnHuhkchkcmHnHu hkchkcj.hdh6,U hdh6,h6,hphpH*hp hzNH hphz hz6]h6,hz0J>hhzjhdh6,U1>@`bprxz8:XZbv   /1FTYɻдۇwsn hX ]huChOZhX 0J* hOZhX hX 6] hX 0J> hX NHhkchkcmHnHuhkchkcmHnHu hkchkcjhqhhX U hqhhX jhqhhX Uh6,hX 0J>hX hzaJhzjhzUmHnHu+Y`bds{|    " < = S T d e          ƻ~ hX NH\hkchkcmHnHuhkchkcmHnHu hkchkcj(hdhX U hdhX jhdhX U hX \ hX NH hX ]hOZhX 0J* hX 6] hOZhX hX hX 0J>h6,hX 0J>.    2 : z           > P R          j x ]fkruvչ}}x huC\h6,hX 0J>h@hX 0J>NHh"hX \ hX 0J>hkchkcmHnHuhkchkcmHnHu hkchkcjhqhhX U hqhhX jhqhhX U hX \]hX hX NH\ hX \h@hX 0J>. "BPemqtu{"#):;\c  !"23 &( *,6нЁxЁhjhX 0J* hX NHhkchkcmHnHuhkchkcmHnHu hkchkcj"hdhX U hdhX jhdhX UhIhX H*hX hX NH\hTOhX 0J>NH huC\ hX \] hX \hTOhX 0J>/68HZ\^lnrBCv~89[\ľ̵ hsJ0J>hrkhsJ0J>hsJhjh rd6]h rd hM60J>hrkhz0J> hrkNHhrkhhz hX 0J>hrkhX 0J* hjhX hX 6]hrkhX 0J>hX hX NH8 ^`hpTZdhvxz0>JLxz÷÷{tfjhqhhrbU hqhhrbjhqhhrbUhrbjhdhrkUhIhjH*hjhjh rd0J> h rdNHh rdhkchkcmHnHuhkchkcmHnHu hkchkcjhdhrkU hdhrkjhdhrkUhrkhz(z$246`lnr|JZv¼¸°°¸{͐unn hhz hzNHjhdhWU hdhWjhdhWUhIhjH*hzhrkhj0J>hjhWhrk hNHh hhhkchkcmHnHuhkchkcmHnHu hkchkcjhqhhrbU hqhhrb+   <`r~(^` Ttxܿܿ{whIhjh20J* h26]hjh20J>h2 hzNH hrb0J>hjhV\0J>hrb hV\NHhV\hWhz hhhhzNHhkc hkchjhhzUjhhzU hhz.(*JLZ\bd    $ * + , 0 1 3 4 6 7 9 : L N k t  ʾ}}}yqmdmhIhrb0J>hrbhIhjH*hj hj0J>hjh20J>NHhjh20J* h26]hjhrb0J>hjh20J>h2 hzNHhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdhWU hdhWjhdhWU$                     !!""" ""8"\"^"""""###"#&#*#:#<#N###õʮِي݂݂|shIh20J* h2NHh2huC hzNHhV\hkchkcmHnHuhkchkcmHnHu hkchkcjhdhWU hdhWjhdhWUhWhzhIhIH*hIh[hIhrb0J>hrb hrbNH-####!$+$,$4$:$G$L$V$|$~$$$$$$$$$$$$$$$$%%%%%%5%N%k%l%m%%%%%%%&&ùùֵ֪hkchkcmHnHu hkchkcj hqhh[U hqhh[jhqhh[UhrbhIh[0J>>*hIh[0J> hINHhIh[ h26]hIh20J>h2hIh4W0J*hIh20J*/&& & &8&L&N&b&&&&&&&'$'L'd'f'l'n''''''' ((((((&('(+(/(m(q(t(u((((((((((())))+)4)>)৷৷֞ҚҚhWhz0J>hzhIhrb0J>hIh90J>>* h9NHhIh90J>hIh4W0J*hIh90J*hIhrb h[NHh9h[jhqhh[UhkchkcmHnHu9>)@)H)I)_)`)p)q)*** * *F*T*V*X*v********$+&+F+H+V+X+^+`+d+++++++++++ ,̾̾|tphrbh`h`H*hIhI0J> h!!0J>h!!jhdh9Uh`hV\hWhz0J>hzhkchkcmHnHuhkchkcmHnHu hkchkcjhdh9U hdh9jhdh9Uh9h9h9H*, X*..37<<?????@C JVJKygdI , ,",n,p,,,,,,,,,,-- --"-l-v----------.$.'.1.2.8.l.............//ƾƷ꫓hWhz hV,h9hIhId0J> hV,NHhIhV,0J>hIdhV,h!! h>yh!!h>yh`H* h>yh`hIh`0J>hkc hkcjh~wUh>yjh>yU h`NHh`2/ /6/7/G/H/000 0>0@0r0|0000000000000&1(1H1J1X1Z1`1b1r1t1|1111111ʾʾ􇺇~uh>yh!!0J>h>yhV\0J>hV\j{hdhWUhWh>yh>yH*h>y hMK{hzh!! hzNHhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdhWU hdhWjhdhWU*11111122 2)2*2/21292:2>2T2X2Y2a2e2l2222222222 3333,3C3D3E3[3\3l3ȿȻȥșȍșȉ~wiwjhdhWU hdhWjhdhWUhW hzNH hV\NHhV\h>yh>yH*h>yh>yh>y0J*h>yh;;0J*h;;hWhz0J>hzhV\NHmHsHhV\mHsHh>yhV\0J>h>ymHsHh>yh!!0J>h!!mHsH)l3m3444 45F5H5V555555566666#6$6(6*6+6/6I6K6O6[6`6~66»ϫynyf^VhzmHsHh`mHsHh;;mHsHh>yh>y0J*NHh>yh>y0J*h>yh;;0J*h>yh;;6]h;;h>yh;;0J>h>yh`0J>h>yhz0J> h>yhz hz6]hWhz0J>hhyhzhkchkcmHnHuhkchkcmHnHu hkchkcjhdhWU!66666666666666666666666677777#7)7]777777777777໯৞|tkkkh~wh;;0J>h~wmHsHh~wh~w0J*NHh~wh~w0J*hyEmHsH h~w0J>h~whId0J>hIdmHsHh~wh;;6NH]h~wh;;6]h>yh;;0J*hzmHsHh>yh;;6]h;;h;;mHsHhWhz0J>hWhz0J>NH*7777777788 88808283888;8<8I8M8P8U8888888888888888888888888&9(9;9Q9R9U99ϿǢǢǚ׉h~wh0J>h#mHsHhmHsHh~wh@0J>hWhz0J>NHhWhz0J>hzmHsHh@mHsHhV\mHsHhmHsHhyEmHsHh~wh;;0J>h~wmHsHh;;mHsH599999999999999999A:B:J:L:M:N:Y:Z:d:m:t:|:::::::::::::::::ߩߩߥۥ|qhxh~wmHsHh~wh 0J*NH h~w6]h~wh 0J*h~wh6]mHsHh h~wh 0J>>*h~wh0J*hh~wh0J>h~w0J*6]h~wh~w0J*h~wh~wh 0J>h~wmHsHhmHsHh mHsH,::::::;;;;;%;);7;^;_;j;t;u;;;;;;;;;;?<J<K<M<N<R<W<`<<<<<<<<ȾȾګڔڔګڔڔڔ{sh_;mHsHh mHsHh< mHsHhuMh< 0J>huMh0J>hxhuM6]mHsHhuMmHsHh~wh~wmHsHh~wh~w0J>>*h~wh~w0J>h~wh0J>hmHsHjh~wh0J>Uh~wh 0J>h~wmHsH*<<L====>>>'?3?8?\?v???????????@)@7@E@N@T@U@\@]@a@@@@@@@@@AAAA"A&A8A:A hrE\ hvzhzhPy h^ENHh^EhvhMhn hzNHh3-AhrEhzhs`cheR\h:\mHsHh#"mHsHh4[mHsHhuMmHsHh@ mHsHhWmHsHh< mHsHh9mHsH1:AAAAAAAAAA B#BOBPBBBBBBBBBBBBBBBBBBBC CCCCC&C?CFC[C\CbCsCtCuCCCCCɼhszh2hzNHhH{h^E hzNHh~ h0ZNHh0Z h2hzhz h'vFhz h'vFh0Z h'vFNHh'vFhKhzNHhPyhrEhM hKhz hc\ hM\ h0ZNH\ h0Z\2CCCCCCCCDDD D&D.D:DIbIdIfIIIIIIIIIIIIIJJJJJ JTJUJVJJJJJK KK@KSKdKpKqKuK{KܿܳܯԜԜh9 h2hXhXh0Z hA&hzh~hc hPlNH h4WNHh4Wjh$U hPyNHhH{hPyhPlhkc hkcjh$Ujih$Uh$7{KKKKKKKKKKKKKK L2LILQLRL^LeLfLLLLLLLLMMEMrMsMzMMMMM"N&NQQQXQ_QcQfQqQQQQQQQQQR4R5RoRpRqRxRRRS&S\SjSoSuSvSSSSɾɾɷɷͱͷŭɷjh*Uh*h hFNH hIhH{ hghH{h'hH{hFhX h*NHh*h$hkc hkcjh$Ujch$U>SSTTT T TTTTTBT`TpTrTTTTU@UUUUUUUUUVV V%V.V/V1VKVVVvVVVVVVVV W W)W+W/W1WLWoW}WWWWWWWWмеееееееƱееббh2hH{NH h'vFhH{ h'vFh'vFh'vF h2hH{hhH{6] hH{NHh.}hH{h hIhH{h*hkc hkcjh*UjhGU:hh8whzjh8UmHnHuh,=hh4W h h4W h h* h h.} h h5 h h$ h.}NHh.}h$h* hZNHh5hZ2xY&^ f|mqty?| ~U~~ր;s8JH!!!!!!!! !!!!! & F/gdFgdj[ gdFgdcgdIgd~gd>egd`gdqgdqgdGgdogd1gd50_2_4_`_b____________``` `&`(`.`T`\`^````````````aa$a˿˿w˿qmehWdhz]hz hWdNHjQhh/ nU hh/ njhh/ nUh/ njԪhhWdUjhhWdU hhWdhWdhkchkcmHnHuhkchkcmHnHu hkchkcjWhhUjhhU hh'$a0a2a4a6aZa\aaaaaaaaabb b bb.b@bBbbbbbbbPcRcZczcccccccccpdddddd׻x hh1jhh1Uh'hWdhz0J>h4W hzNHhkchkcmHnHuhkchkcmHnHu hkchkcjΫhhU hhjhhUhh0J>:hh5h1huhz.ddee"e$e*e,e.eJeeef ff ffffffffffffffff ggggmgtggggggʾߺ}t}ijhhUhh0J*hhhz0J>hhz0J*NHhhz0J*h' hzNHhzhsz hGNHh1hGhkchkcmHnHuhkchkcmHnHu hkchkcjhh1U hh1jKhh1U)ggggghhh h0hnhhhDiFiiijj|j~jjjjjjjjjjkkkkhk|kkkk˿໷{˿tkbhohz0J*hohz0J> hWdhzjEhhoU hhojhhoUhhz0J>hhhz0J* hzNHh1hzhkchkcmHnHuhkchkcmHnHu hkchkcjhhUjȬhhU hh%kkkmm4m6mhmjmlmnmtmvmzm|mmmmnn$n%n&n[n\nnnnnnnooo@oAoBoXoYoiojoppp pֺuhkchkcmHnHuhkchkcmHnHu hkchkcj?hhoU hhojhhoUho hyDhz h2hz hzNHh1hWdhzhkc hkcj­hHUjhHU hHNHhH. p pJpfppppp qq>qLqqqr r$r%r>r[rrrrrrrr's(sYswsxsys{s|ssssssssssVtWttttt u\u]u^utuuuuuvغį hkchkcjhhoU hhojhhoU h1NHhuhoh2hzNH h2hzh1hsz hzNHhohz0J>hohz0J*hzhWd:vvv vvvvvvvvvvvw*wNwbwvwwwwwwxxx x/x0x3x4xBxGxUxpxvxxxxxxxxyyMyWyYybynyoyyyyyyyyyyξηױ׭קέק׭ΜקטױקɱξhO hthzh;CS hzNHh1 hWd0J> hpAhzhohz0J>NH hz6hohz0J>hzjhhoUhkchkcmHnHuhkchkcmHnHu|?|O|Q|l|q|t|u|||||||4}ֶֽ֫֫֫֞hzV"h;CShzNH h;CShzh} h;CSh;CSh' hFNHh;CS6H*] h;CS6]hF hGNHhG h;CSNHh;CShz hz6hohz0J><4}=}E}K}Q}e}o}y}}}}}}}}} ~Z~g~~~~~~~~~~)23ceDEko|ӀԀՀր h+hO h+6] h+NH hq6]h` he!6] hqNHhe!h'h+hqhzV"hzV"0J*hzV"hzV"hzV"0J>C :>@Alp݁./8:;rsjk01[:@AD.TmĽ޹ަަަ޵޵ަަޢhch2hzNH h{pNHh{phI h2h~h~ h'vFNHh'vFh}hzhF h2hzh2h>eH* h2h>eh>eh`h;yq@Ćֆ׆߆LJ̇)*678Jmvx<=Ίъ؊يڊ쿻hkc hkcj9hR8hFUjhR8hFU hR8hF hj[ NH hFNHhj[ hF h'vFhzh'vFh'vFNH h'vFh'vFh'vFheh>e hzNHhz h2hzh2hzNH4  FH\^Ȍʌ "(*,8uv=>WX 0246@B髶闶j-hFUhkc hkcjhFUjhFU hFNHj3hR8hFUh8 h8jhR8hFUhF hR8hFjhR8hFU;J|>Ӽ^2b<=PD ! ! !!!gd&J9gdub^gd[mgdgdO{"]"gdFgd/gdgd=BgdF1gdFBƕȕ"$&(.0xz,.68:<BDnؚښ @BbdflnȽȯȽȽȀȽڽ|hj[ j!hchFUhkchkcmHnHuhkchkcmHnHu hkchkcjhchFUjhchFU hchFj'hFUhkc hkcjhFUjhFU hFNHhF034BXx|~̞Ξܞޞ>Pnp2ؠڠ  *,.@ڡܡõڐõ|w h=BH* h=BNHh=B h'NHh'j *hc7hFUhj[ hFH*hkchkcmHnHuhkchkcmHnHu hkchkcj *hc7hFU *hc7hFj *hc7hFU hc7hF hFNHhF.΢ҢӢ *,LNvxܦ  8:Z\jlrtڧܧĨ̾y̾j *hc7hFU *hc7hFj *hc7hFU hIhF hghF hFNHhkchkcmHnHuhkchkcmHnHu hkchkcj *hc7h=BU *hc7h=Bj *hc7h=BUhFh=B-Ĩƨ*+ͪΪ!"ӫԫ(*,.46>@yzͮή#$ @{|ƺ۶~~h=Bjh=BUmHnHuhkc hkcjhFUjhFUh' hFNHhFhkchkcmHnHuhkchkcmHnHu hkchkcj *hc7hFU *hc7hFj *hc7hFU0аѰ&'wx~46hjlnxzabijƵǵij $68dfʹ̹hkchkcmHnHuhkchkcmHnHu hkchkcj  *hc7hFU *hc7hFj *hc7hFUh}hFH*hkc hkcjhFUjhFU hFNHhFhZ5ֺ׺ܻݻ45   TVpz>/0mn}HI_`ŷ hF6] hNH hhh+Fh h+bNHh+bhkchkcmHnHuhkchkcmHnHu hkchkcj *hc7hFU *hc7hFj *hc7hFU hFH* hFNHhF3  9:67{}~  #$()/7=>?@AGH[mvwx~ŨŞhchF6]hchF0J>>*h{hF0J>hchF0J*NHhchF0J*hchF0J> hF0J>hahF0J>hKghF6] hFNHh hF6]hFhuC:~xy23  xyTVj *hc7hFU *hc7hFj *hc7hFUh'hchF0J>hhuC hFNH h&phF hF0J>hahF0J>hchF0J*hF>'(HJKLIN^c  3567fgοοΩΥοΜοΜΒΉΜΒΉΜΒΉh;?hF0J*h;?hF6]h;?hF0J>hhchF0J> h&phF hF0J>hahF0J> hFNHhFj *hc7hFUhkchkcmHnHuhkchkcmHnHu hkchkc68@RX89jkbp23  '<EF_` hkc hkcjh/Ujh/Uh/hDahuCh-zhFH* hdhF hFH* h&phF hF0J>hahF0J>h'h;?hF0J> hFNHhF; @Vr   RTVX^`bdjz|.VW%456LM]^¾孺hkchkcmHnHu hkchkcjzhhFUjhhFU hhF hFNHhO{hkc hkcj *hhDaUhj *hUhF h/NHhDah/jh/U2  &pr .0>@FHJLNX ǹβ貤وقzozjfzhkc hkcjhFUjhFU hFNH hFH*hhFNHjthhFUhkchkcmHnHu hkchkcjh}_hFU h}_hFjh}_hFUhFhO{ hhFjhhFUhkchkcmHnHu(<ZJZ\/o  '(z{+,56=> DE[˽jh-hFU h-hF hFH*jh8UmHnHuhkchkcmHnHuhkchkcmHnHu hkchkcjnh}_hFUjh}_hFU h}_hFhO{ hFNHhF2[\lm @Bdf(*028PRXjDGHLMjkSʾʾꛖꛈh[mhO{ hFNHh'h hFH*hFjhh-hFUh-hFH*h-hFNHhkchkcmHnHuhkchkcmHnHu hkchkcjh-hFU h-hFjh-hFU2ST"^_bc  FHtvXZ@BDPr>Vۿԗۿ򐉐 h'NHh' h[mh[m h[mhFjbh-hFUhkchkcmHnHuhkchkcmHnHu hkchkcjh-hFU h-hFjh-hFU h[mNH hFNHhFh[m h[mH*1.:NP|~Jpr&lmqv󯧯ߧۯ hub^NHjVhub^Ujhub^Uhub^jپhFU hFNHj\h[mUjh[mUh[mh'hkc hkcj߽hFUhFjhFU8TU 8:R&(HJXZ`b NR7lmۿԗۿ hophF hFH*h&J9jPh-hFUhkchkcmHnHuhkchkcmHnHu hkchkcjӿh-hFU h-hFjh-hFU hFNHhFhub^jhF0J<U3,-EF_` 04jvx&(l&(HJXZ`bn~Ѽxleah hMK{hzhkchkcmHnHuhkchkcmHnHu hkchkc hhjJhhHUjhhHU hhHhhz6 hhzhH hzNHh+hzhkc hkcjh0oUjhOUhO hIhFh{p&- Vq Zr r 3t v v ~  Ѐ  !!!!!!!!gd2;gdbgdffgd8gdffgdZgdZgd`,/gdIgd Igd2;gdIgd0ogdOgdgd{pCDEs*>?  !mnrt%@OPghǻǻ񷩷 hNHjDhhU hzNHhEhhzhkchkcmHnHuhkchkcmHnHu hkchkcjhhU hhjhhUhO:hv   D T ~           "     * , < @ Ϳ~qj h0ohzhmuhz6\]aJ hzH*hkc hkcj>h0oUjhOU hzNHh' hONHhOhkchkcmHnHuhkchkcmHnHu hkchkcjhhU hhjhhUhhzhuC'@ B D p r                      "   Htv(*,XZ˿˿wijh0oh0oUjh0oh0oUhzh0oh0oNHh0o h0oh0oh0ohzNHj8h0ohU h0ohE h0ohzhkchkcmHnHuhkchkcmHnHu hkchkcjh0ohUjh0ohU h0oh)Zz|  "#+0)*2[\hjvxɿɿɸɿɿɜɘjhHUmHnHuhEhjhzH*j2h0oh0oU h0oNH h)hzh0o hzNHhzhkchkcmHnHuhkchkcmHnHu hkchkcjh0oh0oU h0oh0o424XZ\-QUjkVW`acmw (6CFSTǹ㤛h'hIh mHnHuh0omHnHuhzmHnHuhkchkcmHnHuhkchkcmHnHu hkchkcjh0oh0oU h0oh0ojh0oh0oUh0o hzNHhzh<6T%&*+24RWXY[BDFHNPRTV  *,2BD깵ĭ hrNHh&h`,/jhrUjhrUhrh{p hmuhhkc hkcj,hUjhU hNHh2;hhhI hzNHhz: @BDFLN`vJNT\f p p p p _p lp mp wp p p p +q ,q 1q Aq sq |q q q q q q q Rr Sr Zr qr r r r r r r r r r 's -s ƾƺƺƺƺưƾƺƾƬƺƬƬƨhyphZh`h`,/hM hINHhI hNHUhh I hw|NHhw|hkc hkcj&hrUjhrUhrhX h&>etitions held in 2007 and 2008. In these two years of the Morpho Challenge, systems competed in a Linguistic Evaluation of up to five languages: English, German, Finnish, Turkish, and Arabic; and in a Task-Based Information Retrieval (IR) Evaluation in up to three languages: English, German, and Finnish. The scores from both the Linguistic and Task-Based Evaluations of the 2008 competition are directly comparable to scores from the 2007 Challenge because: 1. The Linguistic and Task-Based IR Evaluations of the 2007 and 2008 Morpho Challenge competitions used the same methodology; and moreover, 2. The more recent 2008 challenge scored systems over the same corpora and against the same answer keys as the 2007 competition. While comparing systems from 2007 and the 2008 Morpho Challenge competitions is reasonable, comparing the scores of even the same system across different languages is meaningless. Each language has radically different morphological structure and the Linguistic Evaluation uses different morphologically annotated answer keys for each language. An early prototype of the ParaMor morphology induction algorithm competed in Morpho Challenge 2007 (Monson et al., 2008a). And the fully developed ParaMor algorithm participated in the 2008 Challenge (Monson et al., 2008b). Only the improved performance from the 2008 Challenge is reported for ParaMor in this thesis. In the Morpho Challenge held in 2008, ParaMor took part in all five language tracks of the Linguistic Evaluation, and in all three language tracks of the Task-Based IR Evaluation. The next two sub-sections detail ParaMors performance in the Linguistic and Task-Based Evaluations respectively. Linguistic Evaluation Results from Morpho Challenge 2007/2008  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 summarizes ParaMor s performance in the Linguistic Evaluation of Morpho Challenge 2008 and places ParaMor s results in the context of the best performing systems from the 2007 and 2008 Challenges.  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 contains the precision (P), recall (R), and F1 scores of nine individual unsupervised morphology induction algorithms for the five languages of the Linguistic Evaluation. Six of the nine systems in  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 competed in Morpho Challenge 2008, while three systems participated in the 2007 challenge. Of the six systems from  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 which competed in the 2008 challenge, three are systems that I submitted (Monson et al., 2008b), while three are systems built and submitted by others. The three systems which I entered in Morpho Challenge 2008 are: 1. The ParaMor system alone, in the 2nd column of  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 2. An instance of the unsupervised morphology induction system Morfessor (Creutz, 2006) which I trained myself, 3rd column, and 3. A joint ParaMor-Morfessor system which combines the analyses of 1. and 2., in the 1st column of  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5. The joint ParaMor-Morfessor system developed to leverage the complementary strengths of the ParaMor and Morfessor systems. As uncovered in Section  REF _Ref216858117 \r \h  6.3: ParaMor excels at identifying inflectional morphology, while the Morfessor system discovers the most regular and frequent morphological structures of a language, whether inflectional or derivational. As discussed in Section  REF _Ref216926683 \r \h  6.1.1, the Morpho Challenge competition permits a system to submit more than one, ostensibly ambiguous, analysis of a single word. The ParaMor-Morfessor system joins analyses from the two systems by simply adding Morfessors segmentation of each word to ParaMors as a separate (ambiguous) analysis. The ParaMor algorithm has several free parameters that control the paradigm discovery phase. These parameters were set to values that produced reasonable Spanish paradigms. The parameters were then frozen before learning the morphology of the languages in the Morpho Challenge. In the experiments which combine ParaMor and Morfessor analyses, Morfessors single free parameter was optimized for F1 separately for English, German, and Turkish. To optimize Morfessors parameter for these three languages, morphological answer keys were constructed from pre-existing morphological data and tools. The source for the English and German morphological answer keys was the Celex database (Burnage, 1990); while in Turkish, a hand-built morphological analyzer provided by Kemal Oflazer was used as the basis of a morphological answer key. Having limited access to morphologically annotated Finnish and Arabic data, Morfessors pa-  rameter was not directly optimized for these languages. Instead, as Finnish and Arabic both have rich morphological systems, Morfessors segmentations for these two languages were generated using the parameter value which performed best on the morphologically complex Turkish language. The six systems which appear in  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 that were prepared by groups other than Monson et al. (2008b) are those systems with the top competing performances in the Linguistic Evaluation of the Morpho Challenge competitions from 2007 and 2008. The system labeled Morfessor MAP is a second instance of the same Morfessor algorithm that I trained, submitted, and joined with ParaMor. But the Morfessor MAP submission was prepared by Kurimo and Varjokallio (2008) and likely uses a different parameter setting for each language from the settings used in my Morfessor submissions. A change in parameter setting can sometimes result in quite different performance for Morfessor, c.f. Finnish. The remaining five systems, found in the right-most columns of  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5, bear the names of their principle authors. If multiple versions of a single algorithm competed in the 2007 and/or 2008 Morpho Challenge, then the scores reported in  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 are from the algorithm variant which attained the highest F1 score. Because the number of morpheme-sharing word pairs that were used to calculate the precision and recall scores in the Linguistic Evaluation was quite large, English used the fewest word-pairs, 10,000, most score differences between systems in  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 are statistically significant. All F1 differences of more than 0.5 between systems which competed in Morpho Challenge 2007 were statistically significant (Kurimo, Creutz, and Varjokallio, 2008); and similarly, all F1 differences of more than 0.5 among the Morpho Challenge 2008 systems are also statistically significant (Kurimo and Varjokallio, 2008). To begin the examination of the results of the Linguistic Evaluation of Morpho Challenge, turn to the 2nd column of  REF _Ref217013137 \h  \* MERGEFORMAT Figure  6.5 and examine the precision, recall, and F1 scores that the ParaMor algorithm achieves alone. In all language tracks but Arabic, ParaMor holds its own against the state-of-the-art unsupervised morphology induction systems which competed in Morpho Challenge. Against the six best-performing systems from the 2007 and 2008 Morpho Challenge competitions that were not prepared by Monson et al. (2008b), and looking at F1, the ParaMor algorithm places first in Turkish with 46.5%, second in English with 52.8%, third in Finnish at 39.5%, and fourth in German with 44.5%. Although ParaMors recall scores are consistently lower than precision for each language (in English precision is 58.5% vs. a recall of 48.1%, in German precision is 53.4% and recall 38.2%, etc.), the precision and recall scores of the lone ParaMor algorithm are often more balanced than the scores of other systems. ParaMors balance in precision and recall is particularly noticeable in Turkish, where the morpheme recall of other unsupervised systems is anomalously low. In Turkish, the spread between ParaMors precision, 56.7%, and recall, 39.4%, is 17.3 percentage points. In contrast, the smallest spread in any competing system is 47.0 percentage points in the Zeman (2008) system. ParaMors focus on a recall-centric morphology induction procedure (see Chapter 3) and on a segmentation procedure that can propose more than one morpheme boundary in a single word (see Chapter 5) pays strong dividends when analyzing the highly agglutinative Turkish language. Although ParaMor alone performs respectably, it is when ParaMors analyses are adjoined with Morfessors that ParaMor shines. In German, Finnish, Turkish, and Arabic, the combined ParaMor-Morfessor system achieves the highest F1 of any system which competed in the 2007 or 2008 Challenges. And in English, the joint ParaMor-Morfessor system places a strong second. It is the precision score of the joint ParaMor-Morfessor system that drags the English F1 under that of the first place system, Bernhard (2008). In Finnish, the Bernhard systems F1 is likely not statistically different from that of the ParaMor-Morfessor system. The joint ParaMor-Morfessor system attains its higher F1 scores by balancing precision and recall. The trend for precision to be above recall that was noted for ParaMor is even more pronounced in the Morfessor system: Morfessors lowest precision and highest recall scores both occur in German: 67.2% precision and 47.6% recall. Because of the elevated precision scores in both ParaMor and Morfessor, a boost in the harmonic mean of precision and recall can be achieved by sacrificing precision for a decent gain in recall. This tradeoff of precision for recall is exactly accomplished by adjoining the morphological segmentations that are proposed by the ParaMor and Morfessor systems. Combining the analyses of any two systems will increase the total number of morphemes in the analysis of each word likely lowering precision but possibly increasing recall. But Section  REF _Ref216858117 \r \h  6.3 suggests that the morphological analyses that the ParaMor and Morfessor systems propose are particularly suited for joining: the mostly inflectional morphemes that ParaMor identifies differ substantially from the mix of inflectional and derivational morphemes that Morfessor induces. Hence, although combining ParaMors analyses with those from Morfessor almost always hurts precision, in all five languages, the improvement in recall significantly boosts F1 over that of either ParaMor or Morfessor alone. The performance of the ParaMor-Morfessor system is particularly striking in Turkish and in Arabic. In Turkish, the recall of the joint ParaMor-Morfessor system is double that of all non-ParaMor Turkish systems. This high recall leads to an improvement in F1 over the next best system, Morfessor alone, of 13.5% absolute or 22.0% relative. Contrasting with ParaMors strong performance at Turkish, the language that the lone ParaMor algorithm performs most poorly at is Arabic. New to Morpho Challenge in 2008, Arabics morphology is distinctly different from that of the other four languages in the challenge. Arabic morphology differs most notably in possessing templatic morphology, where a consonantal root is interleaved with vowels to produce specific surface forms. Equally important, from ParaMors perspective, is that Arabic is the only language in Morpho Challenge with significant prefixation: Not only does Arabic verbal morphology include inflectional prefixes, but Arabic orthography also attaches a number of common determiners and prepositions directly onto the written form of the following word. These attached function words act as prepositions in text. With no strategy for identifying prefixes, let alone templatic morphology, the ParaMor algorithm recovers just 8.5% of the morphemes in the Arabic words. Such low recall pulls ParaMors F1 down to 15.4%. It is some small consolation that no lone unsupervised morphology induction system recovered even a quarter of the morphemes in the Arabic answer key. Despite ParaMors own poor performance at uncovering the morphological structure of Arabic, when ParaMors analyses are presented in combination with Morfessors the increase in recall between the two systems is practically additive. Morfessors Arabic Recall is 21.0%, ParaMors is 8.5%, and the recall of the joint ParaMor-Morfessor system is 27.5%. This significant jump in recall implies very little overlap between the morphemes which the ParaMor and Morfessor systems identify. When recall scores are depressed, an increase in recall implies an increase in F1. And indeed, the ParaMor-Morfessor system receives the highest F1 of any system which analyzed Arabic morphology. In the near term, since prefixes are the mirror image of suffixes, a simple augmentation could allow ParaMor to analyze prefixation. The ability to identify prefixes would not only improve morpheme recall in Arabic, but help identify German verbal prefixes, and English derivational prefixes as well. The Task-Based Evaluation of Morpho Challenge 2007/2008 The Task-Based Information Retrieval (IR) Evaluations of the 2007 and the 2008 Challenge covered three languages: English, German, and Finnish. To measure the impact that a morphology analysis system has on an IR system, the Morpho Challenge competition replaced all words in all documents and queries from the IR corpus with the morphological analyses the morphology system suggests. In both years of the Morpho Challenge the same information retrieval corpus and query set were used, making results from 2007 comparable with results from 2008. The Morpho Challenge organizing committee did not measure the statistical significance of average precision scores in the IR Evaluation. Additional details on the procedure used in the Task-Based IR Evaluation of Morpho Challenge are given in Section  REF _Ref216926690 \r \h  6.1.2 and in Kurimo and Turunen (2008) and Kurimo, Creutz, and Turunen (2007). Figures 6.6 and 6.7 contain the results of the Task-Based IR Evaluation of Morpho Challenge 2007 and 2008.  REF _Ref217122670 \h  \* MERGEFORMAT Figure  6.6 contains the average precision IR scores for the eight best performing systems from the 2007 and 2008 challenges; while  REF _Ref217290313 \h  \* MERGEFORMAT Figure  6.7 contains average precision scores for four baseline metrics: 1. No Morphology: IR experiments are run over the raw documents and queries; 2. Snowball (Porter): All words in each document and query are stemmed using the Snowball package of language stemmers. In the case of English, the Snowball stemmer is the Porter stemmer; 3. Answer Key: Document and query words are replaced with their morphological analyses from the answer keys that were used in the Linguistic Evaluation of Morpho Challenge. The answer keys used in the Linguistic Evaluations contain only a subset of the full set of word types found in the IR Evaluation; and 4. Two-Level: Each word is replaced with the morphological analysis provided by a hand-built rule-based morphological analysis system. No hand-built morphological analysis system was evaluated for German. In the IR Evaluation of Morpho Challenge the, combined ParaMor-Morfessor system placed first in English and German, and fourth in Finnish. The IR Evaluation is a black-box experiment, and so it is not completely clear why the ParaMor-Morfessor system fared worse in the Finnish track. The most likely explanation is that replacing each word in each document and query with both the ParaMor and the Morfessor analyses is inappropriate for a language with complex morphology such as Finnish. It is unfortunate that Morpho Challenge did not conduct IR experiments for the morphologically complex Turkish and Arabic. It would be particularly interesting to see ParaMors IR performance on Turkish, which, like Finnish, is agglutinative. Alone, the ParaMor system placed third in English, behind the joint ParaMor-Morfessor system and the system developed by Bernhard (2008). The ParaMor systems also perform well in comparison to the baseline algorithms of  REF _Ref217290313 \h  \* MERGEFORMAT Figure  6.7. Most notably, in all languages, both the lone ParaMor algorithm and the joint ParaMor-Morfessor system improve on the average precision scores the IR system achieves when no morphological analysis is performed. With no morphological analysis, the IR system scores 32.9% average precision for English, 35.1% for German and 35.2% for Finnish; while the joint ParaMor-Morfessor system scores 39.9% for English, 47.3% for German, and 46.7% for Finnish.  The best performing unsupervised systems, including the ParaMor-Morfessor system, also outperform the baseline Answer Key scenario: demonstrating that imperfect morphological analysis performed by unsupervised morphology induction systems can trump perfect analysis of a subset of words. ParaMor and the other unsupervised system face stiffer competition in the two hand-built morphological baselines that have some generalization capacity, Snowball (Porter) and Two-Level. The Porter stemmer has the best average precision of any method against English; but unsupervised systems, the joint ParaMor-Morfessor system among them, outperform the Snowball rule-based stemmers for both German and Finnish. And finally, although the hand-built two-level morphological analyzer improves average precision more than any unsupervised induction method does in Finnish; the joint ParaMor-Morfessor system edges out the hand-built English morphological analysis system. Conclusions and Future Work ParaMor, the unsupervised induction system developed for this thesis automatically discovers the morphological paradigms of a natural language from unannotated text in a three step process. First, ParaMor lays out a space of candidate partial paradigms and applies a recall-centric search strategy to that space (Chapter 3). Second, ParaMor merges candidate paradigms that likely describe portions of the same true paradigm ( REF _Ref190686481 \r \h  4.3). And third, ParaMor culls out the least likely candidates (Sections  REF _Ref190686436 \r \h  4.2 REF _Ref190686503 \r \h  4.4). With a firm grasp on the paradigmatic structure of a language, ParaMor then segments individual words of that language exactly when there is paradigmatic evidence that a particular word adheres to a discovered paradigm (Chapter 5). The paradigmatic structures that ParaMor identifies organize inflectional morphology by grouping mutually substitutable suffixes into paradigm-like structures. With its focus on inflectional paradigms, ParaMor contrasts with other unsupervised morphology induction systems, such as Morfessor (Creutz, 2006), which seek to identify all morpheme types whether inflectional or derivational. With their emphasis on very different aspects of morphology, ParaMors morphological analyses are largely complementary to those of a system like Morfessor. And this thesis leverages the individual strengths of the general purpose morphology induction system Morfessor and the inflection specific system ParaMor by combining the analyses from the two systems into a single joint analysis. The combined ParaMor-Morfessor system competed in the 2007 and 2008 Morpho Challenge competitions, which evaluated unsupervised morphology induction systems at morpheme identification. In four of the five language tracks of the 2007/2008 Morpho Challenge competitions, in German, Finnish, Turkish, and in Arabic, the ParaMor-Morfessor system achieved an F1 score for morpheme identification at or above that of all other systems which competed. The primary reason for the combined ParaMor-Morfessor systems success at morpheme identification is its strong morpheme recall. In Turkish, the combined systems recall, at 52.1%, is twice that of the next highest system. Results from the Morpho Challenge competitions also suggest that the ParaMor morphology analysis system is helpful in higher-level natural language processing tasks. Augmenting an Information Retrieval (IR) system with the morphological analyses that are proposed by the ParaMor system alone significantly improves average precision over a morphologically nave baseline. ParaMors improvement over the baseline IR scores occurs in all three languages of the Task-Based IR Evaluation at Morpho Challenge. Furthermore, the IR average precision scores of the joint ParaMor-Morfessor system place first among all competing systems in English and German. Despite ParaMors successes, this thesis, naturally, could not exhaustively investigate all areas of the ParaMor algorithm. Nor could this thesis explore every aspect of unsupervised morphology induction in general. The next two sections outline particular suggestions for extending ParaMor and specific recommendations to all who will pursue unsupervised morphology induction in the future. Improving the Core ParaMor Algorithms Developing ParaMors large suite of algorithms necessitated prioritizing implementation. The foremost priority for this thesis was to create a complete system that could both identify paradigms and then segment word forms. Unfortunately, focusing on the full ParaMor system required limiting the investigation of both the paradigm identification algorithm as well as the segmentation algorithm individually. Beginning with the paradigm identification algorithm and moving toward the segmentation algorithm, five specific pieces of ParaMor s current algorithms warrant further work. First, although ParaMor s scheme-clustering algorithms of Section  REF _Ref190686481 \r \h  4.3 significantly reduce the fragmentation of true paradigms across ParaMors paradigm models, the final scheme-clusters remain disjointed. For example, among the 42 scheme-clusters that ParaMor outputs over a training corpus of 50,000 Spanish types, nine separate clusters model portions of the ar inflection class of Spanish verbs. But ParaMor was unable to merge these nine clusters because ParaMors csuffix based discriminative restriction on clustering, described in Section  REF _Ref215840852 \r \h  4.3.1, prevents any cluster from containing a pair of csuffixes that are not mutually substitutable on at least one cstem. The verbal csuffixes ndoles and arme, for example, which are each built of a non-finite verb suffix in combination with a pronominal clitic, did not both attach to any single cstem in this Spanish corpus. Although both the ndoles and the arme csuffixes belong to the ar inflection class, any one ParaMor cluster is prevented from containing them simultaneously. In failing to merge models which clearly describe portions of the same paradigm, ParaMor does not capture the full generality and power of paradigms. In a similar vein, ParaMors clustering restrictions prevent distinct inflection classes of the same paradigm from coalescing. For example, in Spanish, ParaMor keeps separate those scheme-clusters that model the ar verbal inflection class from those clusters that model the er inflection class of verbs, which are separated again from models of the ir inflection class. On the one hand, it was an intentional choice to prevent ParaMor from merging distinct inflection classes of the same underlying paradigm: Although all verbs in Spanish inflect for the same morphosyntactic features, verbs which adhere to distinct inflection classes use distinct suffixes to mark the same feature sets, and sometimes even use the same surface suffix to mark distinct sets of features (see Chapter 4). While it is clearly important, then, to somehow distinguish between distinct inflection classes, ParaMors separate models prevents the segmentations of the 3rd Person Singular Present Indicative Spanish word forms bebe she drinks and habla she speaks, as beb + e and habl + a respectively, from containing matching suffixes. The second area of ParaMor that warrants further work is a scaling up of ParaMors paradigm identification algorithms from the relatively small vocabulary size of 50,000 unique types to the orders of magnitude larger vocabularies used in the Morpho Challenge competitions. While ParaMors segmentation algorithms can segment arbitrarily large corpora, in this thesis ParaMor included at most 50,000 types of any corpus in the paradigm induction vocabulary. There is no significant barrier to scaling up ParaMors paradigm identification algorithms to larger vocabulary sizes. Each step in ParaMors paradigm induction pipeline will scale with larger vocabularies: ParaMors on-demand instantiation of the morphology scheme network allows the scheme search procedure to scale reasonably with vocabulary size both in time and space complexity. And bottom-up agglomerative scheme clustering is cubic in time with the number of initial schemes. The primary reason this thesis did not scale up ParaMors paradigm identification algorithms is that ParaMors parameters were set over the smaller vocabulary size, and time did not permit empirical adjustment of the parameter settings to accommodate larger vocabularies. In contrast to scaling up ParaMors paradigm identification phase, the third ParaMor-specific question this thesis does not answer is how ParaMors segmentation algorithms scale down to smaller vocabulary sizes. The larger the vocabulary that ParaMor segments, the more likely ParaMor will be to find evidence that a word belongs to a particular paradigm. The vocabularies of the languages in Morpho Challenge are very large. But an unsupervised morphology induction system might be most useful for languages with limited machine readable data available. Thus, it will be important to scale ParaMors segmentation algorithms down to smaller vocabulary sizes. In ParaMors current morphological segmentation algorithm, a word-final string, f, of a particular word form, w, must be mutually substitutable with another csuffix of a paradigm before ParaMor will place a morpheme boundary in w before f. Perhaps in scenarios where more limited data is available, ParaMors morphological segmentation algorithm should require less evidence that a stem adheres to a particular paradigm before segmenting. A fourth area that this thesis did not fully investigate is the effect that ParaMors five free parameters have on ParaMors ultimate morphological segmentations. ParaMors free parameters directly control the induced paradigms. ParaMors first parameter is the cutoff in the morphology scheme network search on when a parent scheme is deemed likely to correctly extend the coverage of the current scheme (Chapter 3). The second parameter is the lower boundary on the string length of the word forms used in the paradigm induction corpus (Section  REF _Ref190686436 \r \h  4.2). The third parameter decides when to discard a scheme-cluster based on the number of word types a scheme-cluster covers (Section  REF _Ref190768256 \r \h  4.4.1). And the fourth and fifth parameters govern the suffix-internal and stem-internal morpheme boundary error filters (Section  REF _Ref190771388 \r \h  4.4.2). ParaMor s five free parameters were set not by examining word-to-morpheme segmentations, but by examining candidate paradigms, i.e. schemes and scheme-clusters, that ParaMor created. The chronological development of ParaMors core algorithms largely followed the order of ParaMors processing pipeline. In particular, ParaMors word-to-morpheme segmentation algorithm was only developed after ParaMors paradigm identification algorithms were in place. Hence, no experiments directly measured changes in word segmentation quality as ParaMors parameters varied. As described in Section  REF _Ref216858117 \r \h  6.3, ParaMor s free parameters were intentionally set to maximize the recall of inflectional morphemes. Only empirical experimentation could determine if a more restrictive search parameter or more aggressive filtering of scheme-clusters would significantly improve ParaMors precision at the segmentation task without severely hurting recall. Finally, this thesis leaves to future work the development of a more satisfactory approach for combining the segmentations of ParaMor and Morfessor. The current algorithm does not attempt, for any particular word, to merge the morphological segmentations from the two systems. Instead, the procedure described in this thesis simply suggests each systems segmentation as an alternative analysis. A more sophisticated system could select from among the morpheme boundaries that are suggested by the two systems, and produce a single unified segmentation. In fact, a general purpose solution to the problem of combining segmentations that come from multiple morphology analysis algorithms could be used to combine segmentations produced by additional unsupervised morphology induction systems beyond ParaMor and Morfessor. The Future of Unsupervised Morphology Induction This section looks beyond specific extensions to the ParaMor algorithm to the broader picture of the steps future unsupervised morphology induction systems must take. The field of unsupervised morphology induction is still in its infancy. Although unsupervised morphology induction is a large and complex problem, implemented systems almost universally focus on a narrow slice of the most simple forms of morphology. Three areas where very little work has been done on morphology induction from an unsupervised perspective are: 1. Morphological processes other than suffixation, 2. Morphophonemics, and 3. Mapping from morphemes to morphosyntactic features. The next three sub-sections examine these underserved areas in turn. Beyond Suffixation This thesis joins most other unsupervised morphology induction work in addressing the most prevalent morphological process, suffixation. While other morphological processes play an important role in many languages, suffixation plays a significant role in the morphology of nearly all the worlds languages, including those that also employ other morphological processes (Dryer, 2008). Still, suffixation is only the most common morphological process among many, including: prefixation, infixation, reduplication, vowel and consonant mutation, templatic morphology (as in Arabic), as well as suprasegmental changes such as tone or stress. The next step in the field of unsupervised morphology induction will be to address these additional morphological processes. To date, the most promising work on unsupervised induction of non-concatenative morphology is Wicentowski (2002), which explicitly assumes an underlying morphology that is not purely concatenative. Like suffixation, prefixation is a concatenation operation, and is thus the most readily addressed process beyond suffixation. A number of unsupervised systems have already extended their reach to prefixation. Morfessor (Creutz, 2006), for example, agnostically discovers any concatenative morphological processes. Indeed, ParaMor could be easily extended to analyze prefixation by treating initial substrings of words as candidate prefixes in paradigmatic relationships and final substrings as the syntagmatic candidate stems. Morphophonological Change Not only does ParaMor restrict morphological analysis to suffixation, ParaMor also requires that suffixation to be purely concatenative. But in fact, suffixation is often not strictly concatenative. Suffixation, like all morphological processes, can be accompanied by phonologic changes. Even Englishs limited morphology is rife with examples of morphophonology: many nouns voice a final fricative before the plural, i.e. wolf becomes wolves not *wolfs. But more can change than just the word final consonant: both Finnish and Turkish, evaluated in the 2007 and 2008 Morpho Challenge competitions, have vowel harmony where vowels in a suffix change to match vowels in the stem. Although Wicentowski (2002), again ahead of its time, and Goldwater and Johnson (2004) begin to approach learning morphophonology, most current generation unsupervised morphology induction systems do not address morphophonology. In particular, neither ParaMor nor Morfessor (Creutz, 2006) address morphophonology in a systematic fashion. A logical next step for the ParaMor algorithm would be to extend the definition of a scheme from a strict concatenation of cstems and csuffixes to allow for phonological change when a stem and affix are joined. Mapping Morphemes to Features Finally, isolating the stem of a word by identifying and separating off inflectional suffixes, as ParaMor does, constitutes only the first step of full morphological analysis. A full analysis must also specify the set of morphosyntactic features marked by each identified affix. Probst (2003) is the only prior attempt I am aware of to automatically associate induced suffixes with morphosyntactic features such as Person, Number, or Tense. Being able to map morphosyntactic features onto discovered morphemes would be particularly useful for syntactic transfer based machine translation, where unification rules over morphosyntactic features restrict the set of applicable translation rules. Deriving a mapping from surface morphemes to morphosyntactic features requires knowledge of the grammatical features that exist in the specific language. Ongoing work in the Language Technologies Institute at Carnegie Mellon University (Clark, Frederking, and Levin, 2008) plans to acquire this knowledge of grammatical features through a process of feature detection.  