ࡱ>  5ibjbj S) ""###D$$$$%d$W%:!&!&!&!&8Z99U9 WWWWWWW$Ze]2W-#e9s7l8e9e92W"|{#X!&!&R_W.???e9&#!&#!& W?e9 W??6QS!&@T$;nRVW<W S]=]8SS]#The9e9?e9e9e9e9e92W2W?e9e9e9We9e9e9e9]e9e9e9e9e9e9e9e9e9 !: Data Extraction from Web Tables: the Devil is in the Details George Nagy Electrical, Computer, and Systems Engineering Rensselaer Polytechnic Institute Troy, NY, USA 12180 nagy@ecse,rpi.edu David W. Embley, Spencer Machado Computer Science Department Brigham Young University Provo, UT, USA 84602 embley@cs.byu.edu, admiralmachado@gmail.com Sharad Seth, Dongpu Jin Computer Science and Engineering Department University of Nebraska Lincoln Lincoln, NE, USA 68588 seth@cse.unl.edu, jindongpu@hotmail.com Mukkai Krishnamoorthy Computer Science Department Rensselaer Polytechnic Institute Troy, NY, USA 12180 mskmoorthy@gmail.com Abstract Humans can usually discover the relationships of table headers to data cells at a glance, but determining them automatically is far from trivial. We describe common table configurations that require special consideration, including some that require access to semantic knowledge. Clicking on one or two critical cells per table, through a simple interface, is sufficient to resolve most of these problem tables. We show that header paths, a purely syntactic representation of visual tables, can be readily transformed (factored) into several existing representations of structured data, including category trees, relational tables, and RDF triples. From a random sample of 100 web tables from ten large statistical web sites, the system processed 86 tables completely automatically, processed 3 more automatically after clicking on a critical cell, and generated 186 relational tables and 20,262 subject-predicate-object RDF triples. Keywords-visual table, relational table, RDF, header-paths Introduction In the first decade of table processing, researchers concentrated on finding the underlying grid structure of scanned tables from rulings or text alignments [,], and of ASCII tables in email []. In the second decade the emphasis was on locating and bounding HTML pages in web tables. Reviews of earlier work can found in [,]. The next step in table analysis, to which we attempt to contribute here, is efficient extraction of the relations of header cells to content cells and the representation of these relations in appropriate data structures. The resulting multi-dimensional indexing is a prerequisite for understanding individual tables and combining the contents of many tables into a queryable database or a populated ontology. Similar goals are addressed in [,]. The foundations of syntactic table analysis were laid by X. Wang in her 1996 dissertation []. Although she was interested primarily in reformatting tables for various media and page sizes, her definition of categories is equally suitable for analysis. Simple tables have only two categories defined by their row and column hierarchies, but more complex tables, such as her prime grade book example of Year, Term and Mark, require multidimensional indexing. Below we show that the extraction of paths (introduced in []) through the header hierarchy to content cells, and the decomposition of these paths into orthogonal categories, leads to a representation suitable for transforming tables meant for human reading into relational database tables [] accessible by formal languages like SQL [] for relational tables or SPARQL [] for RDF triple stores []. Of course, many of the tables available at large sites of statistical informationour primary focusare generated from databases. Often, however, no direct public access is provided to the databases themselves. Individual users must therefore reconstruct fragments of the database of each source, or possibly of databases of multiple sources, by harvesting and analyzing individual tables. Our starting point is a collection of tables, selected randomly from ten large sites []. HTML tables can readily be exported to Excel. The resulting Comma Separated Value (CSV) files are easy to parse. The transformation from HTML to CSV loses some formatting information like typeface, indentations and merged cells. Nevertheless the standard CSV format retains sufficient information for a broad range of applications, including automated table processing. In Section II we discuss common table formatting conventions that must be accommodated by automated table analysis. Section III describes our heuristics for finding header paths given these conventions. Section IV explains the extraction of Wang category trees using mathematical software designed for the synthesis of logic circuits. Section V demonstrates