From fbf5e108977365302d26dfa20478c01fe398d8ea Mon Sep 17 00:00:00 2001 From: James Ravenscroft Date: Sun, 3 Jan 2021 14:53:09 +0000 Subject: [PATCH] Update blog post --- .../assets/simple_classifier.xml | 2 +- .../simple_classifier_with_generator.xml | 1 + .../simple_classifier_with_generator.png | Bin 0 -> 15636 bytes .../images/yu_2019_figure_1.png | Bin 0 -> 25799 bytes .../2021-01-02-nlp-model-rationale/index.md | 49 +++++++++++++++--- 5 files changed, 43 insertions(+), 9 deletions(-) create mode 100644 brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier_with_generator.xml create mode 100644 brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/simple_classifier_with_generator.png create mode 100644 brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/yu_2019_figure_1.png diff --git a/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier.xml b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier.xml index 9253ca9..940d0cb 100644 --- a/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier.xml +++ b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier.xml @@ -1 +1 @@ -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 \ No newline at end of file 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 \ No newline at end of file diff --git a/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier_with_generator.xml b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier_with_generator.xml new file mode 100644 index 0000000..db543b0 --- /dev/null +++ b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/assets/simple_classifier_with_generator.xml @@ -0,0 +1 @@ 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 \ No newline at end of file diff --git a/brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/simple_classifier_with_generator.png b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/simple_classifier_with_generator.png new file mode 100644 index 0000000000000000000000000000000000000000..0cd512ab5f4de691d7e571fd076470ffdb2716cb GIT binary patch literal 15636 zcmch8cRZVI+kYvws`^%v!BcwW*plic(4`vD#8+JCwXve?*2N8!zh>8P0lBN<8yppOCvcQM5ELcKX!t(F_cFt(8e+H5Uivt4) zAc;Mpz>}VOb>&@h)P(hJK1^J10xap1fa3L zt^iT~BGG?MP5g%+4uk%?)e#-w>Hw?*!Yc^_NbBfo=Zv-U1iHBXLp=>?Gqij_pkpxF zH&8bKVk(7mwf%c^FFRlZe+_K}5Kt6YDRB^Ka~Df{UI5N0GL;f}O$g6nG;s#?f7 zU>qTSCeCs`dPoU7jD@EuLRBJ2!VLzO_d^GGnYy`S3{*|Aju=FclaZyCpKq{aP>`B4 z+Q7rXUB(ZmC+Di}>}}?)?xUmX>Y%HtE^eR-L(2I$NI~F%_Hdl4nvsROgTJn~DnMNg zZ(nyG7)IR1SlrA!$N~p<*46aJVZ7agp)dmmcG4- zpSL#_Sa@I{FhHQbEZi_SP)!al9}p}jts@nL)3igFW2G$uTumW*7LL9qE~;{7KzB6@ zn6EBQ+*I4rTU^ybPu5Y-4GSCGC1F92fF6S+0 z>F4g@?CN0dCS|JYtnTR~hw#T3NqGlYq9Hzhh9Gic19Ma4p^+CEY=X|=Ire&Zw&0gUfRdlSr_JE5Ts)Pg-T-NG_?%PrTnA= zy*|Z!