REF _Ref195279807 \h  \* MERGEFORMAT Figure  7.1 pictures the general process of feature detection, where a bilingual informant aligns the words in pairs of translated of sentences. In this example, a bilingual informant translates the two English sentences  The tree fell and  The trees fell into Spanish. The feature detection system then compares sets of features that are associated with the English sentences. In  REF _Ref195279807 \h  \* MERGEFORMAT Figure  7.1, the feature detection system would find that the morphosyntactic feature structures associated with the pair of sentences are identical except for one feature valuethe Subject Number. To discover whether Subject Number is marked in Spanish, the feature detection system compares the Spanish words which are aligned to the head, to the dependent, and to the governor of the sentence subject. During these comparisons the feature detections system discovers that Subject Number is marked in all three possible locations, i.e. the head rbol vs. rboles, the dependent, El vs. Los, and on the governor cay vs. cayeron, because each of these surface string pairs differ. And now with knowledge of what features are marked in Spanish, csuffixes like es on rboles, and on rbol() can be associated with Singular and Plural Number. ParaMor: A Successful Morphology Induction Algorithm This thesis has empirically shown that the paradigmatic organization of morphemes can be leveraged to induce the morphological structure of natural languages in an unsupervised fashion. With algorithms specifically tailored to the unique structure of natural language, like the scheme-search and scheme-clustering procedures of Chapters 3 and 4, the ParaMor system is able to recover models of the paradigmatic relationships that exist between the inflectional morphemes of a language. The quality of the discovered paradigmatic models is validated both by direct comparison of the induced paradigm models against hand-crafted paradigm descriptions (Chapter 4) and by ParaMors state-of-the-art performance in the Morpho Challenge competitions held in 2007 and 2008 (Chapter 6).  Bibliography Altun, Yasemin, and Mark Johnson. 2001. Inducing SFA with T-Transitions Using Minimum Description Length. Finite State Methods in Natural Language Processing Workshop at ESSLLI, Helsinki, Finland. The American Heritage Dictionary of the English Language. 2000. Fourth Edition, Houghton Mifflin Company. Angluin, Dana. 1982. Inference of Reversible Languages. Journal of the ACM, 29.3, 741-765. Baroni, Marco. 2000. Distributional Cues in Morpheme Discovery: A Computational Model and Empirical Evidence. Ph.D. Thesis in Linguistics, Los Angeles, University of California, Los Angeles. Baroni, Marco, Johannes Matiasek, and Harald Trost. 2002. Unsupervised Discovery of Morphologically Related Words Based on Orthographic and Semantic Similarity. ACL Special Interest Group in Computational Phonology in cooperation with the ACL Special Interest Group in Natural Language Learning (SIGPHON/SIGNLL), Philadelphia, Pennsylvania, 48-57. Baroni, Marco. 2003. Distribution-driven Morpheme Discovery: A Computational/Experimental Study. Yearbook of Morphology. Beesley, Kenneth R., and Lauri Karttunen. 2003. Finite State Morphology. CSLI Publications, Palo Alto, California. Bernhard, Delphine. 2008. Simple Morpheme Labelling in Unsupervised Morpheme Analysis. Lecture Notes in Computer Science: 8th Workshop of the Cross-Language Evaluation Forum, CLEF 2007, Revised Selected Papers, Budapest, Hungary, Springer, 5152/2008, 873-880. Biermann, A. W., and J. A. Feldman. 1972. On the Synthesis of Finite-State Machines from Samples of Their Behavior. IEEE Transactions on Computers, C-21.6, 592-597. Bordag, Stefan. 2008. Unsupervised and Knowledge-Free Morpheme Segmentation and Analysis. Lecture Notes in Computer Science: 8th Workshop of the Cross-Language Evaluation Forum, CLEF 2007, Revised Selected Papers, Budapest, Hungary, Springer, 5152/2008, 881-891. Brent, Michael. 1993. Minimal Generative Explanations: A Middle Ground between Neurons and Triggers. Cognitive Science Society, University of Colorado-Boulder, Lawrence Erlbaum Associates, Inc., 28-36. Brent, Michael R., Sreerama K. Murthy, and Andrew Lundberg. 1995. Discovering Morphemic Suffixes: A Case Study in MDL Induction. The Fifth International Workshop on Artificial Intelligence and Statistics, Fort Lauderdale, Florida. Burnage, Gavin. 1990. CelexA Guide for Users. Springer, Centre for Lexical information, Nijmegen, the Netherlands. Casella, George, and Roger L. Berger. 2002. Statistical Inference. Second Edition, Duxbury, Thomson Learning Inc. Chan, Erwin. 2006. Learning Probabilistic Paradigms for Morphology in a Latent Class Model. ACL Special Interest Group on Computational Phonology, New York City, New York, 69-78. Clark, Jonathan H., Robert Frederking, and Lori Levin. 2008. Inductive Detection of Language Features via Clustering Minimal Pairs: Toward Feature-Rich Grammars in Machine Translation. Workshop on Syntax and Structure in Translation (SSST) at the Association for Computational Linguistics (ACL), Columbus, Ohio, 78-86. Creutz, Mathias, and Krista Lagus. 2002. Unsupervised Discovery of Morphemes. ACL Special Interest Group in Computational Phonology in cooperation with the ACL Special Interest Group in Natural Language Learning (SIGPHON/SIGNLL), Philadelphia, Pennsylvania, 21-30. Creutz, Mathias. 2003. Unsupervised Segmentation of Words Using Prior Distributions of Morph Length and Frequency. The Association for Computations Linguistics (ACL), Sapporo, Japan. Creutz, Mathias, and Krista Lagus. 2004. Induction of a Simple Morphology for Highly-Inflecting Languages. ACL Special Interest Group in Computational Phonology, Barcelona, Spain, 43-51. Creutz, Mathias. 2006. Induction of the Morphology of Natural Language: Unsupervised Morpheme Segmentation with Application to Automatic Speech Recognition. Ph.D. Thesis, Computer and Information Science, Report D13, Helsinki, University of Technology, Espoo, Finland. DeGroot, Morris H. 1989. Probability and Statistics. Second Edition, Reading, Massachusetts, Addison-Wesley Publishing Company. Djean, Herv. 1998. Morphemes as Necessary Concept for Structures Discovery from Untagged Corpora. New Methods in Language Processing and Computational Natural Language Learning (NeMLaP/CoNLL) Workshop on Paradigms and Grounding in Language Learning, Ed. Powers, David M. W., 295-298. Demberg, Vera, Helmut Schmid, and Gregor Mhler. 2007. Phonological Constraints and Morphological Preprocessing for Grapheme-to-Phoneme Conversion. Association for Computational Linguistics (ACL), Prague, Czech Republic. Dryer, Matthew S. 2008. Prefixing vs. Suffixing in Inflectional Morphology. In: Haspelmath, Martin, David Gil, and Bernard Comrie (Eds.) The World Atlas of Language Structures Online. Munich: Max Planck Digital Library, Chapter 26. Available online at . Accessed on December 17, 2008. Francis, W. Nelson. 1964. A Standard Sample of Present-Day English for Use with Digital Computers. Report to the U.S. Office of Education on Cooperative Research Project No. E-007, Brown University, Providence, Rhode Island. Gaussier, ric. 1999. Unsupervised Learning of Derivational Morphology from Inflectional Lexicons. Unsupervised Learning in Natural Language Processing an ACL'99 Workshop, University of Maryland. Goldsmith, John. 2001. Unsupervised Learning of the Morphology of a Natural Language. Computational Linguistics, 27.2, 153-198. Goldsmith, John. 2006. An Algorithm for the Unsupervised Learning of Morphology. Natural Language Engineering. 12.4, 335-351. Goldwater, Sharon, and Mark Johnson. 2004. Priors in Bayesian Learning of Phonological Rules. ACL Special Interest Group in Computational Phonology, Barcelona, Spain, 35-42. Goldwater, Sharon, and David McClosky. 2005. Improving Statistical MT through Morphological Analysis. Empirical Methods in Natural Language Processing, Vancouver, Canada. Gordon, Ronni L., and David M. Stillman. 1999. The Ultimate Spanish Review and Practice. Passport Books, Chicago, Illinois. Hafer, Margaret A., and Stephen F. Weiss. 1974. Word Segmentation by Letter Successor Varieties. Information Storage and Retrieval, 10.11/12, 371-385. Hammarstrm, Harald. 2006a. A Naive Theory of Affixation and an Algorithm for Extraction. ACL Special Interest Group on Computational Phonology, New York City, New York, 79-88. Hammarstrm, Harald. 2006b. Poor Mans Stemming: Unsupervised Recognition of Same-Stem Words. Information Retrieval Technology: Proceedings of the Third Asia Information Retrieval Symposium, AIRS 2006, Singapore, 323-337. Hammarstrm, Harald. 2007. Unsupervised Learning of Morphology: Survey, Model, Algorithm and Experiments. Thesis for the Degree of Licentiate of Engineering, Department of Computing Science, Chalmers University of Technology and Gteborg University, Gteborg, Sweden. Harris, Zellig. 1955. From Phoneme to Morpheme. Language, 31.2, 190-222, Reprinted in Harris (1970). Harris, Zellig. 1967. Morpheme Boundaries within Words: Report on a Computer Test. Transformations and Discourse Analysis Papers. Department of Linguistics, University of Pennsylvania, Reprinted in Harris (1970). Harris, Zellig. 1970. Papers in Structural and Transformational Linguists. Ed. D. Reidel, Dordrecht. Jacquemin, Christian. 1997. Guessing Morphology from Terms and Corpora. SIGIR '97: Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. Philadelphia, Pennsylvania, ACM Press, 156-165. Johnson, Howard, and Joel Martin. 2003. Unsupervised Learning of Morphology for English and Inuktitut. Human Language Technology Conference / North American Chapter of the Association for Computational Linguistics (HLT-NAACL), Edmonton, Canada. Kurimo, Mikko, Mathias Creutz, and Ville Turunen. 2007. Unsupervised Morpheme Analysis Evaluation by IR Experiments Morpho Challenge 2007. Working Notes for the CLEF 2007 Workshop, Budapest, Hungary. Kurimo, Mikko, Mathias Creutz, and Matti Varjokallio. 2008. Morpho Challenge Evaluation Using a Linguistic Gold Standard. Lecture Notes in Computer Science: 8th Workshop of the Cross-Language Evaluation Forum, CLEF 2007, Revised Selected Papers, Budapest, Hungary, Springer, 5152/2008, 864-872. Kurimo, Mikko, and Matti Varjokallio. 2008. Unsupervised Morpheme Analysis Evaluation by a Comparison to a Linguistic Gold Standard Morpho Challenge 2008. Working Notes for the CLEF 2008 Workshop, Aarhus, Denmark. Kurimo, Mikko, and Ville Turunen. 2008. Unsupervised Morpheme Analysis Evaluation by IR Experiments Morpho Challenge 2008. Working Notes for the CLEF 2008 Workshop, Aarhus, Denmark. Kurimo, Mikko, Ville Turunen, and Matti Varjokallio. 2008. Unsupervised Morpheme AnalysisMorpho Challenge 2008. , Accessed on December 10, 2008. Lang, Kevin J., Barak A. Pearlmutter, and Rodney A. Price. 1998. Results of the Abbadingo One DFA Learning Competition and New Evidence Driven State Merging Algorithm. ICGI '98: The 4th International Colloquium on Grammatical Inference, Springer-Verlag. Manning, Christopher D., and Hinrich Schtze. 1999. Foundations of Statistical Natural Language Processing. The MIT Press, Cambridge, Massachusetts. McNamee, Paul. 2008. Retrieval Experiments at Morpho Challenge 2008. Working Notes for the CLEF 2008 Workshop, Aarhus, Denmark. Miclet, Laurent. 1980. Regular Inference with a Tail-Clustering Method. IEEE Transactions on Systems, Man, and Cybernetics SMC-10.11, 737-743. Monson, Christian, Alon Lavie, Jaime Carbonell, and Lori Levin. 2004. Unsupervised Induction of Natural Language Morphology Inflection Classes. ACL Special Interest Group in Computational Phonology (SIGPHON), Barcelona, Spain, 52-61. Monson, Christian, Jaime Carbonell, Alon Lavie, and Lori Levin. 2007. ParaMor: Minimally Supervised Induction of Paradigm Structure and Morphological Analysis. ACL Special Interest Group in Computational Morphology and Phonology (SIGMORPHON), Prague, Czech Republic, 117-125. Monson, Christian, Jaime Carbonell, Alon Lavie, and Lori Levin. 2008a. ParaMor: Finding Paradigms across Morphology. Lecture Notes in Computer Science: 8th Workshop of the Cross-Language Evaluation Forum, CLEF 2007, Revised Selected Papers, Budapest, Hungary, Springer, 5152/2008, 900-907. Monson, Christian, Jaime Carbonell, Alon Lavie, and Lori Levin. 2008b. ParaMor and Morpho Challenge 2008. Working Notes for the CLEF 2008 Workshop, Aarhus, Denmark. Monson, Christian, Alon Lavie, Jaime Carbonell, and Lori Levin. 2008c. Evaluating an Agglutinative Segmentation Model for ParaMor. ACL Special Interest Group in Computational Morphology and Phonology (SIGMORPHON), Columbus, Ohio, 49-58. Oflazer, Kemal, and 0lknur Durgar El-Kahlout. 2007. Exploring Different Representational Units in English-to-Turkish Statistical Machine Translation. Statistical Machine Translation Workshop at ACL, Prague, Czech Republic. Ogilvie, Paul, and Jamie Callan. 2002. Experiments Using the Lemur Toolkit. The 2001 Text Retrieval Conference (TREC 2001), 103-108, National Institute of Standards and Tehnology, Special Publication, 500-250. Probst, Katharina. 2003. Using 'Smart' Bilingual Projection to Feature-Tag a Monolingual Dictionary. Computational Natural Language Learning (CoNLL), Edmonton, Canada. Roark, Brian, and Richard Sproat. 2007. Computational Approaches to Morphology and Syntax. Oxford University Press Inc., New York. Robertson, S. E., S. Walker, S. Jones, M. M. Hancock-Beaulieu, and M. Gatford. 1994. Okapi at TREC-3. The Third Text Retrieval Conference (TREC-3). Schone, Patrick, and Daniel Jurafsky. 2000. Knowledge-Free Induction of Morphology Using Latent Semantic Analysis. Computational Language Learning (CoNLL), Lisbon, Portugal, 67-72. Schone, Patrick, and Daniel Jurafsky. 2001. Knowledge-Free Induction of Inflectional Morphologies. North American Chapter of the Association for Computational Linguistics (NAACL), Pittsburgh, Pennsylvania, 183-191. Snover, Matthew G., and Michael R. Brent. 2001. A Bayesian Model for Morpheme and Paradigm Identification. Association for Computational Linguistics (ACL), Toulouse, France, Morgan Kaufmann. Snover, Matthew G. 2002. An Unsupervised Knowledge Free Algorithm for the Learning of Morphology in Natural Languages. Sever Institute of Technology, Computer Science, Saint Louis, Missouri, Washington University, M.S. Thesis. Snover, Matthew G., Gaja E. Jarosz, and Michael R. Brent. 2002. Unsupervised Learning of Morphology Using a Novel Directed Search Algorithm: Taking the First Step. ACL Special Interest Group in Computational Phonology in cooperation with the ACL Special Interest Group in Natural Language Learning (SIGPHON/SIGNLL), University of Pennsylvania, Philadelphia, Pennsylvania, Association for Computational Linguistics, 11-20. Snover, Matthew G., and Michael R. Brent. 2002. A Probabilistic Model for Learning Concatenative Morphology. Neural Information Processing Systems Foundation (NIPS). Sproat, Richard. (Ed.). 1997. Multilingual Text-to-Speech Synthesis: The Bell Labs Approach. Kluwer Academic Publishers, Dordrecht, The Netherlands. Wicentowski, Richard. 2002. Modeling and Learning Multilingual Inflectional Morphology in a Minimally Supervised Framework. Ph.D. Thesis, Baltimore, Maryland, The Johns Hopkins University. Xanthos, Aris. 2007. Apprentissage automatique de la morphologie: Le cas des structures racine-schme. Ph.D. Thesis, Universit de Lausanne. Xu, Jinxi, and W. Bruce Croft. 1998. Corpus-Based Stemming Using Co-occurrence of Word Variants. ACM Transactions on Information Systems (TOIS), 16.1, 61-81. Yarowsky, David, and Richard Wicentowski. 2000. Minimally Supervised Morphological Analysis by Multimodal Alignment. Association for Computational Linguistics (ACL). Yarowsky, David, Grace Ngai, and Richard Wicentowski. 2001. Inducing Multilingual Text Analysis Tools via Robust Projection across Aligned Corpora. Human Language Technology Research (HLT). Zeman, Daniel. 2008. Using Unsupervised Paradigm Acquisition for Prefixes. Working Notes for the CLEF 2008 Workshop, Aarhus, Denmark. Appendix A: A Summary of Common Spanish Suffixes A Summary of Common Spanish Suffixes This appendix provides the basic knowledge of Spanish morphology needed to understand the examples which appear in the main body of this thesis. Spanish has both inflectional and derivational morphology. Since ParaMor is designed to identify the basic organizational unit of inflectional morphology, namely the paradigm, this appendix primarily describes the inflectional morphology of Spanish. However, Figure A.8, at the end of this appendix, lists a few common derivational suffixes of Spanish, including most of the derivational suffixes which occur in examples in this thesis. This guide to Spanish morphological paradigms is based on Gordon and Stillman (1999), an intermediate Spanish textbook. Figures A.1 through A.7 present sets of inflectional suffixes and word-final clitics that together describe the inflectional morphology of Spanish. Figures A.1 through A.3 describe the inflectional morphology of Spanish verbs. Verbal suffixes in Spanish mark combinations of Tense, Aspect, and Mood together with Subject Number, Subject Person, and in the case of past participles a Gender feature. The verbal paradigm has three inflection classes. The inflection classes of the verbal paradigm are morphologically stipulatedneither phonology nor syntax determines which inflection class a particular lexical verb belongs to. Figures A.4 and A.5 give the two phonologically conditioned inflection classes of the Number paradigm on Spanish nouns. These two inflection classes mark Singular verses Plural on nouns, adjectives, and past participles. Figure A.6 contains the four suffixes which constitute the cross-product of the adjectival Number and Gender paradigms. In addition to the inflectional suffixes of Figures A.1 through A.6, Spanish has a system of pronominal clitics which mimic true inflectional morphemes in Spanish orthography. Spanish pronominal clitics are written as a single orthographic word when they occur immediately following a non-finite verb form. There are three sets of pronominal clitics in Spanish which each masquerade as a separate paradigm: one set of clitics marks Accusative Case, another Dative Case, and a third clitic set contains reflexive pronouns. Figure A.7 presents the three sets of Spanish pronominal clitics. Although rare, it is possible for a single non-finite Spanish verb form to attach clitics from each separate set of pronominal clitics. In a Spanish written word that contained a clitic from each pronominal clitic set, the order of the clitics would be: Accusative, Dative, Reflexivethis is the order in which the clitics of Figure A.7 appear. At times the main body of this thesis refers to the set of string unique inflectional Spanish suffixes (see for example Section  REF _Ref190768256 \r \h  4.4.1). The Spanish suffixes which contribute to this set of inflectional suffixes are the surface forms which appear in Figures A.1 through A.7.    Appendix B: Scheme Clustering, a Pseudo-Code Implementation Scheme Clustering, the Pseudo-Code Pseudo-code implementing ParaMor s scheme-clustering algorithm as described in Section  REF _Ref190686481 \r \h  4.3. At base ParaMor employs a bottom-up agglomerative clustering algorithm. However, Sections  REF _Ref215388958 \r \h  4.3.1 and  REF _Ref215388962 \r \h  4.3.2 describe several adaptations to the basic agglomerative algorithm that aid the clustering of scheme-models of paradigms. All of these adaptations are present in the following pseudo-code (next page): //////////////////////////////////////////// // The Data Structure of a Scheme-Cluster // //////////////////////////////////////////// // Each cluster is one of two types: // 1) A leaf cluster is a wrapper for a scheme // 2) A compound cluster is a cluster that results from the merger of // two child clusters. Each child cluster of a compound cluster can // either be a leaf cluster or a compound cluster itself. // // Leaf clusters and compound clusters inherit common properties from // the base Cluster struct. struct Cluster { // For every  small scheme a CompoundCluster contains, that // CompoundCluster must contain one  large scheme (See Section //  REF _Ref215319787 \r \h  \* MERGEFORMAT  4.3.2). Hence, each cluster, whether a Leaf or a CompoundCluster, // keeps track of the number of large and small schemes it // contains. largeSchemeCount = 0; smallSchemeCount = 0; // Each Cluster keeps track of the set of morpheme-boundary annotated // word types that license the Cluster licensingBoundaryAnnotatedTypes = null; }; struct Leaf subStructOf Cluster { scheme = null; }; struct CompoundCluster subStructOf Cluster { childA = null; childB = null; }; ///////////////////////////////////////////////////////////// // The Bottom-Up Agglomerative Scheme-Clustering Algorithm // ///////////////////////////////////////////////////////////// clusterBottomUp(schemes) { clusters = buildLeafClusters(schemes); while (true) { [bestClusterA, bestClusterB] = findClustersToMerge(clusters); // Stop clustering when there is no pair of clusters that are // permitted to merge. if (bestClusterA == null) return clusters; newCluster = merge(bestClusterA, bestClusterB); clusters.removeAll(bestClusterA, bestClusterB); clusters.add(newCluster); } } buildLeafClusters(schemes) { foreach (scheme in schemes) { leaf = new Leaf(); leaf.scheme = scheme; if (scheme.numberOfLicensingWords >= threshold) { leaf.largeSchemeCount = 1; } else { leaf.smallSchemeCount = 1; } foreach (cSuffix in scheme.cSuffixes) { foreach (cStem in scheme.cStems) { leaf.licensingBoundaryAnnotatedTypes.add(cStem.+.cSuffix); } } leafClusters.add(leaf); } return leafClusters; } findClustersToMerge(clusters) { bestClusterScore = null; clusterToMergeA, clusterToMergeB = null; foreach (clusterA in clusters) { foreach (clusterB greaterThan clusterA in clusters) { if (isMergePermitted(clusterA, clusterB)) { score = cosine(clusterA.licensingBoundaryAnnotatedTypes, clusterB.licensingBoundaryAnnotatedTypes); if (score > bestClusterScore) { bestClusterScore = score; clusterToMergeA = clusterA; clusterToMergeB = clusterB; } } } } return [clusterToMergeA, clusterToMergeB]; } merge(clusterA, clusterB) { newCluster = new CompoundCluster(); newCluster.childA = clusterA; newCluster.childB = clusterB; newCluster.largeSchemeCount = clusterA.largeSchemeCount + clusterB.largeSchemeCount; newCluster.smallSchemeCount = clusterA.smallSchemeCount + clusterB.smallSchemeCount; newCluster.licensingBoundaryAnnotatedTypes = union(clusterA.licensingBoundaryAnnotatedTypes, clusterB.licensingBoundaryAnnotatedTypes); return newCluster; } isMergePermitted(clusterA, clusterB) { // There are two restrictions on clustering // Restriction 1: ParaMor discriminatively requires that each pair // of csuffixes in each cluster be able to form a word that // occurred in the paradigm induction corpus by attaching to a // common cstem. foreach (cSuffixA in clusterA) { foreach (cSuffixB in clusterB) { schemeOfTwoCSuffixes = dynamicSchemeNetwork.generateScheme(cSuffixA, cSuffixB); if (schemeOfTwoCSuffixes.cStems.size == 0) return false; } } // Restriction 2: ParaMor requires that there be at least as many // large schemes as there are small schemes in each cluster. if ((clusterA.smallSchemeCount + clusterB.smallSchemeCount) > (clusterA.largeSchemeCount + clusterB.largeSchemeCount)) { return false; } return true; }  It is the asymmetry in the definition of a scheme that implies tC is in T2 C as follows: Because tC combines with all the csuffixes in FC to form a vocabulary word and since F2 C is a subset of FC, tC must combine with all the csuffixes in F2 C, and The asymmetric definition of a scheme states that T2 C must contain all the cstems which combine with all the csuffixes in F2 C  including tC.  Consider, for example, a hypothetical corpus vocabulary containing just three words: ac, abc, and abbc. These three words could give rise to the scheme bc.c containing the stems a and ab. Reconcatenating the stems and suffixes in this hypothetical scheme gives the boundary annotated forms: a+bc, a+c, ab+bc, and ab+cbut a+bc and ab+c are different segmentations of the same licensing word, abc. PAGE 109 PAGE 104  PAGE 99 PAGE 109 PAGE 159  PAGE 161  PAGE 224 Some stuff I didn't get around to incorporating into the thesis before the defense. Points and Nice pieces of data that I want to keep around for later potential use In the detailed subsections I must state that decent schemes are still being selected at rank 2000. placing a morpheme boundary to the right of the linking vowel in Spanish is arguably reasonable, but I want to match morphosyntactic features the 1000th selected scheme partly captures a morphophonologic change, in future iterations of ParaMor I plan on harnessing this fact. If I have time I could clarify how it is unlikely that natural language would produce a scheme that is double licenced by a single word: 1) stems that are identical word-initially, 2) suffixes that are identical word-finally, 3) the material that is different on the stems is the same as the material that is different in the suffixes, 4) the stems must both belong to the relevant paradigm, 5) even if 1-4 are met it will only be for a very small subset of the full set of lexemes in the language. this number may have accidentally excluded enye from the Spanish alphabet. I have not yet created this appendix. But once I do, place the following sentence immediately after this numbered list about Figure 3.3: An additional scheme network covering a portion of two Spanish verbal paradigms appears in Appendix N. The paths figure has been updated to include enye Appendix 1 (or A, or whatever) must explain the ase csuffix in the cluster with the 4th most licensing types. Should I go into any more detailed a description of how most specific schemes are exactly the minimal forward FSA? And the full fledged scheme FSA has states that are unreachable from the start state etc. This might be interesting to Ron, but as I never actually use this fact in the thesis anywhere, I think I will just leave it out completely. Ron wanted FSA's, I tried, but it didn't happen. I have *good* results, not exactly what was imagined in the proposal, but close. Similar text taken from Ch. 7: Future work. But this text is just too detailed for that chapter. (And I don't think this section needs it either. But it is nice text and I couldn't make myself outright delete it). Currently, when morphophonologic changes occur, the best ParaMor can do is incorporate the phonologic alternation into the paradigmatically varying csuffixes. Such is the case with the 21st Spanish scheme cluster of  REF _Ref190770158 \h \* MERGEFORMAT Error! Reference source not found.. This 21st cluster is modeling a group of Spanish verbs where the realization of the stem-final consonants depends on the initial vowel of the suffix. These Spanish stems end in /s/ before the front vowels /i/ and /e/, but in /sk/ before the back vowels /a/ and /o/. In the Spanish orthography of the verb stems which participate in this alternation, /s/ before a front vowel is spelled c, while /sk/ before a back vowel is rendered as zc (Hence, in the orthography, this consonant change appears as a z inserted before the c, while phonologically a /k/ is inserted after an /s/). Since this morphophonologic change affects only the tail of the relevant verb stems, ParaMor can account for the change by incorporating the tail of the verb in the csuffixes. Incorporating the verb tails results in a scheme cluster that contains csuffixes like ce and cimos but zca and zco. Unfortunately, with this approach, ParaMors segmentations of words like conoce she knows, cono + ce, and bebe she drinks, beb + e, fail to identify the shared 3rd Person Singular Present Indicative morpheme. Counts of the number of unique csuffixes for each language after each step of ParaMor's paradigm induction. While interesting, these statistics don't add all that much to what is already in the table of counts of schemes/scheme-clusters. (And I don't want to have to write about them--I'm running out of time to finish my thesis document!!)  # of Unique CSuffixesSpanishEnglishGermanFinnishTurkishArabicSearch98961003813101136391149913055Excluding Short Types9282911012962139871154815692Clustering9282911012962139871154815692Small Cluster Filtering14969151396303738802010Morpheme Boundary Filtering516214543151118151330 Text that provides too much detail: Two specific details of Spanish morphology lead to ParaMors overgeneralization. First, in Spanish, both nouns and adjectives that end in /l/ form their plural by adding es. Second, while the Spanish word hospital hospital is a noun, a number of Spanish adjectives also end in aland some of these adjectives have related verb forms that belong to the ar sub-paradigm, c.f. accidental accidental (Adverb) and accidentar to cause an accident (Verb); original original (Adverb) and originar to originate (Verb). Seeing a set of shared cstems among these adverb-verb pairs (and not knowing that these words do not share a part of speech), ParaMor constructs scheme-clusters that contain both adjectival suffixes and verbal suffixes. Specifically, ParaMor builds two clusters: one cluster contains the verbal suffix ada and the adjectival suffixes al and ales, while a second cluster contains the incorrect suffix-internal segmentations of these same suffixes: da, l, and les. Furthermore, ParaMors morpheme boundary filters fail to remove the cluster that hypothesizes the incorrect suffix-internal morpheme boundary because of ida and il allomorphic variants of the ada and al suffixes. Although unsatifying that ParaMor has mixed verbal and adverbial suffixes in the same cluster, to this point ParaMor has not overgeneralizednearly all the cstems that the mixed clusters contain are able to form words with all csuffixes in the clusters. ParaMors overgeneralization occurs while segmenting the form hospital. In addition to the singular noun hospital, ParaMors segmentation corpus also contained the plural form hospitales. Here, Interesting stuff that is too much detail to include. Furthermore, Section  REF _Ref215980624 \r \h  4.5.2 demonstrated that, in Spanish, the inventory of csuffixes that match suffixes of true Spanish paradigms steadily decreases as the size of the paradigm induction corpus declines. Unexpectedly, however, csuffix inventory does not decrease in English and German when training over a smaller corpus. Instead, ParaMors csuffix inventory rises slightly, from 214 to 224 in English and from 543 to 544 in German. These increasing csuffix inventories for English and German, together with differences in the specific sets of csuffixes that are learned, permit morpheme boundary recall to increase when ParaMor induces paradigm models over smaller vocabulary sizes. Figure  STYLEREF 1 \s  1. SEQ Figure \* ARABIC \s 1 1 Left: A fragment of the morphological structure of Spanish verbs. There are three paradigms in this fragment. Each paradigm covers a single morphosyntactic category: first, Type; second, Gender; and third, Number. Each of these three categories appears in a separate column; and features within one feature column, i.e. within one paradigm, are mutually exclusive. Right: The suffixes of the Spanish inflection class of ar verbs which fill the cells of the paradigms in the left-hand figure. Vocabulary: blame solve blames roams solves blamed roamed roaming solvinglame lames lamed bame ames amed blme mes med blae es ed blam s d blameoams oamed oaming rlame lames bame ames blme mes blae es blam solv s blame solveolve olves olving slame lamed bame amed blme med blae ed blam d blamelames lamed bames amed blmes med blaes ed blams d blames ed ing roame es ing solvlame bame blme blae blam solvs ed roame ing solvlames bames blmes blaes blam solvs blame roam solves ing roames ing solvlamed bames bl romed bla roaed blam roamd blame roameed ing roamng roami solvi blame blames blamed roams roamed roaming solve solves solvinging roam solvg roamin solvin TypeGenderNumberPast ParticipleFeminineSingularMasculinePluralPresent ParticipleInfinitive a d m i n i s t r a n d o d a o s r o & Figure  STYLEREF 1 \s  1. SEQ Figure \* ARABIC \s 1 2: A Finite State Automaton (FSA) representing surface forms of the lexeme administrar. Arcs represent characters; States are character boundaries. States at morpheme boundaries typically have multiple arcs entering and/or exiting, while states at character boundaries internal to morpheme boundaries typically have a single entering and a single exiting arc. Figure  STYLEREF 1 \s  1. SEQ Figure \* ARABIC \s 1 3: A portion of a morphology scheme network generated from 100,000 words of the Brown corpus of English (Francis, 1964). The two schemes which model complete verbal sub-classes are outlined in bold. e ed es ing not declar & ed es ing address reach & s sed ses sing addres mis & e ed es ing not stag & ed es ing declar pric & TypeGenderNumberadaosandoar Figure  STYLEREF 1 \s  2. SEQ Figure \* ARABIC \s 1 1: A hub, left, and a stretched hub, right, in a finite state automaton Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 1: Some of the schemes, arranged in a systematic but arbitrary order, derived from a small vocabulary (top). Each scheme is specified as a space delimited set of csuffix exponents in bold above a space delimited set of cstem adherents in italics e es blam solv e ed blam es blam solv s d blame s blame solve blame blames blamed roams roamed roaming solve solves solving e es ed blam ed blam roam d blame roame d blame s d blame s blame roam solve es ed blam e blam solv me mes bla me med bla mes bla me mes med bla med bla roa mes med bla me bla Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 2: A portion of a morphology scheme network from the small example vocabulary of  REF _Ref69705490 \h  \* MERGEFORMAT Figure  3.1 C-Suffix set inclusion links Morpheme boundary links .ed.ing.ly 6 clear direct fil open present total Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 3: A portion of a morphology scheme network generated from a 100,000 tokens from the Brown Corpus of English. The scheme that most closely matches a true verbal sub-class of English, .ed.ing.s, is outlined in bold. .ed.ly 11 clear correct direct fil present quiet ed.ly 12 bodi clear correct direct fil quiet .ed.ing.ly.s 4 clear open present total .ed.ing 201 aid clear defend deliver demand total d.ded.ding 27 ai boar defen deman recor regar d.ded.ding.ds 19 ad boar defen depen fiel recor .ed.ing.s 106 add clear defend open present total a.as.o.os 899 act cas futur nuev religios ... a.as.o.os.ualidad 3 act cas d a.as.os 964 cas emplead nuev unid ... a.as.o 1124 act futur norteamerican tcnic ... a.o.os 1418 asit1c encuentr numeros religios ... a.as 2200 act emplead norteamerican realizad ... a.os 1605 cas encuentr nuev universitari ... as.os 1116 crtic expuest much unid ... a.o 2325 asitic futur numeros secretari ... as.o 1330 conocid es ni tcnic ... o.os.ualidad 4 act cas d event a 9020 act futur nuev seor ... as 3182 caden emplead ni unid ... os 3847 cas human much saque ... o 7520 asitic es puest suiz ... ualidad 10 act cas event intellect as.o.os 1009 conocid expuest ni puest ... Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 4: A hierarchical scheme search space lattice automatically derived from a Spanish newswire corpus of 50,000 unique types (1.23 million tokens). The scheme matching the productive adjectival sub-class of Spanish, a.as.o.os, is outlined in bold. o.os 2390 cercan human puest religios ... & & ... ... ... strado 15 trado 30 rado 167 an 1784 Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 7: Eight search paths that ParaMor follows in search of schemes which likely model inflectional paradigms. Search paths begin at the bottom of the figure and move upward. All csuffixes appear in bold. The underlined csuffix in each scheme is the csuffix added by the most recent search step. Each scheme gives the number of adherent cstems it contains. Horizontal links between schemes connect sets of csuffixes that differ only in their initial character. n 6039 a 9020 s 10697 a an 1045 a an ar 417 a an ar 355 a ado an ar 318 a ado an ar aron 263 a ada ado an ar aron 206 a ada ado ados an ar aron 172 a ada adas ado ados an ar aron 148 rada radas rado rados rar raron r 33 rada radas rado rados rar 46 rada radas rado rados 53 rada rado rados 67 rada rado 89 ra rada radas rado rados rar raron r 26 ra rada radas rado rados ran rar raron r 23 n 1863 n r 512 do n r 357 do n r ron 272 da do n r ron 211 da do dos n r ron 176 da das do dos n r ron 150 da das do dos n ndo r ron 115 a o 2325 a o os 1418 a as o os 899 s 5513 trada trado 19 trada trado tr 15 trada trado trar tr 13 trada tradas trado trar tr 12 strada strado 12 strada strado str 9 strada strado strar str 8 strada stradas strado strar str 7 es 2750 es 845 r s 281 trada tradas trado trados trar tr 10 trada tradas trado trados trar traron tr 8 rada radas rado rados rar r 42 1st 2nd 3rd 4th 5th 71st 484th 1113th Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 11: Seven of the 20,949 parents of the a.o.os scheme derived from the same Spanish newswire corpus of 50,000 types as  REF _Ref192403460 \h  \* MERGEFORMAT Figure  3.4. The csuffix added by each parent is underlined. a.amente.o.os 173 democrtic mdic presunt segur ... a.e.o.os 105 est grup muert pes ... a.ar.o.os 145 cambi estudi marc pes ... .a.o.os 94 buen fin primer uruguay ... a.es.o.os 28 fin libr muert pas ... a.o.os.ualidad 3 act cas d a.as.o.os 899 act futur nuev religios ... a.o.os 1418 asit1c encuentr numeros religios ... ... ... ... ... ... a.o.osa.o.osualidad3710ualidad1415485754999014184858250000 a.o.osa.o.osar14513031448ar1273472794855214184858250000 a.o.osa.o.osamente 173159332amente1245484234966814184858250000 a.o.osa.o.osas89922833182as519462994681814184858250000 ualidad 10 act cas event intellect ar 1448 cambi habl import reiter ... amente 332 democrtic frenetic marcad tranquil ... as 3182 caden emplead ni unid ... a.o.os 1418 asit1c encuentr numeros religios ... a.as.o.os 899 act futur nuev religios ... a.o.os.ualidad 3 act cas d a.ar.o.os 145 cambi estudi marc pes ... a.amente.o.os 173 democrtic mdic presunt segur ... Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 12: Four parents of the a.o.os scheme, together with the level 1 expansion schemes which contain the single csuffix which expands a.o.os into the parent scheme. Beneath each expansion scheme is a table of cstem counts relevant to a.o.os and that parent. These schemes and cstem counts come from a Spanish newswire corpus of 50,000 types. MetricExplanationFormulaParent ScoreasamentearualidadHeuristicsRatioRatio of parent cstems to current cstems EMBED Equation.3 0.6340.1220.1020.002DiceHarmonic mean of parent-child and parent-extension cstem ratios EMBED Equation.3 0.3910.1980.1010.004Pointwise Mutual InformationPointwise MI between current and expansion schemes EMBED Equation.3 3.324.201.823.40Statistical TestsPearsons 2Nonparametric test for independence of random variables in categorical data EMBED Equation.3  777294027926.8Wald Test for Mean of BernoulliIf the current and expansion schemes are independent, the Wald statistic converges to a standard normal:  EMBED Equation.3  EMBED Equation.3 27.212.58.641.57Likelihood Ratio of BernoulliRatio of likelihood of expansion scheme given independence from current to the likelihood of expansion scheme given dependence EMBED Equation.3  34108031749.57 Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 13: Six metrics which might gauge the paradigmatic unity of parent schemes during ParaMor s search of the vertical morphology scheme network. Each row describes a single metric. The top three metrics are heuristic measures of paradigmatic coherence. The bottom three metrics treat the current and expansion schemes as random variables and test their statistical correla- PECV tion. In the Formula column: C, E, and P are the cstem counts of the Current, Expansion, and Parent schemes respectively, while  EMBED Equation.3  is the size of the corpus Vocabulary. An expansion scheme heads each of the final four columns, which each contain the value of that rows metric applied from the a.o.os scheme of  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12 to that column s expansion scheme. The mini-table at left shows where C, E, P, and V fall in a 2x2 table of cstem counts, a la  REF _Ref192925307 \h  \* MERGEFORMAT Figure  3.12.  Likelihood Ratio Test CStem Ratio Wald Test Pearson s 2 Test Pointwise Mutual Information Dice Always Move Up F1 Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 14: A small-scale oracle evaluation of six metrics at the task of identifying schemes where each csuffix models a suffix in the same true inflectional paradigm. Each bar reports the peak F1 of its metric over a range of cutoffs appropriate for that metric. RankModel ofGoodCompletePartialErrorCSuffixesC-StemsNAdjVerbDerivationStem-InternalSuffix-InternalChancearerir1%%%%2 s5501apoyada barata hombro oficina reo & 2%%%4a as o os892apoyad captad dirigid junt prxim & 3%%%15 ba ban da das do dos n ndo r ra ron rse r rn17apoya disputa lanza lleva toma & 4%%%%2 es847emir inseguridad orador pu ramon & 5%%%15a aba aban ada adas ado ados an ando ar aron arse ar arn 25apoy desarroll disput lanz llev & 10%%%5ta tamente tas to tos22cier direc insli modes sangrien & 11%%%14 ba cin da das do dos n ndo r ron r rn ra16acepta celebra declara fija marca & 12%%%15a aba ada adas ado ados an ando ar aron ar arn e en 21apoy declar enfrent llev tom & 20%%%%6 l ra ras ro ros8a ca coste e ente gi o pu30%%%%11e en ida idas ido idos iendo ieron i a an16cumpl escond recib transmit vend & 40%%%%%7 es idad izacin izado izar mente8actual final natural penal regular & 100%8 a an e ea en i ino9al c ch d g p s t v127%%%9e en er er era ido ieron i a11ced deb ofrec pertenec suspend & 135%%%%10a e en ida ido iendo iera ieron i a12assist cumpl ocurr permit reun un & 200%%4tal tales te tes5acciden ambien continen le pos300%%4o os ual uales7act cas concept event grad man us400%%%%8a aciones acin ados an ar ativas 10administr conmemor estim investig & 1000%%%8ce cen cer cern cido cieron ci ca9apare estable ofre pertene ven & 1592%%%4ido idos ir ir6conclu cumpl distribu exclu reun & 2000%%%%4e en ieron iesen5aparec crec impid invad pud3000%2 zano3li lo man4000%2icho io4b capr d pred5000%%%2egar ega3desp entr ll& & & & &  Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 16: A learning curve  Spanish morpheme recall and the number of paradigm-clusters ParaMor identifies as the vocabulary size varies from 50,000 to 10,000. Achievable Recall is calculated out of the valid Spanish morpheme strings which occur at least twice in the induction vocabulary.  EMBED Excel.Chart.8 \s  Vocabulary Size Recall # of Clusters Achievable Recall Morpheme Recall # of Clusters 50 41 38 59.8 49.4 58.9 67.5 72.0 67.8 44 50K 40K 30K 20K 10K Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 6: Pseudo-code implementing ParaMor s initial search for scheme models of partial paradigms. // Global initializations dynamicSchemeNetwork = initializeDynamicNetwork(corpus); visitedSchemes = emptySet; // Keep track of all schemes in any path // Follow a separate upward search path for each scheme // that contains just a single csuffix. search() { selectedSchemes = emptySet; levelOneSchemes = dynamicSchemeNetwork.getNonNullLevelOneSchemes(); sortedLevelOneSchemes = sortByDecreasingCStemCount(levelOneSchemes); foreach (levelOneScheme in sortedLevelOneSchemes) { selectedScheme = searchOneGreedyPath(levelOneScheme); if (selectedScheme != null) selectedSchemes.add(selectedScheme); } return selectedSchemes; } searchOneGreedyPath(levelOneScheme) { currentScheme = levelOneScheme; while (true) { [bestParentScheme, scoreOfBestParentScheme] = dynamicSchemeNetwork. getBestParentSchemeWithScore(currentScheme); // The two criteria a parent scheme must pass for // ParaMor to move upward if ((scoreOfBestParentScheme < threshold) || ( ! (bestParentScheme.getNumberOfCStems() > bestParentScheme.getNumberofCSuffixes()))) // If the parent doesnt qualify then, as long as this search // path took at least one upward step, return the current, // now terminal, scheme. if (currentScheme.getNumberOfCSuffixes() > 1) return currentScheme; else return null; // Abandon redundant paths if (visitedSchemes.contains(bestParent)) return null; // Loop updates visitedSchemes.add(bestParent); currentScheme = bestParent; } } Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 5: A conceptualization of ParaMor s initial search for partial paradigms. Search Paths (i.e. More Suffixes) Paradigms Increasing Paradigmatic Complexity More C-Stems Fewer C-Stems 0 60K 31 35.6 47.7 // Define a cstem, csuffix pair struct SegmentedWord { cStem; cSuffix; }; // Two structures in which to save the results of // the first and second halves of the algorithm Hash> cStemToCSuffixes; Hash, Set> mostSpecificSchemes; // First Half: Find all csuffixes that can attach to each cstem. foreach (word in corpus.vocabulary) { Set segmentations = corpus.getAllSegmentations(word); foreach (segmentation in segmentations) { cStemsToCSuffixes{segmentation.cStem}.add(segmentation.cSuffix); } } // Second Half: Find all cstems of each csuffix set. foreach (cStem in cStemsToCSuffixes) { cSuffixes = cStemToCSuffixes{cStem}; mostSpecificSchemes{cSuffixes}.add(cStem); } Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 8: Pseudo-code for computing most-specific schemes. Note that this code uses curly braces { } to index into hashes. And as this algorithm is heavily dependent on data structures, this code segment is strongly typed. using pointy brackets < >. computeMostSpecificAncestorsOfAllSingleCSuffixes(mostSpecificSchemes) { foreach (mostSpecificScheme in mostSpecificSchemes) { foreach (cSuffix in mostSpecificScheme.cSuffixes) { mostSpecificAncestors{cSuffix}.add(mostSpecificScheme); } } return mostSpecificAncestors; } Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 9: Pseudo-code to compute  EMBED Equation.3 , the set of all most-specific schemes whose set of csuffixes include  EMBED Equation.3 , where  EMBED Equation.3  ranges over all csuffixes generated by a corpus. The pseudo-code refers to  EMBED Equation.3  as the ancestors of  EMBED Equation.3 : Imagining a network of most-specific schemes mirroring the scheme networks of Section  REF _Ref193859274 \r \h  3.1.2, the set  EMBED Equation.3  contains all the upward, or ancestor, most-specific schemes of the  EMBED Equation.3  node. Figure  STYLEREF 1 \s  3. SEQ Figure \* ARABIC \s 1 10: Pseudo-code to dynamically compute the cstems that belong to the scheme containing the set of csuffixes cSuffixesOfScheme. This algorithms uses the relevant pre-computed sets of most-specific schemes in  EMBED Equation.3  where  EMBED Equation.3  is a csuffix generated by the corpus. As in  REF _Ref214443964 \h  \* MERGEFORMAT Figure  3.9, the  EMBED Equation.3  are saved in the variable mostSpecificAncestors as a hash on csuffixes. computeTheCStemsOfAScheme(cSuffixesOfScheme, mostSpecificAncestors) { // Intersect the pre-computed sets of most-specific scheme ancestors // of each individual csuffix in cSuffixesOfScheme. mostSpecificAncestorsOfScheme = emptySet; foreach (cSuffix in cSuffixesOfScheme) { if (mostSpecificAncestorsOfScheme == emptySet) mostSpecificAncestorsOfScheme. addAll(mostSpecificAncesors{cSuffix}); // Deep Copy else { foreach (mostSpecificAncestor in mostSpecificAncestorsOfScheme) { if ( ! mostSpecificAncestors{cSuffix}. contains(mostSpecificAncesor)) mostSpecificAncestorsOfScheme.remove(mostSpecificAncestor); } } } // Compute the union all c-stems in all the most-specific scheme // ancestors of the scheme we are building. foreach (mostSpecificAncestor in mostSpecificAncestorsOfScheme) { cStemsOfScheme.add(mostSpecificAncestor.cStems); } return cStemsOfScheme; } Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 1: Schemes selected by ParaMor s initial bottom-up search algorithm over a Spanish corpus of 50,000 types. While some selected schemes contain csuffixes that correctly model paradigmatic suffixes, others are incorrect collections of word final strings. RankCorresponds to  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1Licensing Types# ofSchemes Model ofGoodErr.CSuffixesC-StemsNAdjVerbDeriv.Allo.StemSuffixarerir11110553%%%%4 menente mente s5401apoyada barata hombro inevitable segura & 22474023%%%37 ba ban cion ciones cin da das do dor dora doras dores dos miento mos n ndo r ra ran remos rla rlas rle rles rlo rlos rme rnos ron rse r rn r ra ran1805apoya compra disputa genera lanza lleva reclama senti utiliza visite & 33, 1141424%%%11a amente as illa illas o or ora oras ores os976apoyad captad frank guerr negociad & 45 12 (400)190923%%%41a aba aban acion aciones acin ada adas ado ador adora adoras adores ados amos an ando ante antes ar ara aran aremos arla arlas arlo arlos arme aron arse ar arn ar ara aran ase e en ndose 113apoy celebr desarroll genera hall lanz llev public realiz termin utiliz visit 5-109516%%22ion iona ionaba ionada ionadas ionado ionados ionamiento ionamientos ionan ionando ionantes ionar ionario ionaron ionar ionarn ione ionen iones ion in418administrac condic cuest emoc gest les ocas pres reacc sanc ses vinculac & 10107224%%%7ta tamente tas titud to tos tsimo324cier direc gra insli len modes sangrien ...1130 127 (135)72017%%%29e edor edora edoras edores en er erlo erlos erse er ern era eran ida idas ido idos iendo iera ieran ieron imiento imientos indose i a an62abastec aparec crec cumpl desconoc escond ofrec ocurr permit pertenec recib reun sal suspend transmit vend & 17159238411%%%20ida idas ido idor idores idos imos ir iremos irle irlo irlos irse ir irn ir ira iran a an39abat conclu cumpl distribu eleg exclu permit persegu recib reun segu transmit & 21100034418%%%29ce cedores cemos cen cer cerlo cerlos cerse cer cern cera cida cidas cido cidos ciendo ciera cieran cieron cimiento cimientos cimos ci c ca can zca zcan zco26abaste apare agrade compare conven estable fortale ofre pertene recono ven & 100-822%%4eta etas eto etos33atl bol concr libr metrall papel secr tarj & 122(40)582%6 es idad idades mente sima15casual fort important irregular primer & 200-351%5staba stado star staron star7co conte entrevi gu in manife pre300-3015pondrn pone ponen poner puesta6com dis ex im pro su1000-151%3dismo dista distas5bu ovie parti perio perre2000-121%3ral rales ralismo4electo libe neolibe plu3000-81%2unas unos4alg fa oport vac5000-61%2ntra ntrarse3ade ce conce& & & & & &  Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 4: Typical clusters produced by ParaMor over a Spanish corpus of 50,000 types each longer than 5 characters. 19: a aba aciones acin ada adas ado ados an ando ar arlo aron arse ar arn ara ndose Cosine Similarity: 0.369 524 Covered Types Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 5: A portion of a cluster tree ParaMor built from schemes selected during a search run over 50,000 Spanish types longer than 5 characters. This cluster tree was constructed using ParaMor s bottom-up agglomerative clustering algorithm adapted for schemes as described in Sections  REF _Ref215375150 \r \h  4.3.1 and  REF _Ref215375180 \r \h  4.3.2. Each scheme appears in a solid box; each intermediate cluster in a dotted box. The csuffixes of each scheme or cluster of schemes are in bold, cstems are in italics. Each csuffix in a cluster which uniquely originates in one child is underlined. 16: a aba aban ada adas ado ados an ando ar aron arse ar arn ara 17: alcanz, anunci, aplic, celebr, consider, control, declar, entreg, estudi, expres, inici, investig, negoci, otorg, present, realiz, utiliz 272 Covered Types 16: aba aban ada adas ado ados amos an ando ar ara aron arse ar arn ara Cosine Similarity: 0.696 343 Covered Types 18: a aba aban ada adas ado ados amos an ando ar ara aron arse ar arn ara Cosine Similarity: 0.504 461 Covered Types 15: aba aban ada adas ado ados an ando ar aron arse ar arn ara aran 16: alcanz, aplic, apoy, celebr, control, elev, entreg, estudi, form, hall, lleg, llev, ocup, otorg, realiz, utiliz 240 Covered Types 19: a aba aban ada adas ado ados amos an ando ar ara aron arse ar arn ara aran Cosine Similarity: 0.589 505 Covered Types 23: a aba aban aciones acin ada adas ado ados amos an ando ar ara arlo aron arse ar arn ara aran ndose Cosine Similarity: 0.364 842 Covered Types 15: aba aban ada adas ado ados amos an ando ar ara aron ar arn ara 16: acept, alcanz, apoy, consider, declar, entr, estudi, expres, lanz, llam, lleg, llev, marc, ocup, realiz, trat 240 Covered Types 15: aba aban ada adas ado ados an ando ar ara aron arse ar arn ara 19: alcanz, anunci, apoy, consider, declar, entreg, estudi, expres, hall, inici, lanz, llam, lleg, llev, ocup, otorg, present, realiz, trat 285 Covered Types  Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 6: The number of clusters and their recall of unique Spanish suffixes as the scheme-cluster size threshold varies. The graph marker of each function at the threshold value of 37 unique licensing types is larger in size because it is this value of 37 which ParaMor uses in experiments reported in this thesis. r e o s t t r a y i e s i n g i n g s e a t i n g m i n g Table STYLEREF 1 \s  4. SEQ Table \* ARABIC \s 1 1 s s t.ting.ts res retrea t.ting res retrea .ing rest retreat retry roam .ing.s rest retreat roam Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 9: Above: A trie derived from the small vocabulary: rest, rests, resting, retreat, retreats, retreating, retry, retries, retrying, roam, roams, and roaming. Collapsible redundant sections of the trie are circled. Right: A snippet of the scheme network corresponding to the collapsed trie. The box representing each scheme is outlined in the same pattern as the circled trie branches from which it stems. Figure  STYLEREF 1 \s  5. SEQ Figure \* ARABIC \s 1 1: Pseudo-code implementing ParaMor s word segmentation algorithm. Section  REF _Ref190771388 \r \h  \* MERGEFORMAT  4.4.2 Section  REF _Ref190686481 \r \h  \* MERGEFORMAT  4.3 Chapter 3 Section  REF _Ref190686436 \r \h  \* MERGEFORMAT  4.2 Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 15:  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4 Revisited: Typical clusters produced by ParaMor over a Spanish corpus of 50,000 types. Paradigm Model-s .s s s s s t t +t ,t 1t 3t Jt t t t t t t u u %u 'u (u 0u Xu au pu $v v v v v v v v v v v v v x x x x x ľľຯyhkchkcmHnHuhkchkcmHnHu hkchkcjh<hffU h<hffjh<hffUhff h NHh h I h'NHh' h)WpNHh'vFhIh)Wphyp hMNHhMhZ hZNH-x x y y y y y y y y y z z z \z ^z z l{ |{ ~{ { { { { { { { { { { { { | | | | } } 0} 2} @} B} H} J} } } } } } ~zhbjh<hffUjh<h U h<h jh<h U h H*h hkchkcmHnHuhkchkcmHnHu hkchkcj h<hffU h<hffjh<hffUhff hffNH0} ~ ~ ~ >~ R~ r~ v~ x~ ~ ~ ~ ~ ~ 6 H L \ b d            J X Ā Ѐ ր  & 4 < üçڃh h`,/NHhyphkchkcmHnHuhkchkcmHnHu hkchkcjh<h`,/U h<h`,/jh<h`,/UhWqh`,/h`,/H*h`,/hX hffNHhbhjh2;hffh'vFh<2< N f | ā Ɓ   H J h j  * , ^ ` b d j l p 0 2 d f h j ŷ}rjhUhyshkc hkcjhUjh4oUh< h4oNHh4ohkchkcmHnHuhkchkcmHnHu hkchkcjh<hypU h<hypjh<hypUhyphypH*hyphhff,j t v x z | ʅ ) * 8 9 < N [ c d e | ׆ ݆ 3 4   b c i n o q  ڈ  5 H P j ʾضحةإ hsmh<h<h'bhq hH*hhyp hWqNH hffNHhffh2; hNHhh hWqhX!*h h4ohysjh4oUhkc> ' g <  w R α   !!! !!!!!!!!!!gdqgd7jT]gd<gd#gd bgdQ5|gdwk2gd gd'vFgdq gdqgd'bgdq1gdypj k s x ͉ Չ ։ " % 7 ] h y { | ~  ) * { | ‹ ؋ ً תxjh<hffU h<hffjh<hffUhhff hqNHhX jhX UmHnHuh'bhq h<NH hNH h2-hh<h hsmhsm hsmh hsmh'b hsmh<hsmh<NH.   6 D ƌ ڌ       2 : \  & 5 8 9 ? Z _ f h ֎ ׎  * + . ɽŽŽŽŽŽɽůŽŽūŧŧh'b h'vFh h hfh hWq hffNHhffh'vFhhq h<hffjh<hffUhkchkcmHnHuhkchkcmHnHu hkchkc=. / ʏ ˏ ̏    b d ؐ  * R X Z ‘ ʑ  ˽w˽q hG=NHjh<h \U h<h \jh<h \Uh \h'bhG= h2hG= hG=hG=hkchkcmHnHuhkchkcmHnHu hkchkcjh<hffU h<hffjh<hffUhWqhff hffNH*  & ( = N X b { | ے     0 1 A B      > J T V ” Ĕ N n p ~ ǫxq h'vFhG= h'vFh'vF h'vFNHh'vFh2hG=H*hkchkcmHnHuhkchkcmHnHu hkchkcjh<h'bU h<h'bjh<h'bUh'bh'hi}h h2hG=NHhfhG= h2hG= hG=H*)  4 5 F G N d e f g – Ζ Ж ۖ ܖ     \ ^  ̙ ԙ ֙ ɿvqkke h'bNH hz NH hz H*hkchkcmHnHuhkchkcmHnHu hkchkcjh<hz U h<hz jh<hz Uhz hz H*hz h NHh hG= h'vFhG= h'vFh'vF h'vFh'bh'b hi}NH hi}H*hi} h'vFhi}'  a b  Ԛ ښ ߚ   . 3 9 : ; < s t ɛ ʛ 8 9 C I s z {  + 6 7 I Ƿݡ h NHh2hi}NH h2hi} h~NH hNVNHh'h~ hVNHhVhwk2hh hNVH*h'bh'vFhNV hz NHhz hZ?I O \ b c ҝ L M O P ۞   R S T Z % * . / 4 o y Ơ Ƞ ɠ ٠  $ h2h][ h'vFh h'vFh'vF h'vFhQ5|h2hQ5|H* h2hQ5|hQ5|h2hwk2NH h2hwk2h@Qh~H* hC{NHhC{ h][NHh][h~h h hwk2NHhwk2h hV2$ % 9 ] ^ _ d q r v w x { ȡ ɡ ١ ޡ 0 P h i m Ѣ Ң  = أ ٣ L N X Z 2 p ¥ ѿѻջջշշշjh<Ujh<U h<NH hNHh<h h~NHh bh@Qh~H*h~h hwk2 h][NHh2h][NHh][ h2h][h2h][H*<   L M ~     $ - 7 ˧ ̧ a n Ψ $ * J K Q R t  ίί毧ίhffh'hs;h2h` NHh2h` H* h2h` h2hT+ hT+H* hT+NH h bNHh` hNHh h<NHh<h bhT+hkc hkcjh<U8 ~ ǫ ̫ ͫ ѫ     ݬ    C O P Y _ `   R ۮ ߮ L l v { ~ ȯ ԯ կ + , m n ͱ h` hffH*h'h    B P ȶ ž侵袭蒋h'vF h'vFh7jT h'vFh'vFhkc hkcjyh$Ujh7jTUh2h$NH h2h$h$ hn+h7jT hNHhh' h7jTNHh}h7jThY>h<hsmh Z l x z ȷ ʷ ط ڷ   ĸ Ҹ Ը ָ   $ & 4 6 8 : < > D F Ĺ Ķqeeah<hQ6hqmHnHuhQ6hqmHnHujshQ6hqU hQ6hqjhQ6hqUhq hn+hi}hQ6hQ6mHnHuhQ6hQ6mHnHu hQ6hQ6 hn+hn+jhn+hn+U hn+h"jhn+h"U hi}NHhi}hR&Ĺ ޹  % & , / @ A C t u 1 2 K v    ( ) * 6 b c   \ |   ^ b o r ̾ ;   ` a ſ h$h:&Bhi}6] h@Ahi}h2h7jTh} hi}NHhRhi}hhn+hR0J>:hn+hi}0J>:H ,  \ ^ y 3 Y > !!!! ! !!!!! !!gdu"gdZgdzgd& gdLgd<'gdz ] gd4]0gdwfgdQ61gd2gd7jTgd}gdRſ ƿ x ~ < D F r t   C g ~ j k ĽĽݑݑyuݑݑhqjhqUmHnHuh+ h$NHh$hQ6hQ6mHnHuhQ6hQ6mHnHujhQ6hQ6U hQ6hQ6jhQ6hQ6U hQ6hi} hQ6h2hi} h'vFh'vF h'vFh2 h2NHh2 h2h2.   ! A B U B C W Z [ h m v    h q B C V W Z [ \ ] ^ _ u ſ񲮪 hzNHh3-A hd, hzhzhQ6h3 hq hi}hi} hwfNH hUzGNHhUzGhUzGhi}0J>:NHhUzGhi}0J>:hRhwf hi}NHhHmhi}h$hi}0J>:9 8 : ; T U        8   e f j v y ~  * + 8 A | } V W . / m n w hZh{ h"mNHh"mjgh Ujh Uhkc hkcjmh Ujh<'Uh<'hz hzNHD @ _  ' 9 >  . 4 B M N d o p  + = C Z g h һ´ҦŸh& h NHh& h& h& h& h h& h h& hmdL h& hM h& hL h& h"mh& hzNH h& hz hzNHhZ hzH*hzhMhmdLh<'hLh"m6     $ % * M g h k = [ \ v     N X c s  ] ^   r ͽ͹͹͵jhzUhuChL hd, hzh hzNHhzh& h& hzNH h& hmdL h& hz h& h'h& h NH h& h h& hL h& h& h& hM:   O o  ' ) M n    J L V   * 8 B J N ৯hzmHsHhhz0J> hNHhjahLUjhLUhLh*1hLhz0J> hzNHhuC h*1h*1 h*1hzhAA hZhkc hkcjhzUjhzUhz0N h  / \ ` a h q u ( = > O | }     3 9 : H Y \ h i y    x Ľ hu"h' hu"h#hu"hz0J>huChu"hzNH hu"hAA hu"hzhWhAA hhz0J>hh*10J>h*1hZ hzNHhzhzmHsHh*1mHsH9x o p       # , - 5 \ e n o t d @ [ c k ޮ򔐔ޔޔޔޔhu"hzhu"hu"0J> hZ]hu"hZ6hu"hZ0J>hu"hz6hu"hz0J>hZhZ0J* hu"h# hZNHhZ hu"hWhu"hzNH hu"hz hu"hu"8 # > ? { , . 1 2 J [ ^   n  6 ; 5 O P   # $ & Z [   5 ^   ޾jhU=Uh5n h]2hz hU=6] hU=NHhU= hDNHhD hu"NHhu"hz hzNHK  [ * ~  @ g ;  % b f h n   ! !! !!! !!!!!!!!! & Fgdd, gdBgd6~gd+<gd+<gd]|\gdagdzgd2gd0+gdzgd5ngdU= + ,     " l   D F H J T V X \ f l x , . < P R T z     ܳܳܧܣܧܳhr hzNH h5nNHh5nhdjh SU h>hzj[h SUhuChzhkc hkcjhU=Ujh SUhU=@ N O     n p  S T S T n S T ` a   " G r x z {   h0+ hd, hz hhzhuChkc hkcjUh SUjhdUhd hzNHh5nhzH        K L R Q R  * Y Z \ ] ) * a b      $ % m       H W     I J } ~    X Y   ? C L R W ] hz6h02hz6 haNHh(hhaNH h(hhaha h xhzh5n hzNHhzh0+ h0+NHI]       ? @ M N   2 3 6 7 Q `    9 : f g b h $ * , 2 7 < ? @ : ; H I & K P Q U \ k  ·©· hkchkcjhiFhzUjhiFhzU hiFhz hqJhzha\zh]|\hz0JQh]|\ haNHha hzNH h02hzh0+hz@     f       > @  & N X        " B D R T Z \ ^ ` r        " # ) U Y g h o } ͼɲɲԞۗۼɓɏɏɉɀh]|\hz0JQ hzNHh]|\h86 hkchkcjOhiFhzU h+<NHh+< haNHhahqJhzNHhz hqJhz hiFhzjhiFhzUhkchkcmHnHuhkchkcmHnHu2}            G H n ~                    2 3 L T                   ڮڮڮh]|\hz0J* h+<0J>h+<h+<0J>>*h]|\h+<0J>hqJhzNHh]|\hz0J>NHh]|\hz0J> h+<NHh+<h]|\hz haNHha hqJhzh]|\hz0JQ8          $ % 5 L P ] v           2 B M [ ^ c d i }      , -            \ b d ͿŻɷűɫեՒ$jhqJhqUmHnHtHu hBNH h?2NH h6~NHhuChq hNHh6~h?2hhU,hBh+<h.p h;hzhzh]|\hz0J*h]|\hz0J*NH8d f h j n      * , Z \       C u    , - =    2 l m       3 4 ; < ߳߫xxohYehdNHhNPhd6NHhNPhdNHhYehd6NHhYehd6 hYehd hNPhdhNPhd6h|h99hd6NHh99hd6h99hdNHh99hdPJhYe h99hdhdhzh h8hz hBhB+  u    k    Z! B" " (# # % #& & ' ( $) D* !+ \, >- . . / !!!!!!!!!!!!!!!!!!!!!!!!!!Agd|AAX< > ? B C D E S i k n       5 Y [ _ `      V _ ` t     W X \ ! ! ! ! ! ! " ŹŹ~~hYehdNHhYehd6NHhYehy]mH sH hYehy6NHmH sH hYehy6H*mH sH hYehy6mH sH  hYehyh\kh\kNHh\kh\k6] h\kh\k h\kh'-hYehd6hYe hYehd1" " %" B" X" ^" o" p" " " " " " " " # # # # # # # # # $ % % % h% % % % & }v}oh`V`hh,YhR6NHh,Yhd6 h,Yhd hZha\z hZ6]hZhdh|hNPhd6NH]hNPhd6]hNPhdNHhNPhd6 hNPhdh'vFhdNH h'vFhdh'vFhd6mH sH h'vFhd6h'vFhdmH sH  hYehdhYehd6hYehd6NH & & & & & & & & & & F' w' x' {' ' ' 2( ( ( ( ( ( ( ( $) %) h) ) ) ) * * * D* n* {* * * * ӻ{p{c{hihdNHmH sH hihdmHsHhihdmH sH hihd6NHhihd6 hihd hYehdhZhNPhdNHhNPhd6 hNPhdh hdNHh hd6 h hdh,Yhd6NHh,Yhd6h| h,Yhdh,YhdNH&* * * + + !+ v+ w+ + + + + [, \, v, , , , - - -- ;- - - - - - - - - P. Q. V. X. Z. ƹƮƠƔ{rnrrrrg_hYehd6 hYehdh|hihdNHhihd6NHhihd6 hihdhihi6mH sH hihi6]mH sH hihimH sH hihiNHmH sH hihimH sH hihdmH sH ha\z6mH sH hihd6NHmH sH hihd6mH sH "Z. [. q. . . . / S/ T/ _/ / / / / / / 0 0 0 F0 M0 N0 [0 0 0 0 0 0 81 Q1 νΨyk[kyTLh\khd6 h\khdhNPhd6NH]mH sH hNPhd6]mH sH hNPhdmH sH h hd6mH sH h hdNHmH sH h h mH sH h hdmH sH h'vFhd6NHh'vFhd6h'vFhdNH h'vFhdhYehYe6] hYehYe hYehdhYehd6hYehd6NH/ / [0 0 n1 2 2 4 n4 C5 5 6 7 c8 9 b: ; ; < r= = > m? @ A HB 6C !!!!!!!!!!!!!!!!!!!!!!!!Agd99AgdcAgd Agda\zAgdinAgdaAQ1 R1 Y1 n1 1 1 1 1 1 1 2 2 2 }2 2 2 2 3 M3 N3 e3 3 3 3 3 4 94 A4 a4 b4 4 4 5 5 5 5 Y5 5 5 5 5 5 Q6 R6 q6 6 7 7 ̼杕|h hd6 h hdhNPhd6NHhNPhd6 hNPhd hihdhihd6h\khdNHh|h,Yhd6NH]h,Yhd6]h,YhdNH h,Yhd h\khdh\khd6h\khd6NH/7 7 7 7 7 7 7 7 7 7 &8 N8 b8 c8 8 8 8 9 9 I9 J9 R9 S9 W9 9 9 9 ': O: ʾʾʦ~ppdVdJh99h^{6mH sH h99h^{NH]mH sH h99h^{]mH sH h99ha6NHmH sH h99ha6H*mH sH h99ha6mH sH h99haNH]mH sH h99ha]mH sH h99hin6mH sH h99hin]mH sH h99hd]mH sH  h h0Zh hdNH h hdh hd6h hd6NHO: b: : ; ; ; 4; S; U; ; ; ; ; ; ; ; ; ; ; ; ; 1< 2< {< |< < < < < < = F= U= W= f= g= o= ǿǷDzDzDzǞwpi`ih>h99NH h>h99 hYeh99hYehd6NHhYehd6 ha\z6hYehdNH hYehdhZh hc]hchc] hc]h h0Z6h hy]h h0Z]h ha]mH sH h h^{6mH sH h h^{]mH sH h99h^{]mH sH $o= r= = = = = > :> E> F> p> > ? 4? 5? Q? m? ? ? @ X@ Y@ ^@ @ @ @ @ @ @ @ A Ӻ{tmtdtXh99hy6mH sH h99hyNH h99h h99hy ha\zhha\zhd@q6NH]ha\zhd@q6]ha\zhd@qNH ha\zhd@qha\zhd6NHha\zhd6 ha\zhdhYehd6NHhYehd6 hYehdh>h^{]mH sH h>h^{6mH sH  h>h^{ h>hdA A A !A kA lA pA A A A B 5B GB HB IB B B C 5C 6C D zD |D E {rjcWIWh hdNH\mH sH h hd\mH sH  h hdha\zhd@q]ha\z]mH sH ha\z6mH sH hd@q6mH sH hd@q]mH sH hZ]mH sH h^{]mH sH h99h^{6mH sH h99h ]mH sH h99h^{]mH sH h99hy]mH sH h99hy6NHmH sH h99hy6mH sH h99hy6H*mH sH E E E bE E E F F EF iF jF F F G 1G ?G @G FG nG G G /H [H ]H H H ĸġē{tjt^P^IAhYehd6 hYehdh'vFhd6]mH sH h'vFhd]mH sH hfh996] hfh99hNPhdNHhNPhd6 hNPhdhMzhMz5\mH sH hMzNH\mH sH hMz6\]mH sH ha\z6\]mH sH hMz\mH sH hd\mH sH h hd\mH sH h hd6NH\mH sH h hd6\mH sH 6C E F FG G ]H I I J K 6M M qN /O O [P Q Q FR HR IR zR R ]U W !!!!!!!!!!!!!!!!!!! !!gdd, gd{h X^gdJu & Fzd\$^`zgdJuAgd AH H H vI I I I I @J AJ VJ J J J K LK MK 3L mL nL L L L L M M M M M M M M M M M N 0N 1N 7N ^N oN ׿znzg hfhfhfhf6NH]hfhf6]hihf6] hihf hih'-h'-h|h,Yhd6NHh,YhdNHh,Yhd6 h,Yhdh,Yh 6h,Yh NH h,Yh hYehd6NHhYehd6hYehdNH hYehd(oN qN N N N N DO O O O P 5P 6P KP [P P P P P P P Q Q Q Q Q Q Q Q Q R R R R 3R ER FR GR úò򡝑wk` *hi]mH sH h'vFhi]mH sH h'vFh^{6NHmH sH h'vFh^{6mH sH h'vFh^{]mH sH ha\z h'vFh^{hNPh 6NHhNPh 6hNPh NH hNPh hNPhd6NH]hNPhd6]hNPhd6NHhNPhd6 hNPhd hfhf%GR HR IR UR yR zR |R R R R R R R R /S 0S BS CS S S S S S S S S T T )T *T 5T 6T 9T ST gT T T T T T T 9U NU ̾}vrhc h{hh{h h"<6NH h@Bhd, hchd, h"<6h{h hd, NHhd, hJuhd, 5CJHNH\aJHhJu5CJH\aJHhJuhd, 5CJH\aJHhJuh6?5CJH\aJHhJuhJuCJaJhJuhJuB*CJaJphhJuhyh^{*NU OU [U \U ]U eU U U V V qV vV xV ~V V V V V V V V V V V V V V W W W W X X (X .X 7X HX lX tX |X X X X X X X X Y Y Y Y 9Y ZY |Y }Y Y Y 'Z (Z Z Z Z Z Z Z Z Z 1[ s[ [ [ [ [ [ ̾ hd, 6h"<6hd, 0J* hL1hd, h1}hd, 6h"<6hd, 0JQ hd, NHh"<6hd, h{hhc hcNHIW 'Y \ (_ ._ 2_ 6_ 8_ :_ >_ _ ` c c #d Pd }d ~d d d e e e ! !!!!!!!1!1!1!1!1!1!1!1!1Cgd<[HCgd)!gdN X^gdJu  & Fd\$gdAgdd, `gdd, gd)gdd, [ [ \ =\ >\ \ \ \ &] C] L] M] f] g] ^ ^ ^ ^ ^ V^ X^ ^ _ &_ (_ *_ ,_ ._ 0_ 2_ 4_ 6_ 8_ :_ <_ @_ B_ V_ Z_ ~zvrvk huChhoh)!hdhd, hd, mH sH  h!qhd,  *h5Fmhd, 5 *hd, 5jhd, 5UmHnHu h]hd, jhd, UmHnHu hd, NHhkc hkcj *h%hd, Uj *hd, Uhd, h)NHh"<6h)&Z_ f_ h_ _ _ _ _ _ _ _ ` ` ` J` ^` z` ` ` ` ` ` ` a a a Da Fa a a a a a a a b b b ȽȲvqmg\qmjhuNU h1-#NHhkc hkcjIhuNUjhuNUhuNhuCh1-#h)!hhd5CJH\aJHhJu5CJH\aJHh5CJH\aJHhuC5CJH\aJHhh5CJH\aJHhuChCJaJhhB*CJaJphhhB*CJaJph$b b Db Fb Hb Jb Tb Vb ^b b b b b c c "c Hc c c c c c >d Ld ~d d d d xe ze e e f f f 2f 4f 6f 8f :f f f g ſzqqkɰ h6<:0JJhx=huN0JJhx=h`0JO h`0JO h<[H0JOhx=h(~S0JOhx=h<[H0JOhx=huN0JO huC0JOhx=h6<:0JO hJu0JO hN0JOh)!hJuhN hNHhh1-#hkc hkcjChuNUhuNjhuNU*e e f 6f 8f f 8g h Qh _h vh h h h h 'i *i +i Mi ]i `i ai i i i i i !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1Cgd)!Cgdx=Cgd6<:CgdNCgd<[Hg @g Bg tg vg g g g g g g g g g h h h h h h (h 1h :h Dh Kh ]h ^h _h ph qh rh vh h h h h h h h h h i %i 'i vh`h<[H0JJ h`0JOh`h6<:0JOh`h<[H0JOh<[Hhx=hx=0JJhx=huN0JJhuNhx=hx=h6<:0JOhx=h<[H0JO hkc0JO hkc0JO hx=0JOjhx=hZs0JOUjhx=hx=0JOUhx=hx=0JO+'i (i )i +i 2i 6i 7i Ci Ki Mi Ni Ui [i \i ]i ^i `i ai hi xi i i i i i i i i i i i i i i j mj nj }j ~j j j j j j j j j j j j j j j j ǽhZsh)!0JFhZsh(~S0JFhZsh(~S0JJh(~Sh)!hZsh)!0JJhZsh)!0JHh huC0JOhZsh6<:0JOh6<:hZsh<[H0JJhZshuN0JJhZshuN0JhuNh<[Hh`huN0JJ5i i i /j mj nj j j j k k Ck \k xk k k k l l l l 0l Ol dl |l l l l l !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1Cgd6<:Cgd)!j j j j j j j j k k \k ^k bk ok xk {k k k k k k k k k k k k k k k k k l l l l l l l $l %l ,l /l 1l :l Al Dl Kl Nl Ol Vl [l \l `l al bl dl rl sl |l ~l l l ǦǝǦh*h6<:0JFh*h6<:0JJh*h)!0JHh*h(~S0JH hJu0JHh6<:h*h)!0JFh*h)!0JJhZsh>t0JOh>thZsh)!0JFhZsh)!0JJh)!>l l l l l l l l l l l l l l l l l l m m m !m %m (m 1m 7m :m Gm Km tm wm xm m m m m m m m m m m m m m m m m m m m m m m n 'n -n 0n 9n Bn Dn Mn Pn Sn \n en qn zn |n n n n n hPuh*h)!0JHh6<:h*h(~S0JFh*h(~S0JJh(~S hzq0JJhzqh)!h*h6<:0JJh*h)!0JJHl l %m Km m m m m m m m m m n /n Qn n n n n 3o Wo vo o o o o o o !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1CgdPuCgd)!n n n n n n n n n n n n n n n n n o o o o 0o 1o 6o 7o ;o Ao Bo So Vo [o \o mo no zo {o o o o o o o o o o o o o o o o o o o o p p p !p &p 6p 8p Lp Np Op Yp Zp mp ٴh6h0JJhh6h)!0JH h)!NHh6h>t0JFh6h>t0JJh>thPuh)!h6h)!0JJh6h)!0JFh6hPu0JJCo o o o p :p ;p Zp yp p p p *q Yq q q q q q q q r -r rr r r s -s !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1Cgd(CgdzqCgd>tCgd)!mp np yp p p p p p p p p p p p p p q q *q Vq Wq [q `q aq q q q q q q q q q q q q q q q q r r >r Hr Vr br 󻷱yh6hzq0JOh6hPu0JOh6h0JOhhzqhzq5\ hPu0JJh6h)!0JH hJu0JHh)!h6h)!0JJ h>tNHh6h>t0JFhyihzq0J hNHh6h>t0JJh>thzqhh6h0JJ-br gr qr xr yr r r s s s s s s )s ,s /s 8s As Cs Es Fs Ls Os hs is js os vs |s ~s s s s s s s s s s s s s s s s s s s s s s s s s s t zt ؽشǫǫϢˏhzqhzq0JOh$,hzq5\h6h0JJh6h(0JJh6h0JF hNHhPuh(hzqh6h0JJhh6h0JOh6hzq0JOh6hPu0JOh6h(0JO8-s Ps ks s s s s s s 7t {t t t t u u u u u v bw x !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!|!|!|($ & F d^`a$gd'q ($da$gd'q Cgd$,CgdJuCgdzqCgd(Cgd)!zt {t |t t t t t t t t t t t t t t t t t t t t t u u u u u u u u u u u `u au bu v v v 0v 2v 4v ~v v v v Ᾰ󞔊wwwwhk6H*] hk6]hkjhk0J+Uh$,h)!5\h6h)!0JJh6h60JJh6h< h0JJ h<0JJh6h(0JJh( hNH hzq0JJhzqh<hzq0JJh6h0JJhh6h0JO.v v v v v v v v v Pw Tw Vw w w w w w Zx ^x `x zx |x ~x x x x x x 0y 4y 8y >y Jy Ry y y y y y y z z fz jz lz oz pz vz {z z z z z z z z z z z z z z z z z z z z ǽhk0J%mHnHu hk0J%jhk0J%UhuIa hk0J> hkNHjhk0J+U h3"hk hk]hkhk6H*] hk6]Cx z z z z z z z z z z { { { { { { ({ ){ 6{ 7{ !|!!#$a$#$a$gdx_t#h]hgdx_t #&`#$gdx_t#$a$gd #&`#$gd#h]h#&`#$U$a$gd'q z z z z z z z z z z z z z { { { { { { { { { { { { { { "{ #{ &{ '{ ){ *{ 0{ 1{ 4{ 5{ 7{ 8{ 9{ { { { { | | | | )} *} Z} [} \} O P h i       ഭ hkNH *hkH* *hkNH *hk *hk5jhk0JUhuIahk0J%mHnHu hdhkhk hk0J%jhk0J%Uhl0J%mHnHu?7{ 8{ { { { { | G| | [} O  % 0  K gd$ gd!#gd, 1`` / 0 1 p q ށ ߁ ؂ ق u v   … Å م ܅   O P j m      ( ) 1 4 A F J žhZhk6h`hk6 hk6 hk5\jhZhkU hZhkhZhkH* hkNH hVeohk hk5hkjhk0JUBJ M X Z _ d i l q t ш ӈ ܈    @ C D Ċ Ŋ ˊ ̊ ӊ Ԋ ڊ ۊ   % / ; ? ] ^ hd[hkOJQJhkOJQJhd[hk5NHOJQJ\hk5OJQJ\hd[hk5OJQJ\jhk0JUhZhk6H*hZhk6 hk5 hkNHhk hk6: Ɋ ъ ؊ $$If`a$gd$ l3 'kd=$$If֞) A Yr!)  t0449a      $$If`a$gd$ lT$$If`a$gd$ l\  'kda$$If֞) A Yr!)  t0449a ; @ E K Q W ] $$If`a$gd$ l$$If`a$gd$ lT] ^ 'kd$$If֞) A Yr!)  t0449a^ d e i m ɋ ʋ ׋ ؋    э ҍ       ^ ` 4 6 ޕ ˩ˡˡˡˡˡ˩˙ˎ˙˃˃ hkNH hkjhkUjhkUhkmHsHjhk0JUhk5NHOJQJ\hk5OJQJ\hkhd[hkOJQJhkOJQJhd[hk5NHOJQJ\hd[hk5OJQJ\3^ i n s y  $$If`a$gd$ l$$If`a$gd$ l$$If`a$gd$ lT 'kd$$If֞) A Yr!)  t0449a ċ ɋ $$If`a$gd$ l$$If`a$gd$ l$$If`a$gd$ lɋ ʋ 'kd$$If֞) A Yr!)  t0449aʋ  $$If`a$gd$ l\$$If`a$gd$ l   ( '% gdkd$$If֞) A Yr!)  t0449a( ) ב Ғ Ӓ   + @ O a !e!!!! u5 $IfdhgdFgdFgdޕ < = 8 9     @ B D F H       ` a b s u › Û hkCJOJQJaJhk6CJOJQJaJhk5CJOJQJaJhkOJQJhkOJQJ]hk5OJQJ hk0J6hk5\mHnHuhk5\mHnHujhk5U\ hk5\huIahk hkNH2a b s u › @9@9i@9i@9@9@9@9@9@9@9K@9K@9 $$Ifa$fkd$$If!! t044 a › Û Λ Л ٛ ܛ <r3@93@93i@93i@9 $$Ifa$kd5$$Ifֈm >!sU t044 aÛ Λ Л ٛ ܛ     & ( 1 4 ; ? D I M S U V b d n q y } Ĝ Ɯ ќ ֜ ۜ     % * 1 6 7 = ? D J N V Y c e q x } hkCJOJQJaJhkhk6CJOJQJaJhk5CJOJQJaJTܛ    @9@9@9@9@9@9K@9K@9 $$Ifa$  & ( 1 4 <r3@93@93i@93i@9 $$Ifa$kd4$$Ifֈm >!sU t044 a4 ; ? D I M S U @9@9@9@9@9@9K@9 $$Ifa$U V b d n q <r3@93@93i@93i@9 $$Ifa$kd%$$Ifֈm >!sU t044 aq y } @9@9@9@9@9@9@9@9@9@9 $$Ifa$ )r @9 @9 $$Ifa$kd$$If֞m >!s t044 a Ĝ Ɯ М ќ ֜ ۜ i@9i@9@9@9@9@9@9@9@9@9@9 $$Ifa$ )r @9 @9 $$Ifa$kd#$$If֞m >!s t044 a    % * 1 6 i@9i@9@9@9@9@9@9@9@9@9@9@9 $$Ifa$ 6 7 = ? )r @9 @9 $$Ifa$kd*$$If֞m >!s t044 a? D J N V Y c e q x } i@9i@9@9@9@9@9@9@9@9@9@9@9 $$Ifa$ ͝ )r 4@9 4@9 $$Ifa$kd1$$If֞m >!s t044 a ͝ ѝ ۝ ݝ   H J n p ž ƞ Ȟ ̞ Ξ Ҟ Ԟ ؞ ڞ ޞ          " & ( , . 2 4 8 : H J h j l ĹĹhk5\mHnHujhk5U\ hk5\hkOJQJhkCJaJ hkNHhkhk6CJOJQJaJhk5CJOJQJaJB͝ ѝ ۝ ݝ  @9@9@9@9lrj!j^s^$s^.s ,$d$Ifa$kd8$$IfF>!> t0    44 a $$Ifa$   $ 6 H J |s)s's.-ckd$$If4wF ,88 t,    4af4 &$d$Ifa$`kd$$IfhF ,88 t,    4aJ L ` n p )s's-s)s.sckdj$$If4wF ,88 t,    4af4 &$d$Ifa$ s)s.s1`kd$$IfF ,88 t,    4a &$d$Ifa$`kd8$$IfF ,88 t,    4a ž ƞ Ȟ ̞ Ξ Ҟ Ԟ ؞ ڞ ޞ     qqO/      " & ( , . 2 4 8 : ? @ ң ԣ    0 UOO !xxxxsww$a$$xa$$<a$l n p r t @ G H W     @ B D F H r t ԣ    < B D ` x | ~ Ȥ    ùùÞùù݈hkCJaJhkCJOJQJ^JaJhk56\]^Jhk6CJOJQJ^JaJhk56\^J hk^J hkNHhk5\mHnHuhk hk5\jhk5U\hk5\mHnHu50 < @ B D b p x | ~ Ȥ ֤ wwswsDDDDsssss$s.s ,$d$Ifa$$xa$$<a$/$a$      |)'1-ckd$$If4wF ,88 t,    4af4 $$Ifa$`kd$$IfhF ,88 t,    4a      " $ & ( * 2 8 : > D F J X Z x z | ~ ¥ D F P ^ ` ~ ¦ Ħ Ʀ Ȧ R T  hk0J<5\mHnHuhk0J<5\mHnHujhk0J<5U\hk0J<5\ hkNHhk5\mHnHuhk5\mHnHujhk5U\ hk5\hkCJaJ hk6hk4 $ ( * 4 6 8 )'-).ckdy$$If4wF ,88 t,    4af4 $$Ifa$8 : @ B D F ).4`kd&$$IfF ,88 t,    4a $$Ifa$`kdG$$IfF ,88 t,    4aF H J N P Z [ ` e j k p q v w z   PsPsPsPsP;PsPsPsPsPsP;PsPsPsPsss $7$8$H$a$  R U V Y [ ` j k p q v w z          ! " ( ) + < = C D I J L V W ^ _ c d k l p q u v z { hk0J7NHaJhk0J7aJhk0J7CJ aJ hk0J hk0J7 hk6NH hk6hk hk5L Ǩ Ϩ ը ܨ          ! " ( ssssssssPsP;PsPsPsPsPsPsPsPsP;PsPsP;Ps/ $7$8$H$a$( ) + 1 6 < = C D I J L Q V W ^ _ c d k l p q u v z { PsPsPsPsPsP;PsPsPsPsPsP;PsPsP;PsPsP;Pss; $7$8$H$a$ P R ֫ ګ   " sPsPsPsPsP;PsPsP;Ps ) s) ssssssss64$a$ $7$8$H$a$ ǩ     @ B D F H ̪ Ϊ   4 6 D F L N ֫ ګ " $ 0 2 4 R T V X Z \ ^ $ & 첞 hk0J55hkchk0J<hkchk0J<jhkUjhkU hkNHhk5mHnHuhk5mHnHujhk5U hk5 hk0J7hk hk0Jhk0J7CJ aJ 7" $ $ % - 0 6 > E I Q W Y Z [ a d i o w ~  sssssssssssssssssss6gdG&  % - 0 Y Z [ a d î   + . O P [ _ ¯ ï ȯ ʯ ԯ կ ݯ   B D V X f p v Ѻhk0JmHsHhk0JmH sH hk0J6mH sH hk0J7mH sH hkmH sH hk5CJNHaJmH sH hk5CJaJmH sH hkCJaJ hk6 hk0J hk0J7 hk5hk9 î Ǯ ͮ Ԯ ܮ        + . 1 6 < ssssssssssssssssssssss6< B G M O P [ _ c i p u } ȯ ʯ ί ү ԯ sssswswwswswswswswswswswwswswswswswseseeseses6ԯ կ ݯ    $ @ N V X f p Ȱ Ұ ڰ sssssssssssssssssssss6v x Ȱ Ұ $ . 8 j l v ~ б ұ ڱ    ! 1 2 7 ; < I K [ \ ^ c | } ² ò Ų ʲ Ͳ β   , 7 8 ? ¸¸¸ hk5hk0JNHmH sH hk0JmH sH hkmH sH hk0J7mH sH hk0JNH hk0J7hk hk0Jhk0JmHsHhk0JmHsHB   " $ . 8 @ R \ v ~ ȱ б ұ ڱ     sssAsAAsAsAsAsAsssssssssssss6  ! ) , 0 7 ; < I K O S U [ \ ^ c g m r x | } ssssssHsHHsHsHsHs9s99s9s9s9s9ssssss6 ² ò Ų ʲ Ҳ ղ ۲    spsppspspspspsssssssssssssss6 ( , 2 6 7 8 . 0 : D R ^ j | sssss_ ssssjjj :$  Va$6? @ O     @ B D F H  0 : D ^ z ƶ ж Ҷ ֶ ض    0 2 Ŀַַַַַַ֧֧֯֯֠jhk5U\ hk5\hkmHsHhkmH ,sH ,hkCJaJ hk6 hk0J hk0J7 hkNHhkhk5mHnHuhk5mHnHu hk5jhk5U6 Ķ ƶ Ҷ ض ڶ  zv:zj{v:{j{v:{j{v:{js#s#3b{v:{j{|v:$a$ :$ H Ma$ 9$ hHMa$8d :$  Va$ 9$ hVa$ :$  ha$9$ a$2 4 6 8 : < r t v x z  ȹ ɹ ʹ ͹ ι Ϲ ѹ Թ ݹ     ˾˶˾˶˾˶˭˶˥˶˥˶˥˶˥˶˥˶hkmHsHhk>*mHsHhkCJaJhkmH ,sH , hk>*hk hk5 hk5\jhk5U\hk5\mHnHuhk5\mHnHuA ʹ ι Ϲ ݹ  |j|{v:{j{{v:{j|v:|j|v:|jv:jv:jv: :$  Va$ 9$ hVa$ 9$ hHMa$$a$ :$ H Ma$    2 6 7 F X \ ] m ƺ ̺ Ϻ v:jv:v:jv:v:jv:v:jv:jv:v:j|v:|v:|v:|j :$  Va$ 9$ hVa$ $ 2 5 6 7 = A X [ \ ] ^ _ w } ~ ƺ ˺ ̺ κ Ϻ к Ժ ٺ ں ܺ ݺ ޺        ! $ 0 2 4 5 6 7 : > ? A E Ƚ hk>*hk>*NHmH ,sH ,hk>*mH ,sH ,hkNHmH ,sH ,hkmH ,sH ,hkCJaJhkhk>*mHsHhkmHsHHϺ к պ ں ݺ ޺     % 2 5 6 : ? @ A G K L M |v:|v:|jv:v:v:jv:v:v:j|v:|j||v:|j|$a$ :$ H Ma$ 9$ hHMa$ :$  Va$ 9$ hVa$/E F G J K L M O Q V Y Z [ \ e h i l m n o q s  ̻ ϻ ֻ ٻ ڻ ۻ ݻ ޻ ߻             hk>*NHmH ,sH ,hk>*mHsHhkmHsHhkCJaJhkhkmH ,sH ,hk>*mH ,sH ,KM V Z [ \ i m n o  ƻ ֻ ڻ ۻ ߻ {v:{j{v:jv:jv:v:jv:v:jv:v:j|v:|j$a$ :$ H Ma$ 9$ hHMa$           % + / 2 3 |{v:{j{{v:{j{{v:{j{{v:{v:{j{v:{v:{v:{j :$  ha$9$ a$ :$ H Ma$ 9$ hHMa$$a$     ! " + - . / 1 2 3 5 6 ; < ? C F H J K L N O R Y [ \ f h j k l o p s y z ~ ϼ Ѽ Ҽ Ӽ Լ ׼ ۼ ܼ ݼ ޼ ѽhk>*NHmH ,sH ,hk>*mHsHhk>*mH ,sH ,hkNHmH ,sH ,hkmH ,sH ,hkCJaJhkhkmHsHJ3 9 ? D H K L R Y _ d h k l s z } ~ {v:{v:{v:{v:{j{v:{v:{v:{v:{v:{j{v:{v:{jyv:yv:yv:yjzv:zv:zv: :$  ha$:$  ha$9$ a$ Ƽ ̼ Ѽ Ӽ Լ ׼ ܼ ݼ zv:zjzv:zv:zv:zv:zv:zj{v:{j{v:{j{{v:{j{$a$ :$ H Ma$ 9$ hHMa$ :$  ha$:$  ha$9$ a$         : @ C E F G H T U b c d e g h i j k l m n o p r s t u w y z { }  쳢쳢쳢쳢hkCJaJ hk5CJH*OJQJ^JaJhk5CJOJQJ^JaJ hkaJhk>*mHsHhkmHsHhkNHmH ,sH ,hk>*mH ,sH ,hkhkmH ,sH ,hkCJaJ<     ( 5 E G H e h i j k l m n o s t y z  v:v:jv:v:v:jv:j$v:gv:gv:gv: :$  Va$ 9$ hVa$     @ B F H J 0 2 ^ ` Ͷܬܟytt hk>*hkchkmHnHuhkchkmHnHu hkchk *hkj *hkUj *hkUhk5mHnHuhk5mHnHujhk5U hk5hkhk5CJOJQJ^JaJ hk5CJH*OJQJ^JaJ-     $ , 2 6 7 8 A E I N T X \ ] ^ gv:gv:v:v:p sssssssBsBBsBsBsBsBsBs6        , 7 8 : ; A E T ] ^ ` b h l y    & + . hk0J7>* hk0J7hk0JmH sH hkmH sH hk0J7>*mH sH hk0J7mH sH hk0Jhk0JmHsHhk0JmHsHhkmHsHhk0J7>*mHsHhk0J7mHsHhk6^ h l r y ~ ssssssssssssssssssssss6       & +   * < D F N P R Z \ ^ f h sssssssssssssssgsgshshsgsgsgdY6 * D F N P R Z \ ^ f h j r t v ~ $ B D J h n  3 B C F U X ~   " # * / 2 * D 켴hk0JmH sH hkmH sH hk0J7mH sH hk0JmHsHhk0JmHsH hk0J7 hk5 hk0J5hk hk0Jhk0JmHsHCh j r t v ~ gsgsisis1_ $$Ifa$gdY eX1OOO_ $$Ifa$ $+$If]+a$kd$$If\6F  d t064a eX1OOO_ $$Ifa$ $+$If]+a$kd$$If\6F  d t064a    e\1\\\_ $$Ifa$kd$$If\6F  d t064a  " $ 2 @ B e] [RRRR_ $$Ifa$$a$gdYkd$$If\6F  d t064aB D J R \ f e\\\\_ $$Ifa$kd$$If\ 0 d t064af h n x e\\\\_ $$Ifa$kd$$If\ 0 d t064a e\\\\_ $$Ifa$kd$$If\ 0 d t064a e][NEtEzE^ $$Ifa$ $+$If]+a$$a$gdYkd$$If\ 0 d t064a  eXOtOzO^ $$Ifa$ $+$If]+a$kd$$If\v  d t064a     eXOtOzO^ $$Ifa$ $+$If]+a$kd$$If\v  d t064a   # ) / eXOtOzO^ $$Ifa$ $+$If]+a$kd$$If\v  d t064a/ 0 1 2 3 : A B e]f [RRRR_ $$Ifa$$a$gdYkd$$If\v  d t064aB C F J O T e\\\\_ $$Ifa$kd$$If\ 0 d t064aT U X \ b h e\\\\_ $$Ifa$kd$$If\ 0 d t064ah i j o u { e\\\\_ $$Ifa$kd$$If\ 0 d t064a{ | } ~ e][VsQVsVsVsVs6gdYgdY$a$gdYkd$$If\ 0 d t064a      " # * sssssssssssssssssssss6gdYgdY* /   * < D F Z b j v   $ sssssssssssssssssssssgdY6gdYD F J N Z b & ( , 8 D L X Z d f ~    vkkhk5mHnHuhk5mHnHujhk5U hk5hk0JmHsHhk0JmHsHhkmHsHhk0J7>*mHsHhk0J7mHsHhk0JmH sH hkmH sH hk0J7>*mH sH hk0J7mH sH hk0J hk0J7>* hk0J7hk*$ & ( D L b n ~ Z [ b n v ssssssssq s  $$If^`a$ $$Ifa$gdY6gdYgdY    # [  Q R e f g 䶪{sdZjhkEHUjzK hkCJUVaJhkOJQJjhkEHUjEK hkCJUVaJjhkUhkCJNHOJQJaJhkCJOJQJaJhk5CJOJQJaJhk5CJOJQJaJhkCJaJhkCJOJQJaJhk5OJQJhk hk5jhk5U# ntv:n*v:n1v:n v:$$If^`a$ $$Ifa$vkd$$If4P\d5'X2 #  tX24af4 Ou:>5Qs5 v:55tv: $$Ifa$$qT$If]q^Ta$kd$$If4֞d5')-@/X2 9@ tX24af4   *v:1v: v:5Qskd $$If4ִ d5')-@/X2 b 9@ tX2    4af4 $$Ifa$ Q i o u {  v:1tv:*v:1v: v: $$Ifa$g h i       : < ʻߋʆwh^jyhkEHUjXK hkCJUVaJhk5CJH*OJQJaJ hk5hk5OJQJj hkEHUjK hkCJUVaJhkCJNHOJQJaJhk5CJNHOJQJaJhk5CJOJQJaJhkOJQJhkCJaJhkCJOJQJaJhkjhkU# <`33Qs3 v:3T $$Ifa$kd $$If4ִ d5')-@/X2 b 9@ tX2    4af4    tv:*v:1v: v:3Zkd$$If4ִ d5')-@/X2 b 9@ tX2    4af4 $$Ifa$   Q v:Ttv:*v:1v: v: $$Ifa$$qT$If]q^Ta$ 8  5 < i33Qs3 ! 3 $$Ifa$kd$$If4dִ d5')-@/X2 b 9@ tX2    4af4 8         1 2 3 4 5 I J K P Q i   ۽۪ې}sk\ېMjK hkCJUVaJhk5CJNHOJQJaJhkCJaJjYhkEHUjK hkCJUVaJhkjhkU#jhkCJEHOJQJUaJ%jK hkCJOJQJUVaJjhkCJOJQJUaJhkCJNHOJQJaJhkCJOJQJaJhk5CJOJQJaJhkCJOJQJaJ5 : ? D I J tv:*v:1v: v:3 kd$$If4ִ d5')-@/X2  b 9@ tX2    4af4 $$Ifa$J K i     Qs v: tv:*v:1v: v: $$Ifa$       # $ 3     @ B F H J . 0   ƺƯ鞒Ѝ{ hk5>*hk6OJQJ] hk6hC}*hk5CJaJhC}*hkCJaJ hkNHhk5mHnHuhk5mHnHujhk5U hk5hkCJaJhkCJOJQJaJhkjhkUjhkEHU0    < :X2:2X2\K:dgdC}*kd $$If4ִ d5')-@/X2  b 9@ tX2    4af4 jjjjOikd!"$$If\ @`     t064a $$Ifa$ $+$If]+a$ eXOOO $$Ifa$ $+$If]+a$kd"$$If\ @`     t064a eXOOO $$Ifa$ $+$If]+a$kd#$$If\ @`     t064a  e`^NH-\K^^!C>^^ d^`gdC}*$a$kd$$$If\ @`     t064a    # $ 2 3 F G H I d e * , L N \ ^ f h      j l 緩緩j(hkUhkOJQJhk6OJQJ]hkchkmHnHuhkchkmHnHu hkchkj'hkU hk5j%hkEHUjK hkCJUVaJjhkU hkNH hk5>*hk0   * , @ B V X Z f h      H J N P R A B c d hIhk5:OJQJaJ hkH* hkNHhk5mHnHuhk5mHnHujhk5U hk5hk5H*OJQJ^Jhk5H*OJQJ^Jhk5CJaJhk5OJQJ^Jhkj(hkU-  * , @ B f h G / PPSPSPSd$If^gdI$d$Ifa$gd!#  ( , - . 0 2 3 5 6 8 9 : @ D E $ , 6 οsdRdR"h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJhk5CJOJQJaJht=Ohk5CJOJQJaJht=Ohk5OJQJaJhk5:CJOJQJaJht=OhkCJOJQJaJh#hk5:OJQJaJhIhk5:OJQJaJ"hIhk5:CJOJQJaJ     ! ( ) * + , - .  PSPSPS PS Ff3$d$Ifa$gd!#d$If^gdI$d$Ifa$gdI d$Ifgd!#Ff.$d$Ifa$gd!#. / 0 3 6 9 : ; < = > ? @ A B C D E G   PSPSPS,r ,r,r $$Ifa$gd!#$$If]a$gd!#Ff:$d$Ifa$gd!#$d$Ifa$gdI d$Ifgd!#        $ , 6 ~ ,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r$$If]a$gd!#FfPA $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#6 f h ~   . 4    b ųlųYYų$hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsH"h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH (hg_hkCJNHOJQJaJmH sH $hg_hkCJOJQJaJmH sH         " & ( ,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r$$If]a$gd!#FfH $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#( , . 4     ,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r$$If]a$gd!#FfN $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#  b d h j l p r t v z |  P R X ,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,rFf\ $$Ifa$gd!#$$If]a$gd!#FfU $Ifgd!#$:$If]:a$gd!#b d h    P R X v z   , . n p | J P f l ооrИоrrrоИо$hg_hkCJOJQJaJmHsH$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJht=OhkCJOJQJaJ,X Z ^ ` b d f h l n r t v z    ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r$$If]a$gd!#Ffc $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#        | ,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r$$If]a$gd!#Ffj $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# f l ,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,r$$If]a$gd!#Ffp $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# 2 4 : Z `    0 4 R T z ~   F H ߘߘrߘa hg_hkCJNHOJQJaJ$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmH sH "hhk5CJOJQJ\aJht=OhkCJOJQJaJ$hg_hkCJOJQJaJmHsH"h+hk5CJOJQJ\aJhg_hkCJOJQJaJ! 2 4 : < > @ D H J N P T V X Z `  A,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r $$Ifa$gd!