the transformation to relational tables and RDF triples. Section VI shows our experimental results on 100 web tables, explains some causes of failure, and proposes further work for expanded application of table analysis based on header paths. TABLE FORMATTING CONVENTIONS Since the invention of printing many table formatting conventions have been developed and refined for ease of human access to tabulated data (cf. The Chicago Manual of Style or the US Government Printing Office Style Manual). Although tables can be prepared with any text editor, most document preparation systems (e.g. Office, Latex) include elaborate provisions for constructing tables. A table contains a rectangular configuration of data cells, each of which can be uniquely designated by a named row and a named column. More formally, table is a discrete rectilinear tessellation, or a rectangular tiling, based on the partition of an isothetic rectangle into rectangles defined on a rT x cT lattice []. The bottom-right region of a well-formed table (WFT) contains a set of rd x cd content- (aka value-, data-, or delta- in [ NOTEREF _Ref286587916 \h 6]) cells that can be uniquely specified by a column-header path and a row-header path. Fig. 1 explains our notation and shows how a single critical cell can define the segmentation of the table into stub, header, and delta regions. (Although our observations are drawn from a large collection posted on the project website [], we present simplified examples because most real tables are too large for the ICDAR page allotment.) The extraction of header paths must take into account some commonly occurring configurations discussed below. Virtual Headers Wang defines rooted category trees. Category Roots, however, are often missing in real tables. For example in the table of Fig. 1, which has two column categories A and B and one row category C, it is necessary to add a virtual root header (e.g. CH1) to the category A paths. Table Notation. Lower-case letters above and numerals to the left are not part of the table, just Excel-style cell addresses. Cells a1:b3 are the stub header (stub). The delta-cell region is c4:h5, i.e., everything below and to the right of critical cell c4. The column header path to cell f4 (which has cell-content D14) is B-B2-A1. The corresponding row headerpath is C-C1. Headers in the Stub In the simple example of Fig. 2a it is not obvious whether A in cell a1 (the only cell in the stub here) is the root for category B or C. A more realistic example is shown in Figs. 2b and 2c. The appropriate category trees cannot be determined without semantic considerations. As a practical matter, in the large majority of the tables we have seen, the contents of the stub are row headers: we would expect AGE instead of VERMONT in merged-header-cell b1:e1 Multi-row/column Indexing In a table with rd rows and cd columns of delta-cells, there must be rd row-header leaf-cells and cd column-header leaf-cells. However, for unique indexing it may be necessary to consider some data columns as headers. In the table of Fig. 3, one might at first consider State to be the row-root-header, and all but the first column as delta-cells. There are, however, multiple rows with the same entry (AL, MI, MN). Even adding the second column is insufficient because Minnesota Power Inc appears twice. The row-header here consists of the first three columns.  (a)  (b) (c) (a) Row or column header? In 2b and 2c, GENDER and EDUCATION are in the same position (in cell a3). (b) GENDER is a column-root-header. (c) EDUCATION is a row-root-header.  Three columns are required here to index columns. Row/Column Order In relational tables, rows and columns can be permuted without loss. In tables meant for humans, however, order can be important. In Fig. 4. the Rank column may at first appear inferior to the State column as a row-header. But the table designer would have put State on the left if the table were meant to find the rank of various states instead of the states with various ranks (the title of this table was "Top 3 States for Trade via Port Huron, MI: 2008").  