%V1TFd&pay(t{%=o z8ul`1X$v)9Clh^L;u`gE&Mx+W5~ea*a7_zu-#|5ly)Vv5)6vg704)bm^MYZ}c2FM= zc}*XAX-Al}v4@7UzCF(05~B|>k#jb+cXsnMkoU6G(s7pXmiAZ2dPtkYF#6_xmRb(x zQnJ9rp03{7>OL4(jQ}&aos5gFm%Rbd4E4|gJ{)`{v5+8Dc`toQxV^k~kgThWzm~tf zsfQ;#m`G=PDain+B|^s445?|aVPS~~AqUofn1^4zbwv@6*dx&FPFm@iM`W|jZl1Motw7H#^tB$U# zkCd!dps|x=pk|Psn}h){n31H6ri7aZ+TTG>%|sdpF|yY(bv980#1I;w=4}~dA?Km( z=Z8fj>A@w8oU|a){y0xROJ_h4YIp#@ zX0EDsK~6|BsF|LVXOIgTV&LW!U?vABDh*krjI6q2kcp|jp@WPU5+HzhAkNdy31fhD zm6y@7G}dtUk#*5<@imb#Gm_EJaX0q$@d0}2Y6Kcem`SPzIt57ROL+U)TWX5~=LMA{ zia&8qgrzKoC@6Zq7!61;(%;Zh-opp2>EsPHwgZ|yHI31F_VOk;X$PdFdw{yHRv=NE z5`T!S`d^y(FSShU{+Ia!6fA{hrSpXg92d0JRZW9z*79klGEMeBASz@0xJ$q4ugpKu zEqz9x1gU|6xp|}?mvS1jHr*H3XI(JXVS-80W_~6xe9KVBOI^!+0%pi_CNNk+m8ihNs<`JW3-P>Jmb zJ9fM8WlaGLOs95t@UcXyW(hQwI`vlcDhWW7Pa~U?vu7jHQxoBjC@b3f>6^JiUStr= zn4Wz~NrabZz@FAlXad`$WR09XKRYlM$!+wK!Hcoc?L)Q2EA>9uU5pznL1mZz^GAgv-)rwFQJy2JyLYk`GEju5C&z)NHu%D5=+8)D zTp5;o>-SEpDHcxo=JHxJmk|f$PbX^*#<^xT;#6JcyziK|&nC=pi&x*+<0doPjPXwi z#mI0qg)1-)Q!>9-SD0sdP1oIC1(2r*QPrdzjMnrEM;tN_Ox6%G`)PSW-P>$meFR9K zauwzJS8QrL!yiAz@NgVmkf_q&mxQUR=%U8pU!a%$LVjhe%Z0mZpdYn_QW+Z4`fvhEYk zWYYmBEXtNH7?D9?Dzd9ab4hxO#Eo63O#4BNi&%T!w*A1iu5C@nrgF*%3m19AX}fTE zv|3x{JXBEc`i;+INSi;OM|RrJq*LL+)mz{p<@2^bHNktoAh>PEopZ|sHej)dA>oqj z`BfhkzMYYuqP}OwG`f%WlY>wx3|ak|!n#SRL+ad4v zrZd|TTBazyF7=O=K6UAWzP`5__GdABG0Vy;es8lNWIa{Jb0TAPD~(sT)b>rdvAMaE z(yiM;Fa6dg8t-YAS<_GhvqwWA6MM}7sh+a<^Kvp`>66#sn(VxRQ$7z&Usw(^|L&Ls zj*qRgxZ>^K)46YFqGDn_B1aYdej4Q{2G*=Bl2F{>NWQ%thj|VguXA{?yFP7YUX8Rn zQ##q~TAT^l5Z$DAJ&1w=6f8=47;uM3!3vnQb{s!Q)UqMPveB*k2cd{EaSzxodh)er z-H@U3sUNExrbu`!tO+sKsi-pn<@N70a{g-KpFH@3K+K28OhGPM=$)9B{56kO`zXpcsqN2F?z1M~s3fxz z5Vm2%C)RJm!*IJ_G(YxbWqnlzi1{_gofo9cO$VGC8+HrWmG1f3sYCQFNvDSm9^;TS z8rwxPo^sJm1At^VjBS~P@SEheyS6q>Uim0&AMf^b;Ov^z<5X@s55pHx2H9?Fw9*`M+?M0`GWLo8;B9VRTH5!*t)u-w(;fYKIm zb-I_?p$d7$>ZkJITu72 zpa&WaFC-HDiKIO&uHhjtB)m4rS6I4RY11BDbN65xP$08(EHLsKg4y>)4P<)lU+Xj8 zkev1#~AcqkoUpeFfBA+(Y$NRwiKY6tRvCtHjkpMS<1I5TaWBks@&n z$R;mqL<+^%n*J)4yAMNo{~4F@pK-;gGMqj>K?hG(z_}Sk-TrJZE_s6w?;QS!CjIi+ zSb@la3cmSnr9__psreVr_inWBPY7RHer{If`0QMEG_98mK0txVxSSJ&wa08}*aDvq1T7DKMI-VZwU2;O% zeMr9LJ!My73CNQFQq)@0-?H>u8HL-w+`v>V>uiiy7f(*}8E{p8AX8>hL~9&(VCDAv)eqWNr^`hTl-Kt%xLzIIUI zIr*Bk`KmmDE*+UHpKJ-oF*mZ(l)w3#vE^QI@*$0g04EDDJVMnn4K@H!iK7DqMAM}n z`ujOxr`~gxRa<6~A$pJ90d1gslCbEB*w$Ipm}f2}8g~n#apxc!cgo&4!~)T{=l-X0 zH|tpJQ9l1;h(Fz*{MaxTw!Qn!aD(Y%`aRk^!Mkz#YlUaK{H}m0uqg)sj?LUi4^w0( zfMQlV3rj+GKMyf1fIfe1Pp6shu2`XQRm?FqT(XKNF=~!^Iw>t3PR`^e0T?kw6 zFv8YTA@s8)|7 z_u*!`#Ve;iz%cN#$#~FVFMY3ot38o+=i|@AL{Wuf04d?ipZ=IDs|#2_KP!T$#-ClVxRNKs@t*ZFwDr$~%9$;X2#Vej{MaZu{(bBX{`$EgPTg zOppVI@Zhi$yVkOA#rs%b7qWYqa%849nG}Rh8nK7GI)jdNm#<4{t+(xb8Y_XPo0MsR zrxZ>#$HUFJXO9+=m2J*WHzIN)YEk&F>Z^55UJT|l z0h6AK6Gz(vVoB)vS3^pYg^Z0!