#$$If]a$gd!#FfVw $Ifgd!#$:$If]:a$gd!#        $ ( * , . 0 4 z ~ ,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,rFf $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#Ff}   F H P R T V Z \ ^ ,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r$$If]a$gd!#Fft $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#H P n r  & , n r x ~  ̹̹ߪwdQdwd>w$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJht=OhkCJOJQJaJ$h<hkCJOJQJaJmHsH$h<hkCJOJQJaJmH sH h<hkCJOJQJaJ"h<hk5CJOJQJ\aJ^ b d h j l n r        " ,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r$$If]a$gd!#Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#" $ & , x ~  ,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r$$If]a$gd!#Ffژ $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#  T V ^ ` b d f h l n p r t v z ~ ,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,rFf $$Ifa$gd!#$$If]a$gd!#Ff~ $Ifgd!#$:$If]:a$gd!#  T V ^ z ~   , 2 Z ` $ N P f h r ȶtۅ_ȶ(hg_hkCJNHOJQJaJmH sH hg_hkCJNHOJQJaJ$hg_hkCJOJQJaJmHsHht=OhkCJOJQJaJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJ%      Z `  ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r$$If]a$gd!#FfЬ $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $ f h r t v x z ~ ,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r$$If]a$gd!#Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#       $ & ( * ,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,r$$If]a$gd!#Ff@ $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# * . P T     6 8 B ` d l n x | о߫߉߉о߉߉о߉v߉о߉߉Љ߉߉$h+hkCJOJQJaJmHsHh+hkCJOJQJaJ$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH "hhk5CJOJQJ\aJht=OhkCJOJQJaJ"h+hk5CJOJQJ\aJhg_hkCJOJQJaJ.* . P T A,r,r,r ,r,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r $$Ifa$gd!#$$If]a$gd!#Ff $Ifgd!#$:$If]:a$gd!#     6 8 B D F ,r,r ,r,r,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,rFf, $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#FfF J L N P R T X \ ^ ` d x | ,r ,r,r,r,r,r,r,r,r,rA,r,r,r ,r,r,r ,r,r,r ,r,r,r$$If]a$gd!#Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# ,r,r,r,r,r,rA,r,r,r ,r,r,r!,r!,r ,r ,r ,r,r,r,r,r,r$$If]a$gd!#FfV $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#     1 2 3 4 ,r,r,r,r>,r ,r,r2$a$gd!#Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#   " $ & ( ^ ` d f h     + , 1 4 6 7 > ? D G U V uuuhk5NHOJQJ\hk5OJQJ\hk5CJOJQJ\aJjahkUjbоL hkCJUVaJjhkU hkNHhk5\mHnHuhk5\mHnHujhk5U\ hk5\hkht=OhkCJOJQJaJ-4 D E F G U V Y Z ] ^ a b g h m n s t y z  $a$V Y Z ] ^ a b g h m n s t y z      @ B D F H ^ ` b { ĸҸ hk0JH hk\ hk0JF hk0JJ hkNHhk5\mHnHuhk5\mHnHujhk5U\ hk5\hkCJOJQJaJhkhkCJOJQJaJ= 4   R { . c !!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1ICK$a$    + , / 6 7 8 G I _ ` a b t u v    # & 6 7 I N O W Z [ & ( ) + C D N R ^ c t hkNHhk0JE56\]hk0JJ6] hk0JH hk0JF hk0JJhkR  ' H X  $ Z Y    , N !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1CDKI I J X \ ^ _ ` n      = @ A K L ^ _ s z {     @ B D F H p r ¶жhkCJOJQJaJhk5\mHnHuhk5\mHnHujhk5U\ hk5\hk0JJNH hk0JJ hk0JFhk hkNHAN l o q r s $ & : < !1!1!1!!!PSgPS7PSgPSSPSSPS$a$ $da$gdUTk$d^`a$gdUTk$d^`a$gdUTkgdUTkgdUTkIC  4 < F I J }  Q w !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1KC$a$ # 2 3 F G Q X Y Z _ a s t u v x |   & ( P R     4 6 X Z d hk0JF hkNH hk0JM hk0JJhkhkCJOJQJaJU V \ ` b  j   ^   X \ ^       !1!1!1!1!1!1!1!1!1!!1!1!1!1!1!1!1!1! !B! !Bgd(v:KICd f      z {  @ A F G Z ] ^ _ h { }        ( X \ ^ l n hk56\]hk0JHNH hk0JHhk5\mHnHuhk5\mHnHujhk5U\ hk5\ hk0JFhk hk0JJD   . 0 2 4      D E d e x y z { ϸϟϟυϸ{tϟ hk6]jhkEHU hkNHjhkEHUjhkEHUjeVL hkCJUVaJjhkEHUj{K hkCJUVaJjhkUhk hk5\hk5\mHnHuhk5\mHnHujhk5U\+     " $ J L N P      & ( F H J L N P R      4 E  髡随uukhkOJQJ^Jhk5\mHnHuhk5\mHnHujhk5U\ hk5\jhkEHUjeVL hkCJUVaJ hkNHjhkEHUj{K hkCJUVaJ hkjhkUhkjhkUjhkEHU*                " #       > @ B D n p z       {o{ic hk0JH hk^JhkNHOJQJ^JhkOJQJ^J hkNHjUhkEHUhkchkmHnHuhkchkmHnHu hkchkjhkUjhkEHUjPVL hkCJUVaJjhkEHUj{K hkCJUVaJhkjhkU$  ` f   Q R ^ g o q                       ( = ? V W ] ` e l | }                 K L t }               hPhk5\ hk0JN hk0JO hk0JF hkNH hk0JMhk hk0JJO h j  2 ]      a        F s        !1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1!1IKgd!KC @ D F P  . 8 B D Z \ l n p r t v z 333b$<$Ifa$gd!#Ff  :$If^:gd!# $$Ifa$gd1R $$Ifa$gd!#8gd!#gdP  !     @ B D F H L V X @ B D F P p r t v ƹ²¢kY"h1Rhk5:CJOJQJaJ+jh1Rhk5:CJOJQJUaJhk5:CJOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJ h&>hkhhkNHhk hhkhk5\mHnHuhk5\mHnHuhPhk5\jhPhk5U\     ԾuauK8K8K$hk5:CJOJQJaJmHsH*hXhk5:CJOJQJaJmHsH&hXhk5:CJNHOJQJaJ"hXhk5:CJOJQJaJ"hXhk5:CJOJQJaJ%hkchk5:CJOJQJaJ"hkchk5:CJOJQJaJ+jh1Rhk5:CJOJQJUaJ"h1Rhk5:CJOJQJaJ1jh1Rhk5:CJOJQJUaJ . 6 : < l n v    ޠޠޠޠޯqeVhC/|hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJhk5CJOJQJaJht=Ohk5CJOJQJaJhk5:CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hXhk5:NHOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJ!z Ff $$Ifa$gd!# :$If^:gd!#         " & ( , . 0 4 X $Ifgd!#$:$If]:a$gd!# $$Ifa$gd1R$$If]a$gd] Ff $$Ifa$gd!#  0 4 6 8 F H R T X b                      " $ ( * 0 2 6 8 > @ H J q$hC/|hkCJOJQJaJmHsH$hrhkCJOJQJaJmH sH 'hlYhk5CJOJQJaJmHsH(hrhkCJNHOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJhC/|hkCJOJQJaJhkCJOJQJaJ,X b                  ( 2  $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd] Ffo $Ifgd!#$:$If]:a$gd!#J T V ` b h j v x ~                                          ( 2       Z b Ӛ$h6f+hkCJOJQJaJmHsH$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJhkCJOJQJaJhC/|hkCJOJQJaJ<                 Z b   Ff/ $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd] FfB'b        8 9              ȼȋvcNȼ;$hC/|hkCJOJQJaJmHsH(hrhkCJNHOJQJaJmH sH $hhkCJOJQJaJmH sH (h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJhJu}hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH                       Ff6 $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd]  $ ' ( ) * + ,       L T       $$If]a$gd] Ff> $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1R   L T         " & ( 6 8 > @ J L P R X Z h p   ƺƺl\hkCJOJQJaJmHsH$hrhkCJOJQJaJmH sH hC/|hkCJOJQJaJhhkCJOJQJaJhhkCJOJQJaJhkCJNHOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ(h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH            " h p         $$Ifa$gd1R$$If]a$gd] FfpF $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#            2 4 B H     $ & * ʸ{h{X{XCʸ(hrhkCJNHOJQJaJmH sH hkCJOJQJaJmH sH $hC/|hkCJOJQJaJmHsH$h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJhJu}hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH hkCJOJQJaJmHsH          B H $ & , 6 > D F H J L $$Ifa$gd1R$$If]a$gd] FfN $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#* , 4 6 D F ^ d j l t v | ~                                 & ,      °°hhkCJOJQJaJhhkCJOJQJaJhkCJNHOJQJaJ"hRhk5CJOJQJ\aJhC/|hkCJOJQJaJh4hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ4L P T V Z \ ^ d & ,             $$Ifa$gd1R$$If]a$gd] FfU $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#        T Z           & * 0 2 : ̽r]J>̽>hkCJOJQJaJ$hrhkCJOJQJaJmH sH (hrhkCJNHOJQJaJmH sH hkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH (h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJh|hkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJ      T Z               $$Ifa$gd1R$$If]a$gd] Ff5] $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#  $ & * N T                 $$Ifa$gd1R$$If]a$gd] Ffxd $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#: < B D N T             & v x z        , . 0 8 : < Z ^   ĵӢӀĵqӀĵqӀh|hkCJOJQJaJ$h6f+hkCJOJQJaJmH sH hJu}hkCJOJQJaJ% *hRhk5CJOJQJ\aJhhkCJOJQJaJhhkCJOJQJaJ"hRhk5CJOJQJ\aJhC/|hkCJOJQJaJhkCJOJQJaJ&   & x z                 $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd] Ffk $Ifgd!#$:$If]:a$gd!#  . 0 8 < B F H J L N P R T V X Z ^    $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd] Ffr $Ifgd!#$:$If]:a$gd!#          & * \ ^ ` j l n ! ! $! &! 0! 4! T! V! ȵȋȩ|mȩȋȩ|mȩȋȩ|hhkCJOJQJaJhhkCJOJQJaJhC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmHsH$hrhkCJOJQJaJmHsH*                & * ^ ` Ff" $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd] Ffy` j n t x z | ~ Ffe $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1R$$If]a$gd]  ! ! ! ! ! ! ! ! ! ! ! ! 0! 4! V! X! b! f! j! n! $$If]a$gd] FfF $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd1RV! X! `! b! d! f! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! " " " " *" ߢ߱ߢߓylh^hPhk5\hkhhkOJQJaJh|hkOJQJaJht=OhkOJQJaJhhkCJOJQJaJhC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJhhkCJOJQJaJn! p! r! t! v! x! z! |! ~! ! ! ! ! ! ! ! ! ! ! ! ! ! $$Ifa$gd1R$$If]a$gd!#Ff' $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! " " " " $$If]a$gd!#Ff& $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#" " " " " " " " " " " " j# n# p# &$ X$ ~$ $ rgd!#sgd!#8gd!#gdPgd!#Ff $:$If]:a$gd!# $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#*" ," J" L" N" P" R" T" V" " " " " " l# n# # # $ $ $$ &$ V$ X$ ^$ ~$ $ $ $ $ $ $ $ $ $ $ $ $ $ $ $ "' $' V' X' Z' ɻɻɻɴɴɬ~s~jhkUjhkUhk5mHnHuhk5mHnHu hk5jhk5Uh4hk5 h hk h}hk h&>hkhkhk5\mHnHuhk5\mHnHuhPhk5\jhPhk5U\-Z' \' f' h' r' t' ' ' ' ' ' ' ' ' ( ( h( l( }( ( ( ( ( ( ( ( ( ( ( ) ) $) %) /) 0) 7) 8) >) ?) F) G) P) Q) Y) Z) [) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ӡӡӡӡӡӡӡӡӡӡӡӡӡӡӡh9hk6haEshk6hkmHsH h3-hk h<hk h>hk hk>* hk6 hk5 hkNHj hkUjhkUhk hk?$ ( ( ( %) * !* "* m* * * * * + + + ^+ + + + ;, T, g, h, , gd!#sgd!#rgd!# s  V<gd!#ogd!#dgd0) ) * * * !* &* A* E* Z* ^* k* m* * * * * * * * * * * * * * + + + + + L+ M+ [+ ^+ h+ i+ o+ p+ u+ v+ }+ ~+ + + + + + + + + + + + + + + + + + + + + + + + + + + ɹh_#hk6haEshk6hkmHsH h_#hkh_#hk>*h9hk>* h9hk h hkhkh9hk6 hk6D+ + , , , , , , 3, 8, 9, :, ;, S, T, W, g, r, v, w, ~, , , , , , , , , , , , , , , , , , , , , - - - H- J- S- T- [- \- a- b- k- l- s- t- u- - - - - - - - - - - - - - - - - þh9hk6 hk6haEshk6 h9hk h hkh6f+hk>*mH sH h}hkhk h_#hkh}hk>*H, , - - J- - - - l. / / / / / V1 W1 Y1 Z1 \1 ]1 _1 `1 b1 c1 e1 f1 $a$gd!#gdgd!# s  Vgd!#sgd!#rgd!#- - - - - - - - - - - - - . T. V. h. l. . . . . . . . . . . . . . . . . . . / / / / "/ $/ ./ 0/ 2/ J/ R/ T/ V/ Z/ b/ d/ p/ r/ / / / / / / / / / / / / / žjhk5Uhithk5 h@Whkjh@WhkUhaEshk6hkmHsH h9hk h hkhkh9hk6 hk6B/ / / 0 0 0 0 0 0 0 !0 "0 #0 $0 .0 >0 N0 d0 f0 n0 0 0 0 0 0 0 21 31 X1 Y1 Z1 \1 ]1 _1 `1 b1 c1 e1 f1 h1 i1 k1 l1 n1 o1 q1 r1 t1 u1 w1 x1 z1 {1 }1 ~1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ƹƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲƲ hphk h&)hk hkNHhk h~}hkhithk5hk5mHnHuhk5mHnHujhk5U hk5Ff1 h1 i1 k1 l1 n1 o1 q1 r1 t1 u1 w1 x1 z1 {1 }1 ~1 1 1 1 1 1 1 1 1 1 1 1 1 $a$gd!#1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 gd!#$a$gd!#1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 >2 @2 B2 D2 H2 L2 P2 T2 X2 j2 l2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 ,3 .3 L3 N3 ĻĻĻĻ hk5jhk5Uhw=hk5hUghk0JhUghk0J7 hk0J7 hkCJhkmHnHuhkmHnHujhkUhk hphk=1 1 1 1 1 F2 H2 L2 N2 P2 T2 V2 X2 l2 t2 2 2 2 2 2 2 2 2 2 2 2 2 2 gd!#$a$gd!#gd!#2 3 3 3 3 ^5 _5 6 6 D7 F7 7 7 7 7 F8 H8 9 9 9 &(:$a$gd F8gdO gd+gdmgd FgdRdgdJ^$a$gd!#N3 P3 R3 T3 V3 X3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 $4 &4 ,4 -4 04 24 74 94 @4 B4 J4 L4 P4 R4 W4 ]4 `4 a4 d4 4 4 4 4 5 5 ]5 ^5 _5 f5 g5 v5 6 Ϳͧ͝jh7hk5U\h7hk5\ h-BhkhN>hk6NH hN>hkhN>hk6 hkNHhkhw=hk5 hk5jhk5Uhk5mHnHuhk5mHnHu56 6 6 6 6 6 @6 B6 D6 F6 H6 6 6 6 7 7 27 47 67 @7 B7 D7 F7 V7 X7 7 7 7 7 7 7 7 7 7 7 7 7 7 8 ʻyʻcyʻʻ+j hbhkCJOJQJUaJhkCJOJQJaJhkCJOJQJaJ+jhbhkCJOJQJUaJ%jhbhkCJOJQJUaJhbhkCJOJQJaJhkh7hk5\jh7hk5U\hk5\mHnHuhk5\mHnHu&8 8 88 :8 <8 B8 D8 F8 H8 V8 X8 v8 x8 z8 |8 ~8 8 8 8 8 8 8 8 8 8 8 8 9 9 $9 ǺڪxxmfXfmQ hkchkjh+hkU h+hkjh+hkUhk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\hkhkCJOJQJaJhkCJOJQJaJ%jhbhkCJOJQJUaJhbhkCJOJQJaJ+jhbhkCJOJQJUaJ$9 &9 ,9 .9 9 9 9 : &(:<JLjlnprtvXZ\lnķק{thhbhk5OJQJ hg_hk hkNHhk5mHnHuhk5mHnHu hk5jhk5Uhg_hk5hkOJQJh6.hkOJQJUh6.hk5OJQJ h&>hkhkjh+hkUhkchkmHnHuhkchkmHnHu$s: Scheme-Clusters Raw Text Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 14: The steps of ParaMor s data processing pipeline: Together scheme search, clustering, and filtering transform raw natural language text into models of inflectional paradigms consisting of clusters of schemes. Filters RankCorresponds to  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1Licensing Types# ofSchemes Model ofGoodErr.CSuffixesC-StemsNAdjVerbDeriv.Allo.StemSuffixarerir11110553%%%%4 menente mente s5401apoyada barata hombro inevitable segura & 22474023%%%37 ba ban cion ciones cin da das do dor dora doras dores dos miento mos n ndo r ra ran remos rla rlas rle rles rlo rlos rme rnos ron rse r rn r ra ran1805apoya compra disputa genera lanza lleva reclama senti utiliza visite & 33, 1141424%%%11a amente as illa illas o or ora oras ores os976apoyad captad frank guerr negociad & 45 12 (400)190923%%%41a aba aban acion aciones acin ada adas ado ador adora adoras adores ados amos an ando ante antes ar ara aran aremos arla arlas arlo arlos arme aron arse ar arn ar ara aran ase e en ndose 113apoy celebr desarroll genera hall lanz llev public realiz termin utiliz visit & 5-109516%%22ion iona ionaba ionada ionadas ionado ionados ionamiento ionamientos ionan ionando ionantes ionar ionario ionaron ionar ionarn ione ionen iones ion in418administrac condic cuest emoc gest les ocas pres reacc sanc ses vinculac & 10107224%%%7ta tamente tas titud to tos tsimo324cier direc gra insli len modes sangrien ...1130 127 (135)72017%%%29e edor edora edoras edores en er erlo erlos erse er ern era eran ida idas ido idos iendo iera ieran ieron imiento imientos indose i a an62abastec aparec crec cumpl desconoc escond ofrec ocurr permit pertenec recib reun sal suspend transmit vend & 17159238411%%%20ida idas ido idor idores idos imos ir iremos irle irlo irlos irse ir irn ir ira iran a an39abat conclu cumpl distribu eleg exclu permit persegu recib reun segu transmit & 21100034418%%%29ce cedores cemos cen cer cerlo cerlos cerse cer cern cera cida cidas cido cidos ciendo ciera cieran cieron cimiento cimientos cimos ci c ca can zca zcan zco26abaste apare agrade compare conven estable fortale ofre pertene recono ven & 100-822%%4eta etas eto etos33atl bol concr libr metrall papel secr tarj & 122(40)582%6 es idad idades mente sima15casual fort important irregular primer & 200-351%5staba stado star staron star7co conte entrevi gu in manife pre300-3015pondrn pone ponen poner puesta6com dis ex im pro su1000-151%3dismo dista distas5bu ovie parti perio perre2000-121%3ral rales ralismo4electo libe neolibe plu3000-81%2unas unos4alg fa oport vac5000-61%2ntra ntrarse3ade ce conce Suffix-Internal Morpheme Boundary Scheme Clustering Scheme Search Data Cleanup: Exclude Short Types Stem-Internal Morpheme Boundary Licensing Type Count Filter Section  REF _Ref190768256 \r \h  \* MERGEFORMAT  4.4.1 Section  REF _Ref190771388 \r \h  \* MERGEFORMAT  4.4.2 Figure  STYLEREF 1 \s  7. SEQ Figure \* ARABIC \s 1 1: A General schema for feature detection from word-aligned translated sentences and associated feature structures. fell tree The N Det V VP NP S Subject Number marked in 3 places: 1. on N head with =sg, es=pl, 2. on dependent Det with El=sg, Los=pl, and 3. on governing V with =sg, eron=pl rboles cayeron Los fell trees The N Det V VP NP S (SUBJ ((NUM pl) (PERSON 3sg) ...))) ((TENSE past) (LEXICAL-ASPECT activity) ... (SUBJ ((NUM sg) (PERSON 3sg) ...))) ((TENSE past) (LEXICAL-ASPECT activity) ... rbol cay El Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 7: Average precision scores of four reference algorithms for the Information Retrieval (IR) Evaluation of Morpho Challenge. No MorphologySnowball (Porter)Answer KeyTwo- LevelEnglish32.940.837.339.6German35.138.733.5-Finnish35.242.843.149.8 Inflectional OnlyInflectional & DerivationalEnglishGermanEnglishGermanPRF1PRF1PRF1PRF1ParaMor40.270.151.00.937.666.548.00.857.248.852.70.954.138.945.20.6Morfessor53.347.049.91.338.744.241.20.873.634.046.51.166.937.147.70.7 Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 4: ParaMor segmentations compared to Morfessor s (Creutz, 2006) evaluated for Precision, Recall, F1, and standard deviation of F1, , in four scenarios. Segmentations over English and German are each evaluated against correct morphological analyses consisting, on the left, of inflectional morphology only, and on the right, of both inflectional and derivational morphology. Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 3: An ablation study when inducing paradigm models over 20,000 unique word types. ParaMor s precision, recall, and F1 scores (and their standard deviations in parentheses) at word-to-morpheme segmentation using four configurations of ParaMor s scheme search, clustering and filtering algorithms. (See also  REF _Ref216175094 \h  \* MERGEFORMAT Figure  6.2). EnglishGermanSearchClusterFilterPRF1PRF1(18.1(0.6)75.1(1.3)29.1(0.7)16.2(0.5)72.7(1.1)26.5(0.6)((17.9(0.6)75.5(1.3)28.9(0.7)16.2(0.5)73.0(1.1)26.5(0.7)((58.0(1.0)48.4(1.2)52.8(0.8)55.4(1.0)36.0(1.0)43.6(0.7)(((57.6(1.1)50.6(1.2)53.8(0.8)44.0(0.9)45.4(1.0)44.7(0.6) InfinitiveerPresent ParticipleiendoSingularPlural Past ParticipleFeminineidaidasMasculineidoidos1st Person Singular2nd Person Singular3rd Person Singular1st Person Plural2nd or 3rd Person PluralPresent IndicativeoeseemosenPast Indicative PerfectisteiimosieronPast Indicative ImperfectaasaamosanFuture IndicativeerersereremosernConditionaleraeraseraeramoseranSubjunctive PerfectaasaamosanSubjunctive Imperfectieraierasierairamosieran Figure A.1: The suffixes of the ar inflection class of Spanish verbs Figure A.2: The suffixes of the er inflection class of Spanish verbs InfinitivearPresent ParticipleandoSingularPlural Past ParticipleFeminineadaadasMasculineadoados1st Person Singular2nd Person Singular3rd Person Singular1st Person Plural2nd or 3rd Person PluralPresent IndicativeoasaamosanPast Indicative PerfectasteamosaronPast Indicative ImperfectabaabasababamosabanFuture IndicativeararsararemosarnConditionalaraarasaraaramosaranSubjunctive PerfecteeseemosenSubjunctive Imperfectaraarasararamosaran InfinitiveirPresent ParticipleiendoSingularPlural Past ParticipleFeminineidaidasMasculineidoidos1st Person Singular2nd Person Singular3rd Person Singular1st Person Plural2nd or 3rd Person PluralPresent IndicativeoeseimosenPast Indicative PerfectisteiimosieronPast Indicative ImperfectaasaamosanFuture IndicativeirirsiriremosirnConditionalirairasirairamosiranSubjunctive PerfectaasaamosanSubjunctive Imperfectieraierasierairamosieran Figure A.3: The suffixes of the ir inflection class of Spanish verbs SingularPlurals SingularPlurales MasculineFeminineSingularoaPluralosas Figures A.4 and A.5: The suffixes of the two inflection classes of the Spanish paradigm for Number on nouns and adjectives Figure A.6: The suffixes of the cross-product of the adjectival Gender and Number paradigms. AccusativeDativeReflexive1st Person Singularmememe2nd Person SingularteteteMasculine3rd Person SingularloleseFemininela1st Person PluralnosnosnosMasculine2nd or 3rd Person PluralloslesseFemininelas Figure A.7: The pronominal clitics which appear in Spanish orthography as three separate paradigms of suffixes. Derivational SuffixForms that Appear in Examples in this ThesisMeaningadorador, adores, adora, adoras Verb ! Noun, Agentive of ar verbsidoridor, idores, idora, idoras Verb ! Noun, Agentive of er and ir verbsacinacin, cin, sin, aciones, etc. Verb ! Noun, Abstract nounamenteamente, mente Adjective ! Adverb idadidad, idades Adjective ! Noun izarizacin Noun ! Verb  Figure A.8: A few of the most frequent derivational suffixes of Spanish. Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 3: An excerpt from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1. Five schemes licensed by the word apoyadas. RankCorresponds to  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1Licensing Types# of SchemesModel ofGoodErr.CSuffixesC-StemsNAdjVerbDeriv.Allo.StemSuffixarerir22474023%%%37 ba ban cion ciones cin da das do dor dora doras dores dos miento mos n ndo r ra ran remos rla rlas rle rles rlo rlos rme rnos ron rse r rn r ra ran1805apoya compra disputa genera lanza lleva reclama senti utiliza visite & 45 12 (400)190923%%%41a aba aban acion aciones acin ada adas ado ador adora adoras adores ados amos an ando ante antes ar ara aran aremos arla arlas arlo arlos arme aron arse ar arn ar ara aran ase e en ndose 113apoy celebr desarroll genera hall lanz llev public realiz termin utiliz visit & 10107224%%%7ta tamente tas titud to tos tsimo324cier direc gra insli len modes sangrien ... RankModel ofGoodCompletePartialErrorCSuffixesC-StemsNAdjVerbDeriv.StemSuffixChancearerir1%%%%2 s5501apoyada barata hombro & 2%%%4a as o os892apoyad captad dirigid junt & 3%%%15 ba ban da das do dos ...17apoya disputa lanza lleva & 5%%%15a aba aban ada adas ado ...25apoy desarroll disput lanz & 12%%%15a aba ada adas ado ados ...21apoy declar enfrent llev &  Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 8: An excerpt from  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4. The rank 2 cluster hypothesizes a morpheme boundary internal to true suffixes; the rank 4 cluster correctly models morpheme boundaries that immediately follow verbal stems; while the rank 10 cluster models boundaries internal to stems. Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 7: An excerpt from  REF _Ref215319995 \h  \* MERGEFORMAT Figure  4.4: Six scheme-clusters that are removed by ParaMor s small-cluster filter. Each removed cluster is licensed by fewer than the threshold number of licensing types, 37. The third column gives the number of types that license each scheme-cluster. Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 2: An excerpt from  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1. Six schemes selected by ParaMor s initial bottom-up search algorithm that model er and/or ir verbal suffixes. RankModel ofGoodCompletePartialErrorCSuffixesC-StemsNAdjVerbDeriv.StemSuffixChancearerir12%%%15a aba ada ... arn e en 21apoy declar enfrent llev & 30%%%%11e en ida idas ido idos ...16cumpl escond recib vend & 127%%%9e en er er era ido ieron ...11ced deb ofrec pertenec & 135%%%%10a e en ida ido iendo iera ...12assist cumpl ocurr permit & 1592%%%4ido idos ir ir6conclu cumpl distribu reun & 2000%%%%4e en ieron iesen5 aparec crec impid invad pud RankCorresponds to  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1 REF _Ref190681505 \h \* MERGEFORMAT Error! Reference source not found.Licensing Types# of SchemesModel ofGoodErr.CSuffixesC-StemsNAdjVerbDeriv.Allo.StemSuffixarerir200-351%5staba stado star staron star7co conte entrevi gu in ...300-3015pondrn pone ponen poner puesta6com dis ex im pro su1000-151%3dismo dista distas5bu ovie parti perio perre2000-121%3ral rales ralismo4electo libe neolibe plu3000-81%2unas unos4alg fa oport vac5000-61%2ntra ntrarse3ade ce conce // the substring(String, startIndex, endIndex) method works as follows: // substring(abcde, 0, 2) yields ab // substring(abcde, 2, 5) yields cde // segmentWords( words, schemeClusters, paradigmInductionCorpus, corpusToSegment) { foreach (word in words) { morphemeBoundaryIndexes = emptySet; wordLength = word.length(); for (charIndex = 1; charIndex <= wordLength-1; charIndex++) { stem = substring(word, 0, charIndex); suffix = substring(word, charIndex, word.length()); foreach (schemeCluster in schemeClusters) { foreach (cSuffix in schemeCluster) { if (cSuffix == suffix) { foreach (cSuffixPrime in schemeCluster) { if (cSuffixPrime != cSuffix) { possibleWord = stem + cSuffixPrime; if (paradigmInductionCorpus.contains(possibleWord) || corpusToSegment.contains(possibleWord)) { morphemeBoundaryIndexes.add(charIndex); } } } } } } } // end for (charIndex = 1... // segmentedWords is a hash on each word to an array which holds // the morphemes of the word. startIndex = 0; foreach (boundaryIndex in morphemeBoundaries) { morpheme = substring(word, startIndex, boundaryIndex); segmentedWords{word}.add(morpheme); startIndex = boundaryIndex; } // And add the final segment of the word as a final morpheme morpheme = substring(word, startIndex, word.length(); segmentedWords{word}.add(morpheme); } return segmentedWords; } FilterSuffixInternalBoundaryErrors(schemeClusters) { foreach (schemeCluster in schemeClusters) { countOfErrorSchemes = 0; countOfNonErrorSchemes = 0; foreach (scheme in schemeCluster) { if (likelyModelOfAMorphemeLeftEdge(scheme)) countOfNonErrorSchemes++; else countOfErrorSchemes++; } if (countOfErrorSchemes >= countOfNonErrorSchemes) { schemeClusters.remove(schemeCluster); } } return schemeClusters; } // Measure the entropy of a schemes leftward trie-style links likelyModelOfAMorphemeLeftEdge(scheme) { foreach (cStem in scheme.cStems) { stemFinalChars{cStem.finalChar()}++; } if (entropy(stemFinalChars) > threshold) return true; else return false; } Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 10: A pseudo-code implementation of ParaMor s algorithm to filter out scheme-clusters that likely hypothesize morpheme boundaries for their licensing types that fall internal to true suffixes. Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 11: The first two rows:<Z\lnx(DV`jlFfӶ :$If^:gd!# $$Ifa$gd` $$Ifa$gd!#$a$gd FgdInxʒ̒(޹|i|WCW&hXhk5:CJNHOJQJaJ"hXhk5:CJOJQJaJ%hkchk5:CJOJQJaJ"hkchk5:CJOJQJaJ1jh1Rhk5:CJOJQJUaJ"h1Rhk5:CJOJQJaJ+jh1Rhk5:CJOJQJUaJhk5:CJOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJ(,2>@DV^bd“ʓГؓړ "$ǵǣǔwwwwgZgZgZgZghk5CJOJQJaJht=Ohk5CJOJQJaJhk5:CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hXhk5:NHOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJ$hk5:CJOJQJaJmHsH*hXhk5:CJOJQJaJmHsH#“ΓГړ Ff $$Ifa$gd!# :$If^:gd!#$<$Ifa$gd!# "$(,8<@DFHJNPTVX $$Ifa$gd`$$If]a$gd] Ff $$Ifa$gd!#$(*8:X\^`npz|ܔޔ "(*24@BFHJLPRXZ^`fq$hC/|hkCJOJQJaJmHsH$hrhkCJOJQJaJmH sH 'hlYhk5CJOJQJaJmHsH(hrhkCJNHOJQJaJmH sH $hrhkCJOJQJaJmH sH hC/|hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ,X\ޔ  P $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff` $Ifgd!#$:$If]:a$gd!#fhpr|~ĕƕЕҕؕڕ "(*.068<>DFPZ"($hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJhC/|hkCJOJQJaJhkCJOJQJaJ=PZ  "(ԗ $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff3 $Ifgd!#$:$If]:a$gd!#(ԗ֗ڗ &t\^bdȵ۩xcxPx;ȵ۩(hrhkCJNHOJQJaJmH sH $hhkCJOJQJaJmH sH (h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJhJu}hkCJOJQJaJhkCJOJQJaJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmHsHԗ֗ڗ  &\^Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff^bfpvxz|~̛ԛjlrxFf $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] ~̛ԛXZhjlrvʜ̜МҜ؜ڜȳۧ{ۧllllllllY$hrhkCJOJQJaJmH sH hC/|hkCJOJQJaJhhkCJOJQJaJhhkCJOJQJaJhkCJNHOJQJaJhkCJOJQJaJ(h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH "hRhk5CJOJQJ\aJ$hC/|hkCJOJQJaJmHsHxJLRltz$$If]a$gd] Ffa $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`HJLRbhjžȞxͺ~kXkHkHhkCJOJQJaJmH sH $hC/|hkCJOJQJaJmHsH$h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJhJu}hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH hkCJOJQJaJmHsHhkCJOJQJaJmHsHz|~žȞğƟȟ $$Ifa$gd`$$If]a$gd] Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#xzğƟޟ (*.0<>FHPR\^fhnpxzųvvvvvvvvvvvvvvvvvvvvhC/|hkCJOJQJaJh4hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH (hrhkCJNHOJQJaJmH sH .ȟʟ̟Пԟ֟ڟܟޟLNT^flnprv $$Ifa$gd`$$If]a$gd] Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#>@JLNRT^Ԣڢ&(6Ӕr]rJ:JhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH (h6f+hkCJNHOJQJaJmH sH $h6f+hkCJOJQJaJmH sH h|hkCJOJQJaJ% *hRhk5CJOJQJ\aJhhkCJOJQJaJhhkCJOJQJaJhkCJNHOJQJaJ"hRhk5CJOJQJ\aJhC/|hkCJOJQJaJhkCJOJQJaJvx|Ԣڢvx $$Ifa$gd`$$If]a$gd] Ff&  $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#68vx~£ģΣԣ,.068:>@BbfųvvvvvgXIIvhJu}hkCJOJQJaJhhkCJOJQJaJhhkCJOJQJaJhC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH (hrhkCJNHOJQJaJmH sH Σԣ.08BHLPRTVXZ\^ $$Ifa$gd`$$If]a$gd] Ffi $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#^`bf  $& $$Ifa$gd`$$If]a$gd] Ff $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#f&*fjڥޥ"LNVXZ\|ϢϢۀmZϢۓϓϓ% *hRhk5CJOJQJ\aJ$hrhkCJOJQJaJmHsH$hrhkCJOJQJaJmHsHhC/|hkCJOJQJaJh|hkCJOJQJaJhhkCJOJQJaJhhkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH #&*fj¥ƥȥʥ̥ΥХҥԥ֥إڥޥ $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff $Ifgd!#$:$If]:a$gd!#"LNX\bfhjlnprtvz|ަ $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff% $Ifgd!#$:$If]:a$gd!#ܦަ$&6:hjlvxzԧ֧ا  !#$&'13579ⵦⵦⵦht=OhkOJQJaJ% *hRhk5CJOJQJ\aJhC/|hkCJOJQJaJh|hkCJOJQJaJhhkCJOJQJaJhhkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJ4ަ 6:jlFfV4 $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`$$If]a$gd] Ff-lvz~֧اFf7; $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`$$If]a$gd]  !#$&'$$If]a$gd!#FfB $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# $$Ifa$gd`'()*+,-./013579:;<=>?@A$$If]a$gd!#FfI $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#9:;>JKLN]opqè֨רݨިƹƩƩƩƹƩyfy%jhbhkCJOJQJUaJhbhkCJOJQJaJ#h6.hk5CJNHOJQJaJh6.hkCJOJQJaJh6.hk5CJOJQJaJhk5CJOJQJaJhkhhkOJQJaJh|hkOJQJaJht=OhkOJQJaJhhkCJOJQJaJABCDEFGHIJKLMN^pq$a$gd FgdO FfN$:$If]:a$gd!# $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#Ĩ֨ר  ,gdx_tgdz 7$8$H$gdzgdzgd Fgd!#$a$gd F) "$VXxz|ªĪƪǺڪڔǺڪshhbb hkNHhk5mHnHuhk5mHnHujhiFhk5UhiFhk5+jRhbhkCJOJQJUaJhkhkCJOJQJaJhkCJOJQJaJ%jhbhkCJOJQJUaJhbhkCJOJQJaJ+j}RhbhkCJOJQJUaJ&     '(.5HLMNRX]^cdinowzً}ًhx>hkNH hk6h{hk6)h{hkB*CJOJQJ^JaJph)hx>hkB*CJOJQJ^JaJph%hx>hkB*CJOJQJ^Jphhx>hkCJ hx>hk)hx>hkB*CJ"OJQJ^JaJ$phhk hiFhk1    DEfǬȬ 7$8$H$gdz$  a$gdzgdzgdx_t $7$8$H$a$gdzĬƬǬȬˬϬЬѬլ֬۬ܬ)hx>hkB*CJOJQJ^JaJph%hx>hkB*CJOJQJ^Jphhx>hkCJhx>hk>*h{hk>*)hx>hkB*CJ"OJQJ^JaJ$ph h{hk hk6h{hk6mHsHh{hk6 hx>hkhk*ȬЬѬլ֬۬ܬ#$PQ 7$8$H$gdz $7$8$H$a$gdzgdz  #$FGPQ^`vwϭ汨汔ssehk5\mHnHujhsmhk5U\hsmhk5\hx>hkCJh{hk>*mHsHh{hk>*h{hkNH h{hk-h{hkB*CJNHOJQJ^JaJph,h{hk>*B*CJOJQJ^JaJphhk)h{hkB*CJOJQJ^JaJph%Qvw>@BvƯ $$Ifa$gd3 gdQ6gdzgdz 7$8$H$gdz @BDFJ,.>@Bln|~Ưȯد24Dlnpvг|k|k|k|`h3 hkCJaJ h3 hkCJOJQJ^JaJh3 hkCJaJ-hn+hk5:CJNHOJQJ\^JaJ)hn+hk5:CJOJQJ\^JaJ&h3 hk5CJOJQJ\^JaJh3 hkNHhk h3 hkhsmhk5\jhsmhk5U\hk5\mHnHu#ƯȯدZNNNNN $$Ifa$gd3 kdwS$$If r vP  m m m m 44ap2$.2ZNNNNN $$Ifa$gd3 kd|T$$Ifr vP& m&m &m&m44ap224DNXblZNNNNN $$Ifa$gd3 kdU$$Ifr vP& m&m&m&m44ap2lnprtvxZUSSS=$$Ifa$gdx_tlcgdQ6kdV$$Ifr vP& m&m&m&m44ap2vx԰ְذ"$46FHXZ`brz|Ʊȱбұڱܱ  ḩኩ᩸ḩ{ph8ohkCJaJh0ohkCJOJQJaJ"hrhk5CJOJQJ\aJhk5CJOJQJaJhrhkCJOJQJaJhkCJOJQJaJ"hrhk5CJH*OJQJaJhzhkCJaJhrhk5CJOJQJaJhrhkCJOJQJaJ*x԰ְذ}ggggg$$Ifa$gdx_tlclkd-X$$IfF!d t    4ap$$Ifa$gdx_tlc &*.28<mWWWWWWWWW$$Ifa$gdx_tlckdX$$IfYr ^!bbb t4ap <@DJNRV\`br|ȱұ$$Ifa$gdrlc$:$If]:a$gdx_tlcFf\$$Ifa$gdx_tlcұܱ   *4>FPZdlv$$Ifa$gdrlc\$:$If]:a$gdx_tlc\Ffc$$Ifa$gdrlc 4>lv̲β.02468γξήxxj^xxx^xWW hXX@hkhk5\mHnHuhk5\mHnHujhHhk5U\hHhk5\hkhrhkCJaJh0ohkCJOJQJaJh0ohk5CJOJQJaJhrhkCJOJQJ\aJhrhkCJOJQJaJhrhk5CJOJQJaJ#hrhk5CJNHOJQJaJ "jlnpr$$If`a$gd.#%l)gdFgdHgdx_tgdzFfEi$$Ifa$gdrlc\γг248:np  "02PRTVXZ\~ַطlnԸָ(*JLõõõõږڏvjlh}_hkUjh}_hkU h}_hk hkH*hk5\mHnHuhk5\mHnHujh/phk5U\h/phk5\ hb;}hk hkNHhkhXX@hkH*hXX@hk5H* hXX@hkhXX@hk5.LZ\bdhlԹֹ  ɹɹɹɃyl]Q]yl]Q]yhkCJOJQJaJh*hkCJOJQJaJh-hkOJQJaJhkOJQJaJ$ jh-hk5OJQJ\aJ!h-hk5H*OJQJ\aJ"h-hk5NHOJQJ\aJh-hk5OJQJ\aJhk h}_hkjh}_hkUhkchkmHnHuhkchkmHnHu hkchkP66666$$If`a$gd.#%l)kdCm$$If4r Y ! L   t04aP6$$If`a$gd.#%l$$If`a$gd.#%l\kdIn$$If4lr Y ! L   t04aʹιҹعܹ !($If`gd0"lFfq$$If`a$gd.#%l\$$If`a$gd.#%l !#&(,-/2489;>@ABCDFJKMPRVWY\^bcehjnoqtvz{}ʷʷhk5OJQJ\aJh-hk5OJQJ\aJ$ jh-hk5OJQJ\aJhkhkOJQJaJhkCJOJQJaJh*hkCJOJQJaJh-hkOJQJaJ<(-49@ACEFKRW^cj~~~$If`gd0"l$$If`a$gd.#%l$$If`a$gd.#%lFfw$If`gd0"l$$If`a$gd*ljov{~~$If`gd0"lT$$If`a$gd.#%lTFf~$$If`a$gd*l$If`gd0"l$$If`a$gd.#%lºĺɺ˺κкԺպ׺ںܺݺ޺ߺ˾˾ᝊzzz˾mhk5OJQJ\aJh-hk5OJQJ\aJ$ jh-hk5OJQJ\aJhkhebhk5OJQJ\aJhebhkOJQJaJh-hkOJQJaJhkOJQJaJhkCJOJQJaJh*hkCJOJQJaJh0"hk5OJQJ\aJ*ĺɺкպܺݺߺ~$If`gd0"l$$If`a$gd.#%lFf$$If`a$gd*lT$If`gd0"lT$$If`a$gd.#%lT $+,-./012$$If`a$gd.#%lFf1$$If`a$gd*l$If`gd0"l$$If`a$gd.#%l #$&)+,/0123456789:;>I񧗊{{{{{{i#h9hk5CJOJQJ^JaJh-hkCJOJQJaJh-hkOJQJaJh-hk5OJQJ\aJhkh0"hk5OJQJ\aJhk5OJQJ\aJhebhkOJQJaJhkCJOJQJaJhebhk5OJQJ\aJh*hkCJOJQJaJ"23456789:;<=>ILMv:v:\v:$ $If` a$gd9lgdFFfj$$If`a$gd.#%lIMNZ[ahijkop{}~λл  !./4CD\qrͼμҼӼ̸̸̸̸̸̸̸̸̸̸̤̤̤̤̤̤&h9hk5CJH*OJQJ^JaJ'h9hk5CJNHOJQJ^JaJ#h9hk5CJOJQJ^JaJ h9hkCJOJQJ^JaJ h9hkCJOJQJ^JaJ@MNaghnv:nv:n\v:$ $If` a$gd9lvkdٕ$$IfT<F$  x t6    4apThijkt{}nv:nv:nv:nv:n#v:$ $If` a$gd9lTvkdۖ$$IfT8F$  x t6    4apT}~lRv:Rv:Rev:Rfv:Rv:$ $If` a$gd9lkdݗ$$IfTrl s ( t64ap TkQQv:Qev:Qfv:Qv:$ $If` a$gd9lkdߘ$$IfT4;rl s ( t64ap TĻͻѻػkQ$ v:Q}v:Q}v:Q}v:Q}v:Q}v:Q}v:Q~v:$ $If` a$gd9lTkd1$$IfT4;rl s ( t64ap T  ~v:~v:}v:}v:}v:}v:}v:}v:$ $If` a$gd9lT !469;@C`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd$$IfTֈ$  ( ~~~ t64aTCD\^cfkq`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkdv$$IfT,ֈ$  ( ~~~ t64aTqr`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lTkdٝ$$IfT,ֈ$  ( ~~~ t64aTȼͼ`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd<$$IfT,ֈ$  ( ~~~ t64aTͼμڼ߼`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd$$IfT,ֈ$  ( ~~~ t64aTӼڼ #$3QY[af{}ý #%&129:?IJKOPU_`ab˳ h)hkh6?hk0J>hbhk5 hk5hk'h9hk5CJNHOJQJ^JaJ h9hkCJOJQJ^JaJ h9hkCJOJQJ^JaJ#h9hk5CJOJQJ^JaJ: `F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lTkd$$IfT,ֈ$  ( ~~~ t64aT38>CKQ`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkde$$IfT,ֈ$  ( ~~~ t64aTQRSTUVWX`J$ ^J}^J}^J~^J}^J}^$$Ifa$gdd, lkdȤ$$IfT,ֈ$  ( ~~~ t64aTXYZ[b] [V [V [gdd, gdd, kd+$$IfTֈ$  ( ~~~ t64aT v:v:\v:nTv:Tv:T\v:$ $If` a$gd9lTvkd$$IfT<F$  x t6    4apT$ $If` a$gd9l#%nv:nv:nv:nv:n#v:$ $If` a$gd9lvkd$$IfT8F$  x t6    4apT%&6?CHIlRv:Rv:Rev:Rfv:Rv:$ $If` a$gd9lkd$$IfTrl s ( t64ap TIJKUY^_kQQv:Qev:Qfv:Qv:$ $If` a$gd9lTkd $$IfT4;rl s ( t64ap T_`aeluykQ$ v:Q}v:Q}v:Q}v:Q}v:Q}v:Q}v:Q~v:$ $If` a$gd9lkdr$$IfT4;rl s ( t64ap T bdvxȾɾ־׾ܾ-.3LMYZ_xy}~ǿȿοϿ޿ڵڤڏzڤڤڤڤ(h9hkCJOJQJ^JaJmHsH(h9hkCJOJQJ^JaJmHsH h9hkCJOJQJ^JaJ h9hkCJOJQJ^JaJ'h9hk5CJNHOJQJ^JaJ#h9hk5CJOJQJ^JaJ&h9hk5CJH*OJQJ^JaJ0Ⱦ~v:~v:}v:}v:}v:}v:}v:}v:$ $If` a$gd9lȾɾܾ޾`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkdĬ$$IfTֈ$  ( ~~~ t64aT  `F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lTkd$$IfT,ֈ$  ( ~~~ t64aT37<@GL`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd$$IfT,ֈ$  ( ~~~ t64aTLM_chlsx`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd}$$IfT,ֈ$  ( ~~~ t64aTxy`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lTkd$$IfT,ֈ$  ( ~~~ t64aTĿǿ`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkdC$$IfT,ֈ$  ( ~~~ t64aTǿȿ޿`F v:F_v:Fnv:Fov:Fnv:F_v:$ $If` a$gd9lkd$$IfT,ֈ$  ( ~~~ t64aT޿  $+,-.23>@ALMTUZdefjkpz{|}ȷȣȷȷȣȷȣȣȷڷȣȷڷȏȏȏȏȣȏȏȣȷ&haoohk5CJH*OJQJ^JaJ'haoohk5CJNHOJQJ^JaJ haoohkCJOJQJ^JaJ#haoohk5CJOJQJ^JaJhk h9hkCJOJQJ^JaJ h9hkCJOJQJ^JaJ8`F$ v:F}v:F}v:F~v:F}v:F}v:$ $If` a$gd9lTkd $$IfT,ֈ$  ( ~~~ t64aT b] [Av:Av:A[v:$ $If` a$gdaoolgdd, kdl$$IfTֈ$  ( ~~~ t64aT$*+nv:nv:n[v:$ $If` a$gdaoolvkd[$$IfT<F$ q t6    4apT+,-.7>@nv:nv:nv:nv:n"v:$ $If` a$gdaoolTvkd]$$IfT8F$ q t6    4apT@AQZ^cdlRv:Rv:Rev:Rfv:Rv:$ $If` a$gdaoolkd_$$IfTrl s " t64ap TdefptyzkQQv:Qev:Qfv:Qv:$ $If` a$gdaoolkdi$$IfT4;rl s " t64ap Tz{|kQ$ v:Q}v:Q}v:Q}v:Q}v:Q}v:Q}v:Q~v:$ $If` a$gdaoolTkdü$$IfT4;rl s " t64ap T ~v:~v:|v:|v:|v:}v:}v:}v:$ $If` a$gdaoolT`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolkd$$IfTֈ$  $ }}~|} t64aT!&).4`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolkd$$IfT,ֈ$  $ }}~|} t64aT45JKPdeqrw$()>@bcdmopwy|뼰 h hkh6?hk0J>hbhk5hbhk5mHsHhk5mHsH'haoohk5CJNHOJQJ^JaJ#haoohk5CJOJQJ^JaJhk haoohkCJOJQJ^JaJ;45PSWZ`d`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolTkd$$IfT,ֈ$  $ }}~|} t64aTdew{`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolkd$$IfT,ֈ$  $ }}~|} t64aT`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolkdQ$$IfT,ֈ$  $ }}~|} t64aT`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolTkd$$IfT,ֈ$  $ }}~|} t64aT`F v:F_v:Fnv:Fov:Fmv:F_v:$ $If` a$gdaoolkd#$$IfT,ֈ$  $ }}~|} t64aT`J$ sJ}sJ}sJ~sJ|sJ}s$$Ifa$gdd, lkd$$IfT,ֈ$  $ }}~|} t64aTcdmb] [Vv[<v:$$If`a$gdd, lgdd, gdd, kd$$IfTֈ$  $ }}~|} t64aTmopwyv:oUv:Uv:$$If`a$gdd, lvkd$$IfJ0= t064a$$If`a$gdd, lyz{|hv:hv:$$If`a$gdd, lgdd, vkd$$IfJ0= t064aov:ov:$$If`a$gdd, l\vkd.$$IfJ0= t064ah=v:hv:h>v:$$If`a$gdaoolgdd, vkd$$IfJ0= t064av\'v:\v:\v:$$If`a$gdaoolkdr$$IfJF=# a => t06    4av\'v:\v:\v:$$If`a$gdaoolkd$$IfJF=# a => t06    4aEFvqt ojoj oPv:Pv:$$If`a$gd+lgdd, gdd, kd$$IfJF=# a => t06    4a &,DEFLQq h+hkCJOJQJ^JaJhk5CJOJQJ^JaJ&h+hk5CJH*OJQJ^JaJ'h+hk5CJNHOJQJ^JaJ#h+hk5CJOJQJ^JaJ h%hkh6?hk] hk]h6?hk0JQ hkNHhkhbhk5 hk5,v: v:v:v:v::kd$$If֞r ,<($  tp4a$$If`a$gd+lv:v:mv:v:v:v:v:$xx$If`a$gdaool$$If`a$gd+lT:v:v:$xx$If`a$gdaool$$If`a$gd+lkd$$If֞r ,<($  tp4a  #$(3478>EFHJKNTVW[ijnouw},ٰ١ٰ١ٰٰ١١ĜĎ hkNHhbhk5 hk5hk5CJOJQJ^JaJ&h+hk5CJH*OJQJ^JaJhk h+hkCJOJQJ^JaJ#h+hk5CJOJQJ^JaJ'h+hk5CJNHOJQJ^JaJ8 mv:v:v:v:v::kd$$If֞r ,<($  tp4a$$If`a$gd+l (+,/03v:v:mv:v:v:v:v:$xx$If`a$gdaool$$If`a$gd+l34=J0v:$$If`a$gd+lkd$$If4w֞r ,<($  tp4ap =>ABCDEmv:v:v:$$If`a$gd+l$xx$If`a$gdaoolEFGJ0v:$$If`a$gd+l\kdk$$If4w֞r ,<($  tp4ap G[_`deiv:mv:v:v:v:v:$$If`a$gd+l\$xx$If`a$gdaool\ijtT:v: v:$$If`a$gdaoolT$$If`a$gd+lTkd$$If֞r ,<($  tp4amv:v:v:v:v:$$If`a$gd+lTJ4v:$$Ifa$gdd, l$$Ifa$gd+lkd4$$If4w֞r ,<($  tp4ap mv:v:v:$$Ifa$gdd, l$$Ifa$gdaool+,JEC>Cgdd, gdd, kd$$If4w֞r ,<($  tp4ap ,12YZ\]{|6:HJ04lnƵƵƵ{fQ{Q{Q(h.QhkCJOJQJ^JaJmH sH (h.QhkCJOJQJ^JaJmH ,sH ,(h.QhkCJOJQJ^JaJmHsH#h.Qhk6CJOJQJ^JaJ$h.QhkCJNHOJQJ^JaJ h.QhkCJOJQJ^JaJhkhk5CJOJQJ^JaJ'h.Qhk5CJNHOJQJ^JaJ#h.Qhk5CJOJQJ^JaJ,9@s{|zv:zv: v:(v:S=zv:$$Ifa$gd6?lTkd$$IfF2 z ( t00!    4ap$$Ifa$gd6?lHJ v:6kd]$$IfJF2 z ( t00!    4ap D$Ifgd6?lT$$Ifa$gd6?lTJTzv: v:(v:6kd $$IfJF2 z ( t00!    4ap D$Ifgd6?l$$Ifa$gd6?l2lnzv: v:(v:6kd$$IfJF2 z ( t00!    4ap D$Ifgd6?l$$Ifa$gd6?ln|zv: v:(v:6kd$$IfJF2 z ( t00!    4ap D$Ifgd6?l$$Ifa$gd6?lnFT`j "XZ\^`bx ƸƬԥ랓~pdhkchkmHnHuhkchkmHnHu hkchkjh"thkUjh"thkU h"thk hhkhk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\hbhk5 hk5hk h.QhkCJOJQJ^JaJ%zv: v:6kd?$$IfJF2 z ( t00!    4ap D$Ifgd6?l$$Ifa$gd6?l*FHuv: v:#v:6kd$$IfJF2 z ( t00!    4ap D$Ifgd6?l$$Ifa$gd6?lHJLNPRzs s(s61 gdd, kd$$IfJF2 z ( t00!    4ap <t$Ifgdd, l\$$Ifa$gdd, l\RTHLNX$6@JLbdtvxz|_ Ffe :$If^:gd!# $$Ifa$gdfYa $$Ifa$gd1R $$Ifa$gd!#8gd-gd0Fgdd,  4DFHJLNXxz|~ѿo]J]%hkchk5:CJOJQJaJ"hkchk5:CJOJQJaJ1jh1Rhk5:CJOJQJUaJ"h1Rhk5:CJOJQJaJ+jh1Rhk5:CJOJQJUaJhk5:CJOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJ h&>hk hkhk hk5\hk hkNH  "$6>BDtv~ııġrccccrrSht=Ohk5CJOJQJaJhk5:CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hXhk5:NHOJQJaJhXhk5:OJQJaJ$hk5:CJOJQJaJmHsH*hXhk5:CJOJQJaJmHsH&hXhk5:CJNHOJQJaJ"hXhk5:CJOJQJaJ|~Ff[ $$Ifa$gd!# :$If^:gd!#$<$Ifa$gd!# $$Ifa$gdfYa 6<>@DFLNVXdfjlnptv|~ų$hC/|hkCJOJQJaJmHsHhC/|hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJht=OhkCJOJQJaJht=Ohk5OJQJaJht=Ohk5CJOJQJaJhk5CJOJQJaJ4  $&(,.026 $$Ifa$gd\7$$If]a$gd\7Ff $$Ifa$gd!#6<t~0:@BDHJLPRVXZ`$:$If]:a$gd`>z $$Ifa$gd`>z$$If]a$gd`>zFf $Ifgd\7$:$If]:a$gd\7 "*,24<>DFLNRTZ\`bhjt~$.Z`{$h6f+hkCJOJQJaJmH sH hJu}hkCJOJQJaJ$hrhkCJOJQJaJmH sH hkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJhkCJOJQJaJhC/|hkCJOJQJaJ0RT@Bسxl]]l]l]l]l]l]l]MhkCJOJQJaJmHsHhC/|hkCJOJQJaJhkCJOJQJaJ$h\7hkCJOJQJaJmH sH (hrhkCJNHOJQJaJmH sH $hrhkCJOJQJaJmH sH "hRhk5CJOJQJ\aJ$hhkCJOJQJaJmH sH $h6f+hkCJOJQJaJmH sH (h6f+hkCJNHOJQJaJmH sH ` $Ifgdk'$:$If]:a$gdk' $$Ifa$gdk'$$If]a$gdk'Ff$:$If]:a$gd`>z $Ifgd`>zvxz|~Ff $$Ifa$gdk'$$If]a$gdk'Ffj  $Ifgdk'$:$If]:a$gdk'Btvx 2468Tʻʕn]]RN>hzhk5:OJQJaJhkh\7hkmH sH h\7hkOJQJaJmH sH ' *h\7hkCJOJQJaJmH sH $h\7hkCJOJQJaJmH sH $h\7hkCJOJQJaJmH sH $h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJ$hrhkCJOJQJaJmH sH $hrhkCJOJQJaJmH sH hkCJOJQJaJmHsHFf $Ifgd$:$If]:a$gd $$Ifa$gdfYa $$Ifa$gd$$If]a$gd $$Ifa$gd!#$$If]a$gd!#Ff $Ifgd$:$If]:a$gd $$Ifa$gdfYa $$Ifa$gd  $$If]a$gd!#Ff!$:$If]:a$gd!# $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# "$&(*,.02468BT^d$If^gdz$d$Ifa$gd!#gd\7Ff'$:$If]:a$gd!# $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#T\^  "$&268<\`hrϒqqaN$hg_hkCJOJQJaJmH sH hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJhk5CJOJQJaJht=Ohk5CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJhzhk5:OJQJaJ"hzhk5:CJOJQJaJ^p$d$Ifa$gd!#d$If^gdzkd.+$$If4ִ>, R!B ]  t!    4af4pF  $&(Ff/$d$Ifa$gd!#d$If^gdz$d$Ifa$gdz d$Ifgd!#(*,.02468<@DFHJLPTVXZ\`$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#Ff/6 d$Ifgd!#`hr$$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#Ff<$$If]a$gdz $Ifgd!#$&*HN|$&,`fhln̺̺̺t̺e̺hk5CJOJQJ\aJhkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmH sH hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH "$&*,.2468:<@BFHNFfJ$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#FfC&,fhnprv$$If]a$gd!#Ff}Q$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#vxz|  $$If]a$gd!#FfBX$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# "$&(*,0>@^`bdfhjðӰ}o}oaUo}o}oUo}hk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\hk"hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJhkCJOJQJaJ "$&(*,.0!gd"t8gdgdfYa8gd\7gd`>zgd-Ff ^ $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#*,24$&DFHJLNPprʾ{{mʾjah2UhkUhk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\ h&>hk hkNHhkhkchkmHnHuhkchkmHnHu hkchkj~ah2UhkU h2Uhkjh2UhkU*468:<>@vxz|~ PRfht{ulluh"thk0J> hkNHhkchkmHnHuhkchkmHnHu hkchkjxbh"thkUjh"thkU h"thk hhkhk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\ h&>hkhk* 8jln޽̯qbP"h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJhk5CJOJQJ\aJhk5CJOJQJaJht=Ohk5CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hzhk5:CJOJQJaJhzhk5:OJQJaJ h&>hkhk hkhk (8DZj!3bPSPS PSf PSd$If^gdz$d$Ifa$gd!