ࡱ> 68>?7n&GxD]v*חPNG  IHDRoX|tEXtComment̖tRNSbKGD#2 cmPPJCmp0712`&IDATx^|leSOhu[Ƙ:KZ:kSp/ :_,qD~[9#D ڢZܑY 3h|yw3v88mx}}}[Q[ng4_ wS%~B+[ `/97-r%~cع62< ڭ/Kk@oY{)Ђ7=\@3_ iܳol(Qb {Ѡ.tG7f:c8B~ 0zQi7O80j2@l э!{#Y7o7Mwڃ|Nc;r>U3c7'қ؟ar6Nɡ0/@mk!|'prf8R=|;:C ?mEMtѴF]˳G5~Yd&C$oh~/=XҦƩl]οŏԮ~5~6,ooQ˫;~CϩwG]n58!_kss#R{uQ+ZW&ȋs> !@ sJtwbx>~G%AK.\D 7z:ۮ)gݚh&Nl߹Dacn0?jT>xǽ)mA;h-콹$[2~g9lJx:ZǼ.䠶|f>O~6"}h_R=kVq:z'wm9V~88:]rx:2=! x._ntq=o>m?_l)k2q7b頫WcHD]ĚƑɆ.Rg< |ɳ[>L/S&ַ|IMxs;ddWA"PO@&H*aHwH%nBCe愹P}-9Vѧ7̓ $o/[w\]_n~YI`-CI30 yƙ1%s?=6t`ӯ6?Z?%N;`7_nY5p0Vu 09c ?Bm:~mlR/& A]'_BJzlMw 0ف[w1ES\yǖ#Z EjH†9~fy٤]f&3rV4!>R[ϑ!Pe_pO-g=je.k_>@g[Únma&$k9/ 25_n;UĢ8V,vrxHOȭ_! . V~{F8[\=+R?= ˠa*>4슎ARZ3@qʮnFIqoS^A: j|]}-D2?.5\ay%mlN (pdTW\O&=*O 59a@́"2E)]B6E{J_D7'E Y@>ZOV|;drgJ*DPs-δ41j)Qf+!i13@ŹOC, +mD ֠ADAL+, LQ}~ ll@5y6TjWGl:L#shfEj]U _<,/5ش㓩۞FP3W~OcF. ^5,;vϤE?}ʮr YzKoչxxmk=գs)vX51qv߂Xвw0E;'ze4̯3 9Ue~NjL8svv7B:M|?c][Kx0c$G 9L-NPc vWʁSNwv˳7F.I2p>m^nK+%k̤r-q=)ZOc*{̎bKfB .,ST/꽲 !VS%sEmW [5_UEĽVm76q6\]_;TA}8"X^]\HZNr>[`8κÛ W- /UPljd[gǰ0<ʍ#Qh9Aft8lvPZ)ဴ翎c;&_ Ƌfn;01(~\xĿd O)lf5Ѳa@.qnQџ0LA3Qd"Fz:Nżh,,l4zv"I ^W3Tc˲m.By$ʙvq2lu>[;ͪDŽ(0\X|p'%CTd$ #fڏ)v+0N*`*#hb8ma\DfRjr-On'syɪSi[L5۟¿ISĘ:ߵ'Z:gvoRFwsSPv|.#O]'ea\cWhEiS{ =W372ӍlWdc j}\R'iy3ܢG=4p٥cQ՘JDu znv6)#uaaD1l9ǹIlqZb569Zh #^в%Lq2Ԣ{iLCMjMU/W  %I %#?:֣t[+ܾ|[K:\\ xbg\n_l5C5tvܿ( dh ;'-U`\15-{e/OxZ9BhllgŤlMk < #%ň NQQzHqK2oE+ms}B2T$x>pMAkۏ$ Zcn.e:]iV3ns>J5L2JW#m]u1_JfT$ԟհl޼yMܸD'tGY5"cޟ0lŗqɤAW*yPs_lAaӿA|[59Fooxn"79%vUH6HvG@-G>zyx Lc\vm-zҗx~h[ͦÑn@cL[o8Bdgi;#@ 8~EIH<=G\ueH|NZMd{Ds o<7Ew8_;C;V잘'3fZ6nbDӒ6!8⛟?/cGͰ4>qBCIl$SN{{̋j_Ev-Nba| nA3ΦKe(D˒ݬ̩ f^þ_OGMFM}%RC.x.א]ݧWtlt2?kwS]57iO8K͸9JYE͎at0\QR*%cZ67= YDRuHsq[A-R@Mk}pBl_өphh, %6ҐMtiD{n<V.N_ԍUe݀Ȍ^ʌ$S`5i ,L>Id6G2II̊c;N%^1p4-EpE-t-h1{oz_Cק$\OeQ~Mg8S^Q6 ,˽͸~\aŌ{L 2 "g<U,-E FTptQ mGOΝw\=C Di=,iE2!M&*t8? 2ATgu(2:1=\J{L%\E^6?uT@9s14j4em<ÿj73N- C^pzZe?ք7ޕ* ZJ>o7 '}%?hO6!lS&5u;yc7XjFRT75`iiB$H䁔ƞv h]4xTz'3h;IjOۥB;gk{bZ~٬7m[8?/};|xOn_)3pyYG_|}?l[w>.Y~`2<=uhu"]ڶ,'thw˜5ނ] d6?IaN<c H˅e`E+0*_ODu!WgںNS\ 'otnaOWjc1(\Jh8s/Y7<&UyOla7i۵ v]m={uDuIE#hX:Y 3um'R rKZDQ=*aCGxO аo1 Smµ}=S}ljaJʯ=B"f=s O@^NBrte~+^u{clg)]*QaXDQ܁bCV^zk.7p_POއWh)]C*"c;bQcFH?t={k}:Kq'1|fϯ2-7׏.d2կ~ ZhJq)-ŋ&sLcd;Wle@Y;VO9MKޘl%5x2Ҧ}28|`.vLgRK{ni H!RǠJ6H?H墩5{D:xcUD[&IAI'RiX'%}-RRٌWL8j!:(钊q=RQ*U5U4_=j3[dMUMvPBi}۵+"aA~P)4y}GhS҄9hFL"?c15ݒj[-Q刔Ԕ@D*Ye BO~cN,ȝˣ茷&_1 V;o @Jk+se"Ps2ƅs2~Tbkb/1xl\=˚!+/,o!k]ڋ ʈ m " H8OAnymROtxX."^2υwÜesזjw{>?`=edžr׶Z\bE%dP @iL)@=2E2}P  _@?FH UM"g(#kܰؖrc:?l)yC54n>{}\M[X|Y-p%#&tґrx =Pi^+n?$lZShb[CGk}kn%{gOƩ;|T>Yr"VG)#ѿdT6XF@Eju2J]F،*Pr@?=~-nDACZ;? ϤDaJ@Z12~3EH|<]q>^04ҫ:e{Zx_8ʯw>~68.6b^?rx w3BO]l>*1w:Xݍ~eOr?gjAf?)}p3GY8k*>czjF:m<\B[op /[, z+~/=&欿F{ ߾|xו p_o'8 l͝o|USRG6 @L3~7a,tPc)B1 AM. ԉEم7ƏN!8j5½a4y# S c@.CZ6<5np)1J׎N +4:0%0bQrc,9i~{?ƨܛgҁ#r/%Lmtry$7 njAmq) t/f|H q_ܓ,K7C7_Dz2T9|NDv^lSO5WF /`V2 >RGECg ^}:JFB ;KtKz]1 !