Row order is important here, but column order is less so. If the first column were missing, should we consider adding the row order explicitly before pouring the contents of this data into a relational table? Even if a numerical index satisfies our header path uniqueness requirement, it is often of little value for querying table data. We don't currently add explicit order information but do preserve the original CSV cell addresses as meta-data. One-dimensional Tables and Lists Our fundamental requirement is that each data cell can be indexed uniquely. We call a table-like structure with a single row or column a list because it lacks an explicit index. To avoid trivial tables, we impose the requirement that a table must have rd >1 rows or cd >1 columns of delta cells. We call a table with rd =1 or cd =1 one-dimensional (such a table could have multiple categories). Aggregates Tables often contain rows or columns of aggregate information that could be recovered from the rest of the table. V. Long [] has identified sums in a large collection of Australian reports. Other aggregates, like median, standard deviation, or truncated mean, could also be identified by computer analysis when the span of the aggregation operation is clear. Sums are sometimes identified by simple keywords like Sum or Total which apply to the whole, or part of a, row or column. In other cases more complicated semantic processing may be required. In the Canada Statistics tables, totals for all the provinces and territories appear under CANADA. CANADA, however, also often appears as a row-header root, without any aggregate data. CONSTRUCTION OF HEADER PATHS CSV Version of the HTML table The CSV text file is parsed by a Python program: the description of an earlier version of this program will appear in [ NOTEREF _Ref286587868 \h 7]. Most html formatting information is lost, and the structure is modified because merged cells, like the column-header root containing B in Fig. 5, are unmerged. Therefore one of the tasks of the path extraction program is to copy the cell contents in the atomic cells of the CSV string. For the table of Fig. 1, B, B1, and B2 are copied into all the empty cells, represented by commas, to their right. C is copied into the first empty cell of the fifth row. To simplify subsequent processing, empty rows and columns are deleted, blanks within cell contents are replaced by underscores, and some characters with special meanings in downstream programs (\, +, *, (, ), etc.) are replaced by ASCII strings like backslashtoken. ,,B,,,,, ,,B1,,,B2,, ,,A1,A2,A3,A1,A2,A3 C,C1,D11,D12,D13,D14,D15,D16 ,C2,D21,D22,D23,D24,D25,D26 CSV file and table for the Excel table of Fig. 1. Excel cell addresses changed to x-y coordinates to allow negative indices. Both the header paths and the paths through the delta cell region consist of asterisk-separated-sequences of cell contents in double quotes, with the sequences separated by plus signs. An example is shown in Fig. 6. After the paths are extracted, the original cell coordinates, enclosed by angle brackets <>, are added to each path element. (If C were above C1the customary location of a row header rootthe program creates negative x-coordinates for header-path roots virtually moved from the stub to left of the leftmost row-header column). Path Extraction Heuristics The critical cell is located automatically if either of two conditions holds: (1) the stub header is empty, and (2) the delta cell region consists of numerical information. If neither of these conditions holds, the table is rejected for interactive identification of the critical cell. Under the second scenario, row header roots in the stub are added as roots of the row headers below them. In the table of Fig. 2a, A would be added to the row header paths C1 and C2 (if D11, D12, D21, D22 were numerical). The next section describes the factoring of the path files (which also contain a copy of the original CSV files). rowpaths = (("<0,3>C"*"<1,3>C1") +("<0,4>C"*"<1,4>C2")); colpaths = (("<2,0>B"*"<2,1>B1"*"<2,2>A1") +("<3,0>B"*"<3,1>B1"*"<3,2>A2") +("<4,0>B"*"<4,1>B1"*"<4,2>A3") +("<5,0>B"*"<5,1>B2"*"<5,2>A1") +("<6,0>B"*"<6,1>B2"*"<6,2>A2") +("<7,0>B"*"<7,1>B2"*"<7,2>A3")); Row and column paths for table of Figs. 1 and 5. The delta paths are simimlar but longer. EXTRACTION OF CATEGORY TREES With the asterisks and plus signs added in the row and column header paths, both resemble sum-of-products algebraic expressions. This choice is deliberately made to simplify the process of extracting category trees from these expressions through an algebraic factorization process. As the formulation is completely symmetric for column and row header paths, without loss of generality, we elaborate on only the column header paths in the following description. The algebraic interpretation is based on a covering relation defined between each product term in the header paths expression with the column covered by it. The covering relation can be extended to sums of products by taking the union of individually covered columns. For example, the sum-of-products: ("<2,0>B"*"<2,1>B1"*"<2,2>A1")+("<4,0>B"*"<4,1>B1"*"<4,2>A3") (1) covers the first