#8gd< jlkdb$$If4ִ  v!/ w  t!    4af4pFlnrzPSXPS sPSPSPSFfg$d$Ifa$gd!#d$If^gdz$d$Ifa$gdz d$Ifgd!# A,r,r,r,r,r,r,r,r,r,r,r,r,rs,r$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#Ffm d$Ifgd!#248NT$*,4RTVjnݡݡmZGZm$h<hkCJOJQJaJmHsH$h<hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH *h hk5CJOJQJ\aJmH sH $hg_hkCJOJQJaJmHsHhkCJOJQJaJmH sH $h hkCJOJQJaJmH sH NT,r,r/ ,r,r,r,r,r,r,r,r,r,r,r,r,r,r,rs,r,r $IfgdE$:$If]:a$gdE $$Ifa$gdE$$If]a$gdEFfqt$$If]a$gdz $Ifgd!#*,468:>@BFHLNPRV,r/ ,r,r,r,r,r,r,r,r,r,r,r,r,r,r,rs,r,r,r/ ,r$:$If]:a$gdE $$Ifa$gdE$$If]a$gdEFf8{ $IfgdE$$If]a$gdE068<>@pvxññÐ}m]JÐ}m}ñÐ$hg_hkCJOJQJaJmHsHhkCJOJQJaJmHsHhkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH "h+hk5CJOJQJ\aJhg_hkCJOJQJaJ"hhk5CJOJQJ\aJhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH h<hkCJOJQJaJhkCJOJQJaJ8>vx,r,r,r,r,r,r,r,r,r,r,r,r,r,rs,r,r,r/ ,r,r,rFfԈ$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#Ff,r,r,r,r,r,r,r,r,r,r,r,rs,r,r,r/ ,r,r,r,r,r$$If]a$gd!#Ff $IfgdbN$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# .02RTZ˹y˹cڋcPcPcڋEhshkmH sH $hshkCJOJQJaJmH sH *hshk5CJOJQJ\aJmH sH "hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH hg_hkCJOJQJaJhkCJOJQJaJ"h+hk5CJOJQJ\aJhk5CJOJQJ\aJ$hg_hkCJOJQJaJmH sH $hg_hkCJOJQJaJmHsH"$(*,.2TZ,r,r,r,r,r,r,r,r,r,rs,r,r,r/ ,r,r,r?,r,rR,r,rd,r$$If]a$gd!#Ff`$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# (:,r~,r,r,rx,r,r,rs,r,r,r/ ,r,r! $$Ifa$gdfYa $$Ifa$gd1Rgd<Ff' $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#:<JLRTVڵxexOO@Ohk:CJOJQJ\aJ+jhXhk5:CJOJQJUaJ%hkchk5:CJOJQJaJ"hkchk5:CJOJQJaJ1jh1Rhk5:CJOJQJUaJ"h1Rhk5:CJOJQJaJ+jh1Rhk5:CJOJQJUaJhk5:CJOJQJaJ"hXhk5:CJOJQJaJhXhk5:OJQJaJhk $&(:BFHxzııġrccccrrSht=Ohk5CJOJQJaJhk5:CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hXhk5:NHOJQJaJhXhk5:OJQJaJ$hk5:CJOJQJaJmHsH*hXhk5:CJOJQJaJmHsH&hXhk5:CJNHOJQJaJ"hXhk5:CJOJQJaJ:DNPfhxz|~$<$Ifa$gd!# $$Ifa$gdfYaFf $$Ifa$gd!# :$If^:gd!# $$Ifa$gd$$If]a$gdFf> $$Ifa$gdfYa $$Ifa$gd!#Ff48tx"ųvgTv$hrhkCJOJQJaJmH sH hhkCJOJQJaJ$h6f+hkCJOJQJaJmH sH hC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ"hRhk5CJOJQJ\aJht=OhkCJOJQJaJht=Ohk5OJQJaJht=Ohk5CJOJQJaJhk5CJOJQJaJ "$&(*,.248tx$$If]a$gdFf $Ifgd$:$If]:a$gd $$Ifa$gdfYa $$Ifa$gd"LNX\bfhj $$Ifa$gdfYa$$If]a$gdFf $Ifgd$:$If]:a$gd $$Ifa$gd"LNVXZ\|$&6:hjlvxzȵȋȩ|iȩȋȩ|iȩȋȩ$hrhkCJOJQJaJmH sH hhkCJOJQJaJhC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJ$hhkCJOJQJaJmHsH$hrhkCJOJQJaJmHsH)jlnprtvz| $$Ifa$gdfYa$$If]a$gdFf $Ifgd$:$If]:a$gd $$Ifa$gd 6:jlvz~ $$Ifa$gdfYa$$If]a$gdFfv $Ifgd$:$If]:a$gd $$Ifa$gd $$Ifa$gdfYa$$If]a$gdFf $Ifgd$:$If]:a$gd $$Ifa$gd  !"%123458DEFHIнГбsfsYsfsYUO hk0JhkhhkOJQJaJh|hkOJQJaJht=OhkOJQJaJ$hrhkCJOJQJaJmH sH hC/|hkCJOJQJaJh|hkCJOJQJaJhkCJOJQJaJ% *hRhk5CJOJQJ\aJ"hRhk5CJOJQJ\aJhhkCJOJQJaJhhkCJOJQJaJ !"#$%&'()*+,-. $$Ifa$gd!#$$If]a$gd!#Ff $Ifgd$:$If]:a$gd $$Ifa$gd./0123456789:;<=>?@ABC $$Ifa$gd!#$$If]a$gd!#Ff$:$If]:a$gd!# $Ifgd!#$:$If]:a$gd!#CDEFGHI<W};jCgdCgd7jCgdnMgdFf $Ifgd!#$:$If]:a$gd!#IPY018FKMSWqrs{0689>GUWfjnwʹְʹʹʹʹʹʹʹʹʹʧʹʹʹʹh>jhk0JFh>jhk0Jhwhk0Jhk0JNH hk0J hk0JhBhk0JhBhk0JH hk0JOhwhk0JFhBhk0JOB#%;?W_`lqXYz} !34v hk0Jhk0JNHhwhk0JOhwhk0JF hk0Jhwhk0JQ3q Mm15t%RmCgd\toCgdH)CgdnMCgd7jKgddVCgd>>Cgd!$&/=@NQTUhiú̴̴̺̺̺̺̺̺̺̺̺ hk0JFh\tohk0JJ hk0JEh`hk0JE hk0JJh`hk0JJh`hk0JHhkhwhk0J hk0Jh>jhk0JhdVhk0JF hk0JhdVhk0J; $(_$l4Bbfh^`gdCgd`CgdCgdICgdH)  "#$&(*.BCDE[^qwx"&8DJdj󿹰h`hk0JFh\tohk0JFh`hk0JH hk0JHh`hk0JOh`hk0JE h`hkh`hk0JN hk0JJh`hk0JJ5\hkh`hk0JJ>&(06^bfvx`npJLNz|ͿͿͿͿ͞󕊕|u hkchkjh"thkUjh"thkU h"thkU hhkhk5\mHnHuhk5\mHnHujhPhk5U\hPhk5\hkmHnHuhkmHnHujhkUhkh`hk0JJ. contain schemes first encountered in  REF _Ref215196170 \h  \* MERGEFORMAT Figure  4.1. The final row is a valid scheme not selected by ParaMor. RankModel ofGoodCompletePartialErrorCSuffixesC-StemsNAdjVerbDeriv.StemSuffixChancearerir3%%%15 ba ban da das do dos n ndo r ra ron rse r rn17apoya disputa lanza lleva toma & 10%%%5ta tamente tas to tos22cier direc insli modes & n/a%%15a aba aban ada adas ado ados an ando ar ara aron arse ar arn17apoy disput lanz llev tom ... Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 12: An illustration of ParaMor s morpheme boundary error filters. Suffix-Internal Errors Both of these schemes have little variety in their distributions of cstem-final characters. Indeed both of these schemes have just a single leftward link. A distribution with little variety will have low entropy. Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 13: A pseudo-code implementation of ParaMor s algorithm to filter out scheme-clusters that likely hypothesize morpheme boundaries for their licensing types that fall internal to true stems. Stem-Internal Error Although this scheme has high entropy in its distribution of cstem-final characters (i.e. many leftward links) and so might be mistaken for a morpheme boundary, it has little variety among the characters which occur csuffix initially. Indeed a non-branching rightward path leads to the double-circled scheme which is the left edge of a morpheme boundary. A Morpheme Boundary The double-circled scheme that forms the left edge of this ambiguous morpheme boundary, or stretched hub (Johnson and Martin, 2003), is the morpheme boundary that ParaMor ultimately retains. FilterStemInternalBoundaryErrors(schemeClusters) { foreach (schemeCluster in schemeClusters) { [countOfErrorSchemes, countOfNonErrorSchemes] = [0, 0]; foreach (scheme in schemeCluster) { if (boundaryIsLikelyStemInternal(scheme)) countOfErrorSchemes++; else countOfNonErrorSchemes++; } if (countOfErrorSchemes >= countOfNonErrorSchemes) schemeClusters.remove(schemeCluster); } return schemeClusters; } // If an unbranching rightward path exists from scheme, follow the // path until either 1) the path forks or 2) we reach a scheme that is // likely the left edge of a morphemei.e. we reach a scheme with high // leftward entropy. If we do reach a scheme that is a likely left edge // of a morpheme then return true, that is, return that scheme // models an incorrect stem-internal morpheme boundary. boundaryIsLikelyStemInternal(scheme) { currentScheme = scheme; while (currentScheme.allCSuffixesBeginWithSameNonNullCharacter()) { foreach (cSuffix in currentScheme.cSuffixes) { rightwardCSuffixes.add(cSuffix.removeFirstCharacter()); } currentScheme = dynamicSchemeNetwork.generateScheme(rightwardCSuffixes); if (likelyModelOfAMorphemeLeftEdge(currentScheme)) return true; } return false; } Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 2: An ablation study. ParaMor s precision, recall, and F1 scores (and their standard deviations in parentheses) at word-to-morpheme segmentation using four configurations of ParaMor s scheme search, clustering, and filtering algorithms. A dot in the Search, Cluster, or Filter columns indicates that the relevant step(s) in ParaMors paradigm induction pipeline took part in that configuration. Paradigm models were induced using corpora of 50,000 unique word types. EnglishGermanSearchClusterFilterPRF1PRF1(14.2(0.5)82.8(1.0)24.2(0.7)13.1(0.4)78.6(1.0)22.5(0.5)((14.0(0.5)83.0(1.0)24.0(0.7)13.1(0.4)78.6(1.0)22.5(0.6)((63.0(1.0)47.8(1.3)54.3(0.9)48.4(0.9)41.6(1.0)44.7(0.6)(((57.2(1.1)48.8(1.3)52.7(0.9)54.1(1.0)38.9(1.0)45.2(0.6) Figure  STYLEREF 1 \s  5. SEQ Figure \* ARABIC \s 1 2: Morphological segmentations of some Spanish word forms produced by ParaMor over a corpus of 100,000 unique types. RowWord FormGlossCorrect SegmentationMorphosyntacticIndividual BoundariesParaMors SegmentationCluster RankIllustrative Word Forms1sacerdotepriestsacerdotesacerdotesacerdotesacerdote-2sacerdotespriestssacerdote +ssacerdote +plsacerdote + ssacerdote +s13regularesordinaryregular +esordinary +plregular +esregular +es1224chanchofilthychanch +ochancho +mascchanch +ochanch +o35incgnitasunknownincgnit +a +sincognito +fem +plincgnit +as, incgnita +sincgnit +a +s1, 36descarrilaremoswe will be deraileddescarril +aremosdescarrilar +1pl.fut.indicdescarril +aremosdescarril +aremos47accidentarseto have an accidentaccident +ar +seaccidentarse +inf +reflexaccident +arseaccident +arse48erradoswrong, mistaken (Masculine Plural)err +ad +o +serrar +adj +masc +plerr +ados errad +os errado +serr +ad +o +s1, 3, 49agradezcoI thankagradec +oagradecer +1sg.pres.indicagrade +zcoagrade +zco2110agradecimoswe thankagradec +imosagradecer +1pl.past.indicagrade +cimos agradec +imosagrade +c +imos17, 2111antelacin(in) advanceantel( +)ac1nantelar +cin.Nantel +ac1n antelac +inantel +ac +in4, 512tanteadorstorekeepertante( +)adortantear +ador.Ntante +adortante +ador413vetehe/she should vetovet +evetar +3sg.pres.subjuncvet +evet +e4Randomly Selected 14bambamgnot a Spanish wordbambamgbambamgbambamgbambamg-15clausurarhe/she will concludeclausur +arclausurar +3sg.fut.indicclausur +arclausur +ar416hospitalhospitalhospitalhospitalhospit +al hospital +lhospit +a +l(7), (41)17investidoinvested (Masculine Singular)invest +id +oinvestor +adj +mascinvest +ido investi +do invested +oinvest +i +d +o3, 11, 17, (28)18pacficamentepeaceablypacfic +amentepacficamentepacific +amente pacifca +mentepacific +a +mente1, 3, 12219sabidurawisdomsabidurasabidurasabidurasabidura- Figure  STYLEREF 1 \s  4. SEQ Figure \* ARABIC \s 1 17: The number of schemes or scheme-clusters that ParaMor produces for six languages after each search, clustering, or filtering step in ParaMor s paradigm induction pipeline. # of Schemes / ClustersSpanishEnglishGermanFinnishTurkishArabicSearch8339876710778107741034012528Excluding Short Types69096733985110495961814247Clustering6087639691458797689812744Small Cluster Filtering15019320628025887Morpheme Boundary Filtering42404811610027 Word FormGlossMorphosyntactic AnalysisParaMors Segmentationsacerdotepriestsacerdotesacerdotesacerdotespriestssacerdote +plsacerdote +sregularesordinaryordinary +plregular +esagradezcoI thankagradecer +1sg.pres.indicagrade +zcoagradecimoswe thankagradecer +1pl.past.indicagrade +c +imos Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 5: Results from the Linguistic Evaluation of Morpho Challenge. The unsupe*,.02NVXz  ,026вsfsfsfsfsT"hhk5CJOJQJ\aJhk5CJOJQJaJht=Ohk5CJOJQJaJht=Ohk5OJQJaJht=OhkCJOJQJaJ"hzhk5:CJOJQJaJhzhk5:OJQJaJ h&>hk hkhkhk h"thkjh"thkUhkchkmHnHuhkchkmHnHu `,02<NXjzd$If^gdz$d$Ifa$gd!#8gd,gdn kd$$If4ִ  !/ m  t!    4af4pFƥԥ֥إڥ   "Ff$d$Ifa$gd!#d$If^gdz$d$Ifa$gdz d$Ifgd!#"$&(*,.0268:>@BDFHLNRTZ$:$If]:a$gd, $$Ifa$gd,$$If]a$gd,Ff d$Ifgd!#6TZ¦$BFrx̹ߦq^H*h hk5CJOJQJ\aJmH sH $hk5CJOJQJ\aJmH sH "hhk5CJOJQJ\aJ$h6f+hkCJOJQJaJmH sH hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH $hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsH"h+hk5CJOJQJ\aJhg_hkCJOJQJaJZ¦$&*,.0248:>@BFr $IfgdR$:$If]:a$gdR $$Ifa$gdR$$If]a$gdRFf$:$If]:a$gd, $Ifgd,rx§ħƧȧ̧ΧЧҧاV\$:$If]:a$gdR $$Ifa$gdR$$If]a$gdRFf $IfgdR$9$If]9a$gdRȧʧҧ֧اFHLNTVZ\˵o˵\I\I$h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH hkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsH$hg_hkCJOJQJaJmHsHhkCJOJQJaJmH sH *h hk5CJOJQJ\aJmH sH $hk5CJOJQJ\aJmH sH hg_hkCJOJQJaJ$h hkCJOJQJaJmH sH Ff_$9$If]9a$gdR $IfgdR$:$If]:a$gdR $$Ifa$gdR$$If]a$gdRFf ¨ĨƨȨʨ̨ΨШҨԨ֨بڨܨިFf $IfgdbN$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#$$If]a$gd!#بڨܨި "$&(*,.<>\^`bdױכıı}yrgrgYMghk5\mHnHuhk5\mHnHujhk5U\ hk5\hkhshkmH sH $hshkCJOJQJaJmH sH *hshk5CJOJQJ\aJmH sH $h6f+hkCJOJQJaJmH sH $hg_hkCJOJQJaJmH sH $h,hkCJOJQJaJmH sH *h,hk5CJOJQJ\aJmH sH   $$If]a$gd!#Ff$$If]a$gdz $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!# "$&(*,.+$Kdx^`Ka$gd&m?gd&m?gd,Fff# $Ifgd!#$:$If]:a$gd!# $$Ifa$gd!#dfh%&)*+8?PQ]jkƪǪ  ѸѰykykjhehk5U\hehk5\hxhkCJaJh2Y?hkCJNHaJhkCJNHaJh2Y?hkCJaJhkCJaJhk5>*\aJhxhk5>*NH\aJhxhk5>*\aJhkhk5\mHnHujhk5U\ hk5\$+­Z[o./0cʰ;CaeCgd6.Cgdegd&m?$dx^`a$gdxgde$dx^`a$gd&m? @BFHJ­DEYZ[no-.0:PQ_bdm{~ʽʍʇ~uuuuuh`hk0JJh`hk0JH hk0JHhk5>*\aJh2Y?hkCJaJhkCJNHaJhkCJaJhxhkaJhxhk5>*\aJhkhehk5\jhehk5U\hk5\mHnHuhk5\mHnHu.ɰʰ̰հܰ߰46>BC]_`acegk±ñıűαܱݱޱ²NOw޺ޮޡhk0JONH hk0JO h`hkh xhk0JJ5\h`hk0JNh`hk0JJ5\ hk0JFh`hk0JJhkh\tohk0JJ hk0JEh`hk0JE hk0JJ9ñƱޱ&m@x.ingdFCgd}Cgd`HCgdICgdeCgd xwxγ*-DGHPdgkn~ĴƴʴҴڴɷɷɷɚɱձh xhk0JJ5\h}hk0JJ5NH\hk0JJ5\ hk0JFh}hk0JFh}hk0JJh}hk0JJ5\ hk0JJhkh`hk0JJh`hk0JH hk0JHh`hk0JO8 / @BDFHĸŸ̸͸  +,23ȺȮpp!h-hk5H*OJQJ\aJ"h-hk5NHOJQJ\aJh-hk5OJQJ\aJ hkNH hkH*hk5\mHnHuhk5\mHnHujh/phk5U\h/phk5\hk h}hkh}hk0JJ5\h}hk0JJ, $$If`a$gd.#%l)     P66666$$If`a$gd.#%l)kd&$$If4r Y ! L   t04aP6$$If`a$gd.#%l$$If`a$gd.#%l\kd'$$If4lr Y ! L   t04a&(*-/145789>EJQV]bi$If`gd0"lFf+$$If`a$gd.#%l\$$If`a$gd.#%l34569>EJQV]binuzùȹϹйѹӹԹչڹ ",:<>@BDFHR`jxʺ̽̽̽̽̽̽̽̽̽̽̽̽̽̽̽̽̽̽h0"hk5OJQJ\aJh*hkCJOJQJaJh-hkOJQJaJ$ jh-hk5OJQJ\aJhkh-hk5OJQJ\aJDinuz~~~$If`gd0"l$$If`a$gd.#%l$$If`a$gd.#%lFf(1$If`gd0"l$$If`a$gd*lùȹϹйҹӹչڹ~~$If`gd0"lT$$If`a$gd.#%lTFfg8$$If`a$gd*l$If`gd0"l$$If`a$gd.#%l ",:<@DHR`~$If`gd0"l$$If`a$gd.#%lFf?$$If`a$gd*lT$If`gd0"lT$$If`a$gd.#%lT`jxʺغںܺ޺$$If`a$gd.#%lFfF$$If`a$gd*l$If`gd0"l$$If`a$gd.#%lʺغں ,.02468nprtv/039>C}p`p`h$\hk5CJOJQJaJhk5CJOJQJaJ"h$\hk5CJOJQJ\aJhk5\mHnHuhk5\mHnHujhdVhk5U\hdVhk5\h-hkCJOJQJaJh-hkOJQJaJh-hk5OJQJ\aJhkh*hkCJOJQJaJ$./04CIQ$d$$Ifa$gddlB$$Ifa$gddlBgddVgdFFfL$$If`a$gd.#%lCXY^nx~żƼRST߽()ʽʮʽʽʜʍʜʁʁʁʁʁnnnn$h$\hkCJOJQJaJmHsHhkCJOJQJaJh$\hkCJOJQJaJ#h$\hk5CJNHOJQJaJhk5CJNHOJQJaJhk5CJOJQJaJh$\hk5CJOJQJaJ&h$\hk5CJNHOJQJ\aJ"h$\hk5CJOJQJ\aJ(Q^nżǼѼؼFfU$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFfoQ$d$$Ifa$gddlB&4BOQRSU_ww$d8$Ifa$gd$\l$$Ifa$gdx_tlFfIY$d8$Ifa$gdx_tl\$d8$Ifa$gd$\l\$$Ifa$gdx_tl\ _htȽҽܽ޽$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFf]$d8$Ifa$gdx_tl޽߽%2AFGH$$Ifa$gdx_tlTFfwd$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFf` )*56GHIQR{|¾þľ234Z["  VZrt|@ACཱཱཱཱཱཱཱཱ͋$h$\hkCJOJQJaJmHsH$h$\hkCJOJQJaJmHsHhkCJOJQJaJh$\hk5CJOJQJaJ$h$\hkCJOJQJaJmHsHh$\hkCJOJQJaJ h$\hkCJNHOJQJaJ6HJZn¾þžҾ$d8$Ifa$gd$\l$$Ifa$gdx_tlFf5h$d8x$Ifa$gdx_tlT$d8x$Ifa$gd$\lT Ҿ /1235=`n$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFfk$d8$Ifa$gdx_tlſͿؿFfgs$d8$Ifa$gdx_tl\$d8$Ifa$gd$\l\$$Ifa$gdx_tl\Ffo$d8$Ifa$gdx_tl $<Njw$d8$Ifa$gd$\l$$Ifa$gdx_tlFf!w$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tl $>\|$d8x$Ifa$gd$\l$$Ifa$gdx_tlFfz$d8$Ifa$gdx_tl$d8$Ifa$gd$\l %1=?@ADI\c{$d8$Ifa$gdx_tlT$d8$Ifa$gd$\lT$$Ifa$gdx_tlTFf~$d8x$Ifa$gdx_tlr$d8$Ifa$gd$\l$$Ifa$gdx_tlFf$d8x$Ifa$gdx_tl$d8x$Ifa$gd$\l$$Ifa$gdx_tlFf 57BCHJ/024<=勁ssjh$nhk5U\h$nhk5\hk *h$\hkCJOJQJaJ$h$\hkCJOJQJaJmHsHhwhhkCJOJQJaJ h$\hkCJNHOJQJaJ$h$\hkCJOJQJaJmHsHh$\hkCJOJQJaJhkCJOJQJaJ)*7DFGHKT]p$d8$Ifa$gdx_tl\$d8$Ifa$gd$\l\$$Ifa$gdx_tl\Ff9$d8$Ifa$gdx_tl$d8$Ifa$gd$\l ]foz$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFf$d8$Ifa$gdx_tl\ 0125CM]k{Ffg$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tlFf$d8$Ifa$gdx_tlgd$ngdzFf!$d8$Ifa$gdx_tl$d8$Ifa$gd$\l$$Ifa$gdx_tl @BFHJ*+2TU_iuʻʰʕzʕzʕzʕhkqhkOJQJhd[hk5NHOJQJ\hd[hk5OJQJ\hk5NHOJQJ\hk5OJQJ\hkmHnHu hkNHhkh$nhk5\jh$nhk5U\hk5\mHnHuhk5\mHnHu0 #*$$If`a$gdel3*+27<N444$$If`a$gdel\kd*$$If֞) A Yr!)  t449a<BHNTU4kd$$If֞) A Yr!)  t449a$$If`a$gdelTUuz$$If`a$gdel$$If`a$gdelTN4$$If`a$gdelTkd$$If֞) A Yr!)  t449a$$If`a$gdel$$If`a$gdelN4$$If`a$gdelkdc$$If֞) A Yr!)  t449a),;Ast~ (*,ڻڻ˩si[i[Mhk5\mHnHujh)hk5U\h)hk5\h$\hkCJOJQJaJ&hhk5:CJNHOJQJaJ%hhk5:CJOJQJ\aJ"hhk5:CJOJQJaJhd[hk5NHOJQJ\hkhkqhkOJQJhd[hk5OJQJ\hk5NHOJQJ\hk5OJQJ\$$If`a$gdel$$If`a$gdelN444$$If`a$gdelkd$$If֞) A Yr!)  t449a"&)*4kd$$If֞) A Yr!)  t449a$$If`a$gdel\*+,;Ac~Ukdd$$If\   t222244 a$d$$Ifa$gdxflBgd:\Ovkd@$$If\   t44 a$d8$Ifa$gdx_tl$d8$Ifa$gd$\lOvkd$$If\   t44 a$d8$Ifa$gdx_tl\$d8$Ifa$gd$\l\Ovkd~$$If\   t44 a$d8$Ifa$gdx_tl$d8$Ifa$gd$\l&ZrtOvkd$$If\   t44 a$d8$Ifa$gdx_tl\$d8$Ifa$gd$\l\tOJHHgduvkd$$If\   t44 a$d8$Ifa$gdx_tl$d8$Ifa$gd$\l,.024jlnprXSXTXXX YYYYYYYYZZ%Z&Z@ZAZIZMZOZPZRZ`ZeZfZннгןxg hM6hkCJOJQJ^JaJ#hM6hk:CJOJQJ^JaJ)hM6hk5:CJOJQJ\^JaJ&hEa5hk5CJOJQJ\^JaJhhk5\hhkH*UhhkNH hhkhkh)hk5\jh)hk5U\hk5\mHnHu(rvised morphology induction systems which appear in this table are the nine best systems from the 2008 and 2007 challenges. Systems participated in up to 5 language tracks. In each language track all participating systems were scored at precision (P), recall (R), and F1 of morpheme identification. The ground truth against which Morpho Challenge compares systems is a morphologically analyzed answer key that includes both inflectional and derivational morphology. For each language track, the system or systems which place first at F1 by a statistically significant margin appear in bold. Monson et al. (2008b)Other Authors200820082007ParaMor + MorfessorParaMorMorfessorMorfessor MAPZemanKohonenBernhardBordagPitlerEnglishP50.658.577.282.253.083.461.659.774.7R63.348.134.033.142.113.460.032.140.6F156.352.847.247.246.923.160.841.852.6GermanP49.553.467.267.653.187.949.160.5-R59.538.236.836.928.47.457.441.6-F154.144.547.647.837.013.752.949.3-FinnishP49.846.477.476.858.592.659.771.3-R47.334.421.527.520.56.940.424.4-F148.539.533.740.630.312.848.236.4-TurkishP51.956.773.976.465.893.373.781.3-R52.139.426.124.518.86.214.817.6-F152.046.538.537.129.211.524.728.9-ArabicP79.878.690.490.277.2----R27.58.521.021.012.7----F140.915.434.034.021.9---- Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 1: Full morphosyntactic analyses in the style found in the answer keys of the Linguistic Evaluation at Morpho Challenge, together with ParaMor s Morphological segmentations of five Spanish word forms. ParaMor produced these segmentations when analyzing a corpus of 100,000 unique types. Figure  STYLEREF 1 \s  6. SEQ Figure \* ARABIC \s 1 6: Average precision scores for unsupervised morphology induction systems which participated in the Information Retrieval (IR) Evaluation of Morpho Challenge. The unsupervised morphology induction systems which appear in this table are the eight best systems from the 2008 and 2007 challenges. Systems participated in up to three language tracks. The best performing system(s) for each track appear in bold font. Monson et al. (2008)Other Authors200820082007ParaMor + MorfessorParaMorMorfessorMorfessor MAPMorfessor BaselineMcNameeBernhardBordagEnglish39.939.336.437.138.636.339.434.3German47.336.346.74OZPZQZRZhZvZwZxZyZ~ZZfkdF$$Ift\hL !&h&& & !4ap( $$Ifa$gdEa5gd) fZgZvZwZyZZZZZZZZZZZZZZZZZ[[[![N[O[P[Q[R[S[q[v[[[͹͹عععlX&hEa5hk5CJOJQJ\^JaJ)hEa5hk5CJH*OJQJ\^JaJ hEa5hkCJOJQJ^JaJ*hEa5hk5CJNHOJQJ\^JaJ hk5CJOJQJ\^JaJ&hEa5hk5CJOJQJ\^JaJhhkCJaJ)hn+hk5:CJOJQJ\^JaJ#hM6hk:CJOJQJ^JaJ"ZZZZZZZZZZUkdX$$IfrhL !&h&& & &p !4ap2 $$Ifa$gdEa5 ZZZZZZZZZZZ[ [[[[[[[![&[+[0[5[Ff $IfgdEa5$qq$If]q^qa$gdEa5Ff $$Ifa$gdEa55[:[?[D[I[N[O[P[S[X[][b[g[l[q[v[{[[[[[[[[FfM $IfgdEa5$qq$If]q^qa$gdEa5Ff $$Ifa$gdEa5[[[[[[[[[[[[[[\\\\E\F\G\I\r\s\t\u\v\w\|\\\\\\\\\\\]]]]]] ]2]3]:]<]]]^]a]]]]]]źź֥֑źźź֥֑őźźź֥֑źźź֥&hEa5hk5CJOJQJ\^JaJ)hEa5hk5CJH*OJQJ\^JaJhhkCJaJ hEa5hkCJOJQJ^JaJ&hEa5hk5CJOJQJ\^JaJ)hn+hk5:CJOJQJ\^JaJ9[[[[[[[[[[[[[[[[[[[[[[[[Ff $IfgdEa5$qq$If]q^qa$gdEa5Ff) $$Ifa$gdEa5[[[[\ \\\\\\ \%\*\/\4\9\>\C\E\F\G\I\N\Ff $IfgdEa5$qq$If]q^qa$gdEa5Ff $$Ifa$gdEa5N\S\X\]\b\f\k\p\r\s\t\w\|\\\\\\\\\\\\Ffg $IfgdEa5$qq$If]q^qa$gdEa5Ff $$Ifa$gdEa5\\\\\\\\\\\\\\\\\\\\]]]]Ff- $IfgdEa5$qq$If]q^qa$gdEa5FfC $$Ifa$gdEa5]] ]]]]!]&]+]0]2]3]:]<]A]F]K]P]U]W]Y][]]]^]Ff$qq$If]q^qa$gdEa5Ff $$Ifa$gdEa5 $IfgdEa5^]_]a]f]j]o]t]y]{]}]]]]]]]]]]]]]]]]]gdEa5FfuFf $IfgdEa5 $$Ifa$gdEa5]]]]]]]]]^^^^^ ^@^B^D^F^H^^^T_V_____``````````````aqqqqqqjh3 hk5U\h3 hk5\ hkNHhk5\mHnHuhk5\mHnHujhdVhk5U\hdVhk5\hkhhkCJaJhhkCJaJ hEa5hkCJOJQJ^JaJ&hEa5hk5CJOJQJ\^JaJ*]]```ccc3c4cBcCcdkd$$If\i!& &(  i!4ap( $$Ifa$gd3 gdA6 aaaaa:b;bbbbbccccccBcCcEcTcUcWcccccccccccccdιιήo^o^\U h3 hkCJOJQJ^JaJ&h3 hk5CJOJQJ\^JaJ-hn+hk5:CJNHOJQJ\^JaJ&h3 hk5CJOJQJ\^JaJh3 hkCJaJ)hn+hk5:CJOJQJ\^JaJ&h3 hk5CJOJQJ\^JaJh3 hk5\hkh3 hkNH h3 hk$CcEcJcOcTcUcWckcsc}ccccpkd$$If\i!& &(  &i!4ap( $$Ifa$ cccccccccccccccccc !&+Ff)Ff Ff $$Ifa$6.446.643.947.343.1Finnish46.739.746.844.444.349.249.243.1 !:DIJKLMNлаh$,h)!5\hdhkh3 hkCJaJ)hn+hk5:CJOJQJ\^JaJh3 hkCJaJ&h3 hk5CJOJQJ\^JaJ h3 hkCJOJQJ^JaJ +05:?DIJKLMN!1Cgd$,gd3 FfT  $$Ifa$ 0 0 01h/R :pd/ =!"#$%* 01h0/R = /!"#$%0 0 01h/R :p!#/ =!"#$%0 01h0/R :p!#= /!"#$%- 01h/R :p!#/ =!"#$%0 01h0/R :p!#= /!"#$%0 0 01h/R :px_t/ =!"#$%0 01h0/R :ptH= /!"#$%0 0 01h/R :pz/ =!"#$%0 01h0/R :p#(= /!"#$%( 0 0:p$\/ =!"#$%@= ʆ,I:=$<HKAO x[{pU߽9ፂNJR20v0))) IxT0-B"<HM'3ն%< ߷{n͹)/q{ݻgw\) g>!ʤ|b_! !_qlk~w}KvgG|' uA*E'ۀdvَ gRъ:O7]SI>(.o*Dlwv]]h֠{#\P3}Aۺp<6f 3v8Hۋ/6.# hit Hc}!!)8yd7k7i?~Ϸ71.W:ܓ<]zW/}#¨ě-~d'YK%oМ}n9#p.Ֆ#.?EP1L%~*~$&~ډ봦O 3| RK6- ,:{ ]rGLHg8%8kQ.oʺZHbCd.ڗl |:C9#B-BZ|Ho߮}n`N{?Le\W@}憁* h#c:JvMF\C}檁Tc`Ig>=W0d[fH/|b`]4J@ռ7_bqA살;!5 B#ͷ on(Y#@q4nX堳QsbRK6U?@g@2d?Dll>6NX@1fc8G<%_VpP i5, վp"﹣uE_:р/0va-H<:vEl* 죱 t#_9b`%J`=+AGXlX]eue YW`@;8g~ҽ'4ÌSd'Dd'j /1NE>oﵾS8@')1@Qq4Kw]i1L뀰cCeߡ]51FKw1Zd5h~ 'ntS'/(@'^tKww2U7\1~.C:S珣d v\!"4, owJFȫHN/mG:3 WV-Fy$tՐ8_Ȅd jߩ\։P1kTv.kGjMz|YgnS_6$9@Ԏ|k|;35_]j5>Z?1kI$Ԏk||֥`ޔs.Uj5Y/ϗ| MՎB_t !FVMulT"Cka>қr^"CVD˭:x {TȜï>fws$H_pM3R}L:E"ZIwhuu "nNi"z&4kv25IL5F5] |;Ϫ/5P_h{GM!Ix~ODG2!MԺt㆞kL:[Np I.!D,Q^cz'XKt/ci,З;QWjl_"=7}ݾDSDIY7jŇiZGY[uιJKT>vWC1 -s[|;n1qMSfDM uq NOyUÍF,%wm`Z=uYOtߌga0.|ӿ-y7cxz%X1cb<{W;֩(c=UWP.{,L,;qMdabM2V(cm`#abm2VGz-6{XKäq| ݳjs_hkKUY:BηO*LJ_٫u^?5 ~F}#Of|:oog9Q5س/lDcs~^ո=uZnF} Uc?3CF׽6j9O}DyK _Toc217315416}DyK _Toc217315416}DyK _Toc217315417}DyK _Toc217315417}DyK _Toc217315418}DyK _Toc217315418}DyK _Toc217315419}DyK _Toc217315419}DyK _Toc217315420}DyK _Toc217315420}DyK _Toc217315421}DyK _Toc217315421}DyK _Toc217315422}DyK _Toc217315422}DyK _Toc217315423}DyK _Toc217315423}DyK _Toc217315424}DyK _Toc217315424}DyK _Toc217315425}DyK _Toc217315425}DyK _Toc217315426}DyK _Toc217315426}DyK _Toc217315427}DyK _Toc217315427}DyK _Toc217315428}DyK _Toc217315428}DyK _Toc217315429}DyK _Toc217315429}DyK _Toc217315430}DyK _Toc217315430}DyK _Toc217315431}DyK _Toc217315431}DyK _Toc217315432}DyK _Toc217315432}DyK _Toc217315433}DyK _Toc217315433}DyK _Toc217315434}DyK _Toc217315434}DyK _Toc217315435}DyK _Toc217315435}DyK _Toc217315436}DyK _Toc217315436}DyK _Toc217315437}DyK _Toc217315437}DyK _Toc217315438}DyK _Toc217315438}DyK _Toc217315439}DyK _Toc217315439}DyK _Toc217315440}DyK _Toc217315440}DyK _Toc217315441}DyK _Toc217315441}DyK _Toc217315442}DyK _Toc217315442}DyK _Toc217315443}DyK _Toc217315443}DyK _Toc217315444}DyK _Toc217315444}DyK _Toc217315445}DyK _Toc217315445}DyK _Toc217315446}DyK _Toc217315446}DyK _Toc217315447}DyK _Toc217315447}DyK _Toc217315448}DyK _Toc217315448}DyK _Toc217315449}DyK _Toc217315449}DyK _Toc217315450}DyK _Toc217315450}DyK _Toc217315451}DyK _Toc217315451}DyK _Toc217315452}DyK _Toc217315452}DyK _Toc217315453}DyK _Toc217315453}DyK _Toc217315454}DyK _Toc217315454}DyK _Toc217315455}DyK _Toc217315455}DyK _Toc217315456}DyK _Toc217315456}DyK _Toc217315457}DyK _Toc217315457}DyK _Toc217315458}DyK _Toc217315458}DyK _Toc217315459}DyK _Toc217315459}DyK _Toc217315460}DyK _Toc217315460}DyK _Toc217315461}DyK _Toc217315461}DyK _Toc217315462}DyK _Toc217315462}DyK _Toc217315463}DyK _Toc217315463}DyK _Toc217315464}DyK _Toc217315464}DyK _Toc217315465}DyK _Toc217315465}DyK _Toc217315466}DyK _Toc217315466}DyK _Toc217315467}DyK _Toc217315467}DyK _Toc217315468}DyK _Toc217315468}DyK _Toc217315469}DyK _Toc217315469}DyK _Toc217315470}DyK _Toc217315470}DyK _Toc217315471}DyK _Toc217315471}DyK _Toc217315472}DyK _Toc217315472}DyK _Toc217315473}DyK _Toc217315473}DyK _Toc217315474}DyK _Toc217315474}DyK _Toc217315475}DyK _Toc217315475}DyK _Toc217315476}DyK _Toc217315476}DyK _Toc217315477}DyK _Toc217315477}DyK _Toc217315478}DyK _Toc217315478}DyK _Toc217315479}DyK _Toc217315479}DyK _Toc217315480}DyK _Toc217315480}DyK _Toc217315481}DyK _Toc217315481}DyK _Toc217315482}DyK _Toc217315482}DyK _Toc217315483}DyK _Toc217315483}DyK _Ref192490406}DyK _Ref193208912}DyK _Ref211051873}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref192490406}DyK _Ref193208912}DyK _Ref193208912}DyK _Ref192490406}DyK _Ref211051873}DyK _Ref211051873}DyK _Ref211051873}DyK _Ref211051873}DyK _Ref211051873}DyK _Ref192388754}DyK _Ref212623558}DyK _Ref212623577}DyK _Ref212623601}DyK _Ref212623620}DyK _Ref193859274}DyK _Ref190771388}DyK _Ref192388754}DyK _Ref212696175}DyK _Ref212713404}DyK _Ref212696175}DyK _Ref212713404}DyK _Ref212696175}DyK _Ref212713404}DyK _Ref192388754}DyK _Ref193859274Dd b  c $A? ?3"`?2== 5i÷qS`!= 5i÷q:@xcdd`` @c112BYL%bpu @c112BYL%bpu;vv0o8L+KRs@2u(2tA4Ag!!v120elLd Dd Tb  c $A? ?3"`?2U bh[{ߑ1Z`!) bh[{ߑ HXJxcdd``> @c112BYL%bpu @c112BYL%bpu;vv0o8L+KRs@2u(2tA4Ag!!v120elLd Dd Tb  c $A? ?3"`?2U bh[{ߑ1^`!) bh[{ߑ HXJxcdd``> @c112BYL%bpuBǮ t)``!.>BǮ t):@`!xcdd`` @c112BYL%bpu @c112BYL%bpu;vv0o8L+KRs@2u(2tA4Ag!!v120elLdDd T\   c $A ? ?#" `2U bh[{ߑ1d`!) bh[{ߑ HXJxcdd``> @c112BYL%bpu ɬ #qAc ਃƧ;N+KRsXA2u(2tA4T}b@3XjDd J  C A? " 20sjW}ҧ!l n`!sjW}ҧ!l넒P%xcdd``> @c112BYL%bpu1X +8|-_.PHq1>fiը$d*n@$s]9AFhVT|7U#Ž6"D7.%ڽ$8{1@ M-VK-WM#ҽ b r+aMp ńQA8)?zTh6TX !PO?g |^ 0~[V0g2BÓQSksYE鹉05[j`jLjK\QpsE0>;>=fa p] `pĤ\Y\d.P"CXYO`N1hDd pJ  C A? " 28]hVہǽWq`!8]hVہǽ%X":lxcdd``Naa``baV d,FYzP1n:&U! KA?H1Z ˀ깡jx|K2B* Rjvfb KP 27)?K{!>b8gF\uz m#dDSnwBe2 6ӽO%ƽ"w+d@B.FWg\ 2"@ d@Fbo+BJS&:|I D()ᛦBtJԄob*\( ܐbr<"xr+alE?ίBΟ)p8Yܨ8Q[r@C<&GA 0>[P!@fxrB* d䧧|HdhM>L(LT5Xǚ".a6\..,@OmMm r #t`UHA? ?Y 豌%wIvzj´`> =#RpeqIj.E.Y=LmiDd TJ  C A? "2ZNtF`\T21t`!ZNtF`\T21\` hn XJmxcdd``^ @c112BYL%bpu 1ucDd Tb  c $A? ?3"`?2P^j Ԕ_`!P^j Ԕ_ e XJOxcdd``cd``baV d,FYzP1n:&&! KA?H1Z ㆪaM,,He`c&,eBܤR.ÂB 7. _ `*F\u kbm#dn0;M 2 M-VK-WMNP dr@E 3Y|,6#>j`ddE dfmp pA ev0o8ݡ\Y\pd.P"CX,Āf~EDd Tb  c $A? ?3"`?2X>z9,"`G|kL`!cX>z9,"`G|R @2XJ1xcdd``dd``baV d,FYzP1n:&&! KA?H1Z ㆪaM,,He`c&,eBܤ+R\ >1ܘ!+(|-7Nd*F\u ̛# PNXyʬ`! PNXyʬ:YRxcdd`` @c112BYL%bpu @c112BYL%bpudKF&&\ s: @> 1,$MDd |,\  c $A? ?#" `2Q }aM E-`!% }aM E`0xcdd``> @c112BYL%bpu5I؏s;0DA&M`!5I؏s;0DA&M:RHxcdd`` @c112BYL%bpu @c112BYL%bpu @c112BYL%bpubVsi#/ Y@:*!| s%?Y 2!Wb! W&00lv({(38S 5.p2M%#RpeqIj. @ ]` "9e Dd |J  C A? "2X"Dލ6#c^`!VX"Dލ6#c ` 0$xcdd``fed``baV d,FYzP1n:B@?b sC0&dT20$ͤ `[YB2sSRsv\~G e*F\_>tUB L`+!7#|Fnƛ *a|7P@penR~!.kB8nO0Ȉ3\P,#t] `p\021)W2ԡ"b> 1BDeDd h hJ  C A? "2&vЂ(`!&vЂr@@0J|xڥK@߻F28uBB %["*p*:9+I0޻Wt;{!b`0 b,iq?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~                         2 3 4 5 6 7 8 9 : ; < = > ? @ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z [ \ ] ^ _ ` a b c d e f g h i j k l m n o p q s t u v w x y z { | } ~  Root EntryS FRZ` Data O WordDocumentRObjectPoolU"`tV`RZ`} _Tຄs`!} _Tຄs`)0xڝK@]l-i'PtEZ!E lI-ZPD'QZ] w]p}޽<_`zȊ`/¶m(RoQnA&@i$G]O fhh%U,\ZC]ךF(נsm>濛(~[ 0y.IO@(.e_ʿ ^y+D9yz:%܁$jV:dJ5󜥐TuKsV;!"֊;2ݱCg V 0h/ 9ȵU`= 5Dd ,h\ ! c $A? ?#" `2Qfc }(v| -_`!%fc }(v| @|xcdd``> @c112BYL%bpucR ?bl%.``!cR ?bl%.:RHxcdd`` @c112BYL%bpu @c112BYL%bpucR ?bl%.N`!cR ?bl%.:RHxcdd`` @c112BYL%bpu @c112BYL%bpu @c112BYL%bpu @c112BYL%bpu &E…v7Chڒb"RuFbi,@u@@ڈKӹƗe!7#|Fnƕ *a|7P@penR~&G!ǚ d=._T E.pH3m- #RpeqIj.9E.Yݏ`6^Dd hJ + C A)? "(2Q*߱ ltµ|'-`!%*߱ ltµ|'̒@|xcdd``> @c112BYL%bpub9Ӷsi#+@:+!5#dKF&&\- s:  2oNDd ,h\ , c $A*? ?#" `)2Qfc }(v| -p`!%fc }(v| @|xcdd``> @c112BYL%bpu @c112BYL%bpub9Ӷsi#+@:+!5#dKF&&\- s:  2oNDd @J . C A,? "+2tR",R6UP``!HR",R6U hxڕJAƿ?z\%"O6"N &E…v7Chڒb"RuF!=D2x xcdd`` @c112BYL%bpu @c112BYL%bpu`!6TYohEsT[1H xcdd``> @c112BYL%bpu @c112BYL%bpu`!6TYohEsT[1H xcdd``> @c112BYL%bpu`!6TYohEsT[1H xcdd``> @c112BYL%bpu @c112BYL%bpu.zNYvo$Qs`!I>.zNYvo$@"xcdd`` @c112BYL%bpu t7F=[.=S %.pH2M%=#RpeqIj.\E.E`Dd |,\ 7 c $A5? ?#" `62QIx㊅)}-`!%Ix㊅)}`0xcdd``> @c112BYL%bpuS? P.P16V 1y12ps01Lp䂺 CF=0adbR ,.Ie2C 2OO Dd |,\ 8 c $A6? ?#" `72QIx㊅)}-`!%Ix㊅)}`0xcdd``> @c112BYL%bpuS? P.P16V 1y12ps01Lp䂺 CF=0adbR ,.Ie2C 2OO 2Dd \ 9 c $A7? ?#" `82 |11MNd,Jk^`!V |11MNd,Jk*h$xcdd``e 2 ĜL0##0KQ* W,d3H1)fY+PsC0&dT20(0pEe9 u!f0E00XL'!,l&l&3%v5 @Hfnj_jBP~nbÍr>=`0 @yL`6 jn? $b.\&+F\P.zNYvo$Q`!I>.zNYvo$@"xcdd`` @c112BYL%bpu t7F=[.=S %.pH2M%=#RpeqIj.\E.E`=Dd `@\ ; c $A9? ?#" `:2%!_{1i`!a%!_{1:X /xcdd``ed``baV d,FYzP1n:&B@?b u 10 UXRY7S?&meabM-VK-WMc1sC VPZ^&T T?tUOX@8 30Bԣ@ ;Ѓ3 lBN 27)?ύKoG!ǂ d9=z R%414=cdbR ,.Ieԡ"|b@#3X?QiDd ,\ < c $A:? ?#" `;2Xd >ZQ4.`!,d >ZQxcdd``> @c112BYL%bpu @c112BYL%bpu C A<? "=2Py̸B{I,9`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu2PY=m*ڪt њ),'`!$Y=m*ڪt њ)@ |XJxcdd``> @c112BYL%bpu? "?2P%`d~iG1o?,`!$%`d~iG1o?  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu#g)f !0n(P 27)?a מpb#ܞ &g*YF=`dbR ,.IeHԡ"|b@3X?ڜgKDd T,J C C AA? "B2Q ?Gv,IO$-`!% ?Gv,IO$ XJxcdd``> @c112BYL%bpudKF&&\v @ ] @Yv] `%Ki Dd hJ D C AB? "C2}ᡁ @c112BYL%bpu_En|Q:8[5' .-xQS"_W%ϡW5ܿEE~+| dGcaCJZű7;yeW'&u:Q<8qI M?G:x#LyO%Ci2/|~͸WT [ljt-f+Uf˲GWW@=S3Dd hJ G C AE? "F2e;Yro `!e;Yro`@0|xڥ/AƟm]_a8W^8!BJIDKҥI8r_H8H8ppHN v33γ CdG'誕#g̋wGE1 i]B%f!1Lg2:CtXӄ1ӉF V_j U_ pL*'Ry1Ifqr?CW往 Jq|E.9W(J6b[|x\e_ǃ|k8y]?z>k_% wVT>&5ǭJ.aR_Bϖ׻`.WBfq} k_\L^k$|`--#Wy1FX?w]SX4~!RZ-ʺ hxKa@DW ;%:IʎwFL+:Ugq;[٩ Hj5* cDd HTJ H C AF? "G2{޺gsν|!2W`!O޺gsν|!2 @ "XJxcdd``fed``baV d,FYzP1n:&&V! KA?H1Zz h7T obIFHeA*CL N`A $37X/\!(?71aǺXky`+ P.P16z[ oe!7#|FnO3ȄJ_lO = BܤApb#&&*YF=jP ,.Ie`ԡ"|b@3X? A`Dd hTJ I C AG? "H2Q[0ܥ-`!%[0ܥ@ |XJxcdd``> @c112BYL%bpu 1,Lx}DyK _Ref192582713Dd TTJ J C AH? "I2Pʘw^'[,R`!$ʘw^'[  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu 1,LxDd |J L C AJ? "K2cLE0Ie?/`!7LE0Ieኒ`0xcdd``> @c112BYL%bpudKF&&\k s: @> 1,M\Dd TTJ M C AK? "L2qe ']$M0`!Ee ']$  XJXJxcdd`` @c112BYL%bpu',@u@@ڈ4 Ho%ļ?m`bV 71i +ssj9|n=P{C_*a~y.h(rCl;- `p021)W2<ԡ"|b@33X?"fiDd HTJ R C AP? "Q2rwΗ# [$`!SrwΗ#  @ "XJ!xcdd``fed``baV d,FYzP1n:&&V! KA?H1Zz h7T obIFHeA*CL N`A $37X/\!(?71aǺXk@;:]b mQӷb^?cd60Z1Lp@4 {@AI9 \\e > t7F=.='_T 5.p2m-#RpeqIj.;=@0(2t5= d Dd TJ S C AQ? "R2}Ӌ E)^&`!V}Ӌ E)   dXJ$xcdd``fed``baV d,FYzP1n:&&! KA?H1Z ㆪaM,,He`I? 01d++&1X +8|-#T T0~e<@ `sC*a|=P@penR~Apb#Ԟ,F=fs j\e[ F&&\-= @ ]` "8et"Dd 4@J T C AR? "S2L>b`(`!XL>b p &xcdd``ed``baV d,FYzP1n:&&V! KA?H1:X ㆪaM,,He` 01d++&1X +8|-_$vh(.P16J`< V ۀMJ@Jt@^BܤpAp? ; d9=^`"l/=p{ݯW E.pH;7LLJ% W@2u(2t5= hDd TTJ U C AS? "T2Pʘw^'[,+`!$ʘw^'[  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu 1,LxDd ,@J W C AU? "V2bl)&!A)L>.`!6l)&!A)L xcdd``> @c112BYL%bpu @c112BYL%bpu @c112BYL%bpubiXt9FYΒJy 3D2 z.C``ÄI)$5\E. ,s;.Ff gMDd hJ  C A? "Z2PzUkނJFܢ[,6`!$zUkނJFܢ[@|xcdd``> @c112BYL%bpubҒYt9FYΒJy 3D2 z.C``ÄI)$5!d.P"CDHg!!v120ejMDd hJ  C A? "[2PzUkނJFܢ[,8`!$zUkނJFܢ[@|xcdd``> @c112BYL%bpubҒYt9FYΒJy 3D2 z.C``ÄI)$5!d.P"CDHg!!v120ejMDd hJ  C A? "\2PzUkނJFܢ[,:`!$zUkނJFܢ[@|xcdd``> @c112BYL%bpubҒYt9FYΒJy 3D2 z.C``ÄI)$5!d.P"CDHg!!v120ejM}DyK _Ref214350828Dd pJ  C A? "]2%enC\8)=`!enC\8< @Cxڥ+Q=xo捌)I{?@JY>5%7ˮDXelpkTKV7~DQ @c112BYL%bpudKF&&\v @ ] @Yv] `%KiDd ,hJ Z C AX? "_2Q-6[]T2-A`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd ,hJ [ C AY? "`2Q-6[]T2-C`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd @hJ \ C AZ? "a2PsV(. IbK#7?,E`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd @hJ ] C A[? "b2PsV(. IbK#7?,G`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eM}DyK _Ref192582713Dd ,hJ ^ C A\? "c2Q-6[]T2-J`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd @hJ _ C A]? "d2PsV(. IbK#7?,L`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd TTJ ` C A^? "e2Py̸B{I,M`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu= Bܤmn.pb#ܞ`&g+YF=`dbR ,.Ie82C D,Ā'f~eDd HTJ c C Aa? "h2{:!M8.mXWS`!O:!M8.mX @ "XJxcdd``fed``baV d,FYzP1n:&&V! KA?H1Zz h7T obIFHeA*CL N`A $37X/\!(?71aǺXkl\:]b mӷCoG ܞF+f !0؞ {@AI9 T|=p{@F䂆8dC``A)$5!d.P"CXYO`3-d'Dd TTJ d C Ab? "i2Py̸B{I,V`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu= Bܤmn.pb#ܞ`&g+YF=`dbR ,.Ie82C D,Ā'f~eDd ,TJ g C Ae? "l2Q3}#<@-[`!%3}#<@ XJxcdd``> @c112BYL%bpu @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd ,TJ i C Ag? "n2Q31~A+̣-_`!%31~A+̣ XJxcdd``> @c112BYL%bpudKF&&\M s: @> 1,MeDd ,hJ j C Ah? "o2Q-6[]T2-a`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd hJ k C Ai? "p2%ގb<\c`!T%ގb< @ |"xcdd``fed``baV d,FYzP1n:&B@?b = UXRY7&meabM-VK-WMcرsA VpZW m P.P16zӷƗ\ 1y32p{o2Lp |t{@AI9 a\L > F=L #NWrAC `!v 0y{qĤ\Y\0d.P"CXYO`Kg^Dd ,TJ l C Aj? "q2Q31~A+̣-e`!%31~A+̣ XJxcdd``> @c112BYL%bpudKF&&\M s: @> 1,MeDd ,hJ m C Ak? "r2Q-6[]T2-g`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd ,hJ n C Al? "s2Q-6[]T2-i`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd ,hJ o C Am? "t2Q-6[]T2-k`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd @hJ p C An? "u2PsV(. IbK#7?,m`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eM#Dd \ q c $Ao? ?#" `v2sƧ0vzeOo`!GƧ0vzehxcdd`` @c112BYL%bpu @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd |\ s c $Aq? ?#" `x2b D $ͭ=>s`!6 D $ͭ=`h0xcdd``> @c112BYL%bpu @c112BYL%bpudKF&&\u @ ] U`n ZMpDd @hJ  C A? "z2PsV(. IbK#7?,w`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eM#Dd \  c $A? ?#" `{2sƧ0vzeOy`!GƧ0vzehxcdd`` @c112BYL%bpu @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eM2Dd 0TJ  C A? "}2jy 6IX-%p}`!hjy 6IX-%ܚ kXJ6xcdd``ed``baV d,FYzP1n:&f! KA?H1Z gqC0&dT20ʹ_ @2@penR~G.k 1X +8|-ϖT T3[]Iȼ\Acd2jn ,L y%@(8g%2p'\F`s*i.{9AmG``[E.pLơ;6+KRsA2u(2t5B ~`Dd @hJ  C A? "~2Q`|g-=UIV-`!%`|g-=UIV@ |xcdd``> @c112BYL%bpub1iYt9F.NJy 3D2! z.C``ÄI)$53\E.B 0Cbd`N7Dd HhJ  C A? "2}h!|>lT&JY`!Qh!|>lT&J @@"|xcdd``fed``baV d,FYzP1n:&&V! KA?H1Z h7T obIFHeA*CL N`A $37X/\!(?71aǺXk~:]b mџӹb!7#|FnOc3JOð@(\˵N@a#2\%4Ը!6b;LLJ% s:@Dg!t?1 d2Dd 0TJ  C A? "2S6D*0p̃`!hS6D*0 kXJ6xcdd``ed``baV d,FYzP1n:&f! KA?H1Z gqC0&dT20ʹ_ @2@penR~5!>b5i?X u9~{+ y ߘ@FBMπ07?/̲D5ל1l3\2t/4  #|+i8``#RpeqIj. @ ]` bg!t?0EDd TTJ  C A? "2Py̸B{I,`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu?i` 2ob0! W&00ڽN@aۓ2\%4Ը!6b;LLJ%  s:@ĞB ~bgf Dd hTJ  C A? "2Pg`Aq}!,`!$g`Aq}!@ |XJxcdd``> @c112BYL%bpu @c112BYL%bpubiZ,@u@@ڈ XgX%ļ?]2! z.C``ÄI)$5A d.P"CD|bb#3X(K}DyK _Ref214443964;Dd l\  c $A? ?#" `2ɺ4R+Cv{g`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E'sDd ThJ  C A? "2Q\IXfiW)-Ȕ`!%\IXfiW) @XJ|xcdd``> @c112BYL%bpub iwYt9;ΰJy 3DdBZ%\`2M% #RpeqIj.] @ ] U`n e7O#Dd \  c $A? ?#" `2sƧ0vzeO`!GƧ0vzehxcdd`` @c112BYL%bpu @c112BYL%bpub1iYt9F.NJy 3D2! z.C``ÄI)$53\E.B 0Cbd`N7Dd @hJ  C A? "2PsV(. IbK#7?,P`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd J t C Ar? "2)&Fw0߄]>`!&Fw0߄] xڥJPϽgRdj!*EtpR łCڎb6kI|8w(_@|AU-Ū so0#$"aQ E=.AB*}0-0mxT^5!s՝o\27윲Ω(&g8m=RTh\j7dhvm*o{H,܁W/}'`z}n)Xp\zCBѨhkFM0>9 n-+xϹGw#ZЉssҁ/N\7Kf|-a+ 5%[soq'po߷'y@>W)2L"7Okrz3aU67c@,vv<[0YddڪT")8p_ G{{#HDd J u C As? "2"9XeK}>-`!9XeK}>- _xڥK@Ɵ~N*QD\MEѡZXN Сn._OPb ǽo{ A(@ :XZD Qhtw:B3b,w/F]ntq5lf3NX&܊Fn^JkNX0qLw'I@w[Tǭ9bWq+724"zK8bNqZ.S03%sΥ7\ rBܑ 2:uF0t"ȼQhS~ȀйW=uWȸoCb% Ψ󅵭4_ 7w:NVxxeħB\qN-uuuqㅌQDVBi)39Lq}8X[zq gDd TTJ v C At? "2Py̸B{I,Ũ`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu= Bܤmn.pb#ܞ`&g+YF=`dbR ,.Ie82C D,Ā'f~e$Dd hJ z C Ax? "2BqTHb`!ZBqTH@L |(xcdd``ed``baV d,FYzP1n:&&6! KA?H1Zʞ ㆪaM,,He`HI? 01d++&1X +8|-.NHq=`dt0Bg>#шd[%o61 P@penR~pb#ܞkL #c*iFm=adbR ,.Ie C D|b@#3X?IhDd phJ { C Ay? "2RQl(.IS_q0]`!URQl(.IS_q0 @|#xcdd``fed``baV d,FYzP1n:&V! KA?H1Z@eπqC0&dT20$ͤ `[YB2sSRsv\~J+f*F\#Y@:}+a|.IΕg>#d[%a B Ma`.UN@a#2\%4Ը!6b;LLJ% s:@ĞB ~b eIDd 8hJ | C Az? "23xj+Ps6`!3xj+Ps@v |Mxcdd``ed``baV d,FYzP1n:&lB@?b  ㆪaM,,He`Hc&,eBܤN! v\~?J;T TU@Z+ HKy >L #j*g eZZӝs>` 2Ư yq J;80e > F-fg+N `!v 0y{iI)$5(\E.Gv$Dd hJ } C A{? "2'ߘtxУ7b`!Z'ߘtxУ7@L |(xcdd``ed``baV d,FYzP1n:&&6! KA?H1Zʞ ㆪaM,,He`HI? 01d++&1X +8|-J{T T0rt:WG!| 3Dh2Ʒa(P 27)?ZkG!Ȉsp{@|J.h(rClnv0o8n+KRs^aPdk1 /gdDd TJ ~ C A|? "2׏<^tQlƶZ`!׏<^tQlƶZJ  XJhxcdd``^ @c112BYL%bpun@$s@FB=\! %E B\G>v'\FLZ #<& ZZG{nLp@b5u{сd/{cćVV0܋%=f9QQLoȯCjj2&m$pD;8eKxײڟO]5E n6M򙒺+z^+͓sFѮ"CA>Qc*:-uU'Lыܶ]3:ֺNǺuCwro?-t*:&e|σ 8)W{pҟ=ӭ6޻<NhAECYa7[e/8Wgq nʙ<$)Ȋq8 HK5}`݃|yDd TTJ  C A? "2Py̸B{I,`!$y̸B{I  XJXJxcdd``> @c112BYL%bpu @c112BYL%bpu @c112BYL%bpubnUsi#/ ^ 1y ,L y%@;3?m``[u=؇`cn``ÄI)$5\E.B 0Cbd`XP$Dd hJ  C A? "2'ߘtxУ7b`!Z'ߘtxУ7@L |(xcdd``ed``baV d,FYzP1n:&&6! KA?H1Zʞ ㆪaM,,He`HI? 01d++&1X +8|-J{T T0rt:WG!| 3Dh2Ʒa(P 27)?ZkG!Ȉsp{@|J.h(rClnv0o8n+KRs^aPdk1 /gJDd 8hJ  C A? "2gD/b`xSY`!gD/b`xSY@v |Nxcdd``ed``baV d,FYzP1n:&lB@?b  ㆪaM,,He`Hc&,eBܤ\rv\~Jb*F\u Օ̻+e 1SoDd phJ  C A? "2RQl(.IS_q0]`!URQl(.IS_q0 @|#xcdd``fed``baV d,FYzP1n:&V! KA?H1Z@eπqC0&dT20$ͤ `[YB2sSRsv\~J+f*F\#Y@:}+a|.IΕg>#d[%a B Ma`.UN@a#2\%4Ը!6b;LLJ% s:@ĞB ~b e8Dd @J  C A? "2[> Bv2`!n[> BL <xcdd``ed``baV d,FYzP1n:&lB@?b  UXRY`7S&meabM-VK-WMcرsA VpZ/? P.P16Rcft@y w Y@&AmG;!2B]2\y%0Rp/*,Kd ʽ q )ܤl\> ޷`"8΅ơ;6+KRsePdk1 -fDd HhJ  C A? "2}:+AHD(Yj`!Q:+AHD( @@"|xcdd``fed``baV d,FYzP1n:&&V! KA?H1Z h7T obIFHeA*CL N`A $37X/\!(?71aǺXky`c:]b mџӹb!7#|FnOc3JOð@(\ZKiG!G dĹJ.hqCl:.v0o8.+KRsePdk{> 1fQDd phJ  C A? "2`h޼F22O!]`!U`h޼F22O! @|#xcdd``fed``baV d,FYzP1n:&V! KA?H1Z@eπqC0&dT20$ͤ `[YB2sSRsv\~?Jf*F\#Y@:}+a|.IΕg>#d[%a B Ma`xN@a#2|%4Ը!6b;LLJ% s:@ĞB ~b e7Dd @J  C A? "2;05X7fu`!m;05X7fL ;xcdd``ed``baV d,FYzP1n:&lB@?b  UXRY`7S&meabM-VK-WMcرsA VpZS ͍NHq1ʰt:WBS f9QQLoȯCjj2&m$pD;8eKxײڟO]5E n6M򙒺+z^+͓sFѮ"CA>Qc*:-uU'Lыܶ]3:ֺNǺuCwro?-t*:&e|σ 8)W{pҟ=ӭ6޻<NhAECYa7[e/8Wgq nʙ<$)Ȋq8 HK5}`݃|yDd ,hJ  C A? "2Q-6[]T2-`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMpDd @hJ  C A? "2PsV(. IbK#7?,`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd @hJ  C A? "2PsV(. IbK#7?,`!$sV(. IbK#7?@ |xcdd``> @c112BYL%bpubYt9FYNJy 3D2 z.C``ÄI)$5d.P"CDHg!!v120eMDd ,hJ  C A? "2Q-6[]T2-n`!%-6[]T2@|xcdd``> @c112BYL%bpudKF&&\u @ ] U`n ZMp#Dd \  c $A? ?#" `2sƧ0vzeO]`!GƧ0vzehxcdd`` @c112BYL%bpu)cRlBWQpQ #U.l52E! F-<5Σk0 fq*d2SˎϷ.*QKF.[^G};Rq6AH"ux oydXhR%9q?u Ulvwh-~h^do䘫cL9|W_%ȻQ. w9lq rf%+9u0yeLHV7 B"u#Dd \  c $A? ?#" `2sƧ0vzeO`!GƧ0vzehxcdd`` @c112BYL%bpuaE2d~, .@2R xcdd`` @c112BYL%bpu' Ѧٞ|ۙ)`!' Ѧٞ|ۙ):`!Rxcdd`` @c112BYL%bpul Z<  xcdd`` @c112BYL%bpu 1,NyDd X,\  c $A? ?#" `2i?/I?]}E`!=?/I?]} xcdd`` @c112BYL%bpu @ A B C D E F G H I J K L O P Q R S T U V W X Y Z [ \ ] ^ _ ` a b c d e f g h i j k l m n o p q r s t u v w x y z { | } ~  yK _Ref215812860}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215204016}DyK _Ref215196170}DyK _Ref190686481}DyK _Ref190686436}DyK _Ref190686481}DyK _Ref190686503}DyK _Ref190686481}DyK _Ref215820051}DyK _Ref215196170}DyK _Ref215240305}DyK _Ref215196170}DyK _Ref215290439}DyK _Ref215196170}DyK _Ref215196170}DyK _Ref215290439}DyK _Ref215290439}DyK _Ref215290439}DyK _Ref215290439}DyK _Ref215290439}DyK _Ref215196170}DyK _Ref215296727}DyK _Ref215296727Dd b V c $A? ?3"`?2>rLf+TYw&. 1`!rLf+TYw&.:Rxcdd`` @c112BYL%bpuĄ`=[9XX!g..  d  xcdd`` @c112BYL%bpuz08- j`!%k>z08ǒ@|xcdd``> @c112BYL%bpuz08-k`!%k>z08ǒ@|xcdd``> @c112BYL%bpu qEm`!=жc18Z%.4> q` 0 xcdd``$d@9`,&FF(`T)IRcgbR Al@P5< %!  @_@YN`b]F"L 0L !*'] ZZǰc@𵼇#b&#.#.6&ΌJy 3 ps@&UrA} F=cdbR ,.Iehԡ"b> 1 [ MDd @|J ] C A? "2iжc18Z%.4> qEo`!=жc18Z%.4> q` 0 xcdd``$d@9`,&FF(`T)IRcgbR Al@P5< %!  @_@YN`b]F"L 0L !*'] ZZǰc@𵼇#b&#.#.6&ΌJy 3 ps@&UrA} F=cdbR ,.Iehԡ"b> 1 [ M}DyK _Ref191198647}DyK _Ref215471559}DyK _Ref215471559}DyK _Ref191198647}DyK _Ref215484430}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215196170}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215541334}DyK _Ref215571062}DyK _Ref215571062}DyK _Ref215574539}DyK _Ref190686503}DyK _Ref191364304}DyK _Ref190686436}DyK _Ref190686481}DyK _Ref217137097}DyK _Ref191979081}DyK _Ref215204016}DyK _Ref215204016}DyK _Ref191979081}DyK _Ref215980624}DyK _Ref190768256}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215319995}DyK _Ref215319995}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref215989076}DyK _Ref190768256}DyK _Ref215989076}DyK _Ref190771388}DyK _Ref190768256}DyK _Ref213230715}DyK _Ref213230715}DyK _Ref213230715}DyK _Ref217137149}DyK _Ref216410304}DyK _Ref216410304}DyK _Ref216410304}DyK _Ref190768256}DyK _Ref190771388}DyK _Ref190768256}DyK _Ref217137370}DyK _Ref215396639}DyK _Ref216322123}DyK _Ref216323373}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref215319995}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref216326495}DyK Ole CompObjfObjInfoEquation Native 6 FMicrosoft Equation 3.0 DS Equation Equation.39q, PC FMicrosoft Equation 3.0 DS Equation Equation.39q_1266665338FU W`b W`Ole CompObjfObjInfoEquation Native _1266675225Fb W`b W`Ole CompObj fPv 2PC+E=2)PC())PE())PC()+)PE() FMicrosoft Equation 3.0 DS Equation Equation.39qObjInfo!Equation Native  _1266737496H$Fb W`b W`Ole #ŒF$ log 2 )PV)CV())EV()() FMicrosoft Equation 3.0 DS Equation Equation.39qF P 2 )CompObj#%$fObjInfo&&Equation Native '/_12667531551;)Fb W`b W`ECV+E"P() 2 )EV"C()V+C"P() 2 )CV"E()V+V"E"C+P() 2 )V"E()V"C()V"V  ed !   