+8녺؛YwnIF3W$;4w\oy~E#r[3pCIENDB`n h07@8I86PNG  IHDR#c"!{0PLTE΅HVl TKtRNS@fbKGDH cmPPJCmp0712Om IDAThmm|!^CYc5]I?!iifz=0C,K@WJQ3-`FH|:؃EY<Hhy9۷#]fF`R҈飴z'NV3t1 8؆xa՗YK{JKm?G86#DgO!(+~ IgFwqD˘Ư"G?Bpxwb-4D%y1uza"m/#UFrꈄzY He$݇!q}B};˨#ҬI5;Hrwi;$7=]>$?׽y_9H:O =OoAڥN` ڦZ-E)f=JYY""xe3Wĺ uYv Iu–pHf t@"Ү\vLY~$ /m:R9]tpϑv -BRFˡHHZ}1^Z Fhl! x6!ʉ,SQb҉-56^!WNYw*ާ*w0V1ʗV>1RF?EzJ1}73W)oiz߸)I؟@**%W]] ?TgllR,$s}iw o4\ \OMsf)[o34%Eq:$j2J?xHN}FrȮ؆e*HvK>+^Nit]y)P^"z%HIJ5cGkHOĘȟ I"QP{3Cz$j$iH yEA\!p4DJረF)IHLABz6s2 $JjUt? 6KoxDP6#᤹Cr gHf6%n 툔 5Fm0,"WVS-sIBHS!iF⯍mHj$5H}t˞ hRƬ ##)e8C2c% ?%MNQj⑚j$j$d䒔EH.-rVeYɋ ɫ]uH>IaGV!爨mHA肴 ?tHAo sAܸlr?aƜ"R\) 0Tf̘'<=" %$]#&$ D?""-TQgj$s2{0o*l([S YlahE<ڿbTR]EtC6eUΫ᠒hB2]Lh+$'k$'b؉//(D9j iZP =RNK_/r{SΜoG9c;kʉgoQZN7|"3'n/ G '\)yTB6(AsT!'al>#=ت 3ew!1'Ǝ)D鴵PΏײqܮW/!F~OH∦D"g\K+_DD[q}"tzӗ_ H?]OnJ!_WT$lr]l4oc ]ҘPuߋIo2KDF.4NO 4x>TViWؙ5>V [l%.ݜH*~t]/;Eh*,Ś'I.HS $Cc%?Q8 <p/fě :b:t0}xN H o$w«HPq8.6D @ri' ad$F_# @IoD /H˖,)v5IHW}2ҐrP+$23z>\F'F4߮JHH+ xAkB> Ot+Z2kLHZM^k VD&FHb{TM T! }7Dlky\F 'mBRKA:B`\ˮ iMI%4+7*"҈/LeC5%47.#:tFs`$0I׊*4%K*$/{@{^ٟgVtP൱ }5Ij*t/?Ziyoy?u%rzIENDB``!|Xd}ָ8lo.xڕkA$M2Y64MP=\VФncLiЦP/"x_EQG lC/ w~a i²ٙ޼{ @ _ԽX0Brp#9fH/rG'FF!}f=wʽ,|ĵk4X(Rgda.~\oj fs7}|k,}oPJhu{?M7,3(~a[0LjY2oL93ޕ\ )y#!>X ]&$DrBrF{i c8#jPڊ8Hq(sf(~L(Mҟok` *8hECiT^ww=zB s14 ^jR uաqٳ{9{|bl{,W *;sٮz>j,Mv8[Ӯ+Z)~lK{=ܬJ  Ki=N'-ӼoE#SdxQ!S!R{bm4^v(ǩhp 2~}| gw>J2~Ox'.8zR[0>^l_ĺ` 5b<s Z`!JNPaG/xڕkAGmjE1+U/"[ M]c,AE=I"A#޴hΆ- ggޏ7"H0brFm>Ja85~CFy(q֨05°,D'Y`2^M\kkRv̧7~Az?4з#t.qLFމtZ+8|~3Ac|z=!E{7 ^#6KQr:zXK9q,g.1N< Z Z"}ӿ6]/ʷ1_Yw(4P[Z_abKGDH cmPPJCmp0712Om/IDATxI(@@̈qS yb1xsXW GfOaf?gG m2Y,iF KUf-- UYW׌-؏kցWrm/0/0 |38.h&7$2`j?y^1"-V`dl*1C~60.y \q }XD,E8ˤuTï ~Ua`C%Mzn0kWq a]|LulҮ]_ 8Xn)0߼.~]_ ߙ ]ISן/]m__`_?_-_|~y؇~5_z00> X??WGS =M3MC]^;!_-l~~p:QvO6=1ZԚ6G6; l|uSM[7nuFvZϚw~N#& {5KJts]MwF,r}YRVA|+7b^暅_<{ <_ax@|y"鏯fUϲ&P,Î~&K\xt#?Aq[zʊ7h6;}Nb +z[4Gۆ"J팉Wmrձ`ӂ+ҦXY.a Nv RѶL xx0N«F3ҢTur_,CV4RQiO -[Q``g0.urbʨaG8#l Y190]=tb1RO[lu#c [aG0gI g\Byz͠p&+wab$|Ńn~we$Ƃp& 2Fpk&/kp\!0_,"-$̈7aC!jfk?Ķ#8wx 믑{{xF#oc ؿڱ-ܴc lkYMWg{]kG)} g;u[9ror6ʡW7<+!IKQ{zC'߼vtclKzzֺ`tA"Ѕ\~_҉47QS "p[;vz'XWy:wd[joySM[7nuSM[7nuSM[Ν4OŒd ჷ)yޗ3ùar%`eӡ|PF p܅#Pn(>1A\eTg$GOwGvI(Aq !?R֫ ,Ծ,n %żj+][6Gu`pp ?lVMR^aЍS~O^nTKJװ;#`Rx- $@K3r(: oO}~)p"p0]ޢa:gj~^^HU 0}p\2`OF7J %^?&X~zu{j~kze5YhVt!\ z>\w4\z;Z :Cd5lwQj#c>Wyy+pvx#* `*0O%MA"X>w$tI ;W8w&`C2pGrN_< ÎJ" {jܝ6qx/|4||xK@倱. Ý:Ig;1nݲ1V|9<5jxvX ӑ[D}t>&Aua.?z'h#\:pHD&nҬׁ)K'+36ђV_@L#f8%{c.woxX[ u;*ak8&ڰm X o lk;[̜bX?J*ڰ$nC9pA{ ò.7X.ڤ.h8p3f@Xn ܺ)pn ܺ)p Y|vIȆspn ܺ)pF}dZín ܺ)p붘 o`$Y< HG|Z ,//)py P-TOdK>`07Qkp>b=_ӴP H4#>@)$ep?џ  dˣI@rIW05q{4\ Iצg i#N-LH̯g di[vQ>Σ U*a̙3<: | g!Q~G-,E$?q@R]pWˌ7.o \uS`G>,wݗ\tcb}Kkk| XӶ,;`z91`NUm1`.8ؤWL3;0o)z?0Hc#nl\l>,vCغC\ՖEPn̉1qᑏ]w/l2:^ZBME*5jouS- !JﵥsR9eKP7^GU$i*sv$D-Q/G"p)2RKu),2n`z;Y/wE)үե Rw=-n ܺ)pC#5\0J?