and the third columns of the table in Fig. 5. Note that the cell labels <2,0>B and <4,0>B in this example are identical in the original table. To enforce this constraint, we drop the second (column) coordinate in the header paths. As a consequence, the labels <2,2>A1 and <5,2>A1 in the colpaths expression are also treated as being the same, in accordance with the normal conventions of table layout. The covering relation can be further extended to factored forms, as long as the inverse multiplying-out process can recover the original product terms from it. For example, the whole colpaths expression can be factored as: <0>B*[<1>B1+<1>B2]*[<2>A1*<2>A2*<2>A3] (2) from which the category trees for the headers are derived by our Python program, as shown in Fig. 7. Here, the program has stripped the indices and added the virtual root header CH1 to represent the missing header for the second column category.  Wang Categories for the table of Fig. 1 To facilitate the generation of relational tables (see Section V), the program combines the category-tree structures for the row and column headers into a canonical form, as shown below for the example: C*(C1+C2)+B*(B1+B2)+CH1*(A1+A2+A3) (3) It also generates multiple views of the table resulting from the different choices of the category chosen as the attribute. Another output produced by the program are the values of the table features used in verification of the result against visual inspection of the CSV table, as illustrated below for the example: Row categories: 1; Leaf nodes for row categories: 2 Col categories: 2; Leaf nodes for col categories: 2, 3 A benefit of this formulation is that the algebraic factorization problem has been studied extensively in the past in fields ranging from symbolic mathematics [] to logic synthesis [] and we gain leverage from the sophisticated strategies developed in these fields. We map the header paths expression to Boolean algebra and adapt the logic synthesis tool Sis [] for factorization. Interested readers can find further details in our earlier work [ NOTEREF _Ref286587868 \h 7]. RELATIONAL TABLES & RDF TRIPLES Given a factored expression for a table in canonical form along with the tables data indexed by the header paths, we can generate a corresponding relational table and populate it with the data. We can then query the table with SQL and otherwise manipulate it along with other tables in a standard relational database. Our transformation of an expression assumes that one of the terms representing a category in a table provides the attributes for the relational table while the remaining category terms provide key values for objects represented in the original table. Without input from an oracle, we do not know which of the categories would serve best for the attributes. We therefore transform a table with n categories into n complementary relational tablesone for each choice of a category to serve as the attributes. The transformation from a category term in an expression in canonical form to a SQL create-table statement is straightforward. Transforming the remaining terms in the expression and data indexed in header paths to SQL insert-value statements is also straightforward. For the canonical expression (3) with the choice of the term CH1*(A1+A2+A3) for the attributes along with the header paths and data like in Fig. 5, Fig. 8 shows the generated SQL create statement and SQL insert statement for one of the relational tables. In general, we (1) transform each header path of the term chosen to represent attributes into an attribute name (e.g., CH1_A1 in Fig. 8), (2) transform the root node of the remaining header paths into attribute names (e.g., C and B in Fig. 8), (3) declare the attributes of these remaining header paths to be the primary key (e.g., primary key (C, B) in Fig. 8), and (4) insert tuples into the generated table: the key values are a cross product of the header paths below the root from each category (e.g., {C1, C2} ( {B1, B2} for the four tuples in our example), and the data values are inserted as directed by the header paths of each data value in the original table (e.g., D11 for the CH1_A1 attribute of the tuple whose key is {C1, B1} as Fig.8 shows).  