`q5EWy#V2Z-Qm^fxv#t3EA|-T|N8e2hrO(5+x4VN0RE+KFh3)7%Enl#-lQ|aW#*wwYi*{HW+s{I)2+Hg1$RDj!!t|*JOI!uoXGD5 zZ`OoAl3%*UDq*j#cV?v<^d|hY-}?8bt|5t$n6$*u;uoHRr1UBrLb-c@RdX=-mCpPo zmlkKE1;L{SSiZgYYF%>*wd!-yo0p(2jrQ0F;8Ciy`^e)Vmjgh{WS1l>DU3A6GxGOf z-+zB$g)pzR@$8KkRX#f`@$m^>YaAWa%vTJQ*)yIATA-ERA9ILx7RXb#)81kcldnY7 z7E~(3?FT)w3T;p7U90E{{L6TlI>$<;9#r&J`CX&f3)!&*V{#heG2a@84v5=Asm^6n z8O%7N`e;R_Zy3!L4A?}uf4>>>+fL+Kv&LHe0C z_*gZtCWk)&!3CWQ!(QL*F;s_~c*&qLUen7yww@{pKV7xu&vVJZ#Zh*iW|$bz8$5h; zd%Jh+)l3NUMT#p-(2G6USCtM`9?8v9a$cug1k_)K?>tCL@OKzNN95hN_JO=gNq! z(HmK10#^c7$OD_|HG}G>g?^o9(PdTetH8W;6!YdjZx3|c7G(l_CN>{WLwW%WS#iCE z!+IfrnB^Ov4sG#tZ8(L!C82mqTGv)WbZbyKEZ8lE1*t;E8@*IhxNfp3r46$umW zwfb6G)=OV&D^BG)FhptyJqpy zPSWCy)i3}p>NlvK>0|G#cC13_1n`*1me-Y(OyZ<(4!gzD3aUYtP4Z=`FDL9{S70D( z9>7Hq4y0Gi?RW!V>BV)uyTNC}wc(li%yrIqrnB7Hlb(|#^u_tK=X8a69?xYmIEJ24 z!`)YdH{Ve>@cBSz($u;PsdUD-zC9?Kjs?bo>9ve9vDZ+=zeT6{Ibi~^D}ONM`*Bo2Vgy!)44kS_ueSjPrQh+FT9`&3tYXt`5ePug#q~R}FU5Lc5x<*e!|VEV{Zy|vmApCI ztF3eHZPWj$oG^-wEIhXkoJK40PcXsD)0i0+_I?cygz9H`6)dZ#M{TQy(?G}hgaBlf zvhYX|ywA)@-ItTOxBhYAa^uQB;trXIKP zKNw^znglCFK+O`3E?=OO!n*IQBQedPmwcY?{`{f_c&lkIZ`llx^Bn_#?2o0js-R+r z0{xGH*c*GiZ|$;BX7Ch*p3dLGFscZGXPVfhC;IIyJwXLYVtpU0b0h;WK_UC14mi&+ zGC+l1K!m7Z9^|Vhn|OKWZ2~$4`U!w9rRj_9<*5Vti918evxL`07|vtB?~~4=P+Cy5 zO0xzZ%-}h$Gs7!SGKwNZO90U&7y*zQnF~ z9DCtUwg(3AY^qqJ-`@DeEF_PB82(-c3iFQq*%d6}u7AAP)XFB`&sR0n`|=xi9$vh9 z`!nIyThD&0wlBq36n=6nKo`poyN3)X!886(d77^|*ELpdKRB+{w4O zPL@RhpF$E|a?;%SjLltU765*iM8nMK1xNsBdWK6Y*UxprYWyhorm^@74KFO+^S{75 zb9{$B|7Fk^Pr!TjWt`Rg4k`te{L@IL6g)yN6YOpA+9Rj3%V}dN>JTvPmg}tbhRN)T zBcU2j-jvCOfB;<$KJw}SY&`R2eBARqAs9B&BjRxGf&S+!>zC#q4YVf$#J^*!+}f}v zfY+i)RVt$V+wm^+9{fc3_#-W_7zt5z&j74Rqu-xJKSBo%~`Z9Bv({I9EMExTy>$T_WTQFEQnbl8!i=9(;7_WV4?T%KKw-s0{;<$eI#qj9& z@gNHaSF^Ir+moB)d6T_MFkaAjQQ43XDepZXEIX&Sog_QsvhHAniRBX0pd$|yo?;aU zS;v>xCH&x8Y!bk4CKh!RUIkz`c_Q{Z!PK;`TP)x_)^Ge(V(7s_AY`RVyB<5Klk?5x z(T(=WFY@|7mu|@p(B#^tjF?&J?g)KEZYpSnz$W zJeh5rmIz?VfseAYA`fl?d9S_sjbi{*=O#V!h*a0D8%Ik%FH*VL?|0*UA^)e%BGu^7 zgB!|j44&PZ@!E*3iAHz*ucAx6lN){;Wk+X!mJ73MtXrRG{xku6u>ot9J(H$`X;>l) z6-^4^&J4{<$2&8Dvtjz3;xD_K@{IDvS-6B=UWtBuYoce^5m0jVL0k5M!orMc)v5Y_ zewO$$UK)SIZ#%T=#D3+gdh9&uYyCJ`hn~ywYYpE*rF#?z1s`&~&l(~xv&rZgU5OIV z0f$532k*Wg<5G_Af>+HK&A=7B5&Ul0{zvaK+xlhbVj5AXSEJGpk3zd=$LAB`=!UJ5l~WH7yBiffx$Sjth9nc{4U z2yY$qyp3=(uNFkqzOiHpZ!-F2`kb5lfL-v-^Q`tWW&u10YW}f=gX8-+`i?$w(2dih z%gaaU1AaFi)2%qu&Aq|$f&@OtaC74lnr3qz-MQ(UktsQJ+vJ&HVF3s$5qgNJG_6bn zLX2Ms^hH5^$46)o30)S>uY-Mlt+#v9YBp;DaZ~Kn{UCdFzr)~)P58_2wJ2M!*@J+b zMqJjzGy8^z;}H*0Mhm3l=Ut~|-)`oUVIh0Dm5@IJef33sr0=kH)5V`^eLC{+M|$Jn zNi?*b3adWMUD{XAC(%?PX0 zQs&PhB(1Y#t-v9sQ`}XtSrYgkeo^T%G1fqU^yA7krT4Gt^oddTD9BhnT6bmk&6_Dq z7cHLz(^gzA2&2m!iBO<<)^bvEjr|W^ncr04}4aAA)RVSv; zKma@E*ZQ0D@REv!nv_1dHhvkk2UnOG5cI-Nex;(+qO5T6DxUl|PS-v*&IzExf z+251OK}FPu;um!B*y+q|Wc!Cwt39v!7aqCatt<^E) zOMY9q+DhhZ1%fAPJ-mOt_JYqSf|5i3#O+#*8am15iBg!