ȴ[7nuS֍gΝ)M |)@:4{ ;^96gƻOC FO@wVpm'Gp@o l;'(<0ZWjعx45|Dol!vx 8 c 0'VT6Y[⋇4;W|*eu♙7mM~\Zـ&xruS`y^u-T9Pb dEW ԁ3Q@9 X(h׏CPijRS=)'򁲊09@$HC|a4ce(GꇴJ?\EPDz]2XE`RFKS?,?@fv?E5G?J@F}?<( Nn0 %5x k9&-;>E5ݥvAd>)肰<<-w5&eW.-AoQ'I69#HvKNZ uSGT,El o_>)*4lT0@ ޯbŁϩ`$÷>*7QȪԷ* ̑DSEXǁaWRPV3h8p3f@Xn ܺ)pn ܺ)pVWB\Se$;@U*G`* 0Pp}8:y%T@cWue U˩p1q \NHZ gw(:0[QBP(]TʼVkouSM[7n \;?dڇQA5ܺ)pn Oa~*_H1r=鼳ӓ*b%~JDCMCKٜjT,B2z`HIGDI>`~&v |N@-\# P1yxS-תH!C2ϻMd3.O/ٳ* /C,hhdCU#'UhK3$gO Ò f|V A4[HeKυ_~)pn+6| Wu`"n+>lt0(| WoZgc nmk ظv| Wm`|:nkه-v]vV`V`Vu2бuyduzEh-^%|V ha#Иcꨤ.,.#s"Hj 0 p8ـ &pψ@I%PDAs%OrPWK$^a*iF%< EcRRgRXQ45q 鼢5)b ۱^e&Dn5&΁.oU(@^Lx5<,ud.Q މe@L|VћB|@Ń,H4`xpU~Bo>p>uSK MԷ,2lAେKM\~U.87 5ly4=`<,!O$ `SÞARĥH#`PϨĽ|VN`ptq `T o_#:D}KtCtВF p.,$ C۸ [}_hE5J8R|Z[8eB8=lJ<C o$ `nkouSM[7n la }KLkuSM[7nݞ ϋN8 ϫ.8`Jp>}@pڨKH4T]52!1> 8YTY\#(!)0-jNdvlrNK+\22pc ?x['; `ASM[7n_F]Ip_ru[.u$nyd7- F j׿e\WL2]v KXXx `"Ytu:TFDǑPZ7D1 \d$7ҿc@R0 IjC2`VS`;$UX(Q u*I C0aOa;:]Qs2( ^@Bapu@VB#i^=l Ogdր}]@n¨^(?Ԡ_ܱ#um |?D A@l͢`ي )4;xtH@Jźye_-.2%L^Fl{ -pT/ մ0ݲ8 H́Dgb~;2"l2L F9R[Gaa ~p{qlB[!wqgWH]̗%?nek%ơkZ̲Tg9q" 39Npeie|@\d/7.c ܺ)p>d2ѫZ IENDB``!Cj1g ;IF@H hxڥoPiZ?MR !"CBT@jg~ !ErTMDtbU ]YbeAH #KGT 5Q㻼{p~eϔ.!RjEKJGK rvXKs2T$OLK7dgrJpo;JP lv W 6mʡ` "aFޯ,\ȵf2K,՚kYjی2f V{uJES7x[^-j\?{8+J;a5O`qso̎}=Ͻ d*s^ee7/^6(q[rs9L=()w:ufuˏEY?q7娐#B7Lh9:uӊ>s^?WE3醯Nqܝܐ / hDJ~UL&DZB,Őf@!tl# $^)a%O,(V9]. R\)sag3!CHHA:q5 x'<g3ih=8%'q~Y,R[3fA0H]ޒ^ [ˑFEO8 9;f>\NrN%vRvp,YPV* V .al^W0rmĠNTT>x0F|$l *ܒ 3zLk86ZQ(@B-DԶ;#) rG 62 -@Z:T r:+ =3pGr+QCNw!a=CHnMfm9TȝV+%+JW\(^QrzE%+JW\(^QrzE%+JW\(^^J?rbGӅG%)\`㕜ȻcD(;&g̻Gy\tpϖ.DN\y.…4UԔ\<X~ʘ-#i*JKR%]M.$x_QiZTrTyFF&-0+J<76#\!UKnNIU&XACQɕce_`4TS^TrYޕ0N%H*Q[Cj]ɉ,޷^BL}%Mt&w!yȳ]G(ysj%By~%\[<<_[.Q\(^QrzEJI3>XzSrrfAvXw@x{ߡW tm/ ;~_ee.־#afܩ\ַFr&.ny_J2a,!ߧ<#~ih=8 ‡A ny#վZaoHf=N:N3'x7UgAǩ~#k,Зk#Lvgr1YHdУgUGr8 7:909")DsW,"h&ntA ڕɥ69_@ĥuN.PܜxI8;< .ؕH1I"Grsd-r g"yK*B.HYE=1HuK#[9Z{ol1(6!(O: Ll3 MB.RJ#9i[}Vc묊\vl(:7[F{39*p*9&%R1h̷i<'VI%Ӑ92Nhp̽L{|ܶ%L'48%05ї|KTC<wb 3[_"ĽL}Fa_9 lo@nWn'9|t䖀 xINM\yx1^Id ZU'ͺsLRb c>x8H6rK0F :y.ٔ+Ӕ2t,|7PSrW5ΘKK$DM7MLYKܳ K6\>cF.6K繐o2SyNZ7vNɅFiG\3pN.,V쓚 ťU9 č!r Ʌ!rC{ Vf%0,96S=xO~Yd-%,-Dqdp-$$kgʎb&\JG{bӶcٜW$xYskz%כWN_/w(BkzXw7%wX,V}x[v/B;ѷWrvJn[_m+m};^r kH"׈SrJW%wOWr~U"h$ӯ"rlM`ZPyn;[$6rQ D{H*9:b WYL}0f)ng !,®VGǴ*%vrm2vF,7d)ș9C ״1"a-LQ $6Z0C@;)0Zk<7n%y>9HT|m9Qˤjr}ynܸ嚌ty湾z9+9j[`L zcL h!!j /1_<*@``̋}9mJTjxLQY%Ir=7|وK&S(2\_> :_?Lyx-X_ZvlpgmkvldkAcCLamwp{'3HPt1|_s_07^4+/I. ]mgr"Nruqr9MyIZiJ뗯|CQ;;M\? nN.9|g@Cy0qu59+rL = ֕h~yVJ1/0)>xhJn }*hY"ɑU[eji%Õxl3 ;P9^!r!A9V#@%a|Zy~Oz/SK?jeu(\n9Vhozya2;ٌyz&G\X8r6^9޿;@VV6In opZMrdݵɭ4*A'7ꔜjUkL = 踡UJ * |$9uG'7)E0KKXۖe-W(fUrJN]\cE@Vw(dEJJN\ַWrvJn[_m[c;tF+XzSrқh3ZkۺoַܶWrvJn[\Rr5䔜(9Ѫhu;9 D3Zu{ČVNFy?qCƗUx}у9>6{?V0VFSEG \8OD-jF+4ՌV|E3z[hF+G(.sY%tLNn'X0<2 ܄O3Z9y 幂hŒVD&3Za~q%uW稴*ъH5Pъ}PRwqi5 591sA/=dF+Jf*rt󹞓bךɖ% o6%U=d^rWlI0PgkhRֺ5a88z^Y??Jڶ`'VIN\b%nV(WhFxJ&5` <;"ӑF>9UA1|9#tՊ8BВQ+4dьn[s~2p<?uF+SL9LZT6歁{XaޯTw_ 3Za6Q^Ϻ}^ yhuVX!9ԪAfl޿4'=yg L, I6g䰅{<7#m.V2r݌xܳh39)%75خ-mrqٌVY亮H%=2i/Sq~ґECgK<،6GfU=_PQkP[%99Gt.l5p㴦`wL{ E;zh/& lgm% o9u6-%Ysĕ;)kF?w8Ӷ>tv!