CREATE TABLE Fig_1_0(C varchar(2),B varchar(2), CH1_A1 varchar(6),CH1_A2 varchar(6),CH1_A3 varchar(6), PRIMARY KEY (C, B)); INSERT INTO Fig_1_0 VALUES("C1", "B1", "D11", "D12", "D13"); INSERT INTO Fig_1_0 VALUES("C1", "B2", "D14", "D15", "D16"); INSERT INTO Fig_1_0 VALUES("C2", "B1", "D21", "D22", "D23"); INSERT INTO Fig_1_0 VALUES("C2", "B2", "D24", "D25", "D26"); Tuple generration for the table of Fig. 1 For RDF triples we transform each data value in a relational table into a (subject, predicate, object)-triple. Fig. 9 shows the resulting RDF for the first tuple generated for the relational table in Fig. 8. For each data value (e.g., D11), the subject in the triple is a generated object identifier for the tuple (e.g., the object identifier C-B_0 in Fig. 9 obtained as a hyphenated concatenation of the attributes of the primary key along with a subscript 0 for the first tuple, 1 for the second, and so on). The predicate for the triple is the attribute for the value (e.g., CH1_A1 for the value D11 in Fig. 9 and B for B1). And the object in the triple is the value itself (D11 and B1 in Fig. 9).  ... V W b n o ھְڥڥ|t|ְphڥhBmHsHhN]hN]mHsHhN]hN]mHsHh}_h2mHsHh}_h}_mHsHh}_mHsHhhBmHsHhBhN]mHnHsHhhBmHnHsHh.dh2mHsHh2 hR[h2htZh25CJaJh`%5CJaJh25CJaJ(=>?Ky + W o  - N b w x y z { gdBgd2gd2   , - b v x {  * 0       % 7 \ ^ | ,-.TUVWhjm·کڥڝhh%hW4=hC8hX-h h< h<h< h26]hhBmHsHh.dhBmHsHhhBmHnHsHh2hN]mHsHhN]hN]mHsH hR[h2h}_hB4{ .ixLc0,! 1 F!G!H!I!J!K!L!M!N!""<gd+o`gd5Egd2gd2gd2gd2gdW4=mx!":;<d045:\|}~R\#'򾷨Ѡќє~zѨrjjjѨhg h26hC~h26h hhBjhhB0J*H*UhBhX-hPh}_hnjhnhn0J*H*U hnhnjhnh<0J*H*Uhh2 hhjhh0J*H*U hh2 h?Mwh2 hg h2)mnopqrcsbk289ST[~wogwhhn6hh26 hnhn htZh2 h!h2hA h.jh}_jh.jh}_0J*H*Uh}_jh.jh.j0J*H*Uh.jhnhjh.jh0J*H*U h.jh h.jh2jh.jhX800J*H*U h.jhX80hX80h2%[\XYZ^_ij 3EL[ѩѩ晕桁}uh ih i6h4h i h26jhUhjhUh]qh26hIh2H*OJQJ^JjhkDhkD0J*H*Uh2H*OJQJ^JhIh2OJQJ^JhIh26h2 hnhnhh26 hnh2.JKL1 2 y (!*!+!=!>!?!D!F!N!^!_!!!!!!!𷑊{wsfh+oh5EOJQJ^Jh&Mhsh| h1!h5Eh5E h]qh2h2OJQJ^J hN~h4CJOJQJ^JaJhN~h2CJaJ hN~h2CJOJQJ^JaJh4jh+oUmHnHu h*ZhkDjh*ZhkD0J*H*Uh2hX^h ih%!!! """"6"?"L"M"O"c"h"k"l"p"r"""""""""""""##!###^#_#c#d###8$9$W$Z$t$w$$߹߮ɮɧţu hsh2CJOJQJ^JaJhsh26h}_ hN~h2CJOJQJ^JaJh2 h!h5Eh+oh5EOJQJh+oh5E6OJQJhs hmmh5Eh+oh5E6 h|h5E h1!h5Eh+oh5EOJQJ^Jh+ohsOJQJ^J.$$$$$$$$$$$$$$$$$$$%% %$%%%&%%%T&U&_&`&&&&&&&&&&'' ' ''޴޴޴޴똣늆~p~~jhN~UmHnHuhh5Eh'jh'UmHnHuhN~h2CJaJ hN~h2CJOJQJ^JaJh]qh2H*OJQJ^JjhUmHnHuhhOJQJ^Jh]qh2OJQJ^Jh2 hsh2CJOJQJ^JaJ,"$$&&&&&&'''''' ' ' ' ' '''''' x`gdxgd' !$a$gd+o`gdgd2gd2''*'+':'@'E'N'q's'y'z''''''''''''''''''(((((((())W)X)m)n))))))Ⱥ}yhN~h;{haLl6h;{haLlOJQJ^J hzP*haLlCJOJQJ^JaJh}_hm# h!h hhaLljhUmHnHuh hh_Ohsheh_OOJQJ^Jhh_OOJQJ^J h~ h_Oh_Oh5E/''''''''' ()))))))*++O-Z-gd2gd2!gdaLl`gdaLlgdN~gdaLlgdaLl!gd'`gd x`gd))))*****++,M,R,v,,,,,,,,,,,,,,,,,,---- - - - - ----- -M-N-Ϻ۞ۺۺے抅y hh h6 hzP*6h;{h26hzP*hzP*6hh4hh6hh6H*OJQJ^Jhh6OJQJ^J hhzP*hh;{hzP*6hzP* h!h2h2h6yhaLljhN~UmHnHu/N-_-`-----...//"/,///////90;0001 1 1 1 111111-2F2M2P2S2{t{m{ hUh2 h`<h2 hUh2CJOJQJ^JaJhh2:j}hUhjhU h!hmmhN~h2CJaJ hN~h2CJOJQJ^JaJhq0xh26h4hq0xh46jhaLlh*Z0J*H*U haLlh2h6yh2&Z-;0X0v0234 4 4=4Z4[4\4]4^4_4`4467!xgds`gd  ^ `gdN~$d^`a$gdN~ ^`gdN~gd2sgd2gd2gd2gdmmS2T2222233333333333333333!4<4=4>4ȾȾȾȾȾզՉtbtTjhBUmHnHu"hCJOJQJ^JaJmHsH(h.dhCJOJQJ^JaJmHsH8jhXh2sCJOJQJU^JaJmHnHtHu hZh2sCJOJQJ^JaJ hqh2sh2sOJQJ^Jhqh2sOJQJ^Jh2s h2sh2s hUh2CJOJQJ^JaJh2hUh2CJaJ>4Z4[4\4_4444444266676C6E66666677G7M78888888889999۽۬۬۬sjhAGeUmHnHtHu#hN~h26CJOJQJ^JaJhN~h26CJaJhN~h2CJaJ hN~h2CJOJQJ^JaJh6y hR[h2h2shnhnOJQJhnh2h&Mh1whhaLl hXhCJOJQJ^JaJ%78899999999:3:S:s::: ;<> gd2o$`a$gd2ogd2!gd2s$^`a$gdAGe ^`gdAGegd29::::4;5;f;v;%<&<<<==v=w====>>>>>">$>(>*>,>.>6>8>@>B>T>V>X>Z>`>b>f>h>j>l>t>v>z>|>~>ҺҶҲҮҨ~~~~hh2CJOJQJaJhBCJOJQJaJh8wh2CJOJQJaJ h2aJh\ h&hhUh2CJaJ h26ha hR[h2h2%hAGeCJOJQJaJmHnHtHu+hAGehAGeCJOJQJaJmHnHtHu1~>>>>>?H?J?L?N?P?T?^?`?b?d?f?j?`@c@d@g@l@m@n@o@p@r@s@@A A!A*A/AHAWAAAAʹʹʹʹИʎphshaCJOJQJaJhsh*CJOJQJaJhXt`h26aJhNh26aJh8wh2aJhBCJOJQJ^JaJ hAh2CJOJQJ^JaJ h2aJhNh2aJhX&CJOJQJaJh8wh2CJOJQJaJhBCJOJQJaJ'>>@ABBBBCCC*CC+DDiEgdUN`gd* $<`a$gds gdo'!gd\W$a$gd2$a$gd x`gds gd2o`gd2$d<<`a$gdAAAABB!B"B3B6B7BTBbBhBiBjBBBBBBBCC)C*C[CkCCD*D+DFDLDDDDDDEҿΞΓ||ujffhNhh*OJQJ hh*hsh*CJOJQJaJh*hT hR[h\Whjh,UmHnHu$hUh\W@CJOJQJ^JaJ hh\WhhX&h\Wh2hsCJOJQJaJhshX&CJOJQJaJhsh2CJOJQJaJ'EGBGCGDGGGGGGGGGGGGGGIôzrzrh2CJaJhh2CJaJh`jhUhjhU h6hdh h*Zhjh*Zh0J*H*U h*Zh2hh2OJQJhdOJQJhhOJQJh~4OJQJh2hNh2ha)iEEEGGIK%MPPOPPPP Q]QQQ!gdA 7$8$H$gdxXMhTW$h$F8SżLl@IHMd#itaKMq҂.T*Eʵb 7/o2if*^޽s=f լThyRMRDwKzA2'NqM=9hI@>iQ*ڱaj"҇1)eCФV/C T/ǀpQN[]P$0P6iU+lvsq:M(e7ya[ǝʚP((MӷwM@ҥAD4;H_%n(Zi70A?ϛq4h*iK$֐R3Yk/bu?1mQC>zt!yvLO[;]`~J}c (w*Eywb,F}1ύZY0s_́FI|t%kctǤCYT܃mj%;} \e joBOg-/u:=eе2FrJY_@P]Q渋C) @Yz:Hg:dq`S'~mDgV(@'yy!gqco1]  ̳kէfi/ukVi^fcΚ5 5ȚנfI#o5hYX`Κ9zx)Wj>~OCD,}ƞ{Ș5kSA5{Y5KJ<bx )j@S$@~£ HFmR!Pǀ_뵊zx>s` yj ~7ϐW C k(_CK;ק*ԕᆹwk2溆skkH31?yXCna5^jh7?S{}kdZjc8{Mu{ .< ![⋻rew^9:9f:16(Iƒy: ̣YG#syFڕ|h/q4iIx9Ag pN#;o3cRu~£z&.%7HX=c/s~nVŴ h('H)Q`Q @dȋ"+˵@-v~^r鵍J,{&z6}U,7Gz$ܧ=`}+Hs?򮸽0ly%ȣmn_ۼe`H36a~d0.wcD|I?Yh;`+cށ0u1tQkLчM- pBF&YXڀYsf0Iˇ7 SlQ(-Q+I95lYgz#ZF8/l#éQ}z7<}h&߸8൤:?7*Eпld6)'4%\ҔS5[èKZӉ֤wtOӉ[jo+҉I7㫔9]thifsm`1U}:tVۻdp׍d96Jd݋Sl+VVtZIo*'zfīVڽ3 VNFc=d;ܦ[;nþ0n.