$8$AKLyNPgvhTD4SGTaYZ z%)KU;W&U6$AADUZk?HC5G2eStZz;u^O2ecN83cUA;k^1endqp%5sO8(ki|@dLt|-M20$~-}n|aq@Sz*ren;xT#`c_kR6%KD=o=T<1p$qdc?^! z11)7@(&(j9>J`Wp*lX^_%s(WJ4;j)GYj@P5IlTs8)EhWL>ND$p{BnUiQ>a0uB_Rs&!ZdiU+EHb!T9%dB2Y6M++yba8$YflZ9hu8xp78@zaU0Q4!ZL$A@G`-P@9~5 zIjsoD=ENtH%Y;9M3VXpRg`UZyN7=5EY+-(ZH}3A+Yh!wzV|qR0@6<>%{w#iHUS#L$ zK2ZZ#T;zhzwU!qI>ku8pj_9~;cGkppp8L+6AUe^d$Uh70q1}wBi9n{Q7szn9^)mF| zUcFap&!6v?mSa0&0)%NlDnt)%=Z#`XuFx4$vLxJ5zyC*V?t@NOo2mjbz1anqUhoAm z!nsLlxcfc-^3`7yhD!-F?bV)?(meN1F2&HJt&J$I&1D<1I3ZxN?ucHHc`GrX zE>QXk^;79_kn&aRtwHpFmc(sS zQJjJ`yw@qHEz2xP8=)^XfIaKmmtg^36a8X>)Q6S6% zxP}f@AaM5th)!Q8pRf|azwIYweGj?d1-xNSzSKQJg-6%e2F;N#6uf`(^d4Kx#T%Tg z`=7|oX_q=Jc{m&fW%n4YDU8X!=q-hJAJ*cRuB#cPXF3KMDX)&FAu;93DKx(pZA^er zkX~bojd-j$UZ@4g#&o=#@F!+vIN6Dr@vZQi1f|CVQ&s^Ub<3;(nguflr1F_P z(Ml8&9W2ElNC#pYSwe01+v`w%W9`wmr+>H#9-p6w?Vf6GA0m?c3wATEl7( zK2XCqv>9kYXZWR#8qRrDv4_`f{9u5v$^2kou~95r*xj1dil;HKQnWGPPhS`Rw={bE zTjook>jpXS>-6Ber%3RE-~6^%iaI0WcY^*+#xnYM!p z*I+IN&wXg?vlHY6mmL*S7fI2eZ8Qzbt@%;kmi)`Lj)bTV39!he*FPZLHy}z!sRAf8 z7dF)e1<><>9K0eI7bTzq{wa3NX#V{5=~f!e5~(|>MI2}B7N~{uVKcUyk*TMvEO=x2 z$9-lpQgk0V;gU!lgNhDEE7Acu`9t-Vg=D^hKLdmh@xrok>L}`aLHN;%w%BVm4rhFL zYXKNELo+=yerW#zwDgHwM;X*nXL`GB#cqMGdpiEL@T5-Rld#f|r1g5t02IIf%{{5r zbU`R-AT{w*mBK)))oiBEak3<+-b%z94rUTPkgGY~Rj4$td-9_gW<{6%T6m&At6mgq zJALfDFG)hhD>H!u$?cJ7IO4W<2!EtMcCcR=rHDL5|OGUHQq(Cny**`mUS& zfX5xLCCyQwcz&{zv>mtd402T zsL%im2-gnAD=IUrlVV7d$z#9$b`0X>2EB3~l3DoCp&hFI?Dr1p8G&?>jh8i$JHuh3 zWCNc@GtqNIg6||HgKudhLU9y|?(y$pDT}Gb(Rus6TGTqXvvQpaaC{ozm!@|GyMop1 z+QnyQv|Q+Tswt;v>a(sVvin{b?`~NV-DOQ=-O0RsRqIBU%=)1bhxGmxhSR&WEq>7n zwzpa=oQ0GUkL8KkQ&UM!;-Pe@q=@2s{r+8llIw#Xvlhdw&&T)pEKR3%{1+1}5!c*F z(Sy$4{o-7%TN5!5v+Hq=9U1GeMz=J%+tI}HTr9Wb%aHOf9m(CvS&UIej!?IM`-xCvF|E{_}2uqo0E!~ zeOf%M;T}~}v_BqQNEibuWZqz(y0mw?%(UUQf0iU)htY!QP!UF-ln$oGg^l-}g;qD;dhGXZB|It=MeFc}EteUmtjB5I1LDjl}4ENTtH7xW4c2 zqEZZ7GUx_UKa7^(YrAN~f z=wq}t<;}$?1>G;yN4jSod?H_^E-MH`f(GT0zUQ%LzMA(tpYTOP+&@|UWE^Vno}HOO z&ta9;k=2^N!hn=)`uw{VQsIIw(r9r~laOUnXpHogM7G1ZE1JsI%+uA@oVGy1m{`J* z@`z{;TNk;$g^6e55x7hmx9)>SKw3-`FQ7zVOcvnAMh@MHAt?+*>BRL$cRplbsl7)c zRX%u_S#JExbC=ry0}sb1zeCRiVI#LcR7#CdPH4>OG3iqvN&;taA5c#;R2#oU8!ega zU5tH1v(Vf&M)N@+E$1OOBlca8(WborhhYlz<8QcPI~y;lCmmu6??OZR_Q~7Kqy+4{_Fbp~L9A{J8-7|{GW^ja*V=gEQm z*8R308t*IRyik@p7fGfjI5Lpd9IWb!+|@O*!Z=ROr@8)|AMdwz5G#3-nlmzE1U90D zYEh^5)K%7q?dKVb8qb-Hz7j}B+0v?fB-kTPUs4C^G$%$)(pdQU6l$}4lC4&9X4-f>pnr1UetVpAk(1-!aKS?BY`(x{Z&ZvoTnK&V+-VHPv#IJ-ax8UNS2|NgIL-% zKrHPd;RMpKDk>Un*NIAL?Efh3I^+Ny#`%(D#@&AXV_xmGSN)ZcjjMzCilZ?Syl+V1 z{&Ba}A6rn@sHorMA_&pPrZGbkYi!ylcvcb(WkBZ!p?;Mf5dtw6$_kk_%B;@#x+K5E zb_&JJ@P79I%-UtAll&UGy2O`Nj$-`%EPS32-H^n5w%p;VaLp|537S{;<}w6L^un(| z*BROqwaj^*zM?v9Ug+%BF7~6pllap>{p-9dHC1Q`dSOxTfjtEp<1EN#&d0 z1V(=9*a+ga+AP0Xf_o9FpY8!6?qMEI?i;trBeh5D7%H9$OlawD5u#EWD4jA4um*I- z{={T6*8O=1g}ND+o{PfM^=wp--O5aRt9HO>i~nl==LG(nswamcc#aI=g#F!