*uroQIC;!}>WG^ErF%-ʄOLMH?Ԯ{9)]M`oQ&G1۶I3Z.vg#HSk$pBhIO=Ԕ\_ޟВT;QriE3ervUbxЖURh|>J?/w5U-$ ivcvKe{y%xrzU{l*tF^r*9%.I1Y"w ;hOn}Sr8(9%ףoַܶWrv=VgrJe"ޔav܁.uN&gB\ַ JYRlޮ+}T[DHҒeҨ䎑G(y.QrQ>JyӰݚUb[VIa^S~B.<)<"ߺ,a۰Q>/KWñVS{ 9^`C\ϲy-h3*x=F 9c+9YJ <<%wב/Ka0(Nid 9v9;+v:2%~W6rCÄ g;.O'740+r$p0j3#LtrM9˶YCLꎳkd#V;~%䔜(z&'wb/Sm]~r+m};|%oܶސoTH,)bM}Ymy멼8\)};|%oȡ@O䠞ɦi MՐCwf?1bY~h5~G:pkWF/8r+U3 $39aDMELxZ 9&qT_co \z 9E l<mr2s 1ut:U`*єNM 1}&G~yk7r\4-¸R@v>9^"k|ͬ<˷GZ_ozAN}.X'[M\)V!=BKVAUP[M;=.N9r~r&;'a伯"oȝQњU+[YCJ cd ;7}ӗ:jCJDlj MKw-2}a*,zŻ/;ݷ~r/Skz.$7.Mכ+N#؇wvzawaone;ػ7@N@v&KmkJrVɭ[Zki7iV[zfr w- |_ ^ӷWrvJn[_m+m};ɽ!?NNқ;,*rFe4ȩ%+JW\(^QrzE%+JW\(^QrzE%+C␪ N5őO&C OGVYMx/g(NdJ>\\iP)Ɂ4p0,.c4\9֖!]\CL2V@oD9NuA IkL`kw E+RE ǎnRCXLn8PQSp8X5pqrEA@#abqyhK4?[Cmvtx.n)}C^ uH˓зClQrzE%+JW\(^QrzE%+JW\(^QrzE%+#GOfLzEJMN(^QrzEJz&g#>JC\(^Qr"WWVYf{ǯO\xe72#TQZ#ҪynKWrD^iǯQW\(^QrzE%+JW\(^QrzeBnmRZjr*D%+JW\p(^QrRS9&JW\(^QrzE=.C/5U C*JX* % Ji}[ 0hTEؠ~9H{YIW)b,Yqv^0ɩ)Ǚf&;] ۙq*AY{d9%6['h)RteFbWkD3r!G rAǫz%E`qR渒aC?B4q=R$oߠw^]A.q)@,c0qp%2Z(nV*;+M!D+T7"Uu>@T}L^@Ajw"G/I"C0ϤZ͘ +(~tr 9x@#+J3reǤ\-XrD7Wݓx}=(19Z5 Xrw T Y HÖ;Tdm(߉/"Gfr m&xWv+@OXB J6 ENNZsX eK'˶vP \Cv# ~rhwX˫h 4wX%X2Ʃ&|&G5X/imf&Ѫ-qDdb-[ծPLD)6 j^c+2W1V'a8?`Cxknq+nzv9,=|0>J 1rI>2h0mǩ$;@zqetgI'MG{cOEW|HZщ:5eErŎ ~t9vK9Ńr#+ȥgR/,ǝ!١Ƨd"@n܏\d$T8S&ǜgTKr/YE.UyNvfr9@^~$(~,"Gm+ E,T%b۞ܤh_&1iLpJ P\fEr,ŒYnvN.9;@L<3qD?t(b:XB"!T櫉7懝 }i֛*ˍeX/"n=(/+JW\(^QrzE땋s{\%+JW\(^QrzEɁ "Eb?q99\UGra %X#D;ȗ—V/?͊~r'^ Wl=q!:rYFڨLy6&'ZS:3p:V4ePڸ8RLnI G6EӴMXy[9g)b g`̈OI%4݉: KD@^J)zIi i 3ty*zpȓ |V@Nf d?/Nj&"I!r9S[/N}grG8V<˱\ryN:Cms9?dOše1˵XIS~da.H}JYf9b' 2hV=q)9uiV2P2 >WiŵBVIQynri+s8++ndYxPp!*X-zgr8J _1 (i^vUnޑ yذ|W,80?dۭ%޷v8 ㎊g^IriK+]E\(^QrzE%+:aT%+JW\(^QrzE%+pj PIENDB`gr( !  fN/Equation Equation.30,Microsoft Equation 3.0001Equation Equation.30,Microsoft Equation 3.002$2Equation Equation.30,Microsoft Equation 3.0/ 0DTimes New Romanx$bbv 0b( 0DSymbolew Romanx$bbv 0b( 0  DCourier Newmanx$bbv 0b( 0 10DWingdingswmanx$bbv 0b( 0 @DLucida Calligraphybbv 0b( 0 BPDAuntJudylligraphybbv 0b( 0 ` .  @n?" dd@  @@`` xp;  >  7( " $ #&  :?`b$GxD]v*ח&b$h07@8I86 &2$|Xd}ָ8lop12$JNPaK4$b$)Ey 572$j1g ;IFKTb$PBAa '$JW 0e0e A@A5%8c8c     ?1d0u0@Ty2 NP'p<'p@A)BCD|E?@3 g4JdJdv 0bpppp@ <4!d!d` 0bd%b <4BdBd4b 0bb<ʚ;]Z7ʚ;<4dddd4b{ 0 bb2*___PPT9 z? -O =e3Information Extraction with Tree Automata InductionRaymond Kosala1, Jan Van den Bussche2, Maurice Bruynooghe1, Hendrik Blockeel1 1 Katholieke Universiteit Leuven, Belgium 2 University of Limburg, Belgium     )  P=  OutlineIntroduction: information extraction (IE) grammatical inference Approach k-testable and g-testable algorithms Preliminary result Further workN3N3 DB FIE