12Ole 0CompObj(*1fObjInfo+3Equation Native 4D FMicrosoft Equation 3.0 DS Equation Equation.39q(E$ N0,1() FMicrosoft Equation 3.0 DS Eq_1266753003.Fb W`b W`Ole 6CompObj-/7fObjInfo09uation Equation.39qs )PV()")CV()E/V() ))PV1")PV()()V  FMicrosoft Equation 3.0 DS EqEquation Native :_12667422106,3Fb W`b W`Ole >CompObj24?fuation Equation.39q`Fܙ 8"2log e   EV()  P  1 "EV ( )  C " P EV()  E " P  1 "  E V ( )@ ObjInfo5AEquation Native B_12667409168FW`W`Ole Q@@ V " E " C + P   P C ( )8 888P 88818"888P88C8(8)8 888C8"8P   E " P V " C ( )  E " P 1"E"PV"C() pVp"pEp"pCp+pP ( ) FMicrosoft Equation 3.0 DS Equation Equation.39q Ȋ< V !FMicrosoft Excel ChartBiff8   _Ref215319995}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref216326495}DyK _Ref191979081}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref215319995}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216448569}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216326495}DyK _Ref216925536}DyK _Ref216925572}DyK _Ref216858117}DyK _Ref216925572}DyK _Ref216858117}DyK _Ref216925677}DyK _Ref216925677}DyK _Ref190686436}DyK _Ref216926683}DyK _Ref216926690}DyK _Ref216707798}DyK _Ref216326495}DyK _Ref216707798}DyK _Ref216707798}DyK _Ref216707798}DyK _Ref216707798}DyK _Ref216707798}DyK _Ref216858117}DyK _Ref216707798}DyK _Ref216707798}DyK _Ref190686481}DyK _Ref190686436}DyK _Ref190686503}DyK _Ref190768256}DyK _Ref190771388}DyK _Ref190686436}DyK _Ref190686436}DyK _Ref191364304}DyK _Ref216169257}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref214085008}DyK _Ref215396639}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref215396639}DyK _Ref216925677}DyK _Ref216591034}DyK _Ref216175094}DyK _Ref216591034}DyK _Ref215980624}DyK _Ref216175094}DyK _Ref216591034}DyK _Ref216175094}DyK _Ref216175094}DyK _Ref216591034}DyK _Ref215980624}DyK _Ref190768256}DyK _Ref215980624}DyK _Ref215980624}DyK _Ref216591034}DyK _Ref216175094}DyK _Ref216926683}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216955967}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216952908}DyK _Ref216925677}DyK _Ref217017701}DyK _Ref216858117}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref216858117}DyK _Ref216926683}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref217013137}DyK _Ref216858117}DyK _Ref216926690}DyK _Ref217122670}DyK _Ref217290313}DyK _Ref217290313}DyK _Ref190686481}DyK _Ref190686436}DyK _Ref190686503}DyK _Ref190686481}DyK _Ref215840852}DyK _Ref190686436}DyK _Ref190768256}DyK _Ref190771388}DyK _Ref216858117}DyK _Ref195279807}DyK _Ref195279807}DyK _Ref190768256}DyK _Ref190686481}DyK _Ref215388958}DyK _Ref215388962}DyK _Ref215319787"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/ /  / /  / / / 449"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/  / / / / / / 449"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/ / / / / / / 449"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/ / / / / / / 449"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/ / / / / / / 449"$$If!vh5) 555555#v) #v#v#v#v:V t,5) 5555/ / / / / / / 449}DyK _Ref215980624$$If!vh5!#v!:V  t05!/  / 44 $$If!vh55s5555U#v#vs#v#v#v#vU:V t055s5555U/ 44 $$If!vh55s5555U#v#vs#v#v#v#vU:V t055s5555U44 $$If!vh55s5555U#v#vs#v#v#v#vU:V t0,55s5555U44 $$If!vh55s55555#v#vs#v#v#v#v#v:V t055s5555544 $$If!vh55s55555#v#vs#v#v#v#v#v:V t055s5555544 $$If!vh55s55555#v#vs#v#v#v#v#v:V t055s5555544 $$If!vh55s55555#v#vs#v#v#v#v#v:V t055s5555544 $$If!vh5>55#v>#v#v:V t05>5544 $$If!vh55858#v#v8:V h t,,558/  / / / 4 $$If!vh55858#v#v8:V 4w t,+,558/  / / /  / / / 4 f4$$If!vh55858#v#v8:V 4w t,+,558/ /  / /  / / 4 f4$$If!vh55858#v#v8:V  t,,558/ / / / / / / / 4 $$If!vh55858#v#v8:V  t,,558/  / / / / / / / 4 $$If!vh55858#v#v8:V h t,,558/  / / / 4 $$If!vh55858#v#v8:V 4w t,+,558/  / / /  / / / 4 f4$$If!vh55858#v#v8:V 4w t,+,558/ /  / /  / / 4 f4$$If!vh55858#v#v8:V  t,,558/ / / / / / / / 4 $$If!vh55858#v#v8:V  t,,558/  / / / / / / / 4 {DyK  _Ref69705490}DyK _Ref192403460$$If!vh56555d#v6#v#vd:V t065 55d/  / / / / 4 $$If!vh56555d#v6#v#vd:V t065 55d/ / / / / / 4 $$If!vh56555d#v6#v#vd:V t065 55d/  / / / / / 4 $$If!vh56555d#v6#v#vd:V t065 55d/ / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/  / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/ / / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/  / / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/ / / / / 4 $$If!vh55y55c#v#vy#v#vc:V t065 55d/  / / / / 4  $$If!vh55y55c#v#vy#v#vc:V t065 55d/ / / / / / 4  $$If!vh55y55c#v#vy#v#vc:V t065 55d/  / / / / / 4 $$If!vh55y55c#v#vy#v#vc:V t065 55d/ / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/  / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/ / / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/  / / / / / 4 $$If!vh5 555d#v #v#vd:V t065 55d/ / / / / 4 $$If!vh55 55# #v#v #v#v# :V 4P tX2+++,55 55# /  / / / 4 f4<$$If!vh55 55595@5#v#v #v#v#v9#v@#v:V 4 tX2+++,55 55595@5/ / / / / / 4 f4 Dd D\  c $A? ?#" `2]|c۳ʓ O`Jy9W`!1|c۳ʓ O`JyvHxcdd``$d@9`,&FF(`TiF! KA?H1Z @zjx|K2B* R60pE1At 2Bazf Y wfZr؅,L ! ~ Ay +?Obdt f22 f`tJ. `S7#RpeqIj.|E.`v3`Alt$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4 tX2)v+,5 5b5 55595@5/ /  / / /  / / / / 4 f4Dd p\  c $A? ?#" `2n>[ceVwCJ`!B>[ceVwC* @2xcdd``ef``aaV d,FYzP1n:&^! KA?H1  pC0&dT20<`b e-v0ree`怨]8LZ.UNy +ss >1t\> JÈ+х Uy凉p v$s9AFhVn s)."nvUskR΃wl?,cf#!نh?@fIT??X?ǃ-)~ ;B;c 3$37X/\!(?71H0?Z pp>7; oaEbGU^Y6T0>##D=݂Ću;80 pwecb ?# !|.h.y0@``jF&&\5 {:A> 1$V$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4 tX2+,5 5b5 55595@5/  / / / /  / / 4 f4Dd \  c $A? ?#" `2W#zɳ=t3P `!+#zɳ=t ` Pxcdd``.bb``aaV d,FYzP1n:n! KA?H1Xzjx|K2B* R>0pEM`B]F"L O@,a7lv@V3, fw{L@(\t%:V~. _+ 05 ҪQI<Ic)l #4+an]ov]LpIa6'@+Fh\i00Cq qC\n`F!Ȉ~>*5T?G忆ʿU$GHfnj_jBP~nbdc sȄJߌwYQ埃ӥ3?>,y`Pw*G o܉7$7dln)3-e`m `p)Ĥ\Y\2C dօB `kd$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4 tX2+,5 5b5 55595@5/  / / / / / / / 4 f40Dd \  c $A? ?#" `2|,`/ w*EkW\`!T|,`/ w*EkWR "`@{5P"xڥV_HSQsuiSk3) K")"JUBKIADmH=E%$cJ{^ bsumֹtDžwwv9$a B Ћ A#VE5*~R  %L 2zHh=YH$stI&VE扚 i1~(KVlpetn'n\Ri3T9 o~ޘCphiIĪ:S<>z]3JTb_|j;#{,wFȻ4|>HF>C5c%>=/wq;tܹ\vri(3r|9.s7ĝˤ\v=i'xM;~u{}&d죾u}RlE[]3wP'za~ "N&y>1~_G%ZR﫸Y.6dkSBXFOl SA]GM;P16tPKz2+ xX l}0׳hXݟ]|ҬE9f~-X}Efq]/{943Cy/Ϣ_uu_ [_Pr`v~=fb\|އkDcXr3΢Z?Pςl 6?+9t;=,,ӭPhJ3-¸;56HD uV'PB|"#j$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4d tX2)v+,5 5b5 55595@5/  / / /  / / / / 4 f4DDd l@\  c $A? ?#" `2/Bv^MϊЉpY`!h/Bv^MϊЉb 6xcdd``6dd``baV d,FYzP1n:&B@?b 030 UXRY7ֿ `7LY 4A 27)?L+!'i,@@@ڈ+1^JBmZq:fQNj;H{ pjNxゅSNX@<ϼ֕LZfXyTQY՜J9JF冈`t7$c,pM-VK-WM,h2!ÎN̨LQ(E1.#W8bBޜwyP0@G7/tH(Y^Wnif=!p~3s + [[+KRs8@0u(2t5|b@#3X?}YT$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4 tX2+,,,5 5b5 55595@5/  / / /  / / 4 f4Dd Tx\  c $A? ?#" `2_;:ݔv[]KV`!_;:ݔv[]K "04 xڥkAgg_GlmҦM""8('[/JQRo -xDTrQ/TzQVQ0ogg~Ąo6d!D<>D(_QE V KXXYqIJ^1I<殦kDa5qox9Zl1ͽ:жb*㮞FxmiI;kح ]Tb5 CompObj79RfObjInfo:TEquation Native U)_1287573602M=!F@W`0W`6i,2y'c#q{-6>w ͭsjt:KEoʫލ7TT'=wE.6zͽfG/@fE~w -:*fbc?Pw^C臹EXP} x&~pdyy Ϸ ge x&6{3QCL}-2c<X̳8yx5i3??~w}4~:~#}0gA~./LXkt~i-jW`gi,~jD~1,~~~D~&~W|U.>ygڡ<y_;Uؘ?7\Pze[;0߷Ӌj>{UCjȼ2S ) sSȏ|ڟ{w͸8 LdyLeUlmyOpr>R=.;v/=i}mV(Z9Jp$p?M1d6 4;rN=cop%p_K'I<^\ 0_G绫+VDjh/A L"O7R҅=pBIQ"O74>E7H$$If!vh5 5b5 55595@5#v #vb#v #v#v#v9#v@#v:V 4 tX2+,5 5b5 55595@5/  / / /  / / 4 f4$$If!vh5 5 5 5 #v :V t065 /  / / / / 4 $$If!vh5 5 5 5 #v :V t065 / / / / / / 4 $$If!vh5 5 5 5 #v :V t065 /  / / / / / 4 $$If!vh5 5 5 5 #v :V t065 / / / / / 4 Dd \  c $A? ?#" `29V,8yG%`! V,8yG:`!x=N=a}3F%X 'p DMBH@T@u DTk|˼yC#F%1&qTfٿ.'QAX-G Ö*&pd'F:t,]/>55ΰ3]9]wz o<7I/.R[VJ0a`E)^Ba#Eo'j?X]0 vj0ˣoiOL9}DyK _Ref192925307}DyK _Ref192925307Dd 0  # A2 ϻKw#45?HM^(`!ϻKw#45?HM^ 1hdQ<xWkAovLZ"RK@A[^ '^{* zB^'^ o*nvn 7Pp@msPެǎ^U0T~/V[ft=/iʆM/xjRn$:4X\u?bیx~.y7&YM&_ ++J-<}]aadʡ|Y$y|rͯV"VL{R_bHںg 7=OVUh%O|}]̰;ZY_z 3 5u>@h1؎`; v]X%YUj"uޥH䦞Ӻl %؉`'mJ6% 6,6x[VgpAԡHCi,-D~>Sù5l Sp^Ce™jCs_h'3y8ۇ d:ϐOϑKɸcYQ=o0؝~λt8wXvF_u~%"3Mg 5'1S? jT`\pFbqhB/c|fL&445Pg>9'oO=;czL6GTxKD Rc$$If!v h55}55550555 5  #v#v}#v#v0#v#v#v #v :Vb4 t3)v)v)v)v ++++++++ , 55}550555 5  /  / / / /  / / /  / / /  / / /  / 4 af4pFkd,$$If4 =Z w v$&2}0  t3((((4af4pF;$$If!vh55!5!555555 5 5 5 5 5 5 #v#v!#v#v#v #v #v #v #v :Vb4> t3)v)v)v )v +++++++++ + + + + +,,,55!555 5 5 5 5 /  / / /  / / / / / /  / /  / / / / / / / / 4 af4pxkd0$$If4>N =Z w v$&2`!!  t3<<<<4af4px$$If!vh55!5!5 5 5 555 5 5 5 5 5 555 #v#v!#v #v #v #v #v#v#v :Vb4~ t3)v)v)v+++++++ + + + + +++,,,55!5 5 5 5 555 / / / / /  / / /  / / /  / / /  / / / / /  / / /  / / / 4 af4p֌gkd6$$If4~z  =Z w v$&2!!     t3DDDD4af4p֌e$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb t3,55!5 5 5 5 55>5 /  / / / /  / / /  / /  / / / /  / / / / /  / / /  / / / 2 q2q4 ap֪kd=$$Ifz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֪-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kdD$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kdK$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kdgR$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪s$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / / / / / / / / / / / /  / /  / / / / / / / / / / 2 q2q4 ap֪kd)Y$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֪U$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / / /  / / / / / /  / /  / / / / / /  / / / / 2 q2q4 ap֌skd1`$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֌$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skdf$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kdm$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skdEt$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skdz$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌U$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / / / / / / / / / / / /  / /  / / / / / / / / / / 2 q2q4 ap֌skdQ$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֌U$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / / /  / / / / / /  / /  / / / / / /  / / / / 2 q2q4 ap֌skd$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֌-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kd$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kd$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / / Ole VEPRINT<? <CompObjWbObjInfo>AYlCC5. EMF<A'3\KhCF, EMF+@F\PEMF+"@ @ $@ !@ 0@?@     !" !" !  Rp Helveticax%0&xp[wd&%0%%0Arial%&lww \%%I0dv% RpQArialx%0&xp[w%%0%% 0Arial%&lwws \%%I0dv% % % " !%   G % '%   +CC% % % " !% % %   G &% % % % " !% % %      6 I 6I 666% % % " !% % %   G   6S6S 6 SI 6I S6S6S6666 6 < 6< r6r66% % % " !% % %   CC% % % " !% % %   X$% % % " !% % %   X!% ( % ( 6 '%   V,, r, ^ r @h  V,K  s K d  X _ V,U  U   i    % %  V,v =  3v =  8 [ V,6 j L 6 T j E f V,` * `   k*   V, \  Q \  W  V,z  z    % % V, w   y  w    \ V, 8 k  M 8 V k  G ;  V,5 | , 5  q |  @ , w   V, 1  & 1 ,  V, y  y  < h < h % % \ n V,Y ^ _ ^ i Y ~  y  V,  [   [ U  X  V,       V,       P V,M u S u o M r r % % V,m  x   m  , V, IQ   I3 >Q ; CB l V,]  ] s  {  q V,l  v   l  ' V," i - i ^ " c   % 6 '% &% (  V, F6 y==6 % % % " !% % %   X!   V,]  f   f % % % " !% % %   X!   V,= R  F I   F % % % " !% % %   X!< h   V,  < %  < % % % % " !% % %   X!r   V,& D r ; /; r % % % " !% % %   X!   V,[ e  \ d\  % % % " !% % %   X!% % % " !% % %   X$% % % " !% % %   G &% (   66% 6% I 6%I 6%6%6%% % % " !% % %   CC% % % " !% % %   P,% % % " !% % %   P)% ( & % ( 6   6 6 6< 6r6% ( 6 '% V,)  )  . B %   V,       % % V,<<YK$V,+^sV,W Wd ,e  % % V, + . ( + + .*  V,  !    !  V,       < < % % h V,g H g G H h H  V,   (V,(r(rr(rr% % V,ootV,mc^mca}gV,5wW5UwEE% 6 '% &% (  V0 q 6  6h  6 % % % " !% % %   P)   V0}1 % :    : % % % " !% % %   P)   V0<Z  EQ   E% % % " !% % %   P)<   V0  A<  < 8 < % % % " !% % %   P)r  V0_rhr'rh% % % " !% % %   P)  V0ScZ\% % % " !% % %   P)6   + o % % % " !% % %   P)  + % % % " !% % %   P)   +? W% % % " !% % %   P)<   + u % % % " !% % %   P)r  +8 % % % " !% % %   P)E  +m ~% % % " !% % %   P)% % % " !% % %   P,% % % " !% % %   G % % % " !% % %   G % % % " !% % %   G % % % " !% % %   G   TThIK@E@hILP0TX@E@LP20TX  @E@ LP40TX @E@LP60TX?A@E@?LP80T`d}@E@d}LT100% % % " !% % %   G % % % " !% % %   G     TTZ@E@ZLP0Tl@E@LX10000TlJ @E@LX20000Tl  @E@ LX30000Tl-@E@-LX40000Tlb@E@bLX50000Tl!@E@LX60000% % % " !% % %   G % % % " !% % %   G     TXIK@E@ILP0%T``@E@LT20%T` ` @E@ LT40%T`` @E@LT60%T`?`A@E@?LT80%Td}@E@}LT100%% % % " !% % %   G % % % " !% % %   G % ( % ( % " G !  " !  ( ( " F4(EMF+*@$??FEMF+@ Excel.Chart.89q Oh+'0@Hd Christian MonsoneChristian MonsoneMicrosoft Excel@_w;@ A՜.+,0 PX|  Workbook 4SummaryInformation(@BZDocumentSummaryInformation8^D_1266406405EFRW`RW`             GH           ! " # $ % & ' ( ) * + , - . / 0 1 2 3 4 5 6 7 8 9 : ; < = > ? @ A B C D E F G H I J K L M N O P Q R S U V W X Y Z [ \ ] ^ _ ` a b c d e f g h i j k l m n o p q r s t u v w x y z { | } ~  F\pChristian Monson Ba=@= \<X@"1Arial1Arial1Arial1Arial1Arial1Arial1Arial1Arial1Arial1" Helvetica1Arial1Arial1" Helvetica"$"#,##0_);\("$"#,##0\)!"$"#,##0_);[Red]\("$"#,##0\)""$"#,##0.00_);\("$"#,##0.00\)'""$"#,##0.00_);[Red]\("$"#,##0.00\)7*2_("$"* #,##0_);_("$"* \(#,##0\);_("$"* "-"_);_(@_).))_(* #,##0_);_(* \(#,##0\);_(* "-"_);_(@_)?,:_("$"* #,##0.00_);_("$"* \(#,##0.00\);_("$"* "-"??_);_(@_)6+1_(* #,##0.00_);_(* \(#,##0.00\);_(* "-"??_);_(@_) #/1000\K                + ) , *  `WChart2Sheet12Sheet23Sheet3TZR3 A@@    FMPrinter_to_FileS odLetterPRIV0''''\KhC7"d??3` / ` / GS3d23 M NM4 3Q "Morpheme RecallQ ;Q ;Q3_  NM  4E4 3Q 6Achievable MorphemeRecallQ ;Q ;Q3_  NM  4E4 3Q  # of ClustersQ ;Q ;Q3_  NM  4E4D$% M 3O& Q4$% M 3O& Q4FA x 3O3*@N & 43*Y@4@N#M& ! M43d" 44A x 3O3*! M43*??N & 43d" 444e@@@@@@@@@L@L@L@@@@j@j@j@ebX9?I +??@V-? +?C@A`"??F@MbX9?jt?I@S㥛?&1?G@V-? ףp= ?D@e>  F  dMbP?_*+%"??U~ @bX9?I +?~ ?@~ @V-? +?~ C@~ @A`"?P@F@~ L@MbX9?jt?~ I@~ @S㥛?&1?~ G@~ j@V-?R@D@d@@6@@2(  j  0NMM?  Z]`  FMPrinter_to_FileS odLetterPRIV0''''\KhC7"d??3` /` /mg3d23 M NM4 3Q RecallQ ;Q ;Q3_4E4 3Q &Recall AchievableQ ;Q ;Q3_4E4 3Q ClustersQ ;Q ;Q3_ 2 NM  22d4E4D $% M 3O&Q4$% M 3O&Q4FA; 3OW+ t3*#M43*??N #M! M43d"  3O % M3OQ4444A; 3O3*! M43*Y@4@43d" 444e@@@@@@L@L@L@@@@j@j@j@eV-? +?C@A`"??F@MbX9?jt?I@S㥛?&1?G@V-? ףp= ?D@e xp  6NMM?` 1]   FMPrinter_to_FileS odLetterPRIV0''''\KhC7"d??3` /` / zG3d 23 M NM4 3Q "Morpheme RecallQ ;Q ;Q3_  NM  4E4 3Q 6Achievable MorphemeRecallQ ;Q ;Q3_  NM  4E4 3Q  # of ClustersQ ;Q ;Q3_  NM  4E4D$% M 3O&Q4$% M 3O& Q4FA/ 3O3*#M& 43*Y@4@N#M& ! M43d" 44A/ 3O3*! M43*??N & 43d" 444eee >@7 F  dMbP?_*+%"??sU>@7 F  dMbP?_*+%"??sU>@7 Carnegie Mellon University Sheet1Sheet2Sheet3Chart2  WorksheetsCharts FMicrosoft Equation 3.0 DS Equation Equation.39qOle dCompObjDFefObjInfoGgEquation Native hfJ< Monotype Corsiva e~M {f V } FMicrosoft Equation 3.0 DS Equation Equation.39qZx.\ f V_1288525413'JFRW`RW`Ole jCompObjIKkfObjInfoLmEquation Native n6_1288525392OFRW`RW`Ole oCompObjNPpf FMicrosoft Equation 3.0 DS Equation Equation.39qZx.\ f VOh+'0 $0 L X d p| PARAMOR: ARAChristiaObjInfoQrEquation Native s61TableP ڼ SummaryInformation(Tt / / / / /  / / / / / / /  / / 2 q2q4 ap֌skdm$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skd$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌U$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / / / / / / / / / / / /  / /  / / / / / / / / / / 2 q2q4 ap֌skdy$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֌U$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / / /  / / / / / /  / /  / / / / / /  / / / / 2 q2q4 ap֌skdE$$IfFz  =Z w v$&2!!   >  t3DDDD2 q2q4ap֌-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :VbF t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kd$$IfFz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb3 t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kdӽ$$If3z  =Z w v$&2!!   > t3DDDD2 q2q4ap֪$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb3 t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skd$$If3z  =Z w v$&2!!   > t3DDDD2 q2q4ap֌$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb3 t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skd$$If3z  =Z w v$&2!!   > t3DDDD2 q2q4ap֌$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֌skd$$Ifz  =Z w v$&2!!   > t3DDDD2 q2q4ap֌-$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb t3,55!5 5 5 5 55>5 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 ap֪kd'$$Ifz  =Z w v$&2!!   > t3DDDD2 q2q4ap֪$$If!vh55!5!5 5 5 555 5 5 5 5 5 55>5 #v#v!#v #v #v #v #v#v>#v :Vb t3,55!5 5 5 5 55>5 4 ap֪kd$$Ifz  =Z w v$&2!!   > t3DDDD4ap֪c Dd Zb r c $A? ?3"`? " ʆ,I:=$ @= ʆ,I:=$<HKAO x[{pU߽9ፂNJR20v0))) IxT0-B"<HM'3ն%< ߷{n͹)/q{ݻgw\) g>!ʤ|b_! !_qlk~w}KvgG|' uA*E'ۀdvَ gRъ:O7]SI>(.o*Dlwv]]h֠{#\P3}Aۺp<6f 3v8Hۋ/6.# hit Hc}!!)8yd7k7i?~Ϸ71.W:ܓ<]zW/}#¨ě-~d'YK%oМ}n9#p.Ֆ#.?EP1L%~*~$&~ډ봦O 3| RK6- ,:{ ]rGLHg8%8kQ.oʺZHbCd.ڗl |:C9#B-BZ|Ho߮}n`N{?Le\W@}憁* h#c:JvMF\C}檁Tc`Ig>=W0d[fH/|b`]4J@ռ7_bqA살;!5 B#ͷ on(Y#@q4nX堳QsbRK6U?@g@2d?Dll>6NX@1fc8G<%_VpP i5, վp"﹣uE_:р/0va-H<:vEl* 죱 t#_9b`%J`=+AGXlX]eue YW`@;8g~ҽ'4ÌSd'Dd'j /1NE>oﵾS8@')1@Qq4Kw]i1L뀰cCeߡ]51FKw1Zd5h~ 'ntS'/(@'^tKww2U7\1~.C:S珣d v\!"4, owJFȫHN/mG:3 WV-Fy$tՐ8_Ȅd jߩ\։P1kTv.kGjMz|YgnS_6$9@Ԏ|k|;35_]j5>Z?1kI$Ԏk||֥`ޔs.Uj5Y/ϗ| MՎB_t !FVMulT"Cka>қr^"CVD˭:x {TȜï>fws$H_pM3R}L:E"ZIwhuu "nNi"z&4kv25IL5F5] |;Ϫ/5P_h{GM!Ix~ODG2!MԺt㆞kL:[Np I.!D,Q^cz'XKt/ci,З;QWjl_"=7}ݾDSDIY7jŇiZGY[uιJKT>vWC1 -s[|;n1qMSfDM uq NOyUÍF,%wm`Z=uYOtߌga0.|ӿ-y7cxz%X1cb<{W;֩(c=UWP.{,L,;qMdabM2V(cm`#abm2VGz-6{XKäq| ݳjs_hkKUY:BηO*LJ_٫u^?5 ~F}#Of|:oog9Q5س/lDcs~^ո=uZnF} Uc?3CF׽6j9O;Dd l\  c $A? ?#" ` 2ɺ4R+Cv{g`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E'sDd ThJ  C A? " 2Q\IXfiW)-C`!%\IXfiW) @XJ|xcdd``> @c112BYL%bpub iwYt9;ΰJy 3DdBZ%\`2M% #RpeqIj.] @ ] U`n e7ODd ThJ  C A? " 2Q\IXfiW)-2`!%\IXfiW) @XJ|xcdd``> @c112BYL%bpub iwYt9;ΰJy 3DdBZ%\`2M% #RpeqIj.] @ ] U`n e7O;Dd l\  c $A? ?#" ` 2ɺ4R+Cv{g!`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E'sDd ThJ  C A? " 2Q\IXfiW)-\`!%\IXfiW) @XJ|xcdd``> @c112BYL%bpub iwYt9;ΰJy 3DdBZ%\`2M% #RpeqIj.] @ ] U`n e7O}DyK _Ref193859274;Dd l\  c $A? ?#" `2ɺ4R+Cv{g`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E'sDd ThJ  C A? "2Q\IXfiW)-`!%\IXfiW) @XJ|xcdd``> @c112BYL%bpub iwYt9;ΰJy 3DdBZ%\`2M% #RpeqIj.] @ ] U`n e7O;Dd l\  c $A? ?#" `2ɺ4R+Cv{g`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E'sDd ThJ  C A? "2Q4S/<wp9"--`!%4S/<wp9" @XJ|xcdd``> @c112BYL%bpubiZ,@u@@ڈ XgX%ļ?]2! z.C``ÄI)$5A d.P"CD|bb#3X(K}DyK _Ref214443964;Dd l\  c $A? ?#" `2ɺ4R+Cv{g`!_ɺ4R+Cv{Rh-xcdd``dd``baV d,FYzP1n:B@?b 030sC0&dT20ͤ KXB2sSRs:V~. _ӦUsi#&F'fJJ pI2DǸ dBX%It7\FUAmfo~^~ 03rJ.hscа``㔑I)$5a\E.B ~`E's}DyK _Ref215196170$$If!v h55555u555V5 5 5  #v#v#v#v#vu#v#v#vV#v #v #v :Vb4 t2)v)v)v)v ++++++++ + , 55555u555V5 5 5  /  / / /  / / / / /  / / /  / / / /  /  / / 4 f4pPGkd $$If4 rjX y#h%2uV  t2,,,,4af4pP$$If!vh5555555:55 5 5 5 5 V5 55 #v#v#v#v#v#v#v:#v #v V#v #v#v :Vb4` t2)v)v)v )v+++++++++ + + + + ++,,,5555555:5 5 V5 55 /  / / /  / / / / /  / /  / / /  / / /  / / / / 4 f4pւ7kd+ $$If4`drjX y#h%2`:V  t2@@@@4af4pւ$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb4O t2)v)v )v+++++++ + + + + ++++,55555555 5 5V555 / / / / /  / / /  / / /  / / / /  / / / / / / / /  / / / 4 f4p֖kd$$If4O֐rjXZ  y#h%2V  t2HHHH4af4p֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / / / / / /  / / / / / / / / /  /  / / /  / / / / / / / / / / / / 4 pִkdi$$If֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd #$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd*$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd2$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd:$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kdlB$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kdJ$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kdQ$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kd1Y$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 p֖kd`$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / /  /  / / / / / / / / / / / / / 4 p֖kd h$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 p֖kdNo$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kdv$$IfF֐rjXZ  y#h%2V t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / /  /  / / / / / / / / / / / / / 4 p֖kdr}$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb3 t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 p֖kd$$If3֐rjXZ  y#h%2V  t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb3 t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kd$$If3֐rjXZ  y#h%2V t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kdْ$$If֐rjXZ  y#h%2V t2HHHH4ap֖j$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 pִkd$$If֐rjXZ  y#h%2V t2HHHH4apִ2$$If!vh555555555 5 5 5 5 5 V555 #v#v#v#v#v#v#v#v #v #v #v V#v#v#v :Vb t2,55555555 5 5 5 V555 4 p֪kdנ$$IfzrjX Z  y#h%2V t2DDDD4ap֪}DyK _Ref215375150}DyK _Ref215375180 Dd 0 q # A" ]t3|YJ_ ʧ@=W ]t3|YJ;7$% x[}lTEޕ;VŊH[ zR@!`OGh1A1JF1_F%J_ $hRgߛm:ha|t?ffvvvg{' )_d g2@@t# WW`)`c?EkG&##!\Ɨa>)a KdHRq5NpE5C博$YREj/?|ʏ@MR R.ҸvuqE]~ _KKX?>8H(?8@x7tV8RMHˣ]&G?;wuE46!ckcҰVB3܏1LoX V+ 7 Ƒwpd{F\ٺ3ԩ/ YW7*^# 'vq36H?ܿ1>Ce{XY`UggW^>FuJem%3?lo<.lӂC7t췺> $cHC]>(ӻ sd7ƈ) nӃl^IDZmWkj !kZsGhccc0lepme9~{)GcF dv8@@|s".>Ghk^w[sqmV}f{=Y0VoÞGYxrF0ttyF!x"424{:\oH0;~i! g>_gYf^? ;~YfogM8F'[iG^[kvk(6Hl w'Mt?6s flf)( .!қE[/] Ez .t]H{Z/tv{:3YbV3AL+ f1еb` Vkfn0~ӌf*Ưb$'ІQgޡѲ\ldqB6`jV!E7y#SzVbJUl-6r-p^/ O))k=>ُ:@(cHKIFYN3H|ұ}HSHV{bMRG^:>D$1Nr%ȯ4AQ[6HǭX {oŚ-fZHZ奣uGg)9f\i=kruT믗2na8U4ao1A1 E,N%x==3N^K&XsM~C$[W/]o)Nk S.jDU16_k~|W{Wt (O~!W~o y{ g2L~ۄ eg HRWl0/yfc5?*^qW}|PTl]0 1o, {& t^X)LI6|>F|)=SQ'(իu?` OXL3GXye{v`8x^Sпs6_k ~|WWt lMdBV Zq~!g}ebeھX/e.bYb;2aa wQ]wL;6ck}8=n/uA% 2Km޻:ɘuaL] l~e[+l!җHv%ۅ5lrEHV:| { S~x_ƴ2NxRG"@ sm>?q/ӯ֑l?ij_;aJަgi]9OF ؄ugϨ-;߂X<3.=?6$2/֒}DyK _Ref190771388}DyK _Ref190686481}DyK _Ref190686436}DyK _Ref215319995}DyK _Ref215196170$$If!v h55555u555V5 5 5  #v#v#v#v#vu#v#v#vV#v #v #v :Vb4 t2)v)v)v)v ++++++++ + , 55555u555V5 5 5  /  / / /  / / / / /  / / /  / / / /  /  / / 4 f4pPGkd$$If4 rjX y#h%2uV  t2,,,,4af4pP$$If!vh5555555:55 5 5 5 5 V5 55 #v#v#v#v#v#v#v:#v #v V#v #v#v :Vb4` t2)v)v)v )v+++++++++ + + + + ++,,,5555555:5 5 V5 55 /  / / /  / / / / /  / /  / / /  / / /  / / / / 4 f4pւ7kd$$If4`drjX y#h%2`:V  t2@@@@4af4pւ$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb4O t2)v)v )v+++++++ + + + + ++++,55555555 5 5V555 / / / / /  / / /  / / /  / / / /  / / / / / / / /  / / / 4 f4p֖kdؿ$$If4O֐rjXZ  y#h%2V  t2HHHH4af4p֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / / / / / /  / / / / / / / / /  /  / / / / / / / / / / / / / / / 4 pִkdZ$$If֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ $$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd$$IfF֐rjXZ  y#h%2V  t2HHHH4apִ$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kd]$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kd$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kd$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 p֖kd"$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 p֖kd $$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / /  /  / / / / / / / / / / / / / 4 p֖kd$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / /                           ! " # $ % & ' ( ) * + , - . / 0 1 2 3 4 5 6 7 8 9 : ; < = > ? @ A B C D E F G H I J K L M N O P Q R S T U W X Y Z [ \ ] ^ _ ` a b c d e f g h i j k l m n o p q r s t u v w x y z { | } ~  /  / / / / 4 p֖kd?$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kd"$$IfF֐rjXZ  y#h%2V t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :VbF t2,55555555 5 5V555 /  / / / / / / / / / / / / / /  /  / / / / / / / / / / / / / 4 p֖kdc)$$IfF֐rjXZ  y#h%2V  t2HHHH4ap֖$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb3 t2,55555555 5 5V555 /  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 p֖kd0$$If3֐rjXZ  y#h%2V  t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb3 t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kd7$$If3֐rjXZ  y#h%2V t2HHHH4ap֖L$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kd>$$If֐rjXZ  y#h%2V t2HHHH4ap֖j$$If!vh555555555 5 5 5 5 5 5V555 #v#v#v#v#v#v#v#v #v #vV#v#v#v :Vb t2,55555555 5 5V555 /  / / /  / / / / /  / / / /  / / / / / / /  / / 4 pִkdE$$If֐rjXZ  y#h%2V t2HHHH4apִ2$$If!vh555555555 5 5 5 5 5 V555 #v#v#v#v#v#v#v#v #v #v #v V#v#v#v :Vb t2,55555555 5 5 5 V555 4 p֪kdL$$IfzrjX Z  y#h%2V t2DDDD4ap֪}DyK _Ref190768256}DyK _Ref190771388$$If!vh55m5m5m5m#v#vm:V  ,55m/  /  / / /  44p2U$$If!vh55m5m5m5m#v#vm:V ,55m99/ /  /  / / / /  / / / 44p2$$If!vh55m5m5m5m#v#vm:V ,55m99/  / /  / 44p2U$$If!vh55m5m5m5m#v#vm:V ,55m99/ /  /  / / / / / / / 44p2$$If!vh55d5#v#vd#v:Vc t,55d5/ /  / / 4 p$$If!vh555b5b5b#v#v#vb:VcY t,555b/ /  / / / / / / 4 p$$If!vh55555z5555 5 5 5 5 V5 555h#v#v#v#vz#v#v #v #v V#v #vh:VcY t,5555z55 5 5 V5 5h/ /  / / / /  /  / / / / / 4 p֪kdZ$$IfYz0( ~p^\ZX !zVh tDDDD4ap֪$$If!vh55555z5555 5 5 5 5 V5 555h#v#v#v#vz#v#v #v #v V#v #vh:Vc t,,5555z55 5 5 V5 5h/  / / /  / / / /  /  / / / / / / 4 p֪kd>`$$Ifz0( ~p^\ZX !zVh tDDDD4ap֪$$If!vh55555z5555 5 5 5 5 V5 555h#v#v#v#vz#v#v #v #v V#v #vh:Vc t,,5555z55 5 5 V5 5h/  / / /  / / / /  / / / / 4 p֪kdf$$Ifz0( ~p^\ZX !zVh tDDDD4ap֪}DyK _Ref216175094$$If!vh5 5L55 5 #v #vL#v#v #v :V4 t++++,5 5L55 5 / / / /  / 4 $$If!vh5 5L55 5 #v #vL#v#v #v :V4l t++++,5 5L55 5 / / / / /  4 .$$If!v h555555555 #v#v#v#v#v :VB t, 55555 / /  / / /  / / /  / / /  / / /  / /  / /  /  4 pZMkdOo$$IfB  ! t0$$$$4apZ$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / /  / / / /  / / / / /  / / / /  /  / /  / / / / / / / / / / / / / 4 p֖mkds$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ / / / / / /  / / / /  / /  / / /  /  / /  /  / / / / /  / / / / / /  / / / 4 p֖mkdz$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / / / / / /  / /  / / / /  /  / / /  /  /  / / / /  / / / / / / / / / / 4 p֖mkd>$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / / / /  / / / / / /  / / / /  / /  /  / / / / / / / / / / / / / 4 p֖mkd}$$IfN ! t0<<<<4ap֖$$If!vh5 5L55^5 5 5 5 5 d5 5  5  5  5  5 #v #vL#v#v^#v #v d#v #v  :V t,5 5L55^5 5 d5 5  / 4 p֖mkd$$IfN Y 6 C P]j! L^    d      t0<<<<4ap֖$$If!vh555r#v#v#vr:V< t06,55 5x/ / / 4 pT$$If!vh555r#v#v#vr:V8 t06,55 5x/ / / 4 pT$$If!vh55555##v#v#v#v#v#:V t06,,,5555(/ 4 p TP$$If!vh55555##v#v#v#v#v#:V4; t06+,5555(/  / / /  /  / / 4 p TP$$If!vh55555##v#v#v#v#v#:V4; t06+,5555(/ / /  / / /  / 4 p T$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V t06,5( 5~55~5/ 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / /  / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/ / / /  / / / /  / 4 T$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V t06,5( 5~55~5/ 4 T$$If!vh555r#v#v#vr:V< t06,55 5x/ / / 4 pT$$If!vh555r#v#v#vr:V8 t06,55 5x/ / / 4 pT$$If!vh55555##v#v#v#v#v#:V t06,,,5555(/ 4 p TP$$If!vh55555##v#v#v#v#v#:V4; t06+,5555(/  / / /  /  / / 4 p TP$$If!vh55555##v#v#v#v#v#:V4; t06+,5555(/ / /  / / /  / 4 p T$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V t06,5( 5~55~5/ 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / /  / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/  / / / / / / / / 4 Ta$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V, t06,5( 5~55~5/ / / /  / / / /  / 4 T$$If!vh5$ 5}5}5~5}5}#v$ #v}#v~#v}:V t06,5( 5~55~5/ 4 T$$If!vh555q#v#v#vq:V< t06,555q/ / / 4 pT$$If!vh555q#v#v#vq:V8 t06,555q/ / / 4 pT$$If!vh55555"#v#v#v#v#v":V t06,,,55555"/ 4 p TX$$If!vh55555"#v#v#v#v#v":V4; t06+,55555"/  / / /  /  / / 4 p TX$$If!vh55555"#v#v#v#v#v":V4; t06+,55555"/ / /  / / /  / 4 p T$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V t06,5$ 5}5~5|5}/ 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / /  / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / / / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / / / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / / / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / / / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/  / / / / / / / / 4 Tg$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V, t06,5$ 5}5~5|5}/ / / /  / / / /  / 4 T$$If!vh5$ 5}5}5~5|5}#v$ #v}#v~#v|#v}:V t06,5$ 5}5~5|5}/ 4 T$$If!vh55=#v#v=:VJ t6,55=/ /  / / 4 $$If!vh55=#v#v=:VJ t6,55=/ / / / 4 $$If!vh55=#v#v=:VJ t6,55=/ /  / / 4 $$If!vh55=#v#v=:VJ t6,55=/ / / / 4 $$If!vh5=55>#v=#v#v>:VJ t6,5=55>/ /  / 4 $$If!vh5=55>#v=#v#v>:VJ t6,5=55>/ / / /  /  / 4 $$If!vh5=55>#v=#v#v>:VJ t6,5=55>/ / /  / / /  4 5$$If!vh5555 555#v#v#v#v #v#v#v:V t0p,5555 555/ / / 4 a$$If!vh5555 555#v#v#v#v #v#v#v:V t0p,5555 555/  / / /  / / / /  / / / 4 aQ$$If!vh5555 555#v#v#v#v #v#v#v:V t0p,5555 555/ / / /  / 4 at$$If!vh5555 555#v#v#v#v #v#v#v:V4w t0p+++,5555 555/ / / /  / 4 ap t$$If!vh5555 555#v#v#v#v #v#v#v:V4w t0p+++,5555 555/ / / /  / 4 ap Q$$If!vh5555 555#v#v#v#v #v#v#v:V t0p,5555 555/ / / /  / 4 a$$If!vh5555 555#v#v#v#v #v#v#v:V4w t0p+++,5555 555/ / / / / / / / / 4 ap $$If!vh5555 555#v#v#v#v #v#v#v:V4w t0p+++,5555 555/ / / / / / / / / / 4 ap $$If!vh5z5 5(#vz#v #v(:V t0!,5z5 5(/  / 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ / 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ / / / / 4 ap$$If!vh5z5 5(#vz#v #v(:VJ t0!,5z5 5(/ 4 ap}DyK _Ref215196170}DyK _Ref215196170$$If!v h5555 5l5555 p 5 5 f#v#v#v#v #vl#v#v#v#v p #v #v f:Vb4 t!)v)v)v)v ++++++++ + , 5555 5l5555 p 5 5 f/  / / /  / / / / /  / / /  / / / /  /  / / 4 f4pPGkd$$If4  P-! lp f t!,,,,4af4pP$$If!vh5555 555<55 5 5  5 5 5 p 55f#v#v#v#v #v#v<#v#v #v #v #v #v #v p #v#vf:Vb4` t!)v)v)v )v+++++++++ + + + + ++,,,5555 55<55 5 5  5 5 5 p 55f/  / / /  / / / / /  / /  / / /  / / /  / / / / 4 f4pւ7kd$$If4`dXZ H P-! `< p f t!@@@@4af4pւ$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:Vb4O t!)v)v )v+++++++ + + + + ++++,5555 5555 5 5 5 5  5 55p 55f/ / / / /  / / /  / / /  / / / /  / / / / / / / /  / / / 4 f4p֖kd$$If4O֐XZ H P-!  p f t!HHHH4af4p֖<$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:Vb t!,5555 5555 5 5 5 5  5 55p 55f/  / / / / / / /  / / / / / / / / /  /  / / / / / / / / / / / / / / / 4 pִkdN$$If֐XZ H P-!  p f t!HHHH4apִX$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f/  / / / / / / / / / / / / / / / /  /  /  / / / / / / / / / / / / / / / / 4 pִkd=$$IfF֐XZ H P-!  p f t!HHHH4apִ $$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f/  / / /  / / / / / /  / / / / /  /  / / / / / / / /  / / /  / / / / 4 pִkdH$$IfF֐XZ H P-!  p f t!HHHH4apִn$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f4 pִkd $$IfF֐XZ H P-!  p f t!HHHH4apִn$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f4 pִkd<$$IfF֐XZ H P-!  p f t!HHHH4apִP$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f4 p֖kd]$$IfF֐XZ H P-!  p f t!HH HH4ap֖d$$If!vh5555 55555 5 5 5 5  5 55p 55f#v#v#v#v #v#v#v#v #v #v #v #v #v #v#vp #v#vf:VbF t!,5555 5555 5 5 5 5  5 55p 55f4 p֪kdB$$IfF֐XZ H P-!  p f t!HHHH4ap֪\$$If!vh5555 55555 5 5 5  5 5 5p 55f#v#v#v#v #v#v#v#v#v #v #v #v #v #v #vp #v#vf:Vb t!,5555 55555 5 5 5  5 5 5p 55f4 p֪kdO%$$Ifz2Z P-!  p f t!DDDD4ap֪$$If!vh55B55555 5] #v#vB#v#v#v #v] :Vb4 t!)v)v++++++,55B555 5] /  / / / /  / / / / / / 4 f4pF$$If!v h555555555 5 5 5  5 ] #v#v#v#v#v#v#v #v #v ] :Vb4d t!)v)v)v +++++++++ + + + ,,, 5555555 5  5 ] /  / / /  / / / / / /  / /  / / / / / 4 f4pxkd-$$If4d" B>, R!` ]  t!44444af4pxv$$If!vh555555555 5 5 5 5 5  5] #v#v#v#v#v#v#v #v #v  #v] :Vb4& t!)v)v+++++++ + + + + +,,,5555555 5 5  5] / / / / /  / / /  / / /  / / /  / / / / /  / / / / / 4 f4p֌kd2$$If4&NkB>, R! ]  t!<<<<4af4p֌$$If!vh555555555 5 5 5 5 5 5" 5-50 #v#v#v#v#v#v#v #v #v #v" #v-#v0 :Vb t!,5555555 5 5 5" 5-50 /  / / / /  / / /  / /  / / / /  / / / / /  / / /  / / / 2 q2q4 p֪kdL9$$IfzkB>, 0 R!" -0  t!DDDD2 q2q4ap֪N$$If!vh555555555 5 5 5 5 5 5" 5-50 #v#v#v#v#v#v#v #v #v #v" #v-#v0 :VbF t!,5555555 5 5 5" 5-50 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֪kdg@$$IfFzkB>, 0 R!" -0 t!DDDD2 q2q4ap֪N$$If!vh555555555 5 5 5 5 5 5" 5-50 #v#v#v#v#v#v#v #v #v #v" #v-#v0 :VbF t!,5555555 5 5 5" 5-50 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֪kdJG$$IfFzkB>, 0 R!" -0 t!DDDD2 q2q4ap֪N$$If!vh555555555 5 5 5 5 5 5" 5-50 #v#v#v#v#v#v#v #v #v #v" #v-#v0 :VbF t!,5555555 5 5 5" 5-50 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֪kd-N$$IfFzkB>, 0 R!" -0 t!DDDD2 q2q4ap֪0$$If!vh555555555 5 5 5 5 5 5" 5-50 #v#v#v#v#v#v#v #v #v #v" #v-#v0 :VbF t!,5555555 5 5 5" 5-50 /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֌skdU$$IfFzkB>, 0 R!" -0 t!DDDD2 q2q4ap֌P$$If!vh5555 5555 5 5 5  5 5 5 5" 5-50 #v#v#v#v #v#v#v#v #v #v #v #v #v #v #v" #v-#v0 :VbF t!,5555 5555 5 5 5  5 5 5 5" 5-50 2 q2q4 p֌skd[$$IfFz~B>H  0 R!   " -0 t!DDDD2 q2q4ap֌}DyK _Ref215319995}DyK _Ref215319995}DyK _Ref215196170$$If!vh55/55555 5w #v#v/#v#v#v #vw :Vb4 t!)v)v++++++,55/555 5w /  / / / /  / / / / / / 4 f4pF$$If!v h5555l55555 5 5 5  5 w #v#v#v#vl#v#v #v #v w :Vb4d t!)v)v)v +++++++++ + + + ,,, 5555l55 5  5 w /  / / /  / / / / / /  / /  / / / / / 4 f4pxkdd$$If4d" "  v!`l w  t!44444af4pxh$$If!vh555555555 5 5 5 5 5  5w #v#v#v#v#v#v#v #v  #vw :Vb4& t!)v)v+++++++ + + + + +,,,5555555 5  5w / / / / /  / / /  / / /  / / /  / / / / /  / / / / / 4 f4p֌kdpj$$If4&NS"  v! w  t!<<<<4af4p֌x$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :Vb t!,5555555 5 55H5/ /  / / / /  / / /  / /  / / / /  / / / / /  / / /  / / / 2 q2q4 p֪kdp$$IfzS"  v!H/  t!DDDD2 q2q4ap֪2$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :Vb} t!,5555555 5 55H5/ /  / / /  / / /  / / / / / / / / / / /  / / 2 q2q4 p֪kdx$$If}zS"  v!H/ t!DDDD2 q2q4ap֪2$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :Vb} t!,5555555 5 55H5/ /  / / /  / / /  / / / / / / / / / / /  / / 2 q2q4 p֪kd~$$If}zS"  v!H/ t!DDDD2 q2q4ap֪@$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :VbF t!,5555555 5 55H5/ /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֪kd$$IfFzS"  v!H/ t!DDDD2 q2q4ap֪@$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :VbF t!,5555555 5 55H5/ /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֪kdg$$IfFzS"  v!H/ t!DDDD2 q2q4ap֪"$$If!vh555555555 5 5 5 5 5 55H5/ #v#v#v#v#v#v#v #v #v#vH#v/ :VbF t!,5555555 5 55H5/ /  / / /  / / /  / / / / /  / / / / / / /  / / 2 q2q4 p֌skd<$$IfFzS"  v!H/ t!DDDD2 q2q4ap֌P$$If!vh55?55R55d55~5 5 5 x5 5 5 55H5/ #v#v?#v#vR#v#vd#v#v~#v #v #v x#v #v #v #v#vH#v/ :VbF t!,55?55R55d55~5 5 5 x5 5 5 55H5/ 2 q2q4 p֌skdՙ$$IfFz" 0 R v!?Rd~xH/ t!DDDD2 q2q4ap֌}DyK _Ref215196170$$If!v h555555555  5 5 #v#v#v#v#v#v#v#v#v  #v #v :Vb4 t!)v)v)v)v ++++++++ + , 555555555  5 5 /  / / /  / / / / /  / / /  / / / /  /  / / 4 f4pPGkd$$If4 H$< !  t!,,,,4af4pP$$If!vh555555555 5 5 5 5 5  55#v#v#v#v#v#v#v#v#v #v #v #v #v  #v#v:Vb4` t!)v)v)v )v+++++++++ + + + + ++,,,555555555 5 5 5 5  55/  / / /  / / / / /  / /  / / /  / / /  / / / / 4 f4pւ7kd7$$If4`dH$< !`  t!@@@@4af4pւ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:Vb4O t!)v)v )v+++++++ + + + + ++++,555555555 5 5 5 5 55 55/ / / / /  / / /  / / /  / / / /  / / / / / / / /  / / / 4 f4p֖kd$$If4O֐H$<z> !  t!HHHH4af4p֖$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:Vb t!,555555555 5 5 5 5 55 55/  / / / /  / /  / / / /  / / / /  / / / / / / / /  / / / 4 pִkd׳$$If֐H$<z> !  t!HHHH4apִ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 55/  / / /  / / / / /  / / / /  / / / / / / /  / / 4 pִkdd$$IfF֐H$<z> !  t!HHHH4apִ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 55/  / / / / / / / / / / / / / /  /  / / / / / / / / / / / / / 4 pִkd$$IfF֐H$<z> !  t!HHHH4apִ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 55/  / / /  / / / / / /  / / / / /  /  / / / / / / / / /  / / / / 4 pִkdp$$IfF֐H$<z> !  t!HHHH4apִ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 55/  / / /  / / / / /  / / / /  / / / / / / /  / / 4 pִkd'$$IfF֐H$<z> !  t!HHHH4apִ$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 55/  / / /  / / / / /  / / / /  / / / / / / /  / / 4 p֖kd|$$IfF֐H$<z> !  t!HHHH4ap֖d$$If!vh555555555 5 5 5 5 5 55 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v#v #v#v:VbF t!,555555555 5 5 5 5 55 554 p֪kd$$IfF֐H$<z> !  t!HHHH4ap֪\$$If!vh555555555 5 5 5 5 5 5 55#v#v#v#v#v#v#v#v#v #v #v #v #v #v #v #v#v:Vb t!,555555555 5 5 5 5 5 5 554 p֪kd$$IfzH$< z> !  t!DDDD4ap֪}DyK _Ref215196170$$If!vh55/55555 5m #v#v/#v#v#v #vm :Vb4 t!)v)v++++++,55/555 5m /  / / / /  / / / / / / 4 f4pF$$If!v h5555l55555 5 5 5  5 m #v#v#v#vl#v#v #v #v m :Vb4d t!)v)v)v +++++++++ + + + ,,, 5555l55 5  5 m /  / / /  / / / / / /  / /  / / / / / 4 f4pxkd$$If4d" "  !`l m  t!44444af4pxh$$If!vh555555555 5 5 5 5 5  5m #v#v#v#v#v#v#v #v  #vm :Vb4& t!)v)v+++++++ + + + + +,,,5555555 5  5m / / / / /  / / /  / / /  / / /  / / / / /  / / / / / 4 f4p֌kdy$$If4&NS"  ! m  t!<<<<4af4p֌$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :Vb t!,5555555 5 5V55 /  / / / / / / / / / / /  / / / /  / /  / / / /  / / / / / / / / 2 q2q4 p֪kd$$IfzS"  *!V  t!DDDD2 q2q4ap֪$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :Vb t!,5555555 5 5V55 /  / / /  / / / /  / / / / / / /  / / / / / /  / / / / / 2 q2q4 p֪kdS$$IfzS"  *!V  t!DDDD2 q2q4ap֪@$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :Vb} t!,5555555 5 5V55 /  / / /  / / /  / / / / / / / / / / /  / / / 2 q2q4 p֪kdn $$If}zS"  *!V  t!DDDD2 q2q4ap֪$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :VbF t!,5555555 5 5V55 2 q2q4 p֪kdC$$IfFzS"  *!V t!DDDD2 q2q4ap֪$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :VbF t!,5555555 5 5V55 2 q2q4 p֪kd$$IfFzS"  *!V t!DDDD2 q2q4ap֪$$If!vh555555555 5 5 5 5 5 5V55 #v#v#v#v#v#v#v #v #vV#v#v :VbF t!,5555555 5 5V55 2 q2q4 p֌skd$$IfFzS"  *!                          ! " # $ % & ' ( ) * + , - . / 0 1 2 3 4 5 6 7 8 9 : ; < = > ? @ A B C D E F G H I J K L M N \ Q R S T U V W X Y Z [ O ] ^ _ ` a b c d f g h i j k l m n o p q r s t u v w x y z { | } ~  V t!DDDD2 q2q4ap֌P$$If!vh55?55R55d55~5 5 5 x5 5 5 5V55 #v#v?#v#vR#v#vd#v#v~#v #v #v x#v #v #v #vV#v#v :VbF t!,55?55R55d55~5 5 5 x5 5 5 5V55 2 q2q4 p֌skd!$$IfFz" 0 R *!?Rd~xV t!DDDD2 q2q4ap֌$$If!vh5 5L55 5 #v #vL#v#v #v :V4 t++++,5 5L55 5 / / / /  / 4 $$If!vh5 5L55 5 #v #vL#v#v #v :V4l t++++,5 5L55 5 / / / / /  4 .$$If!v h555555555 #v#v#v#v#v :VB t, 55555 / /  / / /  / / /  / / /  / / /  / /  / /  /  4 pZMkd($$IfB  ! t0$$$$4apZ$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / /  / / / /  / / / / /  / / / /  /  / /  / / / / / / / / / / / / / 4 p֖mkdf-$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ / / / / / /  / / / /  / /  / / /  /  / /  /  / / / / /  / / / / / /  / / / 4 p֖mkd4$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / / / / / /  / /  / / / /  /  / / /  /  /  / / / /  / / / / / / / / / / 4 p֖mkd;$$IfN ! t0<<<<4ap֖$$If!vh555555555 5 5 5 5 5 5#v#v#v#v#v#v#v #v #v #v #v #v:V t,5555555 5 5 5 5 5/ /  / / / /  / / / / / /  / / / /  / /  /  / / / / / / / / / / / / / 4 p֖mkdC$$IfN ! t0<<<<4ap֖$$If!vh5 5L55^5 5 5 5 5 d5 5  5  5  5  5 #v #vL#v#v^#v #v d#v #v  :V t,5 5L55^5 5 d5 5  / 4 p֖mkd8J$$IfN Y 6 C P]j! L^    d      t0<<<<4ap֖$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V ta2, 55555o 5q55 /  / /  / / /  222222222 44 pZYkdqO$$If px`'h.a2o q ta2$$$$222222222 44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2)v+, 55555o 5q55 /  / / /  / / /  44 pZkdS$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdW$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdR[$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd _$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdb$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4( ta2+, 55555o 5q55 /  / / /  / / 44 pZ kdf$$If4( px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdBj$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdm$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdq$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdpu$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd*y$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd|$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / / / / / / / /  44 pZkd$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2)v+, 55555o 5q55 / /  / / / /  / / / /  44 pZkd$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdB$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkd$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!v h555555o 5q55 #v#v#v#v#vo #vq#v#v :V4 ta2+, 55555o 5q55 /  / / /  / / 44 pZkdp$$If4 px`'h.a2o q ta2$$$$44 apZ$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/ / / / / /  / /  / /  / /  / / / 449$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/  / / / / /  / /  / /  / /  / / / 449$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/  / / / /  / /  / /  / /  / / 449$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/  / / / /  / /  / /  / /  / / 449$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/  / / / /  / /  / /  / /  / / 449$$If!vh5) 555555#v) #v#v#v#v:V t0,5) 5555/  / / / /  / /  / /  / /  / / 449$$If!vh555 5#v#v#v #v:V t,555 5/  / 222244 $$If!vh555 5#v#v#v #v:V t,555 5/ 44 $$If!vh555 5#v#v#v #v:V t,555 544 $$If!vh555 5#v#v#v #v:V t,555 544 $$If!vh555 5#v#v#v #v:V t,555 544 $$If!vh555 5#v#v#v #v:V t,555 544 $$If!vh5h55 5#vh#v#v #v:V t !,5h55 59/ /  / / / / 4 p(>$$If!vh5h55 5 5p#vh#v#v #v #vp:V  !,5h55 5 5p9/ /  / / / / / 4 p2L$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 3 !, 5h5555855 5 5 /  /  /  /  / / / /  / / / /  / / /   4 pnYkd$$If3 hLT H  !h     8 8       !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kdA$$If4 hLT H  !&h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kd$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kdݹ$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apn|$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 /  / /  / / / / / / / / /  / /  /  / / 4 pn\kd$$If4 hLT H  !&h& &&&8&8&&&&& !,,,,4apn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / /  / / / / / / / / / / /  / / / 4 pn\kd$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / /  /  / / / / /  / / /  / /  / / /  4 pn\kdq$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kd[$$If4 hLT H  !&h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kd)$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kd$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apn|$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 /  / /  / / / / / / / / /  / /  /  / / 4 pn\kd$$If4 hLT H  !&h& &&&8&8&&&&& !,,,,4apn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / /  / / / / / / / / / / /  / / / 4 pn\kd$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / /  /  / / / / /  / / /  / /  / / /  4 pn\kd$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apnn$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !)v+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kdu$$If4 hLT H  !&h& &&&8&8&&&&& !,,,,4apnh$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kdC$$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apnh$$If!v h5h5555858555 5 5 #vh#v#v#v#v8#v#v #v #v :V 4 !+, 5h5555855 5 5 9 / /  / / / / / / / / / / /  / / / 4 pn\kd $$If4 hLT H  !h& &&&8&8&&&&& !,,,,4apn$$If!vh55( 55 #v#v( #v#v :V i!,55( 55 9/  / / 4 p($$If!vh55( 5 5#v#v( #v #v:V i!,55( 5 599/  / / / / / 4 p( $$If!v h555d5555655 #v#v#vd#v#v#v#v6#v#v :V 5i!, 555d5555655 /  /  / / / /  / / / /  / / /   4 pZkd$$If5 ~ i!   d    6   i!$$$$4apZ.$$If!v h555d5555655 #v#v#vd#v#v#v#v6#v#v :V i!, 555d5555655 9 /  / / / / / / / / / / /  / / / 4 pZkd$$If ~ i!& &&d&&&&6&&i!$$$$4apZ $$If!v h555d5555655 #v#v#vd#v#v#v#v6#v#v :V i!, 555d5555655 9 /  /  / / / / /  / /  / / /  /  / 4 pZkd$$If ~ i!& &&d&&&&6&&i!$$$$4apZ.$$If!v h555d5555655 #v#v#vd#v#v#v#v6#v#v :V i!, 555d5555655 9 /  / / / / / / / / / / /  / / / 4 pZkd$$$If ~ i!& &&d&&&&6&&i!$$$$4apZf Char5\^O1^ Thesis Body: Code Reserved WordI5fORf $Thesis Body: Code Reserved Word Char5\RO1R Thesis Body: Code CommentK5bOb Thesis Body: Code Char1OJQJ^J_HmH sH tH fOf %Thesis Body: Code Reserved Word Char15l@@@ NormalCJ_HaJmH sH tH p@p Heading 1, Char Char$ & FH X@&5CJHKH \^JaJ Z@Z  Heading 2$ & FHh@&5CJ \]^JaJV@V  Heading 3$ & Fh@&5CJ\^JaJR@R  Heading 4$ & F@&5CJ\aJTT  Heading 5 & F@&56CJ\]aJNN  Heading 6 & F<@&5CJ\aJ@@  Heading 7 & F<@&FF  Heading 8 & F<@&6]T T  Heading 9 & F<@&CJOJQJ^JaJDA@D Default Paragraph FontVi@V  Table Normal :V 44 la (k(No List FOF  Thesis Body$d\`a$POP  Thesis Body CharCJ_HaJmH sH tH rOr !#Heading 11, Char Char Char&5CJHKH \^J_HaJ mH sH tH $O!$ nfakpe4@4  ;TOC 1 8!@ :@: TOC 2 ! ^<@< TOC 3 8! ^6U@a6  Hyperlink >*B*phN"@N Caption$ dh^ `a$CJ\`O` Thesis Caption$!d\^!`a$ CJOJQJLOL Thesis Figures$d`a$^O^ Thesis Body: Before Bullets`@O@ Node c-stems6CJOJQJhOh Thesis Body: Bulleted List ^`B'@B Comment ReferenceCJaJ<@<  Comment TextCJaJxOx $Thesis Body: After Bullets No Indent$d\a$aJVOV Thesis Body - Justified `aJHH  Balloon Text!CJOJQJ^JaJ4"4 Header " !4 @24 Footer # !@j@ Comment Subject$5\.)@Q.  Page NumbernObn *Thesis Body: Feature Value &`6CJOJQJ]aJROqR Thesis Body Char4CJ_HaJmH sH tH >@>  Footnote Text(CJaJ`O` c$^"Thesis Body: Emphasis + Small caps:lOl &q Thesis Body: Feature Value Char6CJOJQJ]aJ@&@@ Footnote ReferenceH*TOaT QThesis Body: Feature Name,5\`O` Thesis Caption Char1 CJOJQJ_HaJmH sH tH NON Thesis Body: Equation .TO!T Heading 2: Starting a page/L#L Table of Figures0 ^` TOT Thesis Body: After Bullets1O" -Thesis Body: Between Bulleted Lists No Indent2`aJPO2P Thesis Body: Justified 3`lOBl  Tablecaption4$x7$8$`a$CJOJQJ^JaJmH sH VOQV Thesis Caption CharCJ_HaJmH sH tH JObJ Node Adherent Count6CJaJFOqF Node c-suffixes5CJOJQJ\O\ ACL-07 Caption8$pd ^`pa$CJaJ`O` Scheme Suffixes9  h5CJOJQJ^JaJxOx ;Scheme Affixes Char Char: <CJOJQJ^JaJmHsHvOv :!#Scheme Affixes Char Char Char$CJOJQJ^J_HaJmHsHtH `O` Thesis Caption Char2 CJOJQJ_HaJmH sH tH ROR Thesis Body Char1CJ_HaJmH sH tH \O\ Thesis Body: Emphasis5CJOJQJ\]aJnOn Thesis Body: Feature Value Char16CJOJQJ]aJ\O\ Thesis Body: Feature Name Char15\vOv  Bibliography3A$ '(Qd\7$8$^`Qa$^JaJmH sH `O"` Thesis: CodeB h8pCJOJQJ^JaJO2 Thesis Body: CodeCC > CJOJQJ^JaJdO1Bd Thesis Body: Code Function CallD 56]`OQ` Thesis Body: Code CharOJQJ^J_HmH sH tH jORaj $Thesis Body: Code Function Call Char 56]^OAr^ Thesis Body: Code Function DefG5\dObd #Thesis Body: Code Function DeOl %Thesis Body: Code Function Call Char1 56]XOX Thesis Body: Code Comment Char5ROR Thesis Body Char3CJ_HaJmH sH tH ZOZ ,}UThesis Body: Feature Name Char5\nO!n Thesis Body: Feature Value Char36CJOJQJ]aJ\O"1\ Thesis Body: Feature Name Char35\ROAR Thesis Body Char2CJ_HaJmH sH tH BORB 'q Thesis: FootnoteUdJbJ !#EACL Caption TextV CJaJtHFPrF !# Body Text 2WCJPJ aJnHtHjOj Y!#EACL Text Indent Char CharX$`a$ CJPJ tHrOr X!#EACL Text Indent Char Char CharCJPJ _HaJmH sH tHJJ [!#EACL Text Char Char Z`HOH Z!#EACL Text Char Char Charrr !#Abstract,\$$1$7$8$]^a$CJOJQJ^JaJmH sH zz !#Headingabstract!]$ x1$7$8$a$ CJKHOJQJ^JaJmH sH ~~ !#Headingreferences!^$ x1$7$8$a$ CJKHOJQJ^JaJmH sH tt !#Author1_$$ (#1$7$8$`a$CJOJQJ^JaJmH sH R>@R !#Title`$@&`a$5CJKH\aJ b!#EACL Section Char Char+a$ T>>Q^`Qa$5PJ \aJtHpO!p a!#EACL Section Char Char Char"5CJPJ \_HaJmH sH tH^O2^ !#EACL Reference textcx^`CJaJjBj !#EACL References Headingd$xa$5PJ \aJtH$Q$ !#Style1tObt !##Thesis Body: Between Bulleted ListsfdxCJLArL !#Constituent Nodeg`PP !# Text Indenth$ ` a$CJPJ aJtHOa !#-Thesis Body: Between Bulleted Lists No indentix`Oa !#3Style Thesis Body: Between Bulleted Lists No Indentjx`aJhh !#ACL-07 Text Abstractk$hhd ]h^ha$CJaJnOn !# Thesis Bulleted List. Start at 6l & FdpCJ\O\ !#CLEF-07 Text Justifiedm$da$CJaJNN !#CLEF-07 Text Indented n`JOJ !#CLEF-07 Captiono^`rr !#ACL-07 Heading 1/p & F hed^e`CJ\xx !#CLEF-07 Heading 2.q$ Qd@&^`Q5KH \aJO" !#Scheme Affixes for Clusteringr$  Vxa$CJOJQJ^JaJmHsHO2 !#Scheme Suffixes for Clusterings$  Vxa$5CJOJQJ^JaJmH ,sH ,ZRZ !#ACL-07 Text Justifiedt$d a$CJaJLARL !#ACL-07 Text Indented u`pbp !#ACL-07 Heading 2)v & F ed^e` \^JaJ\r\ !#Special Table Formatw$d8a$ 5CJaJ>> !# Text Indented x`LL !#Text Justifiedy$da$CJaJJJ !#Caption LREC-08zL^`Lxx !#Syntactic Tree Label - LREC 08{$da$CJOJQJ^JaJlql }!#Cap tion Text Char Char|$d^`a$CJtHjOj |!#Caption Text Char Char CharCJ\_HaJmH sH tHJ0J !# List Bullet~ hh^h`^O^ !#My Arial 11 Centered$a$CJOJQJ^JaJVV !# References$^`a$CJPJ aJtH.X. !#Emphasis6]|@#| d, Table Grid7:V0$`a$CJ<m1<d1 / 1.1 / 1.1.1 F+0lA0d 1 / a / i F,@nQ@dArticle / Section F-DTbD d Block Textx]^2Br2 d Body Textx>Q> d Body Text 3xCJaJPMqP dBody Text First Indent `HCH dBody Text Indenthx^hTNT dBody Text First Indent 2 `RRR dBody Text Indent 2hdx^hTST dBody Text Indent 3hx^hCJaJ2?2 dClosing ^$L$ dDate<[ < dE-mail Signatureh$ h dEnvelope Address!@ &+D/^@ OJQJ^JN%" N dEnvelope ReturnCJOJQJ^JaJFV1 F dFollowedHyperlink >*B* ph0_A 0 d HTML Acronym:`R : d HTML Address6]0aa 0 d HTML Cite6]>bq > d HTML CodeCJOJQJ^JaJ<c < dHTML Definition6]Fd F d HTML KeyboardCJOJQJ^JaJRe R dHTML PreformattedCJOJQJ^JaJ:f : d HTML Sample OJQJ^JJg J dHTML TypewriterCJOJQJ^JaJ8h 8 d HTML Variable6].( . d Line Number4/ 4 dListh^h`82 8 dList 2^`83 8 dList 38^8`84" 8 dList 4^`852 8 dList 5^`>6B > d List Bullet 2 & F>7R > d List Bullet 3 & F>8b > d List Bullet 4 & F>9r > d List Bullet 5 & F BD B d List Continuehx^hFE F dList Continue 2x^FF F dList Continue 38x^8FG F dList Continue 4x^FH F dList Continue 5x^:1 : d List Number & F!>: > d List Number 2 & F">; > d List Number 3 & F#>< > d List Number 4 & F$>= > d List Number 5 & F%I" dMessage Headerg8$d%d&d'd-DM NOPQ^8` OJQJ^J4^2 4 d Normal (Web)>B > d Normal Indent ^4O4 d Note HeadingDZb D d Plain TextCJOJQJ^JaJ0K0 d Salutation6@ 6 d Signature ^*W * dStrong5\  dTable 3D effects 1m:Vj#j#j#j#j.j.j.j. 55\5B* \`J phB* `J ph6 6 dTable 3D effects 2:Vj.@j#j9jj 45\5\t t dTable 3D effects 3:Vj.@j j j#j9jj44:B*`Jph B*`Jph5\5\br b dTable Classic 1:V0  j#j#j#jj 9B*`Jph6]5\56\]s  dTable Classic 2#:V0  j% j#j0 jjj%  :5\B*`JphB* `J ph5\Tt T dTable Classic 3:V0    jj0  j0  QB* ph5B*\`JphB* `J ph56B*\]`Jphu  dTable Classic 4:V0  jj0 j0 jj X5\B* `J ph56B*\]`JphB* `J ph5\Rv R dTable Colorful 1:V0    j% j% jj%  <B*ph56\]56\]56\]Tw# T dTable Colorful 2:V0 j% jj0  j @56\]56B*\]`Jph56\]4x3 4 dTable Colorful 3:V0j;$ j0 j%  5B*\`JphyC  dTable Columns 1 :V0    j j jjjj#jj4l5\B*`Jph B*`Jph5\5\5\5\5\5\zS  dTable Columns 2:Vj j jjjj% jj45\B*`Jph B*`Jph5\5B*\`Jph5\B*`Jph5\5\{c  dTable Columns 3:V0j j jjj#j% j4h5\B*`Jph B*`Jph5\5\5\B*`Jph5\|s  dTable Columns 4:Vj j jjj% 4LB*`Jph B*`Jph5\5\B*`Jph`} ` dTable Columns 5:V0    j jjj#j#4VB*`Jph B*`Jph5\5\5\56\]0 0 dTable Contemporary:V0j%@ j% j% 4<@B*`JphB*`Jph5B*\`Jph d Table Elegant_:V0j ;B*`Jph~ d Table Grid 1z:V0jj 6]6]  d Table Grid 2:V0jjj#j ,5\5\5\5\ d Table Grid 3:V0  jjj0  5\5\> > d Table Grid 4:V0  jj0 j0  B5B*\`Jph5B*\`JphB*`Jph  d Table Grid 5:V0    jjj# j 5\5\$ $ d Table Grid 6:V0    jj#j#j (5\B*`Jph5\J J d Table Grid 7:V0    jjj#j# j 25\5\5\5\5\ #  d Table Grid 8:V0jjj%  H5B*\`Jph5B*\`Jph5B*\`Jphv3 v d Table List 1:V0  j%@ jj#j0 j4M@B*`JphB*`Jph56B* \]`J ph5\pC p d Table List 2:V0 j%@ jj#j0 j4G@B*`JphB*`Jph5B*\`Jph5\S  d Table List 3:V0  j# j# j 05B* \`J ph6B* ]`J phc d Table List 4w:V0    j0   5B*\`Jphs d Table List 5:V0jj#  5\5\  d Table List 6:V0j%@ j# j#  45\5\  d Table List 7:V0  j%@ j% jjj# j0  4>@B*`Jph5\5\5\5\  d Table List 8:V0j%@ j% jjj#j0 4D@B*`Jph5\5\5\56\] dTable Professionall:V0j%  5B*\`Jpho dTable Simple 1:V0  j#j# rp r dTable Simple 2:Vj#j# j#j# j#j#O5\5\5B*\`Jph5\5\5\q dTable Simple 3l:V0    j%  5B*\`Jphp p dTable Subtle 1 :Vj0@ j# j# j0  j. jj45\5\pp dTable Subtle 2:V0j0  j0  j# j# jj5\5\ll d Table Theme7:V0# d Table Web 1h:V03j B*`Jph3 d Table Web 2h:V03j B*`JphC d Table Web 3h:V03j B*`JphbOQb CZsThesis Body: Code Char3OJQJ^J_HmH sH tH fORaf IZs%Thesis Body: Code Reserved Word Char35~Oq~ z Thesis Body: Feature Value Char2&6CJOJQJ]_HaJmH sH tH \Or\ zThesis Body: Feature Name Char25\bOb zThesis Body: Code Char2OJQJ^J_HmH sH tH fOf z%Thesis Body: Code Reserved Word Char25vOv Ea5(Morpho Challenge 2008 - text - justified$a$CJrOr <'Morpho Challenge 2008 - text - indented `tp{S      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~)vy|XZt   U d r   J ]$U 4d!OtCp %/:HWj *DXf %Bcn|8\ep|#,E|8cQ>r:q!$&&(()))1)O)U)e)i)*2414B4J4Y4v4x444444444444444444445t;; <<*<N<\<k<n<s<u<y<<<<p?@A3D?FIK-VV_WYA[[5\]]#^_``vaya|aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaabbb5bHbgbbRdd eHeSee\f~ffggrrrr sAs`sssttttttttttttOuQuSu\ueujupuwu|uuuuuuuuu v8v@vFvJvLvwwwwxuz|}KeȄ,+u(_IvK3Bʲz57"{˾̾TU-]S<>jklmnopqrstuvwxyz{|}~%/9 !&+9:;<=>?@ABCDEFGHIJKmorstuvwxz{|}~y w u s q p o n l j h f e c b ` _ ^ *+,-01234567898 EFGHIJKLMNPQRSTUVWXYZ     % $               ( ) , - 1 0 9                                                        3 2 1 0 C , * !  B                        { z y u t s m l k j i h e d a ` \ [ Z Y   `R Q                                       J I     a     )vy|XZt   U d r   J ]$U 4d!OtCp %/:HWj *DXf %Bcn|8\ep|#,E|8cQ>r:q!$&&(()))1)O)U)e)i)*2414B4J4Y4v4x444444444444444444445t;; <<*<N<\<k<n<s<u<y<<<<p?@A3D?FIK-VV_WYA[[5\]]#^_``vaya|aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaabbb5bHbgbbRdd eHeSee\f~ffggrrrr sAs`sssttttttttttttOuQuSu\ueujupuwu|uuuuuuuuu v8v@vFvJvLvwwwwxuz|}KeDŽȄ,+u(_IvK3Bʲz57"{˾̾TU-]`      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijkklmnopqrstuvwxyz{|}~Carnegie Mellon UniversityChristian Monson{րQ_Oe ^62SCMU CMXٺ CMU:y CMU CMUQ CMUCMUove CM CM_ CMs CM%   f  &d̦f[:Rf#cUcloOOkGmGtwS2^Ž#UAzAzAzAzAzAzAzAzA zA zA z@ z@ zAzAz@z@z@z@z@zAzAzAzAzAzAzAzAzAzAzAzA zA!zA"zA#zA$zA%zA&zA'zA(zA)zA*zA+zA,zA-zA.zA/zA0zA1zA2zA3zA4zA5zA6zA7zA8zA9zA:zA;zA<zA=zA>zA?zA@zAAzABzACzADzAEzAFzAGzAHzAIzAJzAKzALzAMzANzAOzAPzAQzARzASzATzAUz!VvAWzAXzAYzAZzA[zA\zA]zA^zA_zA`zAaz!bv!cz!dz ez fz!gz!hz!iz!jz!kz!lz!mz!nz!oz!pz!qz!rz!sz!tvAuzAvzAwzAxzAyzAzzA{zA|zA}zA~zAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAz!vAzAzAzAzAzAzAzAzAzAzAzAzAzAzAz!vAzAzAzAzAzAzAz@z@z@z@z@z@z@z@zAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAzAz|w l(.308BgDhD'I$RWchlpF{5 x^ !)25<ENVA[ackgs^vs~Ň  bôBKz x l)j2o2;SFOOXbl:v:,ی-Za{#*s4;kGmGTT}^xh>s|ZÔrۭUW+ b# tws"+68BcKTYaamfpxzEt<԰Nuis %'+5x?DNV$`&`jt4~נǪc[\ij"\Y}) {}%+S|>  Z T  y!U%pIH{E !t"5#$R%&'$()*+6,-.x/0123V4X56s78~9:;<=>?u@@ABC%D}EFG>HtI&J KLMMxN]OPQR_S<TUVWXBYZ[V\5]A^_`abdcGderfPghijsklmnopq8rstuvywxyz{~|x}~Jo1 5S>jT]Y[qfRbPM2AF }q=<-' $:QRd9:;<=tx`u v G H lFmvwL eB*laT UV8 !!!%"""S##]$$6%% &o&&T''((((/+-%0&020C00h11Z22!3|332444=555`667778]88(999M:: ;d;;<s<<8===G>>>J???X@@ AhAA*BBB9CCCfDiDvD HHHH'IL$LwPSyU[_dhilllmqn;ooppsu:|~62MdNq` 5XSpA P1o%i9^b^]pU]i' 6<SpHk R !wn$$&((()+-g/2349<t@/CDuGJKM PS~VVV9YYY]._V_^_`beffffghjlnxpps^uwyy|s~=Ň8H   TgɨSp"-E-($ "(,BKKRK WzF   3  4a -!"% &(+?+ 01i2k2n2o33U7~9;?DJDGILMNOOOOSqW0XCX}XXf\^cigjkm(prtvz|}:,Ї+2gɕ7iD~7,ݩ<ǵY9,  jkIKaG:ZR{ * "O"&)|,&/34!5B7W::;;;M<<[>>:??@jGkGmGHHHMRTTWZ[[5\W\]cexhhFk'o"r>wdw|܄ }ؐ;rͦ+% :UWOPA~#E_a$ <stvwiF d$&&(n/K1O488|9;>@\EGJJOZQyUYZ^aaaane7impqruxx=}}}}!~T~~~~G%3ڛ+l 39:;=>r³F<><8N4l1CA x\, ` S;/x"&)-2668<x?@D(JKKlOOzPQ6TtT#XX YY\$`&`Eae@h kn[qhsv6x|tO ? s$ҟZc$Yeghkx>"v\NWXxUr65 S1=wN&|jIGi)   Z   :>PJz} 6cd     !!A!!!7"E"\"s"t""" ###3#C#F#G#t#######$S$T$o$$$$$)%B%^%r%%%%%%%&5&J&b&&&&&& '1'r'w'{''''''''(7(o(((()=)\)})))))))))) *!*@*_****+?+r++++++++,X,,,,-6-Q-------.a.......//_/00 2!2.2/202=2>2?2K2L2Y2Z2[2h2i2j2w2x222222/303`33$4466t77788\:]:3;4;@AABB B'B/B7B>B?BFBKBQBWB]BcBiBjBBBBBBBBBBBBBBBBBBCC CCCCC5C9C=CACFCKCPCQCRCwCxC&III!J"JLLLOO7OLO[OmOnOOOOOOOOOOOOOOOOOOOOOPPP$P&P'P2P4P=P@PGPKPPPUPYP_PaPbPnPpPzP}PPPPPPPPPPPPPPPPPPPPPPPPPPPPQQQ QQQQ+Q1Q6Q=QBQCQIQKQPQVQZQbQeQoQqQ}QQQQQQQQQQQQQQQQR RRR'R0R1R2RZP[Q[Y[\[b[j[q[u[}[[[[[[[[[[[[[[[[[[[[[[[[[[[\\\\\\#\&\)\.\4\:\@\F\H\I\W\Z\]\b\h\n\s\y\{\|\\\\\\\\\\\\\\\\\\\\\\\\]] ] ]]]]#]'](]/]4]8]>]L]S]W]X]_]d]l]u]}]]]]]]]]]]]]]]]]]]]]]]]]^ ^^^^^^&^,^4^>^B^C^H^M^U^X^\^c^g^h^u^w^{^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^__ _____ _$_*_4_6_7_?_D_L_T_X_^_b_c_d_````````````````````````````````ccc ccccccccc!c"c#c(c-c.c6c;ciDiIiRiViWi^icikiti|iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiijjj j jjjjjj#j)j/j0j1j6jx?x@xBxDxExFxGxHxJxNxSxwxxxzx{x}x~xxxxxxxxxxxxxxxxxxxxxxxxxxxxxyy(y)y+y-y/y0y1y2y3y5y7y8y9y:y;y=yByFyiyjylymynypyqyrysyuyvyxyyyzy{y~yyyyyyyyyyyyyyyyyyy zz1z2z5z6z7z9z:z;zz@zAzCzDzGzvzyzzzzzzzzzzzzzzzzzzz {{{{{{{{{{{{{{!{#{5{7{Q{R{U{V{W{X{Z{\{]{_{`{b{c{d{e{h{{{{{{{{{{{{{{{{{{{{{||!|"|#|$|%|&|'|(|)|*|+|,|.|0|E|G|[|\|`|a|b|c|e|f|g|i|j|l|m|n|o|q|||||||||||||||||||||}} }!}"}#}$}%}'}(})}*},}-}.}0}A}C}b}c}g}h}i}j}k}l}n}o}p}q}r}s}u}w}}}}}}}}}}}}}}}}}}}}} ~ ~~~~~~~~~~~~ ~!~#~H~J~k~l~q~r~s~t~u~w~x~z~{~}~~~~~~~~~~~~~~~~~~~~~~~~~~~~     &',-./012345679;CESTYZ[]^_`abdfghjtv$%56=>LMijkl|}~ÁāȁɁ́΁ҁӁׁ؁܁݁tuȂ Gp{݃#Xׄڄ=M}DžȅONЇ !CadfghABOP^_abfghilmrstuz{Ɖɉʉ-Yъ=h5acdލK%&'123yz"L}ؔ&QyÕƕȕ 8{ʖ̖͖Ζϖ  Wgu~ǘȘɘʘ˘̘͘ΘϘИјҘ՘ؘۘܘݘޘߘBCEGLOPQSTUWXYZ\_GHJPUWXZ[\]_`bcdgΚӚ֚ךؚښۚܚޚߚ   !"$GKxy|47žŞ&)yz}=@#&OPTVY[\]^_`abcefh¡     &(BCHJMOPQRSTUVWYZ\npĢƢȢʢˢ̢͢΢ϢТѢҢԢբע    !ȣʣˣ&?RS8!45()qNgz{]Iijlmoprsuvxy{|~®ĮŮƮȮɮʮ̮ͮήЮѮ   !()04;<BGOUZ[chpuvEFñı;<FGMOPaqr{|  ()*+,-/38?EFKRSTUVWXYZ[\]`cfghijklmnoprtz|~͵εеҵ׵ڵ۵ܵ޵ߵҶӶն۶#HIKY^abcefgijlmnq7;DHҹֹ !#$&'(+º0149=@ABCDFHIKLMP  !$ȼ˼ #%&()*+,-.013EHuvzڽ۽߽56:<?ABCDEFGHIJKMmo;ξӾվؾھ۾ܾݾ޾߾ !"#$%&')*,68IJOQSUVWXYZ[\]_`boq~οϿ"45FST|}~BCDEFGOPXY]^cdjkoprswxz{~+      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~,349:=>?@%0;<DINSXY`ejoqrz  $).37<AFJOTY]^hmrw{hi  !"#$%&'.6=?ADFHKLNOPU\ahmty !(-4579;@GLSX_dkpw| %)09=DMQX_jqxy  %&27=BJPQegjlqtu>?JMNObghijkt{}~ !469;@CD\^cejop 6:?CJOPQRSTUVWXYdghi|"-4;<OQTV[^_wy~ (*-/478NSY^flmnopqrstuv     478;<?@AUXY\]`ak  ;fgm hi!iy      !"%@ADGKMNPQRSUVWYZ\     $-5;FNOPRV[bcdejqxyz{|}~   7:WXZ[\^_`acdfghil     } "*+,.27>?@AFMTUVWXYZ]`cdefghijklmpqrtuvwyz|}~   ?B[\`abceghjkmnops    ">?@ABCDEFGHIJKLMNOPQRSX ()1234568<AHNOT[\]^_`abcdefilopqrstuvwxy}!#&()*+,-./0235HJdejloqrstuvwxy{|~ !"#$%&'()*+,-./0123456789:;<=>1Lr0_(fBbt&*iGbT}+RU     "SV #ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~  ij~=>?r,JRpt5| O =x}$&'56789AHIJKLMNOV^egilnpstvwx} %,18=DIPU\]_achot{mnos!+5?ABCEPXes!,4CVdq%6P_npqrt| 1=ILMNQ]ft &+,-0:FTdp|~  #.CPiv +7CO_opqt$&'()*+  ? G O V ^ f m n u z                       $ ( , 0 4 8 ; < X [ ^ a e i l m n o ~               ! * 7 C D N V p | }          !()138=BGLQV[`abdinsx} "$%&).38=BGLQST\^chmrw| "',16:?DFGHKPUZ_dinsuv}GHI !#89GHJOTYZ\px &+05:?DINOPQT00000000000000000000000000003000v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 0v 00(0(0(0(0(00200200020002000200020002002002002000200200020002000200200200200020002080208020802080208020802080200208020020802002002080208020802080208020802080208020802080208020@020@020@020020020020020@020@020@020@4 00iD0iD 0iD 0iD 0iD0iD 0iDiD0L0L0L0L0L0L0L0L0L /0iDiD0l 0l0l0l 0l /0iDiD0p0p0p0p0p0p0p0p 0iDiD0M 00 000 000000( 00000( 00505050505050505050505 00% 0%0%0%0%0%.0%30%.0%30%0%0%( 0%%0^0^0^0^0^0^0^0^0^0^ 0^ 0^ 0^10^0^0^0^ 0^ 0^0^( 0%%08 000 000000000000008 00n$0n$0n$0n$0n$0n$0n$0n$0n$ 00202020202020202 00J0J0J0J0J 00V0V0V10V0V 0VV( 0._._0V_0V_0V_0V_ 0V_ 0V_ 0V_ 0V_0V_0V_0V_0V_0V_8 0V_V_10xp0xp0xp0xp8 0V_V_0y0y0y0y0y0y( 0._._080808080808 0808080808308 0VV0( 00ɨ0ɨ0ɨ0ɨ0ɨ0ɨ0ɨ8 0ɨɨ0-0-0-0-( 00000000( 00000( 000008 00, 0, 0,0,20,0,10,0,0,8 000008 00z0z0z0z0z0z0z10z( 0008 00008 00 0 8 00%0%8 00+0+0+0+0+ 0+8 0 0o30o30o30o30o38 00D0D0D8 00L0L0L0L0L 0L0L0L 0L 0L 0L10L0L0L0L0L8 00j0j0j0j0j0j0j0j0j0j0j0j8 0000 0VV00 00 007( 0770D0D0D0D0D0D10D8 0DD0ݩ0ݩ30ݩ0ݩ0ݩ8 0DD000000( 0770000 000008 00k0k0k0k0k 00a0a0a0a10a( 0aa0008 009090909090909090909098 008 00"0"0"0"0"0"( 0``0404( 0``0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V:0V: 0V:0V:0V:0V:0V:8 0V:V:04\04\04\04\8 0V:V:0wh0wh0wh0wh8 0V:V:0=w0=w0=w 00ۄ( 0ۄۄ0000000( 0ۄۄ0q0q8 0qq008 0qq03008 0qq0V0V0V0V8 0qq0008 0qq000000000 0000( 00`0`0`0`0`0`0`0` 0`0`0`0`0`( 00&0&0&0&0&0& 0&0&0&0&0&0&0& 00J0J0J0J0J0J0J0J0J 00a0a0a0a 0aa0p0p0p 0aa0x0x0x0x0x0x0x0x0x0x10x0x0x0x0x 0x0x0x0x0x0x0x0x0x 0aa0 0 0 0 0  00=0= 0==0000( 0000000000000( 00303 0==00/ 00/ 00/ 001000000000000@00@00000000000000008 0000w0w0w0w0w0w 0==06060606060606 0==0K0K0K0K0K( 0KK05T05T05T05T105T05T05T 05T05T05T05T05T05T05T05T05T05T05T05T( 0KK00000000100000 000000 000000000 00( 000( 000( 000 00000P0PA0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A0A000000@000000A00000C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0(C0(C0(C0(C0(C0C0C0(C0(C0(C0C0(C0(C0C0(C0C0C0C0C0C0C0C0C0(0 (0 (0U0x0#0i#00#0#00h#00#0#00#0#00h#00#0000000000000101000000000000 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00000000000000000 0 00 00 00 00 00 00 0 00 00 00 00 00 00 0 00 00 00 00 00 0 0 00 00 00 00 00 00 00 0 00 00 00 00 0 00 00 0 00 00 00 00 00 00 00 0 00 00 00 00 00 00 00 0 00 00 00 0 00,0 ,0 ,0 0 &0 &0 &0 0  &0 &0 &0 0 &0 &0 &0 0 &0 &0 &0 0 00`0000 /000{0{00{0{00{0{00{0{0{0{0{0{0{0{0{0{0{0{0{0{0{0{x0{0{x0{0{x0{0{x0{0{0{0{0{0{0{0{0{0{0{0{`0{0{`0{0{0{0{0{`0{0{0{0{0{ /0`0000 /0`00000000000H,0 ,0 ,0 0 0 0 0 0  0 0 0 0 0 0 0 0 0 0 0 0 00000000000000000000000000 0000000000 /0 0T 0T 0T 0T 0T 0T 0T 0T 0T 0T 0T  /0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (0 0 0 0 (0 0 0 0 (0 0 (40 40 0 (0 60 0 0 0 0 0 0 0 (0 0 (0 60 0 0 0 0 0 0 0 0 0 (0 60 0 0 0 0 0 0 0 0 (0 0 0 0 0 0 0 0 (0 60 0 0 0 0 0 0 0 0 00 60 0 0 0 0 0 0 0 0 00 60 0 0 0 0 0 0 0 0 00 60 0 0 0 0 0 0 0 0 00 60 0 0 0 0 0 0 0 060 60 0 0 0 0 00 60 0 0 0 0 0 0 00 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 80 60 0 0 0 0 0 0 80 60 0 0 0 0 0 0 @0 60 0 0 0 0 0 0 @0 60 0 0 0 0 0 0 @0 60 0 0 0 0 0 0 @0 60 0 0 0 0 0 0 0 @0 0 @0 60 0 0 0 0 0 0 @0 0 @0 0 @:0 0 H:0 0 H:0 0 H90 :0 0 H90 :0 0 H90 :0 0 H90 :0 0 H0 80 0 H90 :0 0 0 H90 :0 0 0 H90 :0 0 0 H90 :0 0 H90 :0 0 H90 :0 0 H90 :0 0 H90 :0 0 H90 90 :0 0 P90 90 :0 0 P90 90 :0 0 P90 90 :0 0 P90 :0 0 P90 90 :0 0 P90 90 90 :0  /0P90W90W:0W0WP90W90W90W:0W0WP90W90W90W:0W0WP90W:0W0W0WP90W:0W0W0WP90W:0W0W0WP90W:0W0W0WP90W:0W0WP90W90W:0W0WP90W90W:0W0WP90W90W:0W0WP90W:0W0W0WX90W:0W0W0WX90W:0W0W0WX90W:0W0W0WX90W90W:0W0WX90W90W90W:0W0WX90W90W90W90W:0W0WX90W90W90W90W90W:0W0WX90W90W:0W0WX90W90W90W:0W0WX90W90W90W90W:0W0WX90W90W90W90W90W:0W0WX90W:0W0WX90W:0W0W0WX90W:0W0W0WX90W90W:0W0WX90W90W90W:0W0W`90W:0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W0W`0W60W0W0W0W0W0W0W0W`0W60W0W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0W0Wh0W60W0W0W0W0Wh0W60W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0Wh0W0W0Wp0W0W0Wp0W0W0Wp0W0W0Wp0W0W0Wp0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0Wp0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0Wp0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0Wp0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0Wp0W60W0W0W0W0W0W0Wp0W60W0W0W0W0W0W0W0Wp0W60W0W0W0W0W0W0Wp0W60W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0Wh0W60W0W0W0W0Wh0W60W0W0W0W0W0W0W0Wh0W60W0W0W0W0W0W0W0Wh0W0W0W 0W 0W 0W 0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W 0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W@0W0W@K0WC0WK0WC0WK0WK0WI0WC0WC0WC0WC0WC0WC0WC0WI0WC0WI0WK0WD0WC0WC0WC0WC0WC0WC0WK0WK0WK0WC0WC0WC0WK0WK0WC0WC0WI0WC0WC0WK0WC0WC0WC0WK0WC0WC0WI0WI0W0W0W@0W0W0W@0W0W@0W0W@0W0W@0W0WH0W0WH0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0WhC0WC0WC0WC0WC0WC0WK0WK0WC0WC0WC0WK0WC0WC0WC0WC0WI0WI0WC0WK0WC0WC0WC0WI0W0Wp0W0WpC0WC0WC0WC0WC0WC0WC0WC0W0Wp0W0W0Wp0W0W0WpC0WC0WK0WK0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WI0WI0WI0WC0WK0WK0WC0WC0WI0WC0WC0WI0WC0WC0W0W0W80W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W  0W  0W  0W 0W 0W 0W 0W 0W  0W 0W 0W  0W  0W  0W  0W 0W  0W  0W  0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W80W0W s0Wr0Wr0W0W 0Wo0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W s0Wr0Wr0W0W 0W0W 0W0W 0W0W 0W0W 0W0W 0W0W 0W0W 0W0W 0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W(0W0W0W(0W0W0W(0W0W0W(0W0W0W(0W0W0W0W(0W0W0W(0W0W0W(0W0W0W0W(0W0W0W0W00W0W0W0W0W0W00W0W0W0W0W00W0W@0W0W80W0W80W0W80W0W80W0W80W80W0W80W0W0W80W0W80W0W80W0W80W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W  0W  0W  0W 0W 0W 0W 0W 0W  0W 0W 0W  0W  0W  0W  0W 0W  0W  0W  0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W80W0W0W80W0W80W0W80W0W0W80W0W0W80W0W0W80W0W80W0W0Wp0W0W0W0W0W0W0W0W,0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W  0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W  0W 0W  0W 0W      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~ 0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W80W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W  0W  0W  0W 0W 0W 0W 0W 0W  0W 0W 0W  0W  0W  0W  0W 0W  0W  0W  0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W  0W  0W  0W 0W 0W 0W  0W  0W 0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0WH0W80W0W00W80W0W0W80W0W0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W  0W  0W  0W 0W 0W 0W  0W  0W 0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0WH0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W  0W  0W  0W 0W 0W 0W 0W 0W  0W 0W 0W  0W  0W  0W  0W 0W  0W  0W  0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WK0WK0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0W0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0W0W0W0W0W80W0W0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W  0W  0W  0W 0W 0W 0W  0W  0W 0W  0W  0W  0W 0W 0W 0W  0W  0W  0W  0W  0W  0W  0W  0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W0W0W0W0W0W0W0W0W0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0WC0W0W00W0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 00W0W0W0W 0W 0W 0W 0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0W0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W00W00W0W00W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W  0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0W0WH0WH0W0WH0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W 0W0W0"0 $:QRd9:;<=tx`u v G H lFmvwL eB*laT UV8 !!!%"""S##]$$6%% &o&&T''((((/+-%0&020C00h11Z22!3|332444=555`667778]88(999M:: ;d;;<s<<8===G>>>J???X@@ AhAA*BBB9CCCfDiDvD HHHH'IL$LwPSyU[_dhilllmqn;ooppsu:|~62MdNq` 5XSpA P1o%i9^b^]pU]i' 6<SpHk R !wn$$&((()+-g/2349<t@/CDuGJKM PS~VVV9YYY]._V_^_`beffffghjlnxpps^uwyy|s~=Ň8H   TgɨSp"-E-($ "(,BKKRK WzF   3  4a -!"% &(+?+ 01i2k2n2o33U7~9;?DJDGILMNOOOOSqW0XCX}XXf\^cigjkm(prtvz|}:,Ї+2gɕ7iD~7,ݩ<ǵY9,  jkIKaG:ZR{ * "O"&)|,&/34!5B7W::;;;M<<[>>:??@jGkGmGHHHMRTTWZ[[5\W\]cexhhFk'o"r>wdw|܄ }ؐ;rͦ+% :UWOPA~#E_a$ <stvwiF d$&&(n/K1O488|9;>@\EGJJOZQyUYZ^aaaane7impqruxx=}}}}!~T~~~~G%3ڛ+l 39:;=>r³F<><8N4l1CA x\, ` S;/x"&)-2668<x?@D(JKKlOOzPQ6TtT#XX YY\$`&`Eae@h kn[qhsv6x|tO ? s$ҟZc$Yeghkx>"v\NWXxUr65 S1=wN&|jIGi)   Z   :>PJz} 6cd     !!A!!!7"E"\"s"t""" ###3#C#F#G#t#######$S$T$o$$$$$)%B%^%r%%%%%%%&5&J&b&&&&&& '1'r'w'{''''''''(7(o(((()=)\)})))))))))) *!*@*_****+?+r++++++++,X,,,,-6-Q-------.a.......//_/!2.2/202=2>2?2K2L2Y2Z2[2h2i2j2w2x22222/303`33$4466t77788\:]:3;@AABB B'B/B7B>B?BFBKBQBWB]BcBiBjBBBBBBBBBBBBBBBBBBCC CCCCC5C9C=CACFCKCPCQCRCwCxC&III!JLLLOO7OLO[OmOnOOOOOOOOOOOOOOOOOOOOOPPP$P&P'P2P4P=P@PGPKPPPUPYP_PaPbPnPpPzP}PPPPPPPPPPPPPPPPPPPPPPPPPPPPQQQ QQQQ+Q1Q6Q=QBQCQIQKQPQVQZQbQeQoQqQ}QQQQQQQQQQQQQQQQR RRR'R0R1R2RZP[Q[Y[\[b[j[q[u[}[[[[[[[[[[[[[[[[[[[[[[[[[[[\\\\\\#\&\)\.\4\:\@\F\H\I\W\Z\]\b\h\n\s\y\{\|\\\\\\\\\\\\\\\\\\\\\\\\]] ] ]]]]#]'](]/]4]8]>]L]S]W]X]_]d]l]u]}]]]]]]]]]]]]]]]]]]]]]]]]^ ^^^^^^&^,^4^>^B^C^H^M^U^X^\^c^g^h^u^w^{^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^__ _____ _$_*_4_6_7_?_D_L_T_X_^_b_c_d_````````````````````````````````ccc ccccccccc!c"c#c(c-c.c6c;ciDiIiRiViWi^icikiti|iiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiijjj j jjjjjj#j)j/j0j1j6jx?x@xBxDxExFxGxHxJxNxSxwxxxzx{x}x~xxxxxxxxxxxxxxxxxxxxxxxxxxxxxyy(y)y+y-y/y0y1y2y3y5y7y8y9y:y;y=yByFyiyjylymynypyqyrysyuyvyxyyyzy{y~yyyyyyyyyyyyyyyyyyy zz1z2z5z6z7z9z:z;zz@zAzCzDzGzvzyzzzzzzzzzzzzzzzzzzz {{{{{{{{{{{{{{!{#{5{7{Q{R{U{V{W{X{Z{\{]{_{`{b{c{d{e{h{{{{{{{{{{{{{{{{{{{{{||!|"|#|$|%|&|'|(|)|*|+|,|.|0|E|G|[|\|`|a|b|c|e|f|g|i|j|l|m|n|o|q|||||||||||||||||||||}} }!}"}#}$}%}'}(})}*},}-}.}0}A}C}b}c}g}h}i}j}k}l}n}o}p}q}r}s}u}w}}}}}}}}}}}}}}}}}}}}} ~ ~~~~~~~~~~~~ ~!~#~H~J~k~l~q~r~s~t~u~w~x~z~{~}~~~~~~~~~~~~~~~~~~~~~~~~~~~~     &',-./012345679;CESTYZ[]^_`abdfghjtv$%56=>LMijkl|}~ÁāȁɁ́΁ҁӁׁ؁܁݁tuȂ Gp{݃#Xׄڄ=M}DžȅONЇ !CadfghABOP^_abfghilmrstuz{Ɖɉʉ-Yъ=h5acdލK%&'123yz"L}ؔ&QyÕƕȕ 8{ʖ̖͖Ζϖ  Wgu~ǘȘɘʘ˘̘͘ΘϘИјҘ՘ؘۘܘݘޘߘBCEGLOPQSTUWXYZ\_GHJPUWXZ[\]_`bcdgΚӚ֚ךؚښۚܚޚߚ   !"$GKxy|47žŞ&)yz}=@#&OPTVY[\]^_`abcefh¡     &(BCHJMOPQRSTUVWYZ\npĢƢȢʢˢ̢͢΢ϢТѢҢԢբע    !ȣʣˣ&?RS8!45()qNgz{]Iijlmoprsuvxy{|~®ĮŮƮȮɮʮ̮ͮήЮѮ   !()04;<BGOUZ[chpuvEFñı;<FGMOPaqr{|  ()*+,-/38?EFKRSTUVWXYZ[\]`cfghijklmnoprtz|~͵εеҵ׵ڵ۵ܵ޵ߵҶӶն۶#HIKY^abcefgijlmnq7;DHҹֹ !#$&'(+º0149=@ABCDFHIKLMP  !$ȼ˼ #%&()*+,-.013EHuvzڽ۽߽56:<?ABCDEFGHIJKMmo;ξӾվؾھ۾ܾݾ޾߾ !"#$%&')*,68IJOQSUVWXYZ[\]_`boq~οϿ"45FST|}~BCDEFGOPXY]^cdjkoprswxz{~+,349:=>?@%0;<DINSXY`ejoqrz  $).37<AFJOTY]^hmrw{hi  !"#$%&'.6=?ADFHKLNOPU\ahmty !(-4579;@GLSX_dkpw| %)09=DMQX_jqxy  %&27=BJPQegjlqtu>?JMNObghijkt{}~ !469;@CD\^cejop 6:?CJOPQRSTUVWXYdghi|"-4;<OQTV[^_wy~ (*-/478NSY^flmnopqrstuv     478;<?@AUXY\]`ak  ;fgm hi!iy      !"%@ADGKMNPQRSUVWYZ\     $-5;FNOPRV[bcdejqxyz{|}~   7:WXZ[\^_`acdfghil     } "*+,.27>?@AFMTUVWXYZ]`cdefghijklmpqrtuvwyz|}~   ?B[\`abceghjkmnops    ">?@ABCDEFGHIJKLMNOPQRSX ()1234568<AHNOT[\]^_`abcdefilopqrstuvwxy}!#&()*+,-./0235HJdejloqrstuvwxy{|~ !"#$%&'()*+,-./0123456789:;<=>1Lr0_(fBbt&*iGbT}+RU     "SV #ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~  ij~=>?r,JRpt5| O =x}$&'56789AHIJKLMNOV^egilnpstvwx} %,18=DIPU\]_achot{mnos!+5?ABCEPXes!,4CVdq%6P_npqrt| 1=ILMNQ]ft &+,-0:FTdp|~  #.CPiv +7CO_opqt$&'()*+  ? G O V ^ f m n u z                       $ ( , 0 4 8 ; < X [ ^ a e i l m n o ~               ! * 7 C D N V p | }          !()138=BGLQV[`abdinsx} "$%&).38=BGLQST\^chmrw| "',16:?DFGHKPUZ_dinsuv}GHI !#89GHJOTYZ\px &+05:?DINOPQT00000000000000000000000000003000u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 0u 00(0(0(0(0(00100100100100100100100000000000000000000808080808000080000000800080808080@0@0@0@00000@0@0@0104 00fC0fC 0fC 0fC 0fC0fC 0fCfC0K0K0K0K0K0K0K0K0K /0fCfC0k 0k0k0k 0k /0fCfC0o0o0o0o0o0o0o0o 0fCfC0J 00 000 000000( 00000( 00202020202020202020202 00" 0"0"0"0"0".0"30".0"30"0"0"( 0""0[0[0[0[0[0[0[0[0[0[ 0[ 0[ 0[10[0[0[0[ 0[ 0[0[( 0""08 000 000000000000008 00k#0k#0k#0k#0k#0k#0k#0k#0k# 00101010101010101 00I0I0I0I0I 00}U0}U0}U10}U0}U 0}U}U( 0+^+^0S^0S^0S^0S^ 0S^ 0S^ 0S^ 0S^0S^0S^0S^0S^0S^8 0S^S^10uo0uo0uo0uo8 0S^S^0x0x0x0x0x0x( 0+^+^050505050505 0505050505305 0}U}U0( 00Ƨ0Ƨ0Ƨ0Ƨ0Ƨ0Ƨ0Ƨ8 0ƧƧ0*0*0*0*( 00000000( 00000( 000008 00) 0) 0)0)20)0)10)0)0)8 000008 00w0w0w0w0w0w0w10w( 0008 00008 0008 00$0$8 00 *0 *0 *0 *0 * 0 *8 0 0l20l20l20l20l28 00 C0 C0 C8 00K0K0K0K0K 0K0K0K 0K 0K 0K10K0K0K0K0K8 00i0i0i0i0i0i0i0i0i0i0i0i8 0000 0}U}U00 00 004( 0440A0A0A0A0A0A10A8 0AA0ڨ0ڨ30ڨ0ڨ0ڨ8 0AA000000( 0440000 000008 00h0h0h0h0h 00^0^0^0^10^( 0^^0008 007070707070707070707078 008 00!0!0!0!0!0!( 0^^0303( 0^^0T90T90T90T90T90T90T90T90T90T90T90T90T90T90T90T90T90T9 0T90T90T90T90T98 0T9T902[02[02[02[8 0T9T90ug0ug0ug0ug8 0T9T90;v0;v0;v 00ك( 0كك0000000( 0كك0o0o8 0oo008 0oo03008 0oo0T0T0T0T8 0oo0008 0oo000000000 0000( 00^0^0^0^0^0^0^0^ 0^0^0^0^0^( 00%0%0%0%0%0% 0%0%0%0%0%0%0% 00I0I0I0I0I0I0I0I0I 00`0`0`0` 0``0o0o0o 0``0w0w0w0w0w0w0w0w0w0w10w0w0w0w0w 0w0w0w0w0w0w0w0w0w 0``00000 00;0; 0;;0000( 0000000000000( 00101 0;;0./ 0./ 0./ 0.10.0.0.0.0.0.0.0.0.0.0.0.0.0.0.8 0..0u0u0u0u0u0u 0;;05050505050505 0;;0J0J0J0J0J( 0JJ03S03S03S03S103S03S03S 03S03S03S03S03S03S03S03S03S03S03S03S( 0JJ00000000100000 00~0~0~0~0~ 0~~00000000 0~~0( 000( 000( 000 0~~0!`0!`0!`0!`0`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`A0h`0`0`0`0`0`0`0`0`0`0`0`A0`0`000C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0C0 C0 C0C0C0C0 C0 C0C0 C0 C0C0C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0 C0C0C0C0C0C0C0 C0 C0 C0 C0 C0C0C0C0C0 C0C0C0C0C0C0C0C0C0C0C0 C0 C0 C0 C0 C0 C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0(C0C0C0C0C0C0C0(C0(C0(C0(C0(C0(C0C0(C0(C0C0(C0(C0C0(C0(C0C0(C0(C0C0C0C0C00""000 wOy00=Oy00Oy00#0#0Oy00|=@#0Oy00@=Oy0 0Oy0 0@#0@#0Oy0 0N@#0Oy00M_#0 0@0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @10 @10 @0 @0 @0 @0 @0 Gy00G900HT@0 Gy000@0x @0x @0x @0x @0x @0x @0x @0| @0x @0x @0x @0x @0x @0x @0x @0| @0x @0x @0x @0x @0x @0x @0x @0| @0x @0x @0x @0x @0x @0x @0x @0| @0x @0x @0x @0x @0x @0x @0x @0| @0x @0x @0x @0x @0x @0x @0x @0| @0@0@0@0@0@0@0z0@0z0Oy0J00 0ؾ@00@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@00@,0@,0@,0@0@&0@&0@&0@0`&0@&0@&0@0@&0@&0@&0@0@&0@&0@&0@0@00@00@0 /0@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{0{@0{@0{@0{@0{0{@0{@0{@0{@0{@0{ /0@0@0@0@0 /0@0@0@0@0@00@0@0@0@00@,0@,0@,0@0@0@0@0@0`0@0@0@0@0@0@0@0@0@0@0@0@00@00@00@0@0@00@0@0@00@0@0@00@0@0@00@0@0@00@0@0@0@0@0@0@0@0@0@0 /0@0T @0T @0T 0T @0T @0T @0T 0T @0T @0T @0T  /0@0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 @0 @0 0 @0 0 @40 @40 0 @0 @60 @0 @0 @0 @0 @0 @0 0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 0 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 0 @60 @60 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 @60 @0 @0 @0 @0 @0 @0 0 @0 0 @0 @60 @0 @0 @0 @0 @0 0 @0 0 @0 0 @:0 0 @:0 0 @:0 0 @90 @:0 0 @90 @:0 0 @90 @:0 0 @90 @:0 0 @0 @80 0 @90 @:0 @0 0 @90 @:0 @0 0 @90 @:0 @0 0 @90 @:0 0 @90 @:0 0 @90 @:0 0 @90 @:0 0 @90 @:0 0 @90 @90 @:0 0 @90 @90 @:0 0 @90 @90 @:0 0 @90 @90 @:0 0 @90 @:0 0 @90 @90 @:0 0 @90 @90 @90 @:0  /0@90W@90W@:0W0W@90W@90W@90W@:0W0W@90W@90W@90W@:0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@:0W0W@90W@90W@:0W0W@90W@90W@:0W0W@90W@90W@:0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@90W@:0W0W@90W@90W@90W@:0W0W@90W@90W@90W@90W@:0W0W@90W@90W@90W@90W@90W@:0W0W@90W@90W@:0W0W@90W@90W@90W@:0W0W@90W@90W@90W@90W@:0W0W@90W@90W@90W@90W@90W@:0W0W@90W@:0W0W@90W@:0W@0W0W@90W@:0W@0W0W@90W@90W@:0W0W@90W@90W@90W@:0W0W@90W@:0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~0W@0W@60W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W@60W@0W@0W@0W@0W@0W@0W0W@0W0W@0( @0( @0( @0( @0, `0( `0( `0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@0 0W@0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @00W@0 0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W@0W@0W@0W@0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W@0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@K0W@C0W@K0W@C0W@K0W@K0W@I0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@I0W@C0W@I0W@K0W@D0W@C0W@C0W@C0W@C0W@C0W@C0W@K0W@K0W@K0W@C0W@C0W@C0W@K0W@K0W@C0W@C0W@I0W@C0W@C0W@K0W@C0W@C0W@C0W@K0W@C0W@C0W@I0W@I0W@0W0W@0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@C0W@C0W@C0W@C0W@C0W@C0W@K0W@K0W@C0W@C0W@C0W@K0W@C0W@C0W@C0W@C0W@I0W@I0W@C0W@K0W@C0W@C0W@C0W@I0W0W@0W0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W0W@0W@0W0W@0W@0W0W@C0W@C0W@K0W@K0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@I0W@I0W@I0W@C0W@K0W@K0W@C0W@C0W@I0W@C0W@C0W@I0W@C0W@C0W0W@0W@80W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W`0W`0W`0W@0W@0W@0W@0W@0W`0W@0W@0W`0W`0W`0W`0W@0W`0W`0W`0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@80W0W@s0W@r0W@r0W0W@0W@o0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@s0W@r0W@r0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W@0W0W@0W@0W0W@0W@0W0W@0W@0W@0W0W@0W@0W@0W0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W0W@0W0W@00W@0W0W@0W0W@0W0W@0W0W@0W@80W0W@0W@0W0W@0W0W@0W0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W`0W`0W`0W@0W@0W@0W@0W@0W`0W@0W@0W`0W`0W`0W`0W@0W`0W`0W`0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W0W@0W0W@0W0W@0W@0W0W@0W@0W0W@0W@0W0W@0W0W@0W0W@00W@00W@00W@00W@0,0W@00W@00W@00W@00W@00W@00W@00W@0@0p@0p@0p@0p0W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@00W@0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @00W0W0W@0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@00W@00W@0( @0( @0( @0( @0( @0, @0( `0( `0( `0( `0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0, @0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @00W@00W@00W@0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0, @0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @00W@0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0, @0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0(@0(@0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0, @00W@00W@0( @0( @0, @0( @0( @0, @00W@0( @0( @0, @0( @0( @0, @00W@0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @00W@00W@00W@0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( `0( @0( @0( `0( @0( `0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( `0( @0( @0( `0( @0( `0( @0, @00W@00W@0(@0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @0( @0( @0( @0, @00W@00W@0W@80W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W`0W`0W`0W@0W@0W@0W@0W@0W`0W@0W@0W`0W`0W`0W`0W@0W`0W`0W`0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W@0W@0W@0W@0W`0W`0W`0W@0W@0W@0W`0W`0W@0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@80W0W@0W@80W0W@0W@80W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W@0W@0W@0W@0W`0W`0W`0W@0W@0W@0W`0W`0W@0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W`0W`0W`0W@0W@0W@0W@0W@0W`0W@0W@0W`0W`0W`0W`0W@0W`0W`0W`0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@K0@K0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C0@C00W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W0W@0W0W@0W@80W0W@0W@0W@0W@0W@0W@0W@0W@0W@0W`0W@0W@0W@0W@0W`0W`0W`0W@0W@0W@0W`0W`0W@0W`0W`0W`0W@0W@0W@0W`0W`0W`0W`0W`0W`0W`0W`0W`0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W@0W0W@0W0W@0W@0W0W@0W0W@0W@0W0W@0W@0W@0W0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W@C0W0W@00W@0( @0( @0( @0( @0( @0, @0( `0( `0( `0( `0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@00W@0( @0( @0( @0( @0(@0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0(@0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@00W@00W@0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0, @00W@0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @00W0W@00W@0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, `0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W@00W0W@00W@0( @0( @0( @0( @0, @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @0( @0( @0( @0( @0( @0( @0( @0( @0( @0, @00W0"0+++++++++++++++++++++++++++++++++:IIWWWWWWWWWWWWWWWWWWWWWWeeeh #gk+ s!<",#$$%&u'9(@)D*++,-.s/904\:?_EKU{\Zcihm6rvRfpLR| f!.:7EXcsN2 ݺ9X\0p 5&?,1 :DMQ`Xv\bh*nr ty|s>;ϓcFrfҼHb3i3  ".5A5FvJ MDQTX[`Hf4jmoqux_{ b xXh'BY@\d b` #&*z,x/H25d9:=@(BEpINNERLVY\r_chlbq&w{9·_z8y8 f49a4T  *| h%(+/ 59=@dE*KNRGUXw]aFdVgjmnFpsy{ "!:,k!e>+R8:Y 6z #&>) ,/1l3679:<:ACDH{KZOSW[0_$adgk pvy4}BĨ~ [Sh@ ZT-s x } < j j .  I $ ͱ ȶ Ĺ ſ N x  ]  }  d < " & * Z. Q1 7 O: o= A E H oN GR NU [ Z_ b g 'i j l n mp br zt v z J ^ ޕ Û l   & v ? 2  E  D  g  6 b H  V d      J b   *  :  V! *" Z' ) + - / 1 N3 6 8 $9 n($f(x6f9vγLIӼb޿,n BT"I6dw3ʺC),fZ[]aN     !#$%&'()+,-./012456789:;<=>?@ABCEFGHIJKLMNOPRSTUVWXYZ[\]^_`acdefghijkmnopqrstuvwxyz{|}~      !"#$%&'()*+,-.0123456 !#$%&'()*+,-./123456789;<=>?@ABDEFGHJKLMNOQRSTVWXZ[]^`bcefhjkrz}   #%'),/378:;=>ADEFG{| %)+,-258;<?BFJKMOPfknqtvwyz; F#0=?GHR9 ižb #mƊ C 6~\.[( xY  / 6C W e i l o -s x 7{   ] ^ ɋ ʋ  ( a › ܛ  4 U q 6 ? ͝  J  0  8 F ( " < ԯ   Ϻ M 3 ^ h  B f   / B T h { * $   5 J        !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~  .  (  X   ^ "  * F 4 N  z X      L      ` n! ! " $ , f1 1 1 2 :XPԗ^xzȟv^&ަl'A ȬQƯ2lx<ұ(j2Mh} CqͼQX%I_ȾLxǿ+@dz4dmy 3=EGi,JnHR|6`^(`$vjl:j.C`"Zr+ i`Q_޽HҾ]*<U*tZZ5[[[N\\]^]]Ccc+N "*3DQbl/"0:CIPUY\_adgilmnopqstuvwxy{|~   !"$&(*+-.0124569<?@BCz}~     !"#$&'(*./0134679:=>@ACDEGHILNghijlmoprsux{|}~M  & B D E G g y & B E F H h  Jfijl$@CDFf!KgjkmTpstvUqtuw*FIJLl   ,C_bce <?@Bb$'(*JuJfijl?[^_a2NQRTt=z   -`|2NRSUu=5f3OSTVv 1 5 6 8 X !!!!!A!!!!!!!""""#"%"E"\"x"|"}""""""""#0#L#P#Q#S#s#######:$V$Z$[$]$}$$$$$$$%/%3%4%6%V%%%%%%%%&& & &+&L&h&l&m&o&&&&&&&&1'M'Q'R'T't''''''''((((:(w(((((C0k0w0001h111VVVWWWYYY[\$\\]]N]v]]^^^$aLaXaabb>cfcrcIiqi}iw+w7wwwxyz$z {4{@{EQقނƓ˓ѓi>Y^y:U\?Z_.5ӝ"MTZ!({8=-HOs@[b 3:[v}!<CLgntXsz;V[a|b!}!!363;3A3\3a3XXX] ^^ _&_+_f ggmn!no o%omqqqrArMr s3s?sq|||3~[~g~Sz8_kz΍ GS'*ΐOv$0Ǔ2Ye(̕˖ҖؖCO$KW+S_;GLv{!-ũe ԭ{ ]/Wczoκںfܾ&24;A\cag(4 *R^(4*=X_Lw~Yt{8?GS")= e q Xsz>fr,TatR"m"t""&J&W&111<<<<<=2=Z=g=>> ?~???@)@6@@@@AAAyBBBFFF"JJJWJbOOO'UBUIU>ZfZsZ[[\L^t^^ybbbtggglAlNlppppqq&sNs[sxxx|/|<|~*~7~1LSYt{e*129ؚߚ3NSTot{(P\ =eqţѣ=IEQS{?gsŮѮ6B7C)5%MY:F16%PU (-v/Wc".W 3[g+S_Px#@#L#^###1)Y)e)999::":j======[>>>?0?5?????@@@A@M@_@@@ A3A?AAAAABBFBnBzBBBBB C,C}CCCCCCDDDDE(EsEEEEF#FFFFFG G,GTG`GG&H2HxHHHHHHIAIMIgIIIIIISJ{JJJJJKDKPKKKKKKLXLLLLLLL"M.MXMMMENmNyN`OOOOP(PHPpP|PCQ^QeQQQQQRRCRkRwRRRRaSSSST(T2WZWfWWX)XXXXYYYZZZ|[[[\] ]B^j^v^hhhiiijjj8lSlZlLmgmnmmn n27*ELMho:Fa,8Go{<dp,W\b;V]Nin )Q]SnstN v  Iq}  < H |      8"S"Z"###''',(T(`(,,,,--...o///0001/161P3x33|444666788888:::o===>>>@?h?t?x???4A\AhAvCCCLEtEEKKK L$L)LLLLtTTToUUUkVVVWERETE\EpErEEEEEEEIIIIJJ-KUKaK4VDVGVIVeVgVfWvWyW{WWWXXXXXX `` `"`>`@`aaaabbbbbbbbYdidldndddde ee@eEe[eeeeeeeeeeefffffffggghssssssssssttSvcvfvhvvvxxxxxx|zzzzzz{ |||‰?gs 035QSfđƑّ  <>QyvҖ̡ϡѡȢˢ͢'O[Хӥ ;>:JMOkmѲ>NQSorӾ13\loq`                      :::::::  :    :        :::::::  :::                                                      ( ,"$ʆ,I:=$       A@  A5% 8c8c     ?1 d0u0@Ty2 NP'p<'pA)BCD|E||s " 0e@        @ABC DEEFGHIJK5%LMNOPQRSTUWYZ[ \]^_ `abN E5%  N E5%  N F   5%    !"?N@ABC DEFFGHIJK5%LMNOPQRSTUWYZ[ \]^_ `ab@c4 (    a  3  s"*?`  c $X99? a tB  <D?"0@NNN?N  t   '`' 3  s"*? `  c $X99?  '`'sB  6D?"0@NNN?N i'`j'r$  (j; C  S"   ^X99?. `T`T`T`T"` (j;Q  B ? "6@`NNN?N3(:Q H B #   %--&QH  #  d+-p,QH B #  -d+p,QH  #  ((Q   6? "0@NNN?N- %-&Qf k#%l  3  FP+> +B  B ND? "0@NNN?Nk##lQB   ND? "0@NNN?N#$lQB   ND? "0@NNN?N#%lQz  HW!B 3  #" D(h)B B NDg ? "0@NNN?N HW!JQB B NDg ? "0@NNN?N W!QB  NDg ? "0@NNN?N AO!BQ  s *8c "`A&^!P+Q f Cl 3  P++B B ND? "0@NNN?NlQB B ND? "0@NNN?NlQB  ND? "0@NNN?NClQf % 3  C(g)B B NDg ? "0@NNN?N%'QB  NDg ? "0@NNN?NVQ  s *8c "`A&TP+Q f ;Wl 3  LP++B B ND? "0@NNN?N;+lQB B ND? "0@NNN?N+lQB  ND? "0@NNN?N+lQB   ND? "0@NNN?N+WlQz  HW!B !3  #" : D(8 h)B "B NDg ? "0@NNN?N HW!JQB #B NDg ? "0@NNN?N W!QB $ NDg ? "0@NNN?N AO!BQ % s *% "`Y A&P+Q ` L.#\ &#  _ %` # "\ '#  # "\B (B ND? "0@NNN?N# \QB ) ND? "0@NNN?N !\QB * ND? "0@NNN?N "\Qz  HW!B +3  #" B ,B NDg ? "0@NNN?N HW!JQB -B NDg ? "0@NNN?N W!QB . NDg ? "0@NNN?N AO!BQz / c $/ "`L.#Q ` .# 0#  Up,2` o>! 1#  o>!B 2B ND? "0@NNN?No> QB 3B ND? "0@NNN?N > QB 4 ND? "0@NNN?N >!Qz  HW!B 53  #" vB 6B NDg ? "0@NNN?N HW!JQB 7B NDg ? "0@NNN?N W!QB 8 NDg ? "0@NNN?N AO!BQz 9 c $9 "`~.#>Q z = 3 :# S"n ; c $X99?"`= 3 < 0<"`9$= 3  > 0>"`B !#   0 "`h#B< &k  0z  & ?# S"n @ c $X99?"` & 2 A S  ?"6@`NNN?N u  2 B S  ?"6@`NNN?Nk u  2 C S  ?"6@`NNN?NFu 2 D S  ?"6@`NNN?N!u 2 E S  ?"6@`NNN?Nu[ 2 F S  ?"6@`NNN?Nu6 2 G S  ?"6@`NNN?Nu 2 H S  ?"6@`NNN?Nu 2 I S  ?"6@`NNN?Nhu 2 J c $?"6@`NNN?NHu 2 K S  ?"6@`NNN?N#u 2 L S  ?"6@`NNN?Nu]  2 M S  ?"6@`NNN?N!u8" 2 N S  ?"6@`NNN?N#u$  O HGHI?"0@NNN?N k   P HG8!HI8!?"0@NNN?N F  Q HGHI?"0@NNN?N!  R HGHNHIHN?"0@NNN?N  S HGHI?"0@NNN?N[  T HGX{HIX{?"0@NNN?N6  U HGHI?"0@NNN?N  V HGhHIh?"0@NNN?Nh  W HGlKHIlK?"0@NNN?NH  X HG\HI\?"0@NNN?N#  Y HGjHIj?"0@NNN?N  Z HGlHIl?"0@NNN?N] !  [ HGHI?"0@NNN?N8"# 2 \ c $?"6@`NNN?N]  2 ] c $?"6@`NNN?N!8" 2 ^ S  ?"6@`NNN?N#$ R _ NGHI?"0@NNN?Nt  b ` NGHEI?"0@NNN?NO ! b a NGHEI ?"0@NNN?N*"# b bB NZG`A|Hz~I # ?"0@NNN?NO ! b cB NZG`A|Hz~I`Ed?"0@NNN?N*"# 2 d S  ?"6@`NNN?NPX  R eB NZG}HSI}?"0@NNN?Nt 2 f S  ?"6@`NNN?NPz R gB NZGgjHSIgj?"0@NNN?N 2 h S  ?"6@`NNN?N+z R iB NZGBHIB?"0@NNN?NxZ  j Nj ?"6@`NNN?NI  u   k Nk ?"6@`NNN?N$u   l Nl ?"6@`NNN?Nu   m Nm ?"6@`NNN?N9u   n Nn ?"6@`NNN?Nsu   o N o ?"6@`NNN?Nu    p N p ?"6@`NNN?Nk)u    q N q ?"6@`NNN?NFu    r N r ?"6@`NNN?N!u    s N s ?"6@`NNN?Nu    t Nt ?"6@`NNN?Nu   u Nu ?"6@`NNN?N !u   v Nv ?"6@`NNN?N"P#u   w Nw ?"6@`NNN?N<   x Nx ?"6@`NNN?N M!    y Ny ?"6@`NNN?N ju!(   z Nz ?"6@`NNN?N"M#    { N{ ?"6@`NNN?N"jK#(   | N| ?"6@`NNN?N   } N} ?"6@`NNN?N   ~ N~ ?"6@`NNN?NP     6 ?"6@`NNN?N &  2z  ,% # S" n  c $X99?"