}_d困iqo7kM̜(Xv݇Q\[IW+*ͭi^dyaۡ]%|{݃5k2kZZJ3zPTtpSK{+tx^V_+`!Dku8"z#xH,К+"`x͜mTy.b AXdM],ˋ/,xıQ\թʊ"eEJVin|J`jK4 *d)Q—(B45Q߽sܽeZ=/<;s\s79Wv+Nzm{~|H.Wpo,,Ϲ0Ds eaquUiE'e2:ml,`eI%o}V)w;B w[ RkWk?Jݍ%׫:׆I>s3>5,]Z_#٬MnUR'O$u|7&}ҝoG +]_9֮ ֯X:Aq:)bp- ]tFWg߈;=Ꮋ.*339s_=~pվ#d]7VWٜ{&}گN"6߿tWuI?^dz [ىʇ~W.?E^RG\&1o [6y&>9Õܔƺ&^eOE}?!]#VW4)谇/t+t$!EGW >:c16U7'<.={jD/KK0_Od9-=gϲ'|Eu䅆OϝUp%#sP>% 0} 6ɉAᤔ*ó_茮֜- o2ϖdv-ȖrP;[l9TBhYϷX׬lv݈%ݤp}Q. ֔nН29V+d~!/ɪ)ՙ vЧn^͵~>>睒U/ !gdy b6v,Ik}1xR∷/hl_5Bk{R ?Xa!\[uI;/HFLhԳ@iu kԆ)-!am`>SЧ|-7U/Ч|CkBExQ?g3B&lLi~~Zt>N}:Q'#}r$Xސ]gj"~3$kwa{dUu )3tw<<3x$m!Y! 1`7Yg4N37YA7CI*C&9J{P&1}Ykq@!8#Bdgd#}lAަyTÔpJ9r0Y^ 99Mc4ĿW3Gp$, g !(dE?vP1r B&y\O^B&lIGBdy>e侎OO|E>Oy"#`okdM3Z^* {#9(֓ 즄d?+%kRI7}G!L~7l P',ٴ}1n&+?oƸ){pC0bZ`?/cp -u [xJ@.* \DF_VH6$&@a7">M~#[>[.sZo pr0YI,e cdbض7?\815ʞyWr.dYITy_*kV՚ >aqjT;zW'S]}o}JsU>SwuTj}_EO{qɨ>V@* NT,UaAY*W8'!|JS$".\vaNK~F*|Od% eԹCw!=1?!r^b 5%_ B8/qy}!Wn^W599]PZY^yyGᒰ\%'?WԿ=d^a޼G e8/-./ZmZ[n|R5٘7'b^flPí*Ow|"L5M\OjlYu%_/,nK 7kjn:5n.%ܬ.nh-7YLnֆnq|,n.%ܬ ȧύv7ӥ {`q}WM>E;IInGVǍvN0n.%ܬ vS{/2t)fvqc],n.%ܬ BהvY0n.%ܬ efpvAvp[%bnOM7YܐnƽF;2Qʿύv7dpe;],nF 7vr$>7YܐnvM}n!%(hgqC7JQF,nF 7Vr+?^K0+; gjѮտQzwz{T p1mDOUu&F͓BE8S3'KBGay}|OQ[qT;+RaET Q듅I^bۚ?4^b4W`J0 R{;V9~A^=q'h4i4$ߦƠP,n|/W4xUW$r!Z.S*pIJtsy/yDWQ3tt1a+x~OH4vؿXzey}xdס}]rrswn&^niV[UMzVG;g2ܙh4e˷#ER"T;ŜSmуw#T=ZtIWwݠnMg"ՠfYX-OKjȾ1q)E:"NˣHv6O;{/Ä{7f(θ|-vqĎ&3LfzT3 V[+E߲XMXVG5fTM,Ynǧ5O'+]ǧ]O>>?-3W$OLOXš\⡶Wd,ýr%d>_M\Wd|.FV&5ޭS<+ZR+ !$Qi-?VޏV;z6U_"OO'u% 4Hn&$8|%selBN-=qF r Llg4xKons:ƏπzM̎8:.g*Î]mi2v[vTuYKvvavoU_Pis]x0F{qtK%a?A}ZWlggS|3:S*IvY^Ee~Uj޲eUq!p,Zs=|mXƜ:}omvC) ,6eA渻5.wfL٢>Hmؚ8M~7c|3K|7c-qF38͉Tn/qƱq\f8MhK m-w~,˲5q^5%˶o7Ьwsj9z[4mR۱[4m_Jm_ʱE[4mϦgslѬ^ȱERj{)zk[4mo7slѬɱE)y_xgTvrdKL`! nDu\-kst8f':k -m xZKW3ݚF %5DnȺn( XjJ1!nTTZCy/ jThKS }s B҂3쇚fq=wǝXQ׭J5 ?7;ޟ8)e_:LԴ<o`b\+퉳µY'Fщ(G2'QU2@%@] {Âp'q KE(;ʁ(۝&y6GCY[" cGhVʚW֖jsAo=ju;s %2^^AdC)&Nt`"tHl677u{ 䉯(w j1{<a|d[0\޾mo7ss|ȿyW!vOl.{e6ʘ~N`?Ob2z"NFq:VԼyb [jǟn.Ԫ'/^#J~wVsr_`܏*9fK9s9KL݀,v&3wظɺ4w,{ <Ѩ[6x]XjUȟVO>PDEU't K)/?]1oet 9Y 3Qu ^uJ67x.1EZ=ֹvZg.8ƈb7 d+*aVK$̕ð b < xou +@{L<,Jqu}P-x'qI!UՈ}}?77(;7n uysnv_Ĺ:us\ 6PyU|e:4Is> saļ|EO[NhϪ%[r!wDL@~L£0LK<Ϥn-8D8W^9iuQ e# Ї֏Ax}ԾOj1nzesı9cx噅5Ir!\MR߇ -$yjrOyT";3σrX{DZ˚c^xagMftgZkw֣*a3 L G?U3^HֈAbe/]-M8ZAn2\1r4"s}^xO:!u5үW 0r.~I?ĺ JLuvE =?U#jko^P;(g?@M >*珃:4.EPj /DR`=_vx̬gl?;,ŖN/6ð,[bV /6ð,[b^`؜22?σv=8hewA?= McȾȸc6/A}rBt6/ng⎠GWë,h+ !58kl;M ՠ#{@[bP2 COh bn+y3){ k;Zz7ղkW*ԓF+ 2䫛G=ȣfSH =O..47=qO mjmvR(.fKF>T_AGЄ^_^ZY[]58ACYTnD6nw\OR4+=6tUQSHI0a2umKz)ju$N3~b~FĬUqgm05H 1 a&z;H'FFRv@9DoJZ4Q'J(*jj:<7;z&AMyhԉ@bh阾#2ړ+Bs{9JDx/WgWWD2_{T"BD&H DF '2y<9Ogjy#oN6OV} =VVl}ɳ+mWW+ ][(-TX[ls߈ dg e܃5f8-nQ2H'}sxh ~jAOΥwYꑻR'>ŕ`!i{uʖE:P[yR 7xYMlTU>͔q  D03`N[(!SmJ MdQѕ.\T("0FM7oxco=s=ZKR"7V$e8-hYJx~РEXn/Ja)`5f݀-98]腶CBpAy߈KØl{̂r',`gf,{8P7_K4[W)U7‡ %b(" DU/I^ ڬJ"ہ5m$ǃn%T?0t1 ,zc[urvR~}}nԡrL:#+bd3~+_&ԻR r xǣ?0oRѬ&}3V7z: 2MkMu}먼Dl:1d?.XG|yF~<8}^GП~шFhgƞs(:*}:@#uĭu{ xuw\atЧ|HHBOMg_EugE@eQ~aXa`Nb>5}E>~>g16ٵ)/A_N9] <3kσ}Ѹ>7/jKs:>76Cd|'?f12l.CG"pZ {sCeQ3wbd=i>rMл9jM^ M*kfMdçwt/PvTFFwAqk0/]纘u,zfƋa{̦\^%/&EpMQ̚?\kB6aĺagiN 6qX |T2'.uOjk.`}`yO+CwqT=cӱ:6֔?6Ŧ+8ngl:^`g:!2&x4k?kǍbW匭yq(6{~9cl72ƞ,g2Jc9?yo~SsͿm)l@H.j4m?Y6vɍCA- 1+il Dl(UY6Mo[yu&R/8׶^j4Y R, m/TL;|Ʀ&OM_ƹ vfw=|'B =k\zgL)[Iߩrθ3m8Go&̮o OWfw6q,^M ufZ&SgSx+sj.z$꣛zٚ#䙇R`Xjz|,N Vb-hVj;sescZ7jY팰fmj[fPv(–aY -r_";%`!yL<(3?Vȿ2n(i@BATxݝkl\qR)QtXl9}%Gl=LRԃHKD)" h%m,w)YZPNmඟ(BHчmM84䃁D֩^ʲ<朙99޻k:evEъhmghC|tdp GW>.2~-?JoDh߱(NM902h<->c]jI~6Ə[ٯ.$ "E9]wZ:ړ 1hT {Zչسr!ɰ=-NDZDsĵ975INw'umH4Fv;Ev٘PGH(s {E݊ M4&ֈm\+77gX1(wM!~3hEj+º/C2ϫD'ۡdMIDG Xe:ݵs<*FQ!FqtHyGQن5)׋^v~>:ѓn]GN>N"?VG)㾬ߍwbruz>佬vp΍KӞXy;kOT$z|(/O{_olom› -/|u:`16}ju??nEQd_o{ސG9Z̩:sn''*do|Iuգ蝽 /79._Ү\'&/k*ǢO5Sz-J~j68pE(T1Qu#;U4]:zn=+Dw$ԚTre0~1}2lQ";o}&xȴz(+Oqf#lNh :AXk11ɴr\}ω-t\I~J 6E#y༨F <ŭVy6A*1^%O[׉b lq~TTIƇ&cE%Eϰ1r /l:1&ձښ)~juօˢED_h:ԟeÄxѤ˷˵CLX Q,kdȸ\{B[ǚl ZE286)8rQ3s!gH~ &W J䄺nYO|H洯oJW)Ge>5q3v sMf~ kf::DzVf< [4YiѸ*Rό΋l XkF_ذIe~-?~ݡg9%~"ږt͊70O#KҰv 1, [0_XdPiL>YpY~>㉕oT~=śLK\4Fsx"?#"$ t.~[r7oxYAZ`QVݷ7O<شiMJ+kg|Js.9@>`fأ\)KzKDΡ2^Yx=-|x3/u:Ly+Ңؘ͍[meoGҦbO?ּYxZ veΫmd6&%6W~1XDlx̃gW$"'Y^q'h=]䓓z.=$~vŻYF˛^.e+ Wƅg(?q |/,|NWoA?ㇶQ%zmxrOuo䒿%rLn΂\rh`Umx'6G>wE/s_E!.|_<<0.Z'䋪zdfD ؕ"|AINeIi23ao\TNxOyͽtZ.LT]c Yo0:7He zB /^)ghN8Ԓ|t5l[:GP~sJrlj] ҅(NѨIYGO/TO(͖+-bwIyh臢O3ooyy5(-xEƕuS;2 J\ҸZ2Ro%rlճ`>!b/{?T.񐑅 rFgx'E+:1o1=ZL>|r9-x^FGDbd<δTenVr,!Ei#Fp9+X\%&i+5#uaD|"c[I9mb65?#9 3t뚈CW2jpw)rv & >f/%_'ؿETga,Oft ^aIDa?xϿݽC׷E.ѸaQx7j<"tׄ"6 ԀS d&f[E'<1O+Х`Ie$ }YZyY>3-\b.ykr\qg3gi*pʳ F,m,u }:~,9z&,sSiL!%{ӹE[^ȷ,{N m/N)Bt[ޠI&o5CBOJ6^%or-Yk~m܇ݜEO?]}CB0-T./x'ϊ wߗU{Os>y,5[Oa/xg[22>%16]%9X?^s|o}RYߧ->L 'z]{P6eay ذs`+7s6?k 6aO;Bb0GߛDe<31od$2'֢2}O?.[TUbءY`""K=-xw7, /gGwuVqTp쩑 1eua^/؉ ;Ŵk"a7MzCD2}EO}bc #C O;k{RD 4c~N8fBF'ED/k<ҋ}]e79|.-x*tMI*Pg!O37|BQotww\l߇oU&O>ޝԑ*C*Y[*k4Cex|yZ{ATӳ}Z+_k쿰:vl3Udno}VZ=˘ׂ6.z?K􅶽-mlfmβRbʶwRβRbʶ} 7YXJlRl6YXJlRFν) mT_(eά\l,Z< m#lXJlRFβRbU]ȵ6YXJlKbF;6Km^*ȵ6YXJlmXs.