from unstructured documents  "Extract certain fields of interest from a text Learner is trained with (positive) examples Each learner focuses on a single field marked with  x $Grammatical inference : finite alphabet Regular language L * Given: set of examples (pos. or neg.) Task: infer a DFA compatible with examples Quality criterion: Exact learning in the limit PAC etc. Large body of workZ%ZZ#d%IE with grammatical inference*Mark field  x as special token Infer DFA for the language L = {S over ( x)* | the field to be extracted is marked by x} Only positive examplesA> ., IE from structured documentsPrevious works learn string language XML or HTML data: tree structured Natural extension is to learn a tree language  x has a b-brother Extraction of a field can depend on structural context ju.:u :u ^%Learning process   )Testing process   # Page examples  Why do we need the context?wNot enough to differentiate the fields of interest that depend on the structural context. Can be chosen automatically.xx  Tree automatalRanked alphabet : finite set of function symbols with arities. E.g. = {a(2), b(2), c(0)} Tree ground term over a(c,a(b(c,c),b(c,c))) : tree with depth 3 Tree automaton: M = (, Q, d, F). d is a set of transitions of the form: v(q1, & , qn) q Where v , n is the arity of v , qi and q Q A1]1;* &P7  ExampleBGiven an automaton M with the following transitions: d1 : c q0 d2 : a(q0 , q0) q0 d3 : b(q0 , q0) accept M accepts a tree t t has b-node as the root 5>.!,BH Unranked treesXML/HTML: bib & paper report book paper & The number of children is not fixed by the label Two approaches: 1. Generalize notion of tree automaton to unranked trees. L transition rules: v(e) q , where e is regular expression over Q 2. Encode as ranked tree. AN0 +26A: B*&  Encoding of unranked trees@There are well-known methods of encoding to binary trees, we use: encode(T) = encodef(T) v if F1 = F2 = e vright(encodef(F2)) if F1 = e, F2 e encodef(v(F1), F2) = vleft(encodef(F1)) if F1 e, F2 = e v(encodef(F1), encodef(F2)) otherwise Where: T := v(F), v F := e F := T, F Example: becomes | (B(QIg -B              -  ,L  2%   k-testable tree languagesLanguages in which membership can be checked by just looking at subtrees of length k-1 that appear in the tree. k-roots: k-forks: k-subtrees:  S   x 6   -   !k-testable tree languages (cont.)An example t: html head body title h1 table f2(t): html head body head body title h1 table ZZ  %+     3 2 S )k-testable algorithm [Rico-Juan, et al.]*)TGiven: a set of positive examples T, a positive integer k Q, FS, D = For each t T, Let R = rk-1(t); F = fk(t); S = sk-1(t); Q = Q R rk-1( F ) S ; FS = FS R; " v(t1, & ,tm) S : D = D {dm(v(t1, & ,tm)) = v(t1, & ,tm)} " v(t1, & ,tm) F : D = D {dm(v(t1, & ,tm)) = rk-1(v(t1, & ,tm))}TX"" """"":     r5g-testable algorithm\Idea: generalize state transitions from forks that are not important for the extraction. Important forks are those that contain  x and (possibly) the distinguishing context.67 =-4g-testable algorithmGiven: a set of positive examples T, positive integer k and l FS, D, Ss, CF, OF, OF = For each t T, Let R = rk-1(t); F = fk(t); S = sk-1(t); Ss = Ss S FS = FS R " v(t1, & ,tm) S : D = D {dm(v(t1, & ,tm)) = v(t1, & ,tm)} CF = CF {f | f F, f contains x} OF = OF {f | f F, f does not contain x} For each of OF, of = gen(of,l) If of covers one of CF then OF = OF of else OF = OF of Let F = OF CF Q = Q F rk-1( F ) Ss " v(t1, & ,tm) F : D = D {dm(v(t1, & ,tm)) = rk-1(v(t1, & ,tm))} iZZZ.Z&ZrZ" """""  """""               D@ /   $        =+g-testable algorithm exampleA set of examples T (with k = 2 and l = 1): html html head body head body title h1 x table title h1 x F = f2(T): html head body body head body title h1 x h1 x table FS = {html} OF = {html(head, body), head(title)} CF = {body(h1, x), body(h1, x, table)} OF = {html(*, *), head(*)}.,ZZZsZuZ   s $%MYx  $ s,$g-testable algorithm example (cont.)F = {html(*, *), head(*), body(h1, x), body(h1, x, table)} Transitions from the trees in the subtrees s1(T): d (table) = table d (title) = title d (h1) = h1 d (x) = x Transitions from the trees in the generalized forks F : d (html(*, *)) = html d (head(*)) = head d (body(h1, x)) = body d (body(h1, x, table)) = body Q = {html ; head ; body ; table ; title ; h1; x}nZ;Z8Z^Z2Z9,  41^ * ExperimentTwo benchmark datasets: Internet Address Finder (IAF) and Quote Server (QS). Comparison with: HMM, Stalker, and BWI. The highlights of our method: More expressive. Doesn t require: manual specifications of windows length of the prefix and suffix of the target field (HMM and BWI) special tokens of the delimiters such as  :  > (Stalker and BWI) embedded catalog tree (Stalker) The limitations: The field that can be extracted limited to whole node Slower when extractingvZ"ZZZMZ"  Mv     " Experiment resultsThe results in % are: Dataset: IAF-altname IAF-org QS-date QS-vol Shakespeare Prec Rec F1 Prec Rec F1 Prec Rec F1 Prec Rec F1 Prec Rec F1 -------------------------------------------------------------------------------------------------------------------------- HMM 1.7 90 3.4 16.8 89.7 28.4 36.3 100 53.3 18.4 96.2 30.9 Stalker 100 - - 48.0 - - 0 - - 0 - - BWI 90.9 43.5 58.8 77.5 45.9 57.7 100 100 100 100 61.9 76.5 k-testable 100 73.9 85 100 57.9 73.3 100 60.5 75.4 100 73.6 84.8 56.2 90 69.2 g-testable 100 73.9 85 100 82.6 90.5 100 60.5 75.4 100 73.6 84.8 69.2 90 78.2 Parameters 4 and (5,2) 2 and (3,2) 2 and (3,2) 5 and (6,5) 3 and (4,2) * F1 is the harmonic mean of recall and precision * The results of HMM, Stalker and BWI are adopted from [Freitag & Kushmerick] bZrry  !     