` ,%Z     7 O 2  S  ?"6@`NNN?N :    6?"0@NNN?N e f   6?"0@NNN?N f   6?"0@NNN?Nf  R  HGMHIM?"0@NNN?NF ;  R  HGMHIM?"0@NNN?NF  Z -%   -% 2  S  ?"6@`NNN?N:  2  S  ?"6@`NNN?NW:  2  S  ?"6@`NNN?N8 r 2  S  ?"6@`NNN?N : /#   HGTHIT?"0@NNN?Nf Wg b  HG0*HXI?"0@NNN?Nd f R  HGHqI?"0@NNN?N-;  R  HG[OH{I[O?"0@NNN?N- b  HG0*H>q  6?"0@NNN?N/ > >q  6?"0@NNN?N9W9q  6?"0@NNN?N9 9q  6?"0@NNN?N33q  6?"0@NNN?N=**q  6?"0@NNN?N/ **q  6?"0@NNN?N% %q  6?"0@NNN?N%$%q  6?"0@NNN?N  q  6?"0@NNN?N q% q  6?"0@NNN?NWk lq  6?"0@NNN?Nk lqB  <D?"0@NNN?N3t$3qB  <D?"0@NNN?N"9t$9qB  <D?"0@NNN?Nq>t$>q2)  A&)" C  S"  ^X99?. `T`T`T`T"` A&)"[z  c $: "` ([ :  6? "6@`NNN?NT[  6; ? "6@`NNN?N v&)"[ ;  6? "6@`NNN?NQL[  6? "6@`NNN?N ![T  C  A[` B c $g  1|;[`  c $g  |]0[`  c $g  |[` B c $g  L(<9 [`  c $g  <([`  c $g  Z=" [` B c $g  '=Zo [` B c $g  -=Zp [`  c $g  ,b [` B c $g  ,b [`  c $g  ,b [` B c $g   R8[`  c $g  8R9[`  c $g  8Rd[` B c $g  mR[`  c $g  R/[`  c $g  R[[`  c $g  !Ee"D[` B c $g  E!w[` B c $g  x[`   c $g  x[z   c $<  "`- #E[ <f    3  Zb -B  B NDg ? "0@NNN?N [B  B NDg ? "0@NNN?N: G H[B  NDg ? "0@NNN?Ne  [l   C  B B NDg ? "0@NNN?N [B B NDg ? "0@NNN?N: G H[B  NDg ? "0@NNN?Ne  [z  c $= "`V"x[ =z   3  #" VB B NDg ? "0@NNN?N [B B NDg ? "0@NNN?N: G H[B  NDg ? "0@NNN?Ne  [z  c $> "`$L[ >   C  #" Q$B B NDg ? "0@NNN?N [B B NDg ? "0@NNN?N: G H[B  NDg ? "0@NNN?Ne  [z  c $? "`A R[ ?z 9Y  3  #" Qx #B B NDg ? "0@NNN?N9Y Z [B   NDg ? "0@NNN?N [t ! S @! "` R[ @z   "3  #"  o B #B NDg ? "0@NNN?N [B $B NDg ? "0@NNN?N: G H[B % NDg ? "0@NNN?Ne  [z & c $A& "`5B,[ A   'C  #" c e5B (B NDg ? "0@NNN?N [B )B NDg ? "0@NNN?N: G H[B * NDg ? "0@NNN?Ne  [ + s *B+8c "`,[ Bl 9Y  ,C  /eB -B NDg ? "0@NNN?N9Y Z [B . NDg ? "0@NNN?N [H /B #  Z|[H 0 #  |<[H 1 #  Z= [H 2B #  (< [H 3 #  <(! [H 4B #  E![T 5 C  B[H 6B #  8, [z (8 7# S"n 8 c $X99?"`(8V 9 0C98c"`7 V C : s *D:"`! V D ; 0E;"`zLV E < 0F<"` LV F = 0G="`L]V G > 0H>"`  t'V H ? 0I?"`Z]'V I @ 0J@"`N>]'V J A 0KA"`d Tu'V K B 0LB"`u'V L C 0MC"`#J(V M D s *ND"`s r)=1V N E s *OE"`)E1V O F 0PF"`)!.1V P G s *QG"`M)=E1V Q H 0RH"`E$p)5(1V R I 0SI"`L"V Sz J c $TJ"`1(8V T K s *UK"`n%]'V UB L  R F&JVB MB  O R# VB NB   OLVB O  JOLVB P  OLVB Q  OLVB RB    VB S   \VB T   VB U  JVB V  JFVB W   JVB XB  \VB YB  VB Z  "VB [B  VB \B  FVB ]  "VB ^B  =&F&p)VB _   t' r)VB `   t')VB aB   u'\r)VB b  \u'E)VB cB  u')VB d  u'E)VB eB   ]'r)VB f  ]'S)VB gB  ]'F)VB h  F]'S)VB iB  E]'")VB jB  S]'")V k 0?"0@NNN?N"F&VP l  "`}=&(Vn m C Vm"`MY c  s *d2"`"8@(FY d  s *e2"`'7->Y e  s *f2"`$i-05Y f  s *g2"`$6%0+Y g  s *h@"`$0#Y h  s *i@"`$0Y i  s *j@"`$ 0^Y j  s *k@"`##Y k  s *l@"`i-#5Y l  s *m@"`7#>Y m  s *n"`8@ FY n  s *o"`kH >MY o  s *p"`#Y p  s *q"` #^Y q  s *r@"`a0kHJ6>MY r  s *s2"`T08@<6FY s  s *t2"`R0796>Y t  s *u2"`1i-=5Y u  s *v2"`16%=+Y v  s *w@"`1=#Y w  s *x@"`1=Y x  s *y@"`1 =^Y y  s *z@"`7kH<>MY z  s *{2"`78@<FY {  s *|2"`67<>Y |  s *}@"`=kHC>MY }  s *~"`kH>MY ~  s *"`8@FY   s *"`7>Y H  # @>M@NYH  # 9>M:NYH  # 9F:kHYH  # 9>98@YH  # F3>MU3NYH  # H3FU3kHYH  # F3>H38@YH  # F3577YH  # 7+7i-YH  # 7#76%YH  # 77YH  # 7^7YH  # %>M%NYH  # %F%kHYH  # %>*8@YH  # *5*7YH  # *+*i-YH  # *#*6%YH  # **YH  # *^*YH  # >MNYH  # FkHYH  # >8@YH  # 57YH  # +i-YH  # #6%YH  # YH  # ^YB   aPbPYB   aPbPYB   (CT0CYB   JJYH  # >MNYH  # FkHYH  # >8@Y  s *"`+5Y   s *"`kH>MY   s *"`8@FY   s *"`7>Y   s *"`+5Y H  #  >M NYH  #  F kHYH  #  > 8@YH  #  5 7YH  # 57YB   JJYB   CCYB   ::YB   00YB   #$Y  s *"`)N/RY   s *@"`)kH/>MY H  # ,>M,NY  s *2"`=8@CFY ZB  S DjJw>GBGYH  # +  YH  # *+ * YH  # 7+ 7 YH  # &*+Y  s *"` 6%*Y   s *"` #Y B   ''YB   d e YH  # &#'6%Y  s *@"`6%#+Y ZB  S DjJ7~6;6YZB  S DjJP*G.GYZB  S DjJ s-uYZB  S DjJ g+~ h+Y  0"`+ #" Y   0"`$+ 0" Y   0"`1+ =" Y   0"` @RCSY   0"`O9R;SY   0"`2R5SY   0"`+RG.TY   0"`(%R'TY   0"`R>TY   0"`iRSY   0"`' RSY  z ( ! # S"n  c $X99?"`( !  0?"0@NNN?NA   0?"0@NNN?NA 5   0?"0@NNN?NVA   0?"0@NNN?NA   0?"0@NNN?NA   0 ?"6@`NNN?N(   0"`f EA    0"`hA    0"`A    0"` A    0"`E%"A    0?"0@NNN?NA %  0"`"(A    0"`  A    0?"0@NNN?NA   0"`   6"`A% &%   0?"0@NNN?NsA%2z 7( (# S"n ) c $X99?"`7(R * T* ?"6@`NNN?N7 %R  + T+ ?"6@`NNN?NT7#R  , 6,3f"`$i P(R  - T- ?"6@`NNN?N d#7(R lz F x ~"2( .# S"n / c $X99?"`F x ~"2(c 0 00"`F x u"c  1 01"`J [c  2 02"`F {c  3 03"`J c  4 04"`J c  5 05"`J .,c  6 06"`J yc  7 07"`J =c  8 08"`  c n 9 C 9"`F ' ~"%c z H(4 <# S"+n = c $X99?"`H(4} F 0FA ? ?#" `&)wh   3  "` : ~ !n h   3  "` o h Z 3 Z"`+ z h M 3 M"`= .z  E 0E"`-(4x n G C G"`E)+v t H S H"`%4((#v t I S I"`n4 (#| h K 3 K"`aPz  J 0J"`i{  L 0L"`j Kz  N 0N"`t"z  P 0P"`nZ?z  Q 0Q"`H!z  R 0R"`  !z  S 0S"`&zz  T 0T"`H<z  U 0U"` )\z  V 0V"`7_z  W 0W"`u!z  X 0X"` !z  Y 0Y"`hF#z n   C  "`I'd|)o n   C  "`'|)o n   C  "`'|)o n   C  "`'|)o n   C  "`X's |)o n   C  "`z ' |)o n   C  "`!'3$|)o    0 "` !F#o    0 "` X !o    0 "` gxn z (H  # S"n   c $X99?"`(HfbR   C jJ"` 8fH   #  fH   #  fH  B # Xf\   3 "`( f\   3 "`0&t&(fbR   C jJ"`0\fbR   C jJ"`4#fbR   C jJ"`%(f\   3 "`'(f\   3 "`@'('f\   3 "`'t<((f\   3 "`@''tfdZ 'tN(   'tN(H  # jJ'N(fH  # jJ'('fZ  S jJ't'(f\   3 "`' 'f\   3 "`&'f\   3 "`&(h'f\   3 "`0't'(f\   3 "`&6'tfZ &'   &'Z  S jJ&h'tfH  # jJ'@'fH  # jJ'('fH  # jJ&t'(f\   3 "`%&f\   3 "`Z&('fZ 2&(&   2&(&H  # jJ2&&fZ  S jJ&(&f\   3 "`tZ (f\   3 "`"v#f\   3 "`"("f\   3 "`V"t #(f\   3 "`""tf\   3 "`  f\   3 "`!!f\   3 "`B!(!f\   3 "`.!t!(f\   3 "` !tfZ 8 !   8 !Z  S jJ8 2!tfH  # jJv!!fH  # jJ!(!fH  # jJ2!t!(f\   3 "` f\   3 "`Z! "f\   3 "`:!X! f\   3 "`""Xf\   3 "`! "f\   3 "`"< " f\   3 "`~! 2"< fDZ !< #   !< #H  # jJh"#fH  # jJh"("fH  # jJh"t"(fH  # jJ!h"tfH  # jJ! !fH  # jJ!Xh" fH  # jJP"h"XfH  # jJP" p"fZ  S jJ!< p" f\   3 "`tf\   3 "` f\   3 "`~X2 f\   3 "`L !Xf\   3 "` f\   3 "`(f\   3 "`t(f\   3 "`6f\   3 "`Rtf\   3 "` f\   3 "`X f\   3 "`p($f\   3 "`t(f\   3 "`Zf\   3 "`jtf\   3 "` f\   3 "`VX  f\   3 "`Xf\   3 "` bfZ    H  # jJfH  # jJ(8fH  # jJt8(fH  # jJHtfH  # jJH fH  # jJ@X fZ  S jJ@Xf\   3 "`(Df\   3 "`t(f\   3 "`Nf\   3 "`Rtf\   3 "` f\   3 "`~X2 fZ  P    PH  # jJfH  # jJ(PfH  # jJtP(fH  # jJHtfZ  S jJ Hf\   3 "`(@f\   3 "`tV(f\   3 "`Rf\   3 "`tf\   3 "`r &f\   3 "`X6 f\   3 "` df\   3 "` <  f\   3 "`RXf\   3 "` < f\   3 "` ^ f\   3 "`b  f\   3 "`$ < f\   3 "` < f\   3 "`B f\   3 "` f f\   3 "`l f4Z H    H H  # jJA J fH  # jJ (J fH  # jJHt (fH  # jJHhtfH  # jJh fH  # jJX fH  # jJh  XfH  # jJh ~ fH  # jJD < ~ fH  # jJD < fH  # jJ fZ  S jJ L fZ        H  # jJfH  # jJ(fH  # jJlt(fH  # jJltfH ! # jJ fH " # jJ XfH # # jJ  ffH $ # jJf< l fH % # jJl < fZ & S jJ  fH ' # jJX f\ (  3 "`(f\ )  3 "`vfZ (* *  (*H + # jJ*fZ , S jJ(*f\ -  3 "`t>(f\ .  3 "`>Xf\ /  3 "`( f\ 0  3 "`<  f\ 1  3 "`2 < f\ 2  3 "`%%f\ 3  3 "`)%(%f\ 4  3 "`$tb%(f\ 5  3 "`f$%tf\ 6  3 "`% %f\ 7  3 "`f$X% f\ 8  3 "`$b%XfTZ $X% 9  $X%H : # jJt%%fH ; # jJ%(%fH < # jJ$t%(fH = # jJ$ t%fZ > S jJ$X% fH ? # jJ$t%tf\ @  3 "`#$f\ A  3 "`$($f\ B  3 "`s#t'$(f\ C  3 "`#H$tfdZ #tu$ D  #tu$H E # jJ9$u$fH F # jJ#(u$fZ G S jJ#t#(f\ H  3 "`f\ I  3 "`&(f\ J  3 "`t_(f\ K  3 "`tf\ L  3 "` f\ M  3 "`X f\ N  3 "`XfTZ X' O  X'H P # jJqfH Q # jJ(fH R # jJtD(fH S # jJ5 qfZ T S jJ5X' fH U # jJDqtf\ V  3 "`f\ W  3 "`(f\ X  3 "`t(f\ Y  3 "`Etf\ Z  3 "`6 fZ  [  H \ # jJ6rfH ] # jJo(rfH ^ # jJto(fZ _ S jJtf\ `  3 "`U f\ a  3 "`(f\ b  3 "`bt(f\ c  3 "`stf\ d  3 "` f\ e  3 "`X f\ f  3 "`bXf\ g  3 "`gf\ h  3 "`+(f\ i  3 "`st'(f\ j  3 "`tf\ k  3 "`w +f\ l  3 "` X f\ m  3 "` f\ n  3 "` (f\ o  3 "`'t(f\ p  3 "`9tfdZ tf q  tfH r # jJeffH s # jJ(efZ t S jJt(f\ u  3 "`f\ v  3 "`(f\ w  3 "`ft(f\ x  3 "`tf\ y  3 "` f\ z  3 "`X f\ {  3 "`fXfTZ <X |  xX;H } # jJfH ~ # jJ(fH  # jJt(fH  # jJ< fZ  S jJ<X fH  # jJtf\   3 "`/f\   3 "`(f\   3 "`t(f\   3 "`tf\   3 "`bX fN  3 jJMfN  3 jJM(hfN  3 jJ>th(f`  c $jJ fN  3 jJ>tf\   3 "`f\   3 "`(Kf\   3 "`vt*(f\   3 "`KtfdZ It   t<H  # jJjfH  # jJI(jfZ  S jJItj(f\   3 "`f\   3 "`!(f\   3 "`tZ(f\   3 "`^tf\   3 "` f\   3 "`0X f\   3 "`ZXfTZ X   XH  # jJl{fH  # jJ({fH  # jJt(fH  # jJl fZ  S jJX fH  # jJltf\   3 "` f\   3 "`(fN  3 jJ1mfN  3 jJ (mf\   3 "` B f\   3 "` (Q f\   3 "`" t (f\   3 "`4  tf\   3 "` B f\   3 "` Xy f\   3 "`P  XfN  3 jJ  fN  3 jJ| ( fN  3 jJ| t (f`  c $jJ  fN  3 jJ  tf\   3 "`'  f\   3 "`( f\   3 "` t (f\   3 "` tf\   3 "`'  fZ 6     6  H  # jJE  fH  # jJE (` fH  # jJ6 t` (fZ  S jJ6  tf\   3 "`S  f\   3 "` (C f\   3 "`} t1 (f\   3 "`  tfdZ b t    b t H  # jJ  fH  # jJ ( fZ  S jJb t (f\   3 "` | fZ "    " H  # jJ8X fDZ "    " H  # jJ*fH  # jJ*(@fH  # jJt@(fH  # jJ(tfH  # jJ( 8fH  # jJXfH  # jJ2 fH  # jJ< 2 fZ  S jJ" < flB  B 0DԔ(f\   3 "` vfZ      H  # jJ7fH  # jJ(7fH  # jJt(fH  # jJ( UfH  # jJXU fH  # jJ(tfZ  S jJ IfH  # jJXf\        !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdfghijklmnopqrstuvwxyz{|}~ 3 "`uXfZ    H  # jJfH  # jJ(fH  # jJvt(fH  # jJ gfH  # jJvtfZ  S jJXfH  # jJgX fV2   # "`X  fV2   # "`fV2   # "`X4 fV2   # "`&'DflB  0DԔFGf   6 "`Ff    0 "`_ f H  B # {fH   # {Xf   6 "` f    0 "`Vf V   # CA"` f   0 "`(,f H  B # {r fH   # {&fnR   c $33jJ"` enR   c $33jJ"`n enR   c $33jJ"`  enR   c $33jJ"` Oe\   3 "`z .f\   3 "`~ 2tf\   3 "` X f\   3 "`  tf\   3 "` _ fN  3 jJ  fN  3 jJ ( f`  c $jJ& U fN  3 jJ% U tf\   3 "`  f\   3 "` (k fZ  S jJ t (f\   3 "`k t(f\   3 "` tK (fN  3 jJ t% (f\   3 "` ^eZ  S jJ te   0 "`"b(zd    0 "`zd \R F  3 jJ"`2^z (&6 " # S"n #  c $X99?"`(&6O $  0$ "`(0Q  %  0% "`0( 4O t  %- '  S"n &  c $X99?"` %-d (  0( "` %c  )  0) "` %d t ] }( +  S"n *  c $X99?"`] }(~ ,  0, "`] }( }  -  0- "`] M }(| z + K( . # S"n /  c $X99?"`+ K(x 0  00 "`+ K( x  1  01 "`+ K(x z h(0 Z # S"n [  c $X99?"`h(0A \ 0?"0@NNN?N8A yA ] 0?"0@NNN?NyA A ^  0^ ?"6@`NNN?N%(h.A  _  0_ "`H'A A  `  0` "`0 A A  a 0?"0@NNN?Ny[&A b  0b "`0#([A  c  0c "`H XA  d 0?"0@NNN?N XyA e  0e "`JA  f  s *f "`P1@ A  g 0?"0@NNN?N XH 1A h  s *h "` A  i 0?"0@NNN?N8A L A j  0j "`@h A  k 0?"0@NNN?NA hA l  0l "`$1'A  m 0?"0@NNN?N%[&1A n  6n 3f"` P$A  o  6o 3f"`8 HA  p  6p 3f"`H"A  q  6q 3f"`X 8($A P r   "`(Hh hA s  6s "`0 A P t   "`PA u  6u "`nA P v   "` (A w  6w "`j`A P x   "`$K3'kA y  6y "`e%-&A TB z  C Da ATB {  C D""ATB |  C D"l "ATB }  C Dx#{ATB ~  C D}ATB   C Dc# % ATB   C DS ATB   C DAhz  XY^  # S"n   c $X99?"` XY^@   N ?"6@`NNN?N XYY@    6 "` YY]@  z a#$-  # S" n   c $X99?"`a#$-;   <  "` 0#/; B   gi;B   i;B   hiD;B   h.;B   YWh;   s * "`9&a#$-;    s * "`~.2#;    <  "`S l;    <  "`;    s * "`20#;    <  "` /; B   YW1;   <  "`7 ; B   !  ;B   Y!  ;B   ;B   W!;B   W;   s * "`Pa#U%;    s * "`Q%; >z  m?x<  # S"!n   c $X99?"` m?x<9   0 ?"6@`NNN?N <,9    s * "` -m?":9 Iz  N%m:  # S"$n   c $X99?"` N%m:7   H ?C"` [" #7    H ?C"`? l 7    H ?C"`> *j +7    H ?C"`l7    H ?C"`,7    H ?C"`pZ"#7    H ?C"`Z"0#7    H ?C"`*+7    H  ?C"`%'7     H  ?C"`Z"#7     H  ?C"`(Z"T#7     H  ?C"`Z"#7     H  ?C"`(\$T%7     H ?C"`.%['7    H ?C"`%'7    H ?C"`X%'7    H ?C"`7    H ?C"`,X7    H ?C"`7    H ?C"`f7    H ?C"`7    H ?C"`(T7    H ?C"`7    H ?C"`O{7    H ?C"`7    H ?C"`u !7    H ?C"`H7    H ?C"`O{7    H ?C"`**+7    H ?C"`($*7    H ?C"`*+7    H  ?C"`&*R+7     H! ?C"`*+7 !   H" ?C"`-9.7 "   H# ?C"`O{ 7 #  0?"0@NNN?N ? "7   0?"0@NNN?N "> _+7  0?"0@NNN?Nl kl7   0?"0@NNN?Nkl7  0?"0@NNN?N,k7   0?"0@NNN?N,kl7   0?"0@NNN?N,k7  0?"0@NNN?Nj,k7   0?"0@NNN?NXjk7   0?"0@NNN?Nl p"7   0?"0@NNN?N""7   0?"0@NNN?N0""7   0?"0@NNN?N0"w&7   0?"0@NNN?N"("7   0?"0@NNN?NT""7  0?"0@NNN?N$(w&7   0?"0@NNN?Nw&.x&7  0?"0@NNN?N[u&w&7   0?"0@NNN?Nu&Xv&7  0?"0@NNN?N0?"7   0?"0@NNN?N?(@7   0?"0@NNN?NT?@7  0?"0@NNN?N~O?7   0?"0@NNN?N?O@7   6?"0@NNN?N?O 7   0?"0@NNN?N{?@7   0?"0@NNN?N?u @7  0?"0@NNN?Nj W+_+7   0?"0@NNN?NW+^+7  0?"0@NNN?N*)^+7   0?"0@NNN?N*^+_+7   0?"0@NNN?N^+&_+7   0?"0@NNN?NR^+_+7   0?"0@NNN?N*^+-7"   <jJ?"6@`NNN?N"dv7"   <Ԕ?"6@`NNN?N> 7"   0?"6@`NNN?NW7"   68c?"6@`NNN?NN7"   <jJ?"6@`NNN?N#'7"   <jJ?"6@`NNN?Nq(]i,7"   <Ԕ?"6@`NNN?N'.7"   <jJ?"6@`NNN?Nk"I7"   <Ԕ?"6@`NNN?NZ"U!7"   0?"6@`NNN?NDy #7"   68c?"6@`NNN?Nr#!7   T$0e 8c?3"``, 17 $   N%0e ?3"``M3 u87 %   Z&pG jJ?3"`"M3%u87 &   Z'0e Ԕ?3"`",%17 '   6?"0@NNN?N1 F/!G/7   6?"0@NNN?N 5!57   0?"0@NNN?N#1#A37   0?"0@NNN?N@1AM37  B 0?"0@NNN?Ny1@X27   0?"0@NNN?N@1]K27  B 0?"0@NNN?Nu8@97   0?"0@NNN?N@u8}97  B 0?"0@NNN?N"1#u27   0?"0@NNN?N#1%S27  B 0?"0@NNN?N"u8#u97   0?"0@NNN?N#u8~$97B  B <D?"0@NNN?N /!/7B  B <D?"0@NNN?N "5"#57B  B <D?"0@NNN?N!*6"+67B  B <D?"0@NNN?N4!5`"57B  B <D?"0@NNN?N+6`,67B  B <D?"0@NNN?N4E/`F/7B  B <D?"0@NNN?Nq60E707   <( ?"6@`NNN?N L/4m:7 (Pz  X[R_ 6 # S"n 7  c $X99?"` X[R_> 8  68 ?"6@`NNN?N X[Y>  9  69 "` W[R_> z (7  # S"7n   c $X99?"`(7:t   S g >"`M(4: gn   C h "`l(<: hn   C i "`4(h7: it   S j >"`'(l: jz *)8  # S"8n   c $X99?"`*)8;t   S k >"`9(l; kn   C l "`Hw)o; ln   C l "`l'; lh   3 m >"`:p#; mh   3 n >"`2Zp#; nh   3 o >"`t+0; on   C p "`7#(; pn   C q "``07; qz (-  # S"9n   c $X99?"`(-;z   c $r@ "`'&#; rn   C s "`&#; sT   # t  (); tn   C u "`@)(+; uhz  XX?0  # S"n   c $X99?"` XX?0Z   Nx ?"6@`NNN?N XX?u*Z x   6v "` %X?t*Z vhz  XX?7  # S"n   c $X99?"` XX?77   N| ?"6@`NNN?N XX?u*7 |   6{ "` (X?27 {hz  d?&9  # S""n   c $X99?"` d?&9   N} ?"6@`NNN?N  d?% }   6z "` %d?5 zhz  Xd?]N  # S"#n   c $X99?"` Xd?]N5   Nw ?"6@`NNN?N Xd?&95 w   6y "` 7d?G5 yt  %G+   S"%n   c $X99?"` %G+\   0 "` %"[    0 "` ="%w&Z hz  s X?5  # S"&n   c $X99?"` s X?5/   N ?"6@`NNN?N V?2/    6 "` -X?5/ z b'   # S"'n   c $X99?"`b' 2b  HG"HIR?"0@NNN?NIL2b  HG*H I? "0@NNN?NP! $22   0  ? "6@`NNN?N 2R  HGHqI? "0@NNN?N2} Ns 2R   HG[OH{I[O? "0@NNN?N2N2b   HG0*H NI ?"0@NNN?N!|$2   0 ?"6@`NNN?N4b'2 t -   #  #"  # 2   S  ?"6@`NNN?N:  1R  BGHqI?"0@NNN?N-;  1R   BG[OH{I[O?"0@NNN?N- 1Br    Z!/   0 "`&/ 2   c $ ? "6@`NNN?NC /b   HG0*H NI ?"0@NNN?N L N/t -   # #"  TwD 2   S  ?"6@`NNN?N:  /R  BGHqI?"0@NNN?N-;  /R   BG[OH{I[O?"0@NNN?N- /B  B  LN/B  B  wNCO/t -   #  #" 6& 2   S  ?"6@`NNN?N:  .R  BGHqI?"0@NNN?N-;  .R   BG[OH{I[O?"0@NNN?N- .   0 "` P. H  #    -Br     @ )   0 "`%( H  # @  (z  %.  # S"(n   c $X99?"` %.G   0 "` %C(G    0 "` '%+G pz (;  # S")n   c $X99?"`(;X   N; ?"6@`NNN?NdP+#p,X ;   N: ?"6@`NNN?Nd$#%X :   N9 ?"6@`NNN?N<$ 'X 92   s *?"6@`NNN?N#&X   N8 ?"6@`NNN?N*-X 8   N7 ?"6@`NNN?NX 7   N6 ?"6@`NNN?NX 6   N5 ?"6@`NNN?Nx!"X 5   N4 ?"6@`NNN?N`'x*X 42   s *?"6@`NNN?Nx*H-Xr   s *?"6@`NNN?N#H-X   N2 ?"6@`NNN?N4'(X 2   6?"0@NNN?N%&lX   6?"0@NNN?NH-h.X !  61! ?"6@`NNN?N4(9X 12 "  s *?"6@`NNN?N')X2 $  s *?"6@`NNN?NX2 %  s *?"6@`NNN?NX X &  6?"0@NNN?N&)'*X '  6?"0@NNN?N&&''X (  6?"0@NNN?Nb"c!$X )  6?"0@NNN?N& 'GL$GL$  q  wO >   <=GL$GL$  KdO =   6<  "`= &w()O <t ' "G%@   S"-n   c $X99?"`' "G%@[   B~  ?"6@ NNN?N' "G%?Y ~   B)  ?"6@ NNN?N' ?G%@W )z  !% 6  # S",n   c $X99?"` !% 6+   B  ?"6@ NNN?N !%2.+    B  ?"6@ NNN?N -%0) n   C  "` -%j2* Vz  %  # S"1n   c $X99?"` %w   0f "` % w f   Be  ?"6@ NNN?N O %Vw eVz  H%  # S"0n   c $X99?"` H%   0 "` H%    B  ?"6@ NNN?N % z  f#  # S"/n   c $X99?"` f#`   N ?"6@`NNN?N f#n `    N ?"6@`NNN?N n # ` z  r &4  # S"3n   c $X99?"` r &4#   0 "` & )#    0yZ<yZ< "` w&&4# z H(  # S"4n   c $X99?"`H(8   0 "`H( 8    0 "`-H(8 z H(45  # S"5n   c $X99?"`H(45-   0` "`" !*- `   0_ "`)!,.- _n   -c.  # 3"n   c $X99?"`  -c.   <?"6@`NNN?N  -c.n   -c. # 3"n  c $X99?"`  -c.  <?"6@`NNN?N  -c.n @8A?= # 3"n  c $X99?"`@8A?=+  <?"6@`NNN?Nq$;A?=*n (c. # 3"n  c $X99?"`(c.#  <?"6@`NNN?N  -c.#n )c.  # 3"n   c $X99?"`)c.    <?"6@`NNN?N  -c. t ' G%$   S"n   c $X99?"`' G%$n   -c. # 3"n  c $X99?"`  -c.  <?"6@`NNN?N  -c.B S  ?;/99%/ ABOBCPCDQ DER EFS FGT GHU HIVIJWJKXKLYLMZMN[K_\]`]^a\]b]^cKdeJfgJhi !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQ+/R0S+1T2U 3V 4W&+5X!YZ[\]^+_+`+a&!6b&c&d&e!f!g!hijk l mn o:CLp:9Mq9<Nr;9Os=9PtI9Qu<>RvA<SwB<Tx?;Uy@;Vz>;W{=AX|=?Y}=KZ~IB[I@\IK]CH^>D_>E`ADaAGbBEcBGd?De?Ff@Eg@FhKGiKFjKCk}|{x~~wuvvwvu+ , - . / 0 1 2 3 4 5 6 7 8 9 : ;  <  = > ? @ A B C  D  E  F  G  H I J K L M N O P Q R S T U V W X Y  Z  [   \  ]  ! ^  " _  # `  $ a  % b & c  d  e  f  g   h  ' i) ( + j( - , k. / l/ 0 m0 1 n . o2 3 : p3 4 ; q4 5 < r6 7 = s7 8 > t5 6 ? u@ A E vA B F wB C G xH I P yI J Q zJ K R {L M S |M N T }K L U ~V W \ W X ] X Y ^ Y Z _ ` a a b b c d e e f c d g h h i i j k l j k m n r n o s o p t u v } v w ~ w x  y z z { x y                                   1  f  l        l      e _ \ e ` ] e b a e c d f c g h _ i j ` k l b m                                                               9 : ; < = E F M N O Q S T U W X Y$ %  Z , ["  & \ " ' ]6  ( ^% 6 ) _* $ + b6  5 c% 6 7 d$ % 8 e " 9 f"  : gZ ] h\ ^ i[ _ j a b k ` c l d f m e g nk j n om i o pl i p qh i q rj h r ss l v tt m w uu k x vy s | w{ t } xz u ~ y z { | } ~   =yUd~s( W 4ak2O:)kGTؐ%UA_t8yUxFx&D$`e{}S \ 4|4L64 ! x\Hw4 \ <-24 ~}4^ :/U64:!2 ?!4 ![ @ Ppth! !2@@t@t!(@!1 @ 7!2@ " " !0 pt#)2 ?' ! + <!2 . !! !d Z !- (Y2! .!(# 6 3"  !  !  R2!  T!2 !  !0  !  !2 ( !t"  X !2  Z   h!(  $ !/ @ X1# <!2 ! !  !# G h0(#  Ov  !l  ! O !2  C!0  C!  lC!2 W !2 @# "2  %"2  !0(  _Ref87849420 _Ref88462760 _Toc217315416 _Toc217315417 _Ref99174560 _Ref192390534 _Ref192410500 _Ref192516316 _Toc217315418 _Toc217315419 _Ref195281412 _Ref195281471 _Ref195281532 _Ref212901048 _Ref212901089 _Toc217315420 _Ref192388754 _Ref192489643 _Toc217315421 _Toc217315422 _Toc217315423 _Toc217315424 _Toc217315425 _Ref212623558 _Toc217315426 _Ref212696175 _Toc217315427 _Ref212713404 _Toc217315428 _Ref212623577 _Ref212954957 _Toc217315429 _Ref212891806 _Toc217315430 _Ref212891829 _Toc217315431 _Ref212623601 _Toc217315432 _Ref99513436 _Toc195349908 _Ref212623620 _Ref212891998 _Toc217315433 _Ref191352940 _Toc195349909 _Toc217315434 _Ref87944071 _Ref97712489 _Toc195349910 _Toc217315435 _Ref192582713 _Toc195349911 _Toc217315436 _Ref89494171 _Ref193859274 _Toc195349912 _Toc217315437 _Ref192388888 _Ref192653580 _Ref192755263 _Ref192932146 _Toc195349913 _Ref214085008 _Toc217315438 _Ref214372949 _Toc217315439 _Ref214372965 _Toc217315440 _Toc217315441 _Ref192653564 _Toc195349915 _Toc217315442 _Ref193274112 _Toc195349916 _Ref214336453 _Toc217315443 _Toc217315444 _Ref191352961 _Ref191977252 _Toc195349917 _Toc217315445 _Ref215190588 _Toc217315446 _Ref215204016 _Ref215240305 _Ref215326291 _Ref215396596 _Ref215464223 _Ref215812367 _Ref215820051 _Toc217315447 _Ref215464457 _Ref215812423 _Toc217315448 _Ref190686436 _Toc195349918 _Toc217315449 _Ref190686481 _Toc195349919 _Toc217315450 _Ref215319774 _Ref215375150 _Ref215377056 _Ref215388958 _Ref215840852 _Ref216074389 _Toc217315451 _Ref215319787 _Ref215375180 _Ref215377063 _Ref215388962 _Ref215396791 _Toc217315452 _Ref215396639 _Ref215575354 _Toc217315453 _Ref190686503 _Toc195349921 _Toc217315454 _Ref162423936 _Ref190768256 _Toc195349922 _Toc217315455 _Ref190771388 _Toc195349923 _Toc217315456 _Toc195349925 _Ref215812860 _Ref216169257 _Toc217315457 _Ref191979081 _Toc195349924 _Ref216448569 _Ref216448572 _Toc217315458 _Ref215980624 _Toc217315459 _Ref217137097 _Toc217315460 _Ref191974946 _Toc195349926 _Toc216088692 _Toc217315461 _Ref216322123 _Toc217315462 _Toc217315463 _Toc217315464 _Ref191974173 _Toc195349927 _Toc216088694 _Toc217315465 _Ref216925536 _Toc217315466 _Ref216926683 _Toc217315467 _Ref216926690 _Toc217315468 _Toc216088693 _Ref216925572 _Ref216955967 _Ref217017701 _Ref217137149 _Toc217315469 _Ref216858117 _Toc217315470 _Ref216925677 _Toc217315471 _Toc217315472 _Ref217137370 _Toc217315473 _Toc195349928 _Toc216088695 _Toc217315474 _Ref193688215 _Toc195349929 _Toc216088696 _Toc217315475 _Toc195349930 _Toc216088697 _Toc217315476 _Toc217315477 _Toc217315478 _Toc217315479 _Toc217315480 _Toc195349931 _Toc217315481 _Toc195349932 _Toc217315482 _Toc217315483 _Ref213570732 _Ref192490406 _Ref193208912 _Ref211051873 _Ref95885829 _Ref95885852 _Ref69705490 _Ref69705424 _Ref69705450 _Ref87327235 _Ref192403460 _Ref192516178 _Ref192774275 _Ref192925307 _Ref192997341 _Ref193104570 _Ref213230715 _1286960845 _1286960887 _1286960960 _1286967045 _1286967120 _1286967217 _1287571686 _1287571954 _1287572782 _1287572888 _1287573458 _1287573602 _1287573814 _Ref213917055 _Ref214525226 _Ref214350828 _Ref214443964 _Ref214461298 _Ref215196170 _Ref195258155 _Ref215319995 _Ref190840539 _Ref191194358 _Ref191198647 _Ref216323373 _Ref215989076 _Ref191364304 _Ref195279807 _Ref217290313 _Ref216952908 _Ref216591034 _Ref215296727 _Ref215471559 _Ref215456615 _Ref215290439 _Ref215484430 _Ref215541334 _Ref215571062 _Ref215574539 _Ref216175089 _Ref216175094 _Ref216326495 _Ref216410304 _Ref217013137 _Ref216707798 _Ref217122670w (((((30iDiDiDiDiDiDLLLlpM55%%%^^22JJJJJVVV_._._._V_V_V_s888ɨɨ77DDDDDDDDaaa444444W:W:W:܄܄܄rrraabbb&&JJaaaappx >>RU4411111166KK6T$kkLLRLT V VVYY>Zd_` gelq]v             ݁hd'ϖ!!SvF|@ii ''+ IT  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~@@@@@@@@@@@@@ (B0uDuDuDuDuDuDuDuDuD#L#L#Lllpc WWXhh%&33JKKKKVVVU_U_U_U_]_]_]_sGGG'''''33331ȕhh}}}}}}}}<<<<FFF 5 5 5 5 5 5:::   &&JJaaaaqqx2qqqqkkBBBBBB66KKsTˀџџџџXww/MORTDVVVOYYyZ_8a\glrv             c \ǣ5 {H33bMg (TINumbered Styles 2Numbered Styles 3Numbered Styles 4Numbered Styles 5Numbered Styles 6Numbered Styles 7Numbered Styles 1Drawn Objects and Text Boxes 2 Text Boxes 2Drawn Objects and Text Boxes 3Drawn Objects 138Drawn Objects 139Drawn Objects 140Drawn Objects 141Drawn Objects 142Drawn Objects 143Drawn Objects 144Drawn Objects 145Drawn Objects 146Drawn Objects 147Drawn Objects 148Drawn Objects 149Drawn Objects 150Drawn Objects 151Drawn Objects 152Drawn Objects 153Drawn Objects 154Drawn Objects 155Drawn Objects 156Drawn Objects 109Drawn Objects 110Drawn Objects 111Drawn Objects 112Drawn Objects 113Drawn Objects 114Drawn Objects 115Drawn Objects 116Drawn Objects 117Drawn Objects 118Drawn Objects 119Drawn Objects 120Drawn Objects 121Drawn Objects 122Drawn Objects 123Drawn Objects 124Drawn Objects 125Drawn Objects 126Drawn Objects 127Drawn Objects 128Drawn Objects 129Drawn Objects 130Drawn Objects 131Drawn Objects 132Drawn Objects 133Drawn Objects 134Drawn Objects 135Drawn Objects 136Drawn Objects 137Styles 1Drawn Objects 1Drawn Objects 2Drawn Objects 3Drawn Objects 4Drawn Objects 5Drawn Objects 6Drawn Objects 7Drawn Objects 8Drawn Objects 9Drawn Objects 10Drawn Objects 11Drawn Objects 12Drawn Objects 13Drawn Objects 14Drawn Objects 15Drawn Objects 16Drawn Objects 17Drawn Objects 18Drawn Objects 19Drawn Objects 20Drawn Objects 21Drawn Objects 22Drawn Objects 23Drawn Objects 24Drawn Objects 25Drawn Objects 26Drawn Objects 27Drawn Objects 28Drawn Objects 29Drawn Objects 30Drawn Objects 31Drawn Objects 32Drawn Objects 33Drawn Objects 34Drawn Objects 35Drawn Objects 36Drawn Objects 37Drawn Objects 38Drawn Objects 39Drawn Objects 40Drawn Objects 41Drawn Objects 42Drawn Objects 43Drawn Objects 44Drawn Objects 45Drawn Objects 46Drawn Objects 47Drawn Objects 48Drawn Objects 49Drawn Objects 50Drawn Objects 51Drawn Objects 52Drawn Objects 53Drawn Objects 54Drawn Objects 55Drawn Objects 56Drawn Objects 57Drawn Objects 58Drawn Objects 59Drawn Objects 60Drawn Objects 61Drawn Objects 62Drawn Objects 63Drawn Objects 64Drawn Objects 65Drawn Objects 66Drawn Objects 67Drawn Objects 68Drawn Objects 69Drawn Objects 70Drawn Objects 71Drawn Objects 72Drawn Objects 73Drawn Objects 74Drawn Objects 75Drawn Objects 76Drawn Objects 77Drawn Objects 78Drawn Objects 79Drawn Objects 80Drawn Objects 81Drawn Objects 82Drawn Objects 83Drawn Objects 84Drawn Objects 85Drawn Objects 86Drawn Objects 87Drawn Objects 88Drawn Objects 89Drawn Objects 90Drawn Objects 91Drawn Objects 92Drawn Objects 93Drawn Objects 94Drawn Objects 95Drawn Objects 96Drawn Objects 97Drawn Objects 98Drawn Objects 99Drawn Objects 100Drawn Objects 101Drawn Objects 102Drawn Objects 103Drawn Objects 104Drawn Objects 105Drawn Objects 106Drawn Objects 107Drawn Objects 108 Text Boxes 1 OLE Objects 1 OLE Fields 95 OLE Fields 1 OLE Objects 2 OLE Fields 94 OLE Objects 3 OLE Fields 93 OLE Objects 4 OLE Fields 92 OLE Objects 5 OLE Fields 91 OLE Objects 6 OLE Fields 90 OLE Objects 7 OLE Fields 89 OLE Objects 8 OLE Fields 88 OLE Objects 9 OLE Fields 87OLE Objects 10 OLE Fields 86OLE Objects 11 OLE Fields 85OLE Objects 12 OLE Fields 84OLE Objects 13 OLE Fields 83OLE Objects 14 OLE Fields 82OLE Objects 15 OLE Fields 81OLE Objects 16 OLE Fields 80OLE Objects 17 OLE Fields 79OLE Objects 18 OLE Fields 78OLE Objects 19 OLE Fields 77OLE Objects 20 OLE Fields 76OLE Objects 21 OLE Fields 75OLE Objects 22 OLE Fields 74OLE Objects 23 OLE Fields 73OLE Objects 24 OLE Fields 72OLE Objects 25 OLE Fields 71OLE Objects 26 OLE Fields 70OLE Objects 27 OLE Fields 69OLE Objects 28 OLE Fields 68OLE Objects 29 OLE Fields 67OLE Objects 30 OLE Fields 66OLE Objects 31 OLE Fields 65OLE Objects 33 OLE Fields 63OLE Objects 34 OLE Fields 62OLE Objects 35 OLE Fields 61OLE Objects 36 OLE Fields 60OLE Objects 37 OLE Fields 59OLE Objects 38 OLE Fields 58OLE Objects 39 OLE Fields 57OLE Objects 40 OLE Fields 56OLE Objects 41 OLE Fields 55OLE Objects 42 OLE Fields 54OLE Objects 43 OLE Fields 53OLE Objects 44 OLE Fields 52OLE Objects 45 OLE Fields 51OLE Objects 46 OLE Fields 50OLE Objects 47 OLE Fields 49OLE Objects 48 OLE Fields 48OLE Objects 49 OLE Fields 47OLE Objects 50 OLE Fields 46OLE Objects 51 OLE Fields 45OLE Objects 52 OLE Fields 44OLE Objects 53 OLE Fields 43OLE Objects 54 OLE Fields 42OLE Objects 55 OLE Fields 41OLE Objects 56 OLE Fields 40OLE Objects 57 OLE Fields 39OLE Objects 58 OLE Fields 38OLE Objects 59 OLE Fields 37OLE Objects 60 OLE Fields 36OLE Objects 61 OLE Fields 35OLE Objects 62 OLE Fields 34OLE Objects 64 OLE Fields 32OLE Objects 65 OLE Fields 31OLE Objects 69 OLE Fields 27OLE Objects 70 OLE Fields 26OLE Objects 72 OLE Fields 24OLE Objects 74 OLE Fields 22OLE Objects 75 OLE Fields 21OLE Objects 76 OLE Fields 20OLE Objects 83 OLE Fields 13OLE Objects 84 OLE Fields 12OLE Objects 85 OLE Fields 11OLE Objects 86 OLE Fields 10OLE Objects 88 OLE Fields 8OLE Objects 89 OLE Fields 7OLE Objects 90 OLE Fields 6OLE Objects 91 OLE Fields 5OLE Objects 92 OLE Fields 4OLE Objects 93 OLE Fields 3OLE Objects 95Drawn Objects and Text Boxes 1fff  q q 6 6 a a     , , t t       " " D D k k       EEaa==Z?Z?v?v???xzxzzz{{E{E{{{||||1}1}}}  ??##лл""##??;;..dd((FF::uu[[llxxxxxxxx#y#y7y7yMyMyiyiyyyzzz/T  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHfff  q q 6 6 a a     , , t t       " " D D k k       EEaa==Z?Z?v?v???xzxzzz{{E{E{{{||||1}1}}}  ??##лл""##??;;..dd((FF::uu[[llxxxxxxxx#y#y7y7yMyMyiyiyyyzzz/Tt??u??|lv??lw??lx????|>M?>M?>M?M?N???@BACEDFHGIJKMLNPOQSRTUVWXY[Z\^]_a`bdcegfhjiklmnpoqrsutvxwy{z|~})")-)-)HNNۭls~~2;; jss@NN  CSSJUUaoo(44  &&IYYCLL$$qzz[aa7GGEEbO]]            S c c q                x:x:OOOOtPtPPPOQOQhhT     "!#$%'&()*,+-/.0213546879;:<=>?@BACEDFHGIJKMLNPOQSRTUVWXY[Z\^]_a`bdcegfhjiklmnpoqrsutvxwy{z|~}8O*urn:schemas-microsoft-com:office:smarttagsdateB*urn:schemas-microsoft-com:office:smarttagscountry-region9*urn:schemas-microsoft-com:office:smarttagsState8*urn:schemas-microsoft-com:office:smarttagsCity=*urn:schemas-microsoft-com:office:smarttags PlaceName=*urn:schemas-microsoft-com:office:smarttags PlaceType9*urn:schemas-microsoft-com:office:smarttagsplace 1012172008DayMonthYearOO  Xٺ :y  Q ove  _ s % {ҀIX ^.2/ {րQe ^62//7   *!*"*)*@*D******1-7-//666698;8C8E888AAIBKBGG3O:OkkkkkkkkcynyDzOzTz_z ځ܁ڂۂqu!*ˆ͇ԇǓȓZ[VW[\/0',+1ݟNO"#+09: KRsy~ IJ5@=ERW-2\] 45wx=>hituWX!  !'9?AGZ`6<ek39 ) 1 k s R Z w %7B'rz %NY%0OWQW_j4<}~V ^ ~!!"">"I"j#r#####v$~$$$$$%%''''((`*e*V,\,--..//\2a27383]3^344440565;5C5 6666666677n8t8y8888889999999999; ;r;x;};;;;@@@@@@@@@@@@@AAA}BBBBBBBB/Cqt38df"'W]YZxyuv9:#$tubh06S\%+\bFL ]c')W^      n"o"m&s&&&''((O(V((()),,x-~-.!.../////02"2a2f2o2s22222q4444444;5O5U58899b:h:;;<<<<b=c=>>@?R???1@2@@@AABBFFRJSJJJJJJJKKKK LLLLOOCUDUGXKXO\V\\\\\]]|^}^````bbggggginijjjjIlJlmmppq qxxJU'2́MNuv +,34ٚښOPpqӟܟ<s67MP  nrotvz ³:=>UVefgƴɴ9<PSmsX[23QR)*ek%)/I-4ek79\^wy{}  (*)+02XZ_auw|*/qsy{ INvzFJ /4EJ##?$F$y$$$$$$$$ %%l%u%w%%%%%%%%X&`&T+\+%,(,-,6,d-l-p-y-------99::====1?2??? @ @DDgDjDDDEEHE.F1FhOhiiTlUlhmimnnqqssvvzzzzzz{{B{F{f{o{E|G|H|M|Q|||||~ ~~$~)~6~~~~~ADnw<Bˀ΀!$.fmNSۆ܆~lqOTz֙ڙܙޚ1y}koPTZ^ΤϤPQ|̲Ӳܲ $'dg}ry$+krvM5Jh|wG\GIfh!"UVDE48JM  $ ' 1 4 g  $(*.45:1:(-SU[]kl~x | ##U,V,O.P. G$GGGGG9J:JWWWWW\X\TcZcdchctcxcccccdde"ehhjk$k(kdkhkllIpJp@tCttt>G4=BLBDSZԇ݇߇ɈkuҊۊLUƋ@Gkt*3?I'/48}~Zfu~*+ϗؗ;Cop~DFS\aeҚך      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFJKLMNOPQ^STUVWXYZ[\]_`abcdefghijklmnopqrsuvwxyz{|}~ ajћ؛t~$(*.04:>ҝ՝Xbҟן؟ܟ&'+,./;QU@IǢɢ̢΢ѢҢ :>KYqrw!"&()BCMNU[]klst¤ 47;<AB^_abdeghijklno¥ƥ9RVʦߦ-8aeϧЧLNQZèȨ 23XY7 lm56VWӱԱ:;34*5Q\FGijJL:COXZelvx $(ioFPR[ +MTY_!(XYWXjk  opil}),cfY\_bdg58 (.[aGMeh{  T"U"..//000111669999J:O:::w<|<>>KK%L&LLLUZVZX[Y[!_&_1_8___&`-`+c6cgg,h7hmmtt#$=Ddkyzvwvw Gץޥߥ XZ $GHXY]h <@ELx} "($v}*4jo+0Y_af xrz| S^`f1<>D+3EK=CV[cjpuv!&*5aj -45<&,kqsx~IPbhotio) /   ! % ) /   Z `                     ",:<>C>Cef      !!0!E!T!!!// 2 2!2!202<2?2J2L2L2[2g2j2v2x22224466778888]:]:@@RCRCSCwCxCNDODDD E!EHEIEiEjEGGHHIIISJTJLLLsRtRyRzRSUUUmUoUUUUUUUUUWWWWWWWWWWWWKXMX[X_XnXsXXXXXXXXXXXXXXXXXXXXXXXXXXYY YYYYY ZZQ[X[[[[[[[[[\"\&\(\I\V\|\\\\\\] ] ]](].]M]N]X]}]]]]]]]]]^^^!^"^C^c^h^t^^^^^^^^__7_X_``````````````````c cccc c#c$c%c,c.c5c6c:c?BCR+,78BDJLQSY[celɧϧѧק٧ߧ $%35XYqrҨӨ')`a{}éǩɩͩϩԩ֩ܩީ-.23KLMNfgjkz{ªêЪѪfhnptv~Ыԫ֫ګܫ>?SU[]ackmsu{}ĬƬˬͬԬ֬ܬެƮǮ !')/034:<A[b56{|ܴݴ̵Ҷ#Hq+,7;Dֹ+_`º0$ȼ˼Mmo45GNPWsv,278\]dfv|}jlmcd|*+%@3:Sl  >stu>""SViu "#@?_`nsz$%78YZghtu#$,-NOtv+OQTtwv.. 1 1112222,3-33333C7E7773949::;;R>S>>>sAvAAA5B8B$R0RXX]dedd e+j7jWюݣ X`MRW\""2289998QAQffhhooovpq qzz נ "ѷqt[t'*AFKNNPSY}GJ|uw   6 ;       * / 6 ; i l     8` (""$$'',,x.}.o2s288OON\e\E~P~O#67vz7A]g?Asy%+  R^44DDeDfDmGrGPPPPQ!QTTZZdd\l]luuÍčEH"FKPR`g/925JM  }'-w|~~  88JJS S7UBUzkkchko).=Yklp  238IP> >G@F|S  sCvCRYUY&`-`ww}}|{w~}-x<>xsf 9@Bt-[EMNpV 1h  Wx  TUlm=pq=>pr5%5Y J  4RzSW?Z[MNK=|%@A{}jHIiGNi' B     F     ! Y       :XN~oPT\gk  ! !/!22SCvCxC%I&IILjRkRmRnRpRqRsRtRvRwRyRzR|R}RRRRRRRRRRRRRRRRRRRRRRRRRNUOUjUlUUUUUUUWWWWWWIXJXXXZXfXgXXXXXXXXXXXXXXXXXXXXY Y Y\\\\]] ] ] ]]]]]]X]^]_]c]d]k]l]t]u]|]]]]]]]]]]]]]C^G^H^L^M^T^U^W^X^[^\^b^^^^^^^^^__7_>_?_C_D_K_L_S_T_W_``````````````````````c cccc c#c'c(c,c.c5c6c:c?Q78 !35')pqMNfgy{\]HIjkmnpqstvwyz|}ƮǮʮˮήϮ  ̵Ѷ#Gq6;Cֹ+º/$Ǽ˼MloGNPW^bdi,248\^vmy%?6:Vl   >s!"=fj"RV "#@?` +67BCNT:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::1xw :5=5I:M:`;d;iDvDhHHL$LllmmppMdNq- 5X<pn$$JKVVV_^_fffgxppyy8H(,B4o330XCX a:Z4!59W:5\W\xhh>wdw܄ r+ &&aapq 3Ur1Cx66KKҟ../!2.2L2Y22:2;AAARCLO8OMOQQQQhRjRkRmRnRpRqRsRtRvRwRyRzR|R}RRRRRRRRRRRRRRRRRRRRRRRRRRRRRNUZUeUfUhUtUUUUUUUUUUUUUUVWWWWWWWWWXX X+XGXIXVXXXdXfXsXuXXXXXXXXXXXXXXXXXXXXXY YYYY ZZ&ZKMhl{āǁɁ́΁сӁց؁ہuZلۄ>҇#E`bf@BNP]_`beikmquyɉ6bߍN3{#|*QZ~͖ ˘ҘC_HgKy47Ş&=#OPh¡BC &Q83'Nyjkmnpqstvwyz|}®ĮƮȮʮ̮ή   '):<B[cı:<EGPprzV]εӶ#Iq7;Dֹ+º1P$ȼuvڽ۽Mmo;ξͿϿ35RTGNPWY\^bdiknpqsvxy{}*,248:<;<XYqr LP5;_y?!WY"<t?A`a;gm%ANPaeijx~79:l*,=AEFTZjmnpSX 16_fxyde2Nt3_cn )Ibd-TV "SV#  j~?s0Vov@|%'w49tx]corsACprLN+-~ +Ooq& m n m o       C D | }  !GHNONT/!22LTKurt CoxChristian MonsonKurt CoxKurt CoxKurt CoxKurt CoxKurt CoxKurt CoxKurt CoxKurt Cox&|x-)}# :~Dۯ{89|`|]LLZDa'uaif((uAx@2ZR]eZ +w |DrPQ  2 9+T8f-|8/X[UA1Zoj H7K0$MڰzmNwm;R# dTR.b<pRvhdXR=83UTR3MtWvh] A_h ~@_lNk p2^`.^`.88^8`.^`. ^`OJQJo( ^`OJQJo( 88^8`OJQJo( ^`OJQJo(hh^h`. hh^h`OJQJo(8^`3*GCJsHtHaJ_H.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH.8^`3*GCJsHtHaJ_H.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH.^`o(. ^`hH. L^`LhH. e e ^e `hH. 55^5`hH. L^`LhH. ^`hH. ^`hH. uLu^u`LhH.8*^`3*GCJ_HaJsHtH6o(hH.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH.h]]^]`OJQJo(hHh--^-`OJQJ^Jo(hHoh  ^ `OJ QJ o(hHh  ^ `OJQJo(hHh^`OJQJ^Jo(hHohmm^m`OJ QJ o(hHh==^=`OJQJo(hHh  ^ `OJQJ^Jo(hHoh^`OJ QJ o(hH hh^h`hH. P^`PhH.. ^`hH... x^`xhH....  ^`hH .....  X@ ^ `XhH ......  ^ `hH.......  8x^`8hH........  `H^``hH.........h ^`o(hH.^`o()h pLp^p`LhH.h @ @ ^@ `hH.h ^`hH.h L^`LhH.h ^`hH.h ^`hH.h PLP^P`LhH.^`o(. TT^T`hH. $L$^$`LhH.   ^ `hH.   ^ `hH. L^`LhH. dd^d`hH. 44^4`hH. L^`LhH. hh^h`hH) ^`hH) 88^8`hH) ^`hH() ^`hH() pp^p`hH()   ^ `hH. @ @ ^@ `hH.   ^ `hH. P^`Po(hH Chapter :  @@^@`o(hH. 0^`0o(hH.. `^``o(hH ^`o(hH .... ^`o(hH ..... ^`o(hH ......  `^``o(hH.......  00^0`o(hH........h ^`o(hH.h ^`hH.h pLp^p`LhH.h @ @ ^@ `hH.h ^`hH.h L^`LhH.h ^`hH.h ^`hH.h PLP^P`LhH.^`o(.TT^T`o(. $L$^$`LhH.   ^ `hH.   ^ `hH. L^`LhH.n Monsondhrihri Normal.dotoChristian Monsond2riMicrosoft Word 10.0@Ik@w`@B`@B`EB  FMicrosoft Word Document MSWordDocWord.Document.89q  Carnegie Mellon University39 (+ DocumentSummaryInformation8RCompObj|j dd^d`hH. 44^4`hH. L^`LhH.h88^8`OJQJo(hHh^`OJQJ^Jo(hHoh  ^ `OJ QJ o(hHh  ^ `OJQJo(hHhxx^x`OJQJ^Jo(hHohHH^H`OJ QJ o(hHh^`OJQJo(hHh^`OJQJ^Jo(hHoh^`OJ QJ o(hH{^`{o(. ^`hH. L^`LhH. e e ^e `hH. 55^5`hH. L^`LhH. ^`hH. ^`hH. uLu^u`LhH.hh^h`o(^`o(.0^`0o(..0^`0o(... ^`o( .... X X ^X `o( ..... `'``'^`'``o( ...... -`-^-``o(....... 44^4`o(........0^`0o(. ^`hH.  L ^ `LhH.   ^ `hH. xx^x`hH. HLH^H`LhH. ^`hH. ^`hH. L^`LhH.8*^`3*GCJ_HaJsHtH6o(hH.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH. ^`hH Article . 8^`hH Section . P^`PhH() `p`^``phH() P^`PhH) P^`PhH) ^`hH) P^`PhH. 0p0^0`phH.՜.+,D՜.+,L hp  Carnegie Mellon University39 (+  PARAMOR: Title 8@ _PID_HLINKSAD0_Toc2173154830_Toc2173154820_Toc2173154810_Toc2173154800|_Toc2173154790v_Toc2173154780p_Toc2173154770j_Toc2173154760d_Toc2173154750^_Toc2173154740X_Toc2173154730R_Toc2173154720L_Toc2173154710F_Toc2173154700@_Toc2173154690:_Toc21731546804_Toc2173154670._Toc2173154660(_Toc2173154650"_Toc2173154640_Toc2173154630_Toc2173154620_Toc2173154610 _Toc2173154600_Toc2173154590_Toc2173154580_Toc2173154570_Toc2173154560_Toc2173154550_Toc2173154540_Toc2173154530_Toc2173154520_Toc2173154510_Toc2173154500_Toc2173154490_Toc2173154480_Toc2173154470_Toc2173154460_Toc2173154450_Toc2173154440_Toc2173154430_Toc2173154420_Toc2173154410_Toc2173154400_Toc2173154390_Toc2173154380_Toc2173154370z_Toc2173154360t_Toc2173154350n_Toc2173154340h_Toc2173154330b_Toc2173154320\_Toc2173154310V_Toc2173154300P_Toc2173154290J_Toc2173154280D_Toc2173154270>_Toc21731542608_Toc21731542502_Toc2173154240,_Toc2173154230&_Toc2173154220 _Toc2173154210_Toc2173154200_Toc2173154190_Toc2173154180_Toc2173154170_Toc217315416 hh^h`o(hH P^`Po(hH. ^`o(hH.. x^`xo(hH...  ^`o(hH ....  X@ ^ `Xo(hH .....  ^ `o(hH ......  8x^`8o(hH.......  `H^``o(hH........ P^`Po(hH Chapter :  @@^@`o(hH. 0^`0o(hH.. ``^``o(hH... ^`o(hH .... ^`o(hH ..... ^`o(hH ......  `^``o(hH.......  00^0`o(hH........^`o(. ^`hH. L^`LhH. e e ^e `hH. 55^5`hH. L^`LhH. ^`hH. ^`hH. uLu^u`LhH.^`o(. ^`hH. L^`LhH. e e ^e `hH. 55^5`hH. L^`LhH. ^`hH. ^`hH. uLu^u`LhH. P^`Po(hH Chapter :  @@^@`o(hH. 0^`0o(hH.. ``^``o(hH... ^`o(hH .... ^`o(hH ..... ^`o(hH ......  `^``o(hH.......  00^0`o(hH........ hh^h`hH) ^`hH) 88^8`hH) ^`hH() ^`hH() pp^p`hH()   ^ `hH. @ @ ^@ `hH.   ^ `hH.8^`6o(hH.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH.8^`6o(hH.8 ^`o(hH.8 pLp^p`LhH.8 @ @ ^@ `hH.8 ^`hH.8 L^`LhH.8 ^`hH.8 ^`hH.8 PLP^P`LhH. hh^h`hH) ^`hH) 88^8`hH) ^`hH() ^`hH() pp^p`hH()   ^ `hH. @ @ ^@ `hH.   ^ `hH. P^`Po(hH Chapter :  @@^@`o(hH. 0^`0o(hH.. `^``o(hH ^`o(hH .... ^`o(hH ..... ^`o(hH ......  `^``o(hH.......  00^0`o(hH........4 AAH+w+wHQpRAHA HA,HzmNA8HADHAAPH3MtWNkpA\HAhHAtHAHK0$MZ/dTRj H7~}|9+~@_A_ia'u]eZ]m;RUTUA1dXRf-|D H mHHPTN<K0XfiDVReSnV&OJ^J&&D                         8Y       RKV        RKVRKV                w                 h*&        L         @ @ @ @P%U,.=:U\%7&MYbjdn+ F.pwOk[C~)*;rEz}qV(Wgo>*1<4W$ ? X j[ |e {   x& , O X ] 'q \y z s W & d, AA @ OY I*W&+WXNnY==g6,N`acoV*/?q gR8:J$\ e-/;<:O#i/l;r1$r&+x=G=K"73H{K,dn"t*/ kd^ e!)5n:syy@f*go6~ n+L0FZs'-S' w. < Kk )!'!,!2!qo!Uu!]"""3" 9"(M"zV"5r"1-# $\!$4$<$=$Q$G]$k$s$(z$ %%.#%h/%N%a%&/6&;&\&q&ow&Z'9( ((#()(L(H) )S`)g)q)****X!*L4*N*n*C}*+u+!&+\+,`(,,,n,--&-*-. .3.U.q/`,/ @/f[/e|/4]0m0 1$1>1@2C2 i2wk2 3<3(e3T5Ea5w6686"<6A6B6M6g[6|67=7Q7c788*8068[89999I9&J99;::,:6<:{N:(v: ;;2;':;;;_;<<q=<*U<j<F5=G=U=6>>>i>J??PA?EH?2Y?&m?@@@aE@*@A3-A?ABAB'eB8CXCf[CuCDDWyD}D#E/EEP?E@E^EbEfEpEyE>jF'vFNwFezF G&G@G@GUzG=H<[HtH I9IhRIg JsJKdKVKwK]>L@LbAL LLmdLxLMMMuMLM`M_gMnMzM6NaBNCZNbNjNuNOBOTOYAPOQPQ.Qc6Q%?QCQLdQ tQR1RURrZRnR;CSnS(~S0TMT7jTyTeYewfffCfffhfxfd*g3gV9gEJg\gwgTh!h'`hi7iyi}i=jHj=^j kUTkZk\klElpl"m=m[m_mmin/ nmn$n4oo0oCozLocoaoo\top)Wpspyp{pTqd@qRDqWq;yq rIrIrcrlrrb's^?sZs2fs;msysx_tvyPyRyhyzz`>zMza\zsz{H{^{y{||.|Q5|H|Q|k|w|}.}X}[}i}"n}n}u~~%V?2>IIZ>bo#IP V>tY^ '0+-zG` /@(NR8LM s xU H,=]Iy96?@Oqh9bgi}~.Fy\:42UojpC{cn !'2b u"&-<SHNPGZrbe$6F}Uy GL3bi9:Ayebsmk^eT+V,6p* ;RTnc^i 17!QuZ"x!#1!:d` K^|{-9<'+7Vc"B%1EPt2 D++|0j67*CIbkqgrsuxrkQ- +Q 'I+Zx2{Zqh>0<lVZyX[l4BLkcqs ')DENPKL\Khx YB\H#9_AFI)_tqvVYq{ {$7s;C\0056;cgqVE'`h$s'z4<-K^ah#jBAeFw5+<fWNVbhhPlk379yNR_Juq kjj*15MW[Ead6g6,#r=BG:WXw 1Zy 9F`F,TZfyp#"Uso!!Y>=bV&/(9dEGc( "i(~MPz\~4~{<,Yb$o$,f;YOhijm = Q6\7J^/ppI][s)>NOZ5\0s9'Cmey9>eR d0Ww +N$dVk'2 Irv.02Mj6<H`HRwL|k c%7dmPEMSabeeko\q^E2Glq}0"-2_=HK_x/"$`1 Si-`axX$,MWn~wgLFgG } {/!DV$~KQ(mp *RXa2>jnvB<opw<IFEk9RKgds9?c n-<;?c7jrwqR\WIdZ.HX&pdmtX F 0nFr~+FLGMHuuz  ,1F\whukw~z !#.?OEhGlXZ \x6.hrO{[,-A`fn~dh*]1y3 DyNi,7/Sdf5Y3/5Pj{ 4%J#^Hmqtz7)I6K^`{hzq{Phdy/!22ABB B'B/B7B>B?BFBKBQBWB]BcBiBjBBBBBBBBBBBBBBBBBBCC CCCCC5C9C=CACFCKCPCQCLOmOnOOOOOOOOOOOPP&P'P4P@PKPUP_PaPbPpP}PPPPPPPPPPPPPPPPQ QQ+Q6QBQCQKQVQbQoQ}QQQQQQQQQQR RRR'R0R1R2Rx?x@xBxDxExFxGxHxJxNxSxwxxxzx{x}x~xxxxxxxxxxxxxxxxxxxxxxxxxxxxxyy(y)y+y-y/y0y1y2y3y5y7y8y9y:y;y=yByFyiyjylymynypyqyrysyuyvyxyyyzy{y~yyyyyyyyyyyyyyyyyyy zz1z2z5z6z7z9z:z;zz@zAzCzDzGzvzyzzzzzzzzzzzzzzzzzzz {{{{{{{{{{{{{{!{#{5{7{Q{R{U{V{W{X{Z{\{]{_{`{b{c{d{e{h{{{{{{{{{{{{{{{{{{{{{||!|"|#|$|%|&|'|(|)|*|+|,|.|0|E|G|[|\|`|a|b|c|e|f|g|i|j|l|m|n|o|q|||||||||||||||||||||}} }!}"}#}$}%}'}(})}*},}-}.}0}A}C}b}c}g}h}i}j}k}l}n}o}p}q}r}s}u}w}}}}}}}}}}}}}}}}}}}}} ~ ~~~~~~~~~~~~ ~!~#~H~J~k~l~q~r~s~t~u~w~x~z~{~}~~~~~~~~~~~~~~~~~~~~~~~~~~~~     &',-./012345679;CESTYZ[]^_`abdfghjtv Wgu~ǘȘɘʘ˘̘͘ΘϘИјҘ՘ؘۘܘݘޘߘBCEGLOPQSTUWXYZ\_GHJPUWXZ[\]_`bcdgΚӚ֚ךؚښۚܚޚߚ   !"$GKxy|47žŞ&)yz}=@#&OPTVY[\]^_`abcefh¡     &(BCHJMOPQRSTUVWYZ\npĢƢȢʢˢ̢͢΢ϢТѢҢԢբע     ()*+,-/38?EFKRSTUVWXYZ[\]`cfghijklmnoprtz|~͵εеҵ׵ڵ۵ܵ޵ߵҶӶն۶#HIKY^abcefgijlmnq7;DHҹֹ !#$&'(+º0149=@ABCDFHIKLMP  !$ȼ˼ #%&()*+,-.013EHuvzڽ۽߽56:<?ABCDEFGHIJKMmo;ξӾվؾھ۾ܾݾ޾߾ !"#$%&')*,68IJOQSUVWXYZ[\]_`boq~%0;<DINSXY`ejoqrz  $).37<AFJOTY]^hmrw{ !"#$%&'.6=?ADFHKLNOPU\ahmty !(-4579;@GLSX_dkpw| %9M_xy  %&27=BJPQegjlqtu?JMNObghijkt{}~ !469;@CD\^cejop 6:?CJOPQRSTUVWYdghi|";<OQTV[^_wy~ (*-/478NSY^flmnopqrst    478;<?@AUXY\]`ak  ;fgm !iy      !"%@ADGKMNPQRSUVWYZ\     $-5;FNOPRV[bcdejqxyz{|}~   7:WXZ[\^_`acdfghil    "*+,.27>?@AFMTUVWXYZ]`cdefghijklmpqrtuvwyz|}~   ?B[\`abceghjkmnops    ">?@ABCDEFGHIJKLMNOPQSX ()1234568<AHNOT[\]^_`abcdefilopqrstuvwxy}!#&()*+,-./0235HJdejloqrstuvwxy{|~ !"#$%&'()*+,-./0123456789:;     "SV #ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~6789AHIJKLMNOV^egilnpstvwx} %,18=DIPU\]_achot{nos!+5?ABCEPXes!,4CVq%6P_npqrt| 1=ILMNQ]ft&+,-0:FTdp|~  #.CPiv +O_opqt$&' ? G O V ^ f m n u z                       $ ( , 0 4 8 ; < X [ ^ a e i l m o ~               ! * 7 C D N V p | }       !()138=BGLQV[`abdinsx} "$%&).38=BGLQST\^chmrw| "',16:?DFGHKPUZ_dinsuv}!#89GHJOTYZ\px &+05:?DINOT@@L$012VWXYZ \ ] ^ _ a bncndyiyjyvywyx z { |MMMMMMMMMMMMNNNNNIIIIIIIIIIWWWW ~!~"~2~3XY]^_mnoqrs{|c~cccccccʐʑʒʓʔʕ̖̗̘̙SS[[[ww)))ֺֻ555555555555&&&&&&uuuuOOOOm m \\\\\\ \"\#\%\&12<=?@ABCFGIJNOCUCV0X0YnZn[n\n^n_bdfghujukulupuqVsVtwxy{|2~2MMMM++333XXXmmmmmmͮͯ͵Ͷ111[[H#H$a)a*9:<=>?@ABCDEFGHIJKLMNOPQRSTVWXYZ[]^_hijklmnoqrstuvwxyz|}~ΤΥʯʰ888BBBB]]]]]]]]UUUUUUUUUUU !#$U,U-U.U4U5U69:9J9KPQRUVWXW\W]hilmoptuv6x6y6z6{6|6}6~LLLxxxxxx66qqllllll333333333llWWWWWWWWYYYYYYYYYY~ ~ ~~D D!T"T#T$'(,-./01345678:;=>?@ACDEKLMTUVWXYZ[abdefghist#####yyyyvvXX44444~~ !O/O0O1SJSK M N R SdTdUdVdW,Y,Z,[,],^|_|`|a8g8hgigjklopqrtuv8x8y8z8{8|8}8~88Œœ^^^^^^^^^^^^^^^XXXXXXX??????CC C C S@@@8@:@x@@`@b@@@f@@@j@l@n@@@r@@@v@@@~@@@@@ @@@@,@@@@@@@@@@l@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @@@@@@@@@4@@@<@@X@@@\@@@n@@@@0@@@@@@@@@d@@@@p@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @ @@@@$@@@@8@@@@@@"@$@L@@(@T@@6@p@@:@x@@@@@@D@F@H@J@L@N@@@T@V@@@\@@@b@@@h@@@l@n@p@@@|@@@@@@@@@@ @@@@@8@@@@@H@@@P@@@X@@@`@@@|@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @ @@@@(@@@0@@@@@@@"@H@@&@P@@.@`@@4@l@@D@@@H@J@@@N@P@@@T@@@X@Z@\@@@d@f@@@j@@@r@@@v@@@|@~@@@@@@@@,@@@@@@@@@@@\@@@l@@@@@@@@@@@@@@@@@@@@@@@@@(@@@@@@@ @"@$@&@(@*@,@.@0@2@4@6@8@:@<@>@@@B@D@@@H@J@L@N@P@@@T@V@@@b@@@f@h@j@l@n@@@r@t@v@x@@@|@@@@@ @@@@ @@@ @@@, @@@4 @@@@@ @@@H @@@P @@@X @@@@t @@@ @@@ @@@@ @@@ @@@ @@@ @@@@ @@@ @@@ @@@ @@@ @@@ @@@ @@@@@ @ @@@ @@@@, @@@@@@ @"@$@&@(@*@,@\ @@0@d @@4@l @@8@:@<@>@@@B@D@ @@H@ @@L@N@P@R@ @@`@b@d@f@h@ @@l@ @@p@ @@v@ @@z@|@ @@@ @@@@ @@@@0 @@@< @@@` @@@@t @@@@@ @@@ @@@ @@@ @@@ @@@@ @@@@@@@@ @@@@@@@@@ @@ @ @ @@@( @@@@@@@@ @@"@H @@&@P @@*@,@\ @@0@d @@4@l @@D@F@H@ @@N@ @@T@ @@X@ @@^@`@b@d@ @@h@j@l@ @@p@ @@v@ @@@ @@@@@@8 @@@@D @@@@@T @@@\ @@@h @@@t @@@ @@@ @@@@ @@@ @@@@ @@@ @@@@@@@ @@@@ @@@@@@@ @@@ @ @@ @@@@,@@@@@@x @z @| @~ @ @ @ @ @@ @@@ @$@@ @,@@ @4@@ @L@@ @ @t@@ @@@ @@@ @@@ @@@ @ @@@ @@@ @ @$@@ @4@@D @@@^ @` @b @@@f @@@v @x @@@ @,@@ @4@@ @@@@ @ @ @P@@ @ @\@@ @d@@ @ @p@@ @@@ @@@ @@@ @@@ @@@ @ @@@ @ @ @ @ @ @ @ @@@ @@@ @@@ @@@ @@@ @@@ @ @@ @ @ @ @$@@ @ @ @ @ @ @ @" @$ @& @P@@. @`@@2 @h@@6 @8 @@@@@@@@@@@@@P@@@X@@@@@@@@t@@@@@@@@@@@@@@@@@@@@@@@@@@T!@@@\!@@@p!@@@x!@@@!@@@!@@@"@@^@`@"@@#@Unknown Gz Times New Roman5Symbol3& z Arial;& z HelveticaIArial Unicode MS?5 z Courier NewA& Arial Narrow5& zaTahoma3z Times; BatangG  MS Mincho-3 fg;Wingdings"1eʍfʍffEB(9 EB(9 4d++45 2qX d PARAMOR: Christian MonsonChristian Monson&                         !" # $ %