^$R=7yݩSDU6 ˘&%mn6YXJlRF>βRbU]6YXJlKbF;6Km^*6YXJl-Q_`ko>bAfKsto>*H6sםߐg=߭o.^'B7c=Q!YT?/O:sD"kP3-*:E[Er+ZS\Zr/ע5;]ْkqps7n&B[[G/Fk[Fr+<_YBڷj-~ymE!۞XNNF{Rg8YƇ #|y)&\N루}GpPR!/xlfr0򧋵 qx/ת@t|&Wj5Η'ԙ/m{w i-w=t=kF,ӻ'F~<E-Ӫ>$t/-VC.߯5RmHL|V>U=~&:h%|ڒy=r]>c|\X&k*W1m29/_' #I;-I\@uo-so<{[ڶX)JKZ$ OI’PYۙ]Ղrc_u-lNFYg十f޸ٵUijIz< ]x﷏(7'v+e ;9sTS2axR#W 'IVJw<:u;xe\A-WwT{׹:& N{݉h!{kڙ8+z? 9?(\ 犧 U{@d-uB~||i!񵣻{Wsм-wj#SEm)E@G"4z w߁werl[XuXK Fg3hj 3QPOO"<gk;z=>$t/IlDOlz*|1磏UðLCWn-ճeům}U=ZMx)+L_0>_mop޳cr_I{^5> <*9]ip:wD գK*~"[jtOdYIJD۶z>-76S Ǻ^G\NL7%1@ҋzc]S[߻e8zl5vmޫ z;v' z#CA'CX?>.^{5̷+}wNJ?%+V8!#3OTNG&>x~]jӣlU~Cד}(kәWih ?0#ExxT{S_mGj { זۖckH BG:dX[#;Kv젯ː:ԡh>R)˅nbacHۖU=t%d۽}ov N#DNDŽ$/<#:5Kp6TWO棦+w NXo,|f>2gwKz{O#pS vWcx2S=Q> z\ں}ΪAo^MGjefuR<<^"L#Ljug8e128;;gv&Vl=iv/(ěmf;?+*ra%vS֧ %'w7$ؿD…|a*"g 8r;MOb(%Lf;zCx-<.8uO/1o4$oaS .xrhtgX^?ܓlm}> Tr{ۧC KJܭJ&'Gtw/s;+\x?ym92a|fpQ7N0f9^V;ny뗻UL| Z onױFo`XVƿ`\]sNxYs_ٝ4رl PsߙNƯw{O}\7p3pfq}|Ie5~}|)g8f5~}|)g8f5~}|)g8f5~}|)g8f}%3OT}To[cPfu˕EG32=~n煺5|5Po\WMP7 [)w/-Q&h/oq׬Pi|?[[pO8m`! TVF|!ޝ.Mxf004 xZ]h>3ZY?%1,,[l$~--!$&+SARrYjܟ`BlCCӼ!J):!/܇%y(qA;hy:|sϜY'u"Un-I[O&~KH3&m,Aՠ0.a";ǎR6w!ܢ :0) x''+~P+p]wbnӗIK}tDݶBN0絫PW{ +"+# 5EY #v;_]]f;g.sIvjAAw5Wy{{'(m:dF0`*7i ѷ<B[Ge 5;/rZ`F̟ ,K~ya役3MЉ[gXW)d\h?sSzȣ칗-#X/<]."iGwr#Oz=i3sy=xMm]?xI|y=T%9 W {?)gi.bjjqK{1?0|}2?mk/WxkJ2.(,iyq\?HTY2;5Leaȹ)- *œg]??HTY 2ƞϨcT9'\? *<-un>ٳwcZ*Ybj14We6`3އBz!>D=Q>HzC0o1 IZzXzM'UqqUgqlܾ_ *-yxc<*-CKUY6ZW쵬S 2(#oށxÆ5\U(j$w Ua+WO`b'!+I#}DN|.~#}Nx$|>.` >t2q!;|d,ʉ3FXTKG%drߩV2O6bÊ%'IuܰUS((⽲4ʃR]ܟ++.uԖ9 ^\yQ[}Ĝ{ IPoQ#}5}U7ִi.G}y:7HtMA۫,7|S]MmTi߮R߸NMu97]oo_aH}:7HtYmpo\雮7˷!q4T#}5}UYe+HrjWu;_}&cv}:7եo)4(o2r}:7եou%ϴVA/߿㣠< n\'š:?I.}+>#,E?6s'IGWb EgAdX+!i/xX{7s+<jMm<`jU࿆_pVWy>~k4R]f[]MZ%]kiq5lLbgF7wWs7G|>0?}PE? RkY ! i)r2id8m)\3G}Ùut>3rEǽ tgN/8шF##Ü5ƬFQo4ѱh֢zw5.ｐn'آ|myVO-ijSH5,+&׈oDP Zf/ȿlAbO,?ƿg1~;h,,, daQJYhh4MfQl4A4EhFDYM&}=?Oyu~~t~y%L| /yNr=mzZǻO'nFN0m;h1ߩZ)yrG*ӸL♺4&63ņjk3nWB ;Hfh;aKE׵sRU5^\׶ãfҝ\c@C.?Q#`! P|M>GtQP*C.V xZ]h{w-YH2qlYejd'Wq`C-Ԯ(!PI/q!&ZP >!@S&L4}Jd(R↨7;ghW{Qtg7gf̮SmJ.}IV^ThwSm;ı1*婏 RwkN /C1.;iQO# c?Y)UAɫ_^Pp{pފ~\ ,hDJ^϶K_c1f勤D|JK鯬n}VPw%I5,==v^62V2` T0N!$>X*5gPԳ׈F-4}Ƈ]T` ?1O\Ly21/\Ld٘C&O@.&m`gc Ճ9?6dbՃ1=l*1[]L<`̽ Ճy`A:Uö1#;ikw ; I/B;yWRD碌y'Q^,ty睴=ͅovvv>fX/\QBy1 G'kx6-M}D-v<3U$y 6uYwAIlԝ?&*֧@3p΁ XԸaۘz⼈~fn|.CӦzu^Bz.de*gTx zβlg^iϷi+[OԻiEbQrv >k"APLlY%3ПA@/6qT jf1oDܿ* WlٸA6zkC[Ykmaڸ\ڲ06؟%-6o܅\e㖣5\ʽtAlٸks[&b];|'-kkBb})X 8mԭunx>j;A= _lYcQ9)uC=Sp& rs[|w rV~$/zˈO¿=rVP =߁pVp v2 _lpwAn6v^\D}h.&`:%3UA _lg XUs b^`3@|gz5X7 ./#> 0Gt?YpVOHʟuN|_lঢ়$y~E&F~y* }Pm?rVPmO" ı8"4/+zwq!90ޝfW/ֹͬWm=\PhUu^i4Z3hi.hM`i^ךy[k,m4S,̀&\u:$Ul>_\0r\0Oיqά6(eW0BPX܉В7BF[ţuuq w04nF,qFH!z+ԕk#㱟'ĭԈgvdJO[+Wz$gl3l׳WNPG}o}2!Gv/σ+\ɪ"e[YYTEEUYYTEEUYYTEEma;-7Kpz%jgO{uMzn_"xu)7J[<ᘂ_ c[{tc'ɌXƝb3b^^ψ{-v1]̈{-v9]Έ{-GcpzpYCQ졌XKl/|mL;$=/ݗKxh;Kx(tF,=QtF,=sQ\F,=N[:_쾐=*?0=`!/9TRbCznR xWMh\U>ߛ> 1#N y3C3iiLC X.f#pQЅK", (t&MsBWU{f޼0%F7sϽ7d^A}}lP%|i܋|5*ӺP[<0ҙY0M=uОH»~ۓƷItO;pC7\=,-BfGZ!p+έ13l^(-qph~+7MG좟P^)X|t_hSɹ)c݁솃u&o\cn!vJ,Krkkr~jLF~謁uf흜[c/Rraq%'PK*%h;+[nbr1>C_4[cɛsJ9g췜\WL( nhirτs c椰}*KП2ypK复,{v;Կbq=}xFiyP!Fio-@UXr-yHsN[IHN4Aw|29jsX<Y psX kB+تȘuHcLޏ4VnXޟo^̗k- ]VܵeҼVreҼjotI:oEpAy |oͻqlGa;$rR?;ށG @p</4%\ch@e[RV]g%sۄ 8:fn!׹>mٱ#Y7eNmȭdnʍM=scy8C,7Ϯw#@xg?pn~>̻x>12c@0 c> 8gXg8#Ȉq1ӻkuM,ql{yQ no%|Q^XY^2gj*;%eTREUxFN9 uA<(顄n* [~E-cbc;t6܎=~%0*rӜ-u /_yUףuIVO_fVNǻj~s]tΘs! C GO ʕsK6mglXzv8qӱmg'!=;ckZXzvڀa'@&}GE$}DyK _Ref286587916}DyK _Ref286587868}DyK _Ref286587868b? 0002 0@P`p2( 0@P`p 0@P`p 0@P`p 0@P`p 0@P`p 0@P`p8XV~PJ_HmH nHsH tH>`> Normal$a$_HmH sH tH j@j Heading 1/$$ & F h@P@&`:mHnHu^@^ Heading 2 $$$ & Fx<@&a$6]mHnHuhh Heading 3*$ & F hd@&`a$6]mHnHujj Heading 4,$ & F hh((@&`ha$6]mHnHuP@P= Heading 5 hP@&:mHnHuDA D Default Paragraph FontRiR  Table Normal4 l4a (k (No List ToT Abstract $a$5CJ\_HaJmH sH tH HoH Affiliation$a$_HmH sH tH VoVAuthor$h(a$CJ_HaJmHnHsH tH uFB@"F x4d Body Text$ d` a$@T!2T bullet list# & F h^`RBR equation  d OJQJ^JnoRn#figure caption$ & FPa$CJ_HaJmHnHsH tH uf/bf footnote# & F(A&#$./CJ_HaJmH sH tH lorl key words$ x` a$+56CJ\]_HaJmHnHsH tH uf/fpaper subtitle $xa$#CJPJ_HaJmHnHsH tH u`o` paper title $xa$#CJ0PJ_HaJ0mHnHsH tH ulol references$ & FdL2a$#CJPJ_HaJmHnHsH tH uf/f sponsors# g$dN` CJ_HaJmH sH tH FF table col head5CJ\aJLL table col subhead6CJ]aJV/V table copy$a$CJ_HaJmHnHsH tH u^/^ table footnote$<a$CJ _HaJ mH sH tH n/n& table head $ & F dxa$":CJ_HaJmHnHsH tH u(OQ($1!Figure!*"*'1!Table"^/1^1!figure caption CharCJ_HaJmHnHsH tH u.2A.!1! Figure CharJRJ E References%$ & FPa$CJPJX/aX 1!table head Char":CJ_HaJmHnHsH tH u,bq,"1! Table CharJ+`J )E0 Endnote Text($a$OJQJmHsHN/N (E0Endnote Text CharOJPJQJ^JtH >*`> E0Endnote ReferenceH*<`< ,`<0Header + H$mHsH2/2 +`<0 Header ChartH < `< .`<Footer - H$mHsH2/2 -`< Footer ChartH FOF t No Spacing/$a$CJPJ^JaJB' B 2ayComment ReferenceCJaJH H 22ay Comment Text1CJaJmHsHtHB/!B 12ayComment Text CharCJaJ@j @ 42ayComment Subject35\N/AN 32ayComment Subject Char5CJ\aJP RP 62ay Balloon Text5CJOJQJaJmHsHtHJ/aJ 52ayBalloon Text CharCJOJQJaJ> r> 8 Footnote Text7mHsH@/@ 7Footnote Text ChartH @& @ Footnote ReferenceH*6U 6 gS Hyperlink >*B*ph<0< X& List Bullet ; & Fm$NON>+oStyle1(<$ ]^a$FFzHeading 5 Char:mHnHtH u.B.<+o Style1 CharPK![Content_Types].xmlj0Eжr(΢Iw},-j4 wP-t#bΙ{UTU^hd}㨫)*1P' ^W0)T9<l#$yi};~@(Hu* Dנz/0ǰ $ X3aZ,D0j~3߶b~i>3\`?/[G\!-Rk.sԻ..a濭?PK!֧6 _rels/.relsj0 }Q%v/C/}(h"O = C?hv=Ʌ%[xp{۵_Pѣ<1H0ORBdJE4b$q_6LR7`0̞O,En7Lib/SeеPK!kytheme/theme/themeManager.xml M @}w7c(EbˮCAǠҟ7՛K Y, e.|,H,lxɴIsQ}#Ր ֵ+!,^$j=GW)E+& 8PK!Ptheme/theme/theme1.xmlYOo6w toc'vuر-MniP@I}úama[إ4:lЯGRX^6؊>$ !)O^rC$y@/yH*񄴽)޵߻UDb`}"qۋJחX^)I`nEp)liV[]1M<OP6r=zgbIguSebORD۫qu gZo~ٺlAplxpT0+[}`jzAV2Fi@qv֬5\|ʜ̭NleXdsjcs7f W+Ն7`g ȘJj|h(KD- dXiJ؇(x$( :;˹! I_TS 1?E??ZBΪmU/?~xY'y5g&΋/ɋ>GMGeD3Vq%'#q$8K)fw9:ĵ x}rxwr:\TZaG*y8IjbRc|XŻǿI u3KGnD1NIBs RuK>V.EL+M2#'fi ~V vl{u8zH *:(W☕ ~JTe\O*tHGHY}KNP*ݾ˦TѼ9/#A7qZ$*c?qUnwN%Oi4 =3ڗP 1Pm \\9Mؓ2aD];Yt\[x]}Wr|]g- eW )6-rCSj id DЇAΜIqbJ#x꺃 6k#ASh&ʌt(Q%p%m&]caSl=X\P1Mh9MVdDAaVB[݈fJíP|8 քAV^f Hn- "d>znNJ ة>b&2vKyϼD:,AGm\nziÙ.uχYC6OMf3or$5NHT[XF64T,ќM0E)`#5XY`פ;%1U٥m;R>QD DcpU'&LE/pm%]8firS4d 7y\`JnίI R3U~7+׸#m qBiDi*L69mY&iHE=(K&N!V.KeLDĕ{D vEꦚdeNƟe(MN9ߜR6&3(a/DUz<{ˊYȳV)9Z[4^n5!J?Q3eBoCM m<.vpIYfZY_p[=al-Y}Nc͙ŋ4vfavl'SA8|*u{-ߟ0%M07%<ҍPK! ѐ'theme/theme/_rels/themeManager.xml.relsM 0wooӺ&݈Э5 6?$Q ,.aic21h:qm@RN;d`o7gK(M&$R(.1r'JЊT8V"AȻHu}|$b{P8g/]QAsم(#L[PK-![Content_Types].xmlPK-!֧6 +_rels/.relsPK-!kytheme/theme/themeManager.xmlPK-!Ptheme/theme/theme1.xmlPK-! ѐ' theme/theme/_rels/themeManager.xml.relsPK] :| ~ o iJ%y==C>5`m-V n#i $ ( + ?y{5` B   m[!$')N-S2>49~>AEINSVZ\z^_`ceg5i579:;<=>@BCEFHIKLNOPRSUWXYZ[]{ "'Z-7>iEQZ]/g5i68?ADGJMQTV\( ) )>>>5`HTHTHTv   22$RJ 90/2$Dku8"z#x2$nDu\-kst8 2$H,=ϏU ^2${uʖE:P[yqQ2$yL<(3?Vȿ2$TVF|!ޝ P2$װˏi*t2$P|M>GtQP J2$J|r7\JIk2$/9TRbCz# 0e0e     @  5% 8c8c     ?1 d0u0@Ty2 NP'p<'pA)BCD|E|| ":0e@H:H: H:H: H:H: H:H: H:H: H:H:  H:H: H:@ABC DEFGHIJK5%LMNOPQRSTUWYZ[ \]^_ `abN 5%  N 5%  N    5%    !"?N@ABC DEFGHIJK5%LMNOPQRSTUWYZ[ \]^_ `ab@H 0(  z 1(  J   #" `?V  3 A #" `?V  3 A #" `?V  3 A #" `? ! s * ?"<@ ?NNN?N V $ 3 A #" `?V % 3 A #" `?V ( 3 A #" `? * s * ?"@H:H:H:H:H:H: H:H:H:H:H:H:H:H:H:?NNN?N , s * ?"@H:H:H:H:H:H: H:H:H:H:H:H:H:H:H:?NNN?N V . 3 A  #" `?V 1 3 A  #" `? B S  ?1!+=,19GK5`(Fq t "t$M" 0tu. t%I  tMtq3;t.tIt*t E t1t!Bt, t _Ref286587916 _Ref286587868 _Hlt286587891 >6`@ >6`y$+<>Vo-NTv0 9 XZ]_++++ ,!,{77<<6G=GCGJG[GbGnGuGGGaIfIJJKKKKLLLLLLLLOOOOPPPPSSSSSSSSSSSS T T6`y*+Vo-MNabv 9VW]_ ++, , ,, ,<,>,Y, 66&6(6K6Q666b7d79 9::<<6G>G[GcGKKKKLLLLELGLRLTLLLLLSSSSSSSSSSSS T T8VVVW6`333333333333333333333333+jx!1FY ##O%Z%;(X(+,=,`,./11222 36K6>>LLSSSSSSSSSSSSSS T TzTT*UdUVVxWW{\\P]t]]/^6` v7(;3 N&bi36f7+@O9Ԉr[>`AjJTR~֮X,@l-l !joNJ% ^`OJQJo( 8^8`OJQJo(^`OJ QJ ^J o(o  p^ `OJ QJ o(  @ ^ `OJ QJ o( x^x`OJQJo(H^H`OJ QJ ^J o(o ^`OJ QJ o( ^`OJ QJ o( hh^h`OJQJo( ^`6]^Jo(.^`^J.pLp^p`L^J.@ @ ^@ `^J.^`^J.L^`L^J.^`^J.^`^J.PLP^P`L^J.> ` 56789;<CJH*OJQJS*TX\]^JaJo( ^`^J.pLp^p`L^J.@ @ ^@ `^J.^`^J.L^`L^J.^`^J.^`^J.PLP^P`L^J.  ^`OJQJo( ^`OJ QJ o(o pp^p`OJ QJ o( @ @ ^@ `OJQJo( ^`OJ QJ o(o ^`OJ QJ o( ^`OJQJo( ^`OJ QJ o(o PP^P`OJ QJ o(^`^Jo(;@`789;<B*CJH*OJQJS*TX^JaJo(ph.G h^ `56789;<B*CJH*OJQJS*TX\]^JaJo(ph.G`56789;<B*CJH*OJQJS*TX\]^JaJo(ph)#hv`h56CJOJQJ\]^JaJo()@  ^@ ^Jo(x^^Jo(()H^^Jo(()^^Jo(()^^Jo(()#hh^h`56CJOJQJ\]^JaJo([],56B*CJOJQJ\]^JaJo(ph Figure . ^`^J.pLp^p`L^J.@ @ ^@ `^J.^`^J.L^`L^J.^`^J.^`^J.PLP^P`L^J.#856CJOJQJ\]^JaJo( TABLE . hh^h`CJOJQJo(sH []6f7X,@l&>`A>`A>`A>`AJTR-lO93 >`AX,@lX,@lX,@l!joX,@lX,@lX,@lJTRJTR         `y        F        ,}         \t !;F hl'+#%-+@.9>7&~7h\$^D_m#2s iZ ,"8K_1wbP|,at,#`%'|'y) )AL*zP*.-X-P.o./[$/O/X80~4W7-;%=W4=%?5 DkD5ENREJ&MNUN_OT[EW[W:[:]<_`9cAGe&h.j9laLlmmn@nwtn+o1r6yzN~aC8n-5i24gSV5s%X&|)Aq j\ sk:Y*Z9}_+/h `@Unknown Sharad Seth G* Times New Roman5Symbol3. * Arial7.{ @CalibriA. Arial Narrow7K@CambriaG5  hMS Mincho-3 fg;SimSun[SOO PLucida GrandeArial?= * Courier New;WingdingsA BCambria Math"hUU mG* mG*i584SSD2q ?2! xx$Paper Title (use style: paper title)IEEENagy8         Oh+'0 , L X dpx(Paper Title (use style: paper title)IEEE Normal.dotmNagy5Microsoft Office Word@^в@ |@x mG՜.+,0 hp  IEEE*S %Paper Title (use style: paper title) Title  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~Root Entry FIData 1Table^WordDocument SummaryInformation(DocumentSummaryInformation8MsoDataStoreZ@TZW2A4UT3XKZA==2Z@TItem  PropertiesUCompObj y   F'Microsoft Office Word 97-2003 Document MSWordDocWord.Document.89q