o    Further workMore generalization while using bigger context is achieved, but sometimes the binarisation makes the context far from the field of interest the generalization cannot go very far Work on the algorithm that can work directly with unranked trees.$CN   ` ` ̙33` 333MMM` ff3333f` f` f` 3>?" dd@,|?" dd@   " @ ` n?" dd@   @@``PR    @ ` ` p>> jb(    6` "P  T Click to edit Master title style! !$  0dc "  RClick to edit Master text styles Second level Third level Fourth level Fifth level!     S  0@h "``  X*  0m "`   Z*  0r "`   Z*H  0޽h ? ̙33  blank 0 zr@ (    0 2:   P*    0 h :  R*  d  c $ ?)q    0  7c  RClick to edit Master text styles Second level Third level Fourth level Fifth level!     S  6l W2   P*    6h Wh   R*  H  0 go ? ̙33@ Pp(  p p 0 2:   X*  p 0$O h :  Z*  p 6O W2  O X*  p 6 O Wh  O Z* H p 0 go ? ̙33   % 0((  ~  s *EP   x  c $Fp  n  c FA"L:\img_sdzgr.gif t  c LA(L:\img_logo_LUC.gif` P[H  0޽h ? ̙33     0(   x  c $7W 7   x  c $8  H  0޽h ? ̙33O   ` (  ^  6> x  c $OP  O x  c $O  O  o <O? 0  = Requirement 2 k < O?   = Requirement 1 b <?@p  9 Job Title  W <0 O?`  < Company Name RB  s *D00RB  s *D  RB  s *DRB  s *D  RB  s *D  RB  s *D  RB  s *D  RB  s *D  RB  s *D  H  0޽h ? ̙33   p0(  x  c $h&OW 7  O x  c $P(O O H  0޽h ? ̙33   0(  x  c $.OP  O x  c $.O O H  0޽h ? ̙33&   g&_&= %(   x  c $AOP  O x  c $2O O   00Op " F P    @ T P` p   #  ` p B  B ZD??P`  B   ZD??` p T P` p   #  p B  B ZD??P`  B   ZD??` p T P` p   # @ 0p B  B ZD??P`  B   ZD??` p    T|EO ?? f  1a    `Ǹ ??&`  1a    `ʸ ??@   1b    `QO ??0@   1b    `SO ?? @   1x    `VO ??@   1a    `ZO ??@   1cT P` p   #   B  B ZD??P`  B   ZD??` p B   ZD??  B   ZD??@ @     `˸ ?? FP  1c    ``O ?? 6P  1c    `8cO ??  P  1c !   `fO ?? P  1cF v  b  @ T P` p  c # ``B d B ZD??P`  B e  ZD??` p T P` p  f # B g B ZD??P`  B h  ZD??` p T P` p  i # pB j B ZD??P`  B k  ZD??` p  l  TLlO ??0v 1x m   `oO ??R 2 a n   `0sO ??0, 1b o   `lvO ??@@ 1b p   `zO ??0@ 1a q   `}O ?? @ 1a r   `xO ??@@ 1aT P` p  s # 0@ B t B ZD??P`  B u  ZD??` p B v  ZD??p@p B w  ZD??@  x   `@O ??  1c y   `O ??v  1c z   `p|O ??6  1c {   `dO ??   1cT P` p  | # @@ B } B ZD??P`  B ~  ZD??` p     `O ??  1c    `TO ??v  1cH  0޽h ? ̙33[    !(  x  c $XOW 7  O r  S O O ] 8 p@  !@p@ @ @@  @@`  0@@  <O2 Z*< > & .. & & & . & & . < >  <Opp   DStructured documentH@ P@  ` xB B HDP@xB  HD@xB  B HDPPxB  B HDxB   HD@xB   HDl@    0  ~B B ND@  ~B  ND@  ~B B ND @ ~B B NDp  p~B  NDp p ` ~B  ND ` ZB  s *D0ZB  s *DP    <ĭO@ \  EParsed and annotated  <xO p D  ; Transformed ZB  s *Dp0e@ 0 P`  00  00 ETree automata learnerH  0޽h ? ̙33    #(  x  c $OW 7  O   <O  O 8 p@  #@ T @@ # @@`  0@@  <O# Z*< > & .. & & & . & & . < >   BOpp   DStructured document\T P@  # ` xB  B HDP@xB   HD@xB  B HDPPxB B HDxB  HD@xB  HD\T    # 0  xB B HD@  xB  HD@  xB B HD @ xB B HDp  pxB  HDp p ` xB  HD ` `B  0D0`B  0DP    6,p   6Parsed  BO p j  ; Transformed `B  0DT 0 # ` `  00  0lO0k f6With each text nodes replaced, run the tree automaton77ZB   s *D " <,O2 6OutputH  0޽h ? ̙33      |$ (  |x | c $O    O \ | C 4AL:\html1a.gif P\ | C 4AL:\html2a.gif D2 | 00@,$D02 | 00,$D02 | 0` 0 ,$D0 | 0p  ,$D0 | 0 ,$D0 | 0  ,$D0 | 0 ,$D0 | 0P,$D02 | 0`P ,$D 02 | 0`P@,$D 02 | 0`P` ,$D 02 | 0 `P ,$D 0 | 0 ,$D 0 | 0@,$D0 | 0  ,$D0H | 0޽h ? ̙33z   *"@(  @r @ S OkK  O r @ S pO@ 00 O  @ 0 O @  . |-- tr |-- td |-- td |-- lastupdate (CDATA) |-- td |-- b |-- 12/4/98 |-- td |-- tr |-- td |-- tr |-- td |-- td |-- organization (CDATA) |-- td |-- b |-- ABC |-- td |-- tr |-- td, P*35H @ 0޽h ? ̙33   ^V$(  $x $ c $OW 7  O r $ S O O  F 0v  $ 6` T P` p  $# `B $B ZD??P`  B $ ZD??` p T P` p  $# pB  $B ZD??P`  B  $ ZD??` p   $ TO ??00 1c  $  `f ?? 2 a  $  `f ??0& 1a $  `df ?? @ 1b $  `f ??@@ 1bT P` p  $# @ B $B ZD??P`  B $ ZD??` p  $  ` f ??6  1c $  `f ??&  1cT P` p  $# @@ B $B ZD??P`  B $ ZD??` p  $  `f ??  1c $  `Df ??v  1cH $ 0޽h ? ̙33   801=((  (x ( c $4fW 7  f x ( c $\5f f  F   (  pT P` p  (# `B (B ZD??P`  B ( ZD??` p T P` p  (# @B  (B ZD??P`  B  ( ZD??` p   ( T<_f ??00 1a  (  `|bf ?? 2 b  (  ` ff ??0& 1a (  `\if ?? @ 1a ( Tmf ??@ 1cT P` p  (# @ B (B ZD??P`  B ( ZD??` p  (  `pf ??6  1c (  `sf ??&  1cT P` p  (# B (B ZD??P`  B ( ZD??` p  (  `wf ??F@ 1c (  ``df ??6@ 1chF p& 0 &(   T P` p  '(#  PB ((B ZD??P`  B )( ZD??` p T P` p  *(#   B +(B ZD??P`  B ,( ZD??` p  -( TLWf ??p   1c .(  `Zf ??   2 b /(  `>f ?? f  1a 0(  `}f ??` p 0 1c 1(  `D:f ?? p& 0 1chF   2(  V T P` p  3(# B 4(B ZD??P`  B 5( ZD??` p T P` p  6(# 0B 7(B ZD??P`  B 8( ZD??` p  9( Tt:  4  `rg ??  p  u gen(t,1):6   5  fxg ?? p  u gen(t,2):6 H  0޽h ? ̙33   `$(  r  S |g  g r  S 8g g H  0޽h ? ̙33   pP(  x  c $h  h r  S Th  h   6h0 P `  :> R = r1(T): html Ss = s1(T): table , title , h1, x < 2 2  rF   p0B  ZD??B B ZD??rF   p 0B   ZD??B  B ZD??rF     B   ZD??B  B ZD??rF   B  ZD??B B ZD??B  ZD??rF    p B  ZD??B B ZD??rF    p B  ZD??B B ZD??rF    @p B  ZD??B B ZD??B  ZD??ppB  ZD?? ppp B  ZD??H  0޽h ? ̙33   0(  x  c $th!   h x  c $0h h H  0޽h ? ̙33   *(  x  c $l!h   h r  S ("h`0 h H  0޽h ? ̙33   d@(  dx d c $Ўa8װ~ S =O#ej-Uj? >d?r+rr':| ҧr` `9 Jjl E"S ox z3aܫ=GgSt2 _ēvHH/ 0Ӣm|ejP xe0JH)ӽ\)㸄l$߻2~Cs3ɓ { 剡9~82ȍa:uáD<I"dbhfO"V&&s.lI SL 4F!ƒ~Z!d۶e[c`Ki1헲彼MW_5a:N))vf\5i*ޠ<*+ 3Đ%!7̝<2w;izfo4cSٳMx_`SW~+b3-2bk|"j'2v"W{b"l<<@2QO ZiV˵'<#5'PZONx -sv♠,HҶvqx"(;Kg" BGH+/gĀ47p5p6v ֿCUdV>‡F ی˛O(iO"Jդ.ŧnϩ*!}e8\![+C|Y8rZV,ߊȜ|$ZknߦJxWO\E?3@ T`ktS/&TS hwe-5W֔F41F}2}0M_Lj|2MJ4/BISSMsf^,khLɝߜs|ܹW,Ժ EtBi;@*raUȔo^6_:i|,BL;iVJwlҍGe)/)WdKpyO4c &*o}:ɧrns8d|ʧfpjP *UArNzTLj!6Vm{"?m>vG!G~L˷SL $ U{{j?y"ΦVoLOERz~a㩯fʳOv\}' U +YL rMJkRZ ي