>Zzkq6 zbDfAnrPpr5I&UZVUT}Y}7R^G@J<9qhS{U0(&hwi4M9j`w)?@50w4LF8raZqPb*{Ab z;@ey$$S6-*BEJ?-myXI?sw|xUOU!*=Bc?RpNW9y)?toTiCgYXMPZ&?_E8b!Nhaez z%|hjcZuj-Ly>nHd>J;efjcvls%$0&@H0tMhL6nwECflq=MwTlSIyVT#fg3Z>n->`} zQctRZFhFNK<48IX@HZePX*;+hZ;WTBX9i(aYbc5pQ*-a#X5-OY_@4hQV%rxp5x$Tl z`{V|d$4#C~4P}U08@@Vro(sUP1@dghV?47^MJeeQl7G*8sN}`$T#@H_%RZ;gAaBqG zWGtBncJ;z`Ss9i-Zst8g-DTD?Ro;`xj_HHJpXul7b}kL7r1Cu=%l^9X!VqcFE0+{I zz&4Wo$6;_clV+EYdoROS8K+RenCFq@c{JCEsKh zg19EnS@aKM3P-0mr~ASPn1w}7pLXZE8Aj;qgm3K-&usNwy zYssaToci^h_+rnr%}B-?R0Ay^lFK8`%3!<6W87ZmaxZgOJ%RML>v1>8R?2#Q8c6nP za(IVo*pOk8#RTdDBnNu1sny}tmt(Zhk}-0rOvyaqyNu|Y-E76(e3W+FvT;o*W;f== zfs&B?BWjsl5Q6^~uRQ}8KBFEulifxjz3e_(;|cLeSGn3rf!}JW=FMfp@Vj!XNqT<= z+*z`DS?^+bi2xB%rkG0$PU|Zdc@~~m)uaJA=(j0WBND6ejd6>et1ReIjgvODfO_lf zxdcMK`{;ihO)nKCIRz!pOpFmnvOL{9$I?5BTT&EPZb+#pV3;+`eR}ULMyW|msjt0* zLoE8wVt(Ja6WlOv+u(Q~2>TCWg!ao(#DkGfNhGIXd7OTf>)kQlrhmy5A_-I|eMzCo z0F&{X-*4UMn55L0*_y$>1fn5J^!-RKqbC3)6sXG^bjU0cK{bD-ybHu7$XOpzZRv2T zN%cR~H|}v#nIl|}%`ho*DU8QKvh>0tTOdRR z^ovZk8pwbn41V=><~>BEz;i={8v0!C_|Lyu6?ODmlDcJo5x&LZj<2wI^=vcx)gl1y zWU{gx_oA7zXrMdBZ&Q>o_8*Wh&?ISOpXg*Z+OK2;tv+W?mQYU&Hv0OC7v!f5GVp~4~YTN*=eS29@6qG{U8)dfi(y_ty zTaSRN4OVqHZ{qI$Vw#TvHj;)>M?eQxKkuJ7pK*<=tB*d`#FI(QAb5TQHTDMnLtc7` zm0JDXabR%(V=B#rhJ2ZLI3+#depF>{x{Ka49bbldhQ2{lUC~{>^4Yv4hc87h|BUHI z0n64=yfzn1{^@bw3*Bq}KZ;}bNU9$4iq32yu_Tc@Enqv&ckF@|r@OT7`fXo7#@}Ap zpNY8p`yq|)J<$}33Bt}HDLv|yQxfO%@idwqwOcPAdQ3F*A*q6Qf5L0xO?|sg=03f@ zUiSXnQtn`@wfUAxSXN)vZ+!EtB)%B2Vg{AL=>6m}-;U>bYK%LR%D`;|X0{N@+;pwS4U zFr=-dp1hy)&6K>ma4)jz-krH+T*z|NpCy8l6?V>(F$!|`?FBW=i^hZvlaKn(2vLpw zx9#VmjMQd07zL^WEOp}@YiyPT>WqjP`Gf}eOm}N^0wxu zXo?5s&2#fh?^VgXinl1(2lXJ>cLk|>S$vASlo8KLrh!sKmmyoD_CVg+43wibSo{k% za5sP8q_iY2I$pY%_u zVVAQx47%u-@;XWgPlJ4ixC;rSbIHS`g%M2N29J$8IGLE%8H@5po|_Hq=0^EnT72>b z_dsI)O(J=38jV4RQOo#S%Gv(>%W)*2Sf7&lhD3d;+XA<{x?O6q`GWYjQR&s;zZ^{( zmclVkHSoB)_bSg{>er2A6>N;(e<+pp1qHRMJpIB0lz4A~#v4a_FiBzHbpoZ9}J& zjE;bm@!cGicqL-tR2!>+#a3(xecOj_4=_W?)D|kVL4ZsDDF^rG9c}Q~HM!yYJakks zu-M|u5^wFJjgz7hxZHdV-)qv&&o?!W1p(RTt1&-Kk+fru90AA&FEJ{;FRzae{6t0Z zVVU?><8@kG;B@aD_tXd+;i$Zx7s5)GOw=;X4tMU}oKI|2srzF8(f%^Kf}eOPwfj|G zE393Km1;O02g+~RxS^@^2^*>l;*;3POX9&ngM9K+mW5%2(_ztu9yE^E`gDlzf;I?u z@INMT6tid$P~6yB$K^DyE`~4ydDnlz62=V8xv_fpcrdr3I&Z!`Q?L6{jImq%5xX(Q z#=(9==A{@$MRoeO!wsP^7x-4aX&=l>vUXcYb}C^_6sQ{QgJltl$5;v;X##hMAZkE~ z=ie3#I3$eO|EJp{DQr_z!^SsBNXg88PBwY79*{Nsedi~)v4(^WDEw@N+V;i%`!y06 zK~6ND9=Mq#tOhLH9w>c$kqF#)s)?-N020=J!=uzhj3)Ei1T8*ppRVD6cvpyrG=}z3 z@Rn0~@(~iXSNPMr%=jhBC7zEIxX~f+t$NS`ToU>`77L#}CsB_cX5oscaRF{SU(kjb KsMkR6M*Tn3)WReH literal 0 HcmV?d00001 diff --git a/brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/yu_2019_figure_1.png b/brainsteam/content/posts/2021-01-02-nlp-model-rationale/images/yu_2019_figure_1.png new file mode 100644 index 0000000000000000000000000000000000000000..6405ad193df3696c42a27cc3bac399db99a06f10 GIT binary patch literal 25799 zcmce;cUX@9|36#_SEQ~at6iE(3GFmck_H+?1EG@k-cpn_G&L2{k`(QUrf6s{O5F7! z6xL*9 zkt3-eEV@1vXrANsxhSISBeMT!L^fAlhg*QKv03QR2w~p4*L|4Mgg^+)Xm&t0Y@p4U6|jKtf-`zR7ANq6GO)|je>zsW`= z5_GrXZ^~_HTKG%O@c;bJhst)OUz6lMa>Ts1&~CVne9Jb6!&PHMcP&L33Ga89 z>UkO&nS7-B$4=5mjoyCq=FNFk)h%1L&?(qTu2%&x64mA@Thi~xhj7t6dYPH|{Q2`B;tk>bQue<)a_=#=DwF3^QInDN(#NSS z{rOW-XK!z>k^Gn{MH`=5#1ZrNbyCs^F|kWA)a9gg@-!c7%eZ_2c| zeSN8qiAL$trT5v{As-EgM@H_NnqG|4v)HbajvtJ3U7q<^S$Rf5VGv6g5)$(L`}aQg zzk&+B-)=vbatIv%O{PiuIXBnUKed&WTvAdlKfi>puP%I+zj^a!vE}cMth6-#@$4;x zkE(`P=p&E){l9(vdeUvxoA|x4vF7X7Fg(PwXDzL*`Gth4CptPhPMqSP~Pg(>rvitWQq*>_QIc-PcAt-rPoLXTZy4XD9Xoa$I&?_%URPL1 zNVfa>YNh`^PH*>(m61PxY}0Hg|J`(jzKTlSblK)cv$||+OADKKRUos=+;Hv32iu1n z4j(?Os;atk=g!H=$>-8e3NM>@Rde$mq+{sB9ZUUqK|4h};W#xEa9H|3l9<;$1J?!3*7ztd)-jO@cxQ~0m4^wiW@J*Tb@CSn#p zG&MBNs;D%7|GxX)>;AHTC(OzgTa(p&nO#pvmnz*E7ZkGFn3r|y1P0pOD z{2F!al!C(PaPIzpyIQU|ST%HRe!lj#^1XK4i=bdeZ;=D1_ggGHHvdzj9m}5oZhoJ% zYUuCZzj537tS@r|J`Z$scPl9>nwpp-y?)Kl&%fi?m2%c=*rA@69oJV+s5igz_4QpO zeWYZZ`|2MPW#!rzFJ3%%UsZNY^%9lKK`q;po{_OIJDZiA{WvH{iBM2fROa@#t-HI3 zUhmeECr?~mU5N@jYO$VvI(+0<6S-GynFf79b&HeTwRtVAt+mTRvKwcs78Vw2m%oG` zu3hf`_@FjFGIHMmF)=ZoA)AqUt|3uTQKPk@=~B)i8X6iS*YB^Cje1&J!=HG}e-kru zt$K36$o0=?lTpv-&kvr^vYw26B6KC&`w8QL15amL({+|6+SM;yxS+4^-dE~oChF2( zem{ss=6id)*ohNdvI=jfCnoCLM#m>6ta=(6f}~2#Iz{N|>1iI#_d4{Yrlvlfeer^$ z*j-MJY-#n%l`G;bOiW*y)m2q1XYaPZwW=N(9JE}U>MiP-MI{+>baWJVSZjKzB+kOX zP|1vy5DnxxbjY&0c2OJEs%JJKfsLu5wss)!xwsWYwaU43`9t~5bP6diUU*5BI!x;` zS(up(2YRIPn)l>CeZshZf3TH~j*j)BqdBR*v0c>H*T+)TMG935<>ch-qNbj)*3{gy z67VWDwf08VjJ2)p@u}-x#EUvQtjCRs`tB=pw`^>V>1%6goe&kRsdCU zz`)>mMOW9k!k5l*<+;bk%AYZM{fMxz582tD>MZ(7*OzBG^&^6V_n$D`qaXO>$=X=* zbDRgF!g+3+&kuK%mX=<;co7$`tgIZSd!Lh2xl=+y0_WDKr>g3qS~92i+`m6H%cD&% zhlYl#pB(r+V7XUFzu5321rw`;wbApFZbXGMXU=qXb`Gy{Dn1F0h)^VihlLfpECoJ^ ziHWJJsbTXzr=oI9M5Jc<-SyA-0z?J;((1@_kDC0bs3M=xk= zD-yJ{wM(7mA3kAZWK2&>-u84jz( ziZ>;zKX1=8*lAyvaF5hDJkOT6uRD)@Pg>SaL%o0gyf1`Hh4SQ}p_W!SDucZIN37h- z7cXY)1c~J0jVRF^M$Lz=B^MSIB_}64yfyi*Q(ISulD3;29f*Gae%_hM2Pk-wslcW_3=y&E@Pp!K4-8t-mwY@C^&KSe+>qa7jh-bLyr z#K%~W?v9Q(dPTW~g}TbhV>1Jv>FDTQScDZB)$ZQDeYPYw_8 zlia>oX{U>dirZ-KlU8r7^>GY$YkT{pxZcgX=H^u|PaimV&|c=Sx_U#_&Fb7c-!iV| zwY9X^S~*aBjH_)$V``M{Y^LLT)+b`x#1vTGVwuS(gCYjSeFl-*pTXE(ZYCo(MT z&==fjrIE-M3X1*Iii(O=@iH!j*?eAXEYgm`PIE?6Y!Z(TNnAWCPgDwI^7Hrq%*n># zZRrpe5wX{6Vl}6rAd1N(#Jl?8aNy9-)m7)ItSv`r9-Rt2s(j;p=dfp3#<8s2+^QjQ z>pS)DtA*}G8JU~cyG45a*P_3pZjF46dQdis^JR(#8Xm6x{A6zvRqyGW!!=K3*JidT z)HE~*9X*=W-0c?>BqS(UX|o$mV(YeTrpCtCN?j#Rp8V0(RcKJ=p3g4{ED^{ob%=|r z;W?i6n$w9BIfr%LR#aAwO-u~Q&!E~D-ToHszfZ^<8xP3TaR?_4OI7Bvxsj%o9wBh$ z;?=7zYs*$>sF#w}(5D-kn=k3;M8+~EeE48H|LfbG`q<8VtD$12c}rng*}m#(9$wy2 zJbAQkdI#>vBYK6-3;h0-pI(lma=QPW%xi0FTYi8R)bjoNrxS&=ag`{y49 zzx5&~^UTBz_LRECj0;fqV^d=ddJK}!bI;?yHAh+_qN5K=N3Yd9<($>J$+Nt=IzBoY zEHpniH)^|zuTdW>5gH#KNCXz+Aq=93o##Hr#pT21J2z)ttY&FBGg#?QzkBzMn>Vqs ztgWqQ6tTdpvhL@J6&ve+N4~`<5l@?#jH%16pC(e8>;HMaxt5%mc%O}M?AS3@R#q=I zAc$|@zTtWVF9H=+*1LCX1dbCYT6=nm-@OZF^L?fglvTmz;ruQ`z zUeeWl+;>pQ-t5vP`aZ!tZXqF7M#gSQntyeURJ^T*6w}_1re$Y~O!56Vi>{d@nUhJ_ zwW~69_3vK?^PSq-+CP5$7_K{Zr~X+?465}(AzDUJ>S`byTRXeixtzQ00)m1KaZ-;A zM8(A3UluShFjx!uJaAT7c`GGl{~zpQV`E-<;?}KODdl5Ny6FN^zIj8PDQ(=C@GVA+ zp7Oedg~!~mvKQNKdU`iWc#15O`B2R1i1N5A_?BeMU=Q%)s*US3OarnKF6 zgZ$`u+&9;k2#JDMbJ2Wk{|u2Hz}d6?@!^YL0QgPEn5C&+&gTUWA3h98?=7%-RBJ?f zuxn@4)Mz9m3~?k+o!SyVZ82Ek_x?^p{E53*o`~FB(I54Ayr`{21$1GwO+cZeM~|ZL z0SEkDn)aCKmv?kr_NC(l7>RuTTs`;TB~8r=e6wL(cc?ljGB~(g=Ds& zC7u6_qDYiG)Vb1DiuJiZ;Q2Jvn#jtBFL`W&Y zW`~rdo5m{q=%rRW6A}{8Sv7TZh9ZHM(YbM4KGtF5IljIK2KDxB;Pm4APBSg7*8YA& zYtAD3$?wh0oCa_8iXtj+WObUks^{nCwxL}%H;4B8LT#9h-cHNH1neMBd_=?iy{)av z?e^L3+@Ws+=hgMgOSmp2OpmuyF;VE%g_I~)@X5}T}Vi18ix=2CO0dKYUj?- zr%$iH?|gsz8;977j11L4^{}`&x7oqUvW-=LwdCpEA~9j%ci1fG0FHCRR0^Xj^CM_U z5fKrD#Ny)OK95c2z6xzk&9#|DJ4 zkf>)-Q7B;njABQmrF)v1=y(!dz1p*PZ!UQE_0JF8OzIq?|16H%`hcd!HTL}bQyK6z zUZ!Y^flY4@y3khZVW=2>ZJX^WZm z?@vid@$K8ObLT{N-gJGO)WpODnxpjHHi0QsfXTxA{8(|T!C&8ESXfw=2X5)<#YIH4 z%??%Hym3QXMy9XW=@hZz;VwoT;Rf`Wsh$E7FK3e{;tb^G=7MY-yZ!a@``fCwZ*$8N zfByUl_LTGPT?bGbs=Bp}jgyFmB{ngL8xw*UZELgD#LTg)^ng%w+<%tCa z1#W*Af3kdh{CS|*V^bf^y6=o}oU*=t{M)xjmh8XO)rE(LmlghOLj_TbldQZGn;O{m z8M{)#HAm8A(JYOg?g$SL&!G?LT3&23!=i*lr63k6lYM$mLqk6s3ga-Hw&PfnS1|~xzw2+Nq)^`dF>jL zgeHT6F);1Nj~{Q{y4Czzm6JfhUSTMV2gfNSq^|6oz;pl8wyw+%AN~y2h9)K^qTFH| z>+0&>ta`HYGt%JIt5^EPj#g{Sv*7boxfhQ7#BJv~FN^~>Gl^Ro{`wlVJgCL=gi#DY z(Mt19)84yx@1l^MwD4Uu|5H|6Qes5`G3fInT5;d9jSo3Fp`oFv=cxZkYjjjRmRrr( zG1NJZItH>1fCePf($V3#x#31dMTP#gHpS|(cAp65zq_BM>|95Kc2-xH?p}W%;z3cf z%k7LNTfUf6O!pMftFy_~-*0bhBpy5nY{X*=Y({odT21vxSeFWWH=<}KB_+** zzkLAXVlm7dv9h#$hVuqooomr|62;1Owu=Ge4S+l#Fc9^^%)}%$B}MqyG1U1#Gc&}q zpKb(vI{fnyE$bi?`ehK4gcoseDs-nOnf&#M>y$3^%WEt1ghZ`$Z8sC2*RQ$3M?o#8 z`hg^G*T3p(J1=~xs2Bz{ANE}YdE2s$ znr$S2j!QJAw^DXQrF7+23=U756@R*xXpE};^ew1n4pXm$4>Z32ckQbg(6iXQYv^2c z{R0Bd5GFg{KYGF>VSQRxS5iRW4O%gHI_mF8s_j;VuvUI~;=Uu-$~TqrtEVSZ0QW(h zLPA@E7pHo!>gcp5tx4XRXv;Wp;>4$pZ1KgbNzG&N@$rtGfLdr4d$_kai}IQc4q-`2 zJ-qeGn$Gp>*RilqpEd=vpDO*J29mtmw>>>sEiU`rJ4;*Ju;5@}F|j^~d`qv6^tPbX zT+`E|Q@}M23kptF?2V3zx$`ah1n}n8oqK1BPq?kx&IL+V4Oq^2kMyus??gqDX0<4> z1^`kZG#sp(=UdzRKlmBlVW4woe0uR@hTXx+b?sXHU0C>G@~v|F6W;mx#fVviEDcwZ z$6HmlwdDrb;RJm&6vpz`_!bsAD=2&d?we7uu&}U>es4b6k##dWI~$v_>)rKJJSW{z z*dVcB=VvWk?P+dCOELdZyaI6uXDn9CLSIi$&(JX8;lrw*b&=rH85tRW7bcXnUq~Nr zZ*M0ROI6iLbOopaC?;sRtDsUq#v98+L8v`}`;StHb)WpdVfL(wcO!~o`G z{iS5Cd2N2OYaWE5@A+}KcCDfC9tH+0ZXEH1|DHZQ>tNf*+jS9wgv6WGfjL=$Koq!2 z0VggBsNlBxXLezM>p9zwSYHc@o5PyY(2B=K@9?Nbd*6xW;)GIDxR#lpFaCcNPVKp{dI*Yx!Rb%2jZ<>>O|4p8FRSJpU7IJ9V=l=7|{Yf-gEuV20D zKmQ+Co-w7A=RJHFJu!cwZTIfoT`SdH-QCHrU*`s1(#}v+Qc_V-adC9KR_sXkT;aNO z6q%>zC+^!78M+H|a|#4KARCp#=))W&!MUEk;menUvhLCfHEnI$XVs^9Q?)zIL`!=! zl9QF{)4S9F{Gf#lQEmKbR`*SRnov@*UK`3o<#qP_`48{jl{dGeuT}N*tf8DZM)eT- z!hv#bo0uq{ER92l0KlBFmXj+7(3`P7d)617M;|5YcGhL6E6ZGuK0gf!*+C!}m*l*C z`HS1~_U+e@MSOfJK}lb$KM&-&4EE!0Wp$Km#gm142ct2g`%Kk*2M~rNA0L&_(Dw9~ zNwTO`DvNQ!cI}~L=0z^)8!=l=yHX}5lwfM5! z3ZsDOIIBx2P_v~Qs}n+V><)nR1+LRSN}wdqDZ2d!i>HoMmYn!V)Lidz$Ad%_YHVpa zR9(d09bzf=VN1dUPE;af3H10Q=BGV%(xDkE5uQd$b)X#A@F`er5tBZsgF~UJy31hz zgkPS3;`eqU>Vsl!EB~*dp$EiI4ocmdsS6YBOJmKC?$k#?wFzw}9Sxw-Ut_c1aXld+ zk6&L86btb1JYsd6*Nggg8>&&ry>QaT9Uk_MWR-FDCO!@fR zLCBgD^ygOAWi!#3J9jDA$J1LufD7{T!OhV+LSnzvL-!YgCGaFUV;@=EHXl`KK3-nm zR>D^Zp<0(Nkw7bMK5z{(GQJ!MpSWMaZYM&s!qG$t>&A}8okRiR_f@_W>g?{m^t6b$ zvbuVZkr7=7tvRR6uMZs*h^uzx7f}J_9Gxux{!xXq7cR`A`=TW;|1|Je2SV6$-~diE z8utStmi2~66ug7Fy1Jnv>+|QSW4?i$?4{Yuz(7ZNp9ndl$YtrRmzP&ypc)`ud|F#o z6}Nx@HTx@+oma2CWkddh#emMuW47@bb~S(zSev?$UF6zL2DsWQh@ea9V5HMs)z#HD z4S!)$aDEAsQ{U>7afjZltQDX;COSGTHMO&=3wi+fU%^paWmPHSK8JS+ds@`51Vc{86 z3ebjl(R;dXZZht9rKLY`gwLG01&R4?+nd_2Uk?!mWj9xX9zEKw&n_&ifdyM#m@v?F zf>tmbK%YshH_x2e>hLN%dk$g&^iMO< ztDX`w3Awslgd0sS&Jft4%|nq*KE*1)$LALipnC3{Y40=t&jX;L!40<@kBf=z+qW;z z=9U0#1zBxlbMqs-ylg|$ePtt8j-iga4C(zzG7Zl@3YV;;8T%FDLFLTbQ~;aE9~ZT> ze)@s)`?qG8u}e74^rJmsM??r-UBvn>(*An$_U%T}kn+o>KMMcg`@ z(n6oNwzY8&IXgQ`-wipHLB(+XCJkIhSr{hx6b_PM0(PNfCplQz(eW&qTw)u^I`I^W zhESK+=k8C`EpvCfaDlFM%=c-ekii`DOvq|XIrg}+ipqP)t<34<#CX_V0Ur&|$4R!9 zmwS$1^CT;F`>RcM>YLxu?2@B7fvmC{D5t0kP{w?`y*1(jJLp?JcdLCCj3MbFb{q*K zH-+~T60hc&Ux+)KsSTdcoS8*$8JeUI)TZ(ZMRCxG^lIYxQuRQ&8|4KP^YA zM5K)Qrl`@8^7{4bwuJy=V`FM2@plv5W5K|m(&CZ64yucVuNzT--d-!%&A^azqw+DJ z{Xo+f=Yr|dzf+M0?njRwANGU5=)5yqMJiO;Eo-};E!ccdm|_?%DRHmsIMDmoujlEe zqpbmu@kMY-h4}ebp|02XxBn23mSze%Q~6+(TSZV%kWf12IrFrYf9>JRHf-jam6pOt z{f%3<%*>wLag}mkbHuU6QVNZyYe6l1alee7ZjWf1UZL$6s2ouFfs$g~GlYA+g{>78 z-k%2?ot$#wshA8~bkg%*eG+3PioyjgB&2C!yMabDkNv?PREzL&Ti(~1nL>Pgosw#9 zIgiVJJKHF4eL=31p5Sju_i3`$Mey$32`-fgHt)sXouSa}?IbnNojZ3BD0?D5Uo(A1 zb-VYjx7zai;G8W^WR@*rXJ%zt0ckMNJQ}yHpjl|;57T5j*l47D{x>{9^ia~>&fM}L zDcW#TNQ!oI=r4U_VQUGwtg znj2Y90s>HyQtJKBfy!xVXr#P(1L;jfOl%X3XvXqzNp+l*!%2^g9@3)h5WIiixLnon zC>@&WzI{q|dzoP$eWH(&cFGP5+rw4sO9kf+^o_$CPA&>IYAtGDs@gdL4TBh|C4>LT z0%UwrHDo(zNq_l#ER;{UkyweNaLJ@UFTisTNHeOC*VF4IsXI19OE`j>KRLBiGPM&K z$cVe}_#^{rjjYwPxbUo2S95fpVvR0c8geRJ>;JeVZ7va9MzXJO6i?hTXUFL+cqTfQ z?Jw!M&j-W97e4T%$@sW|K_42D)uzR{vuAk)1(Spi)4_TndGZtDn&}C7o_Ep{4wGop zk)=7Q@>hV(J!KvSIK*Ycmf}F;Ew2yJ;D!dcg-JIgz>U5Irf3S8@EvX!E6U-$pZZ}; z3?p``KQsk+lrZRqB4->KE?my|H8LV}E#DG02#_}+F*ulr$Jf}Xkv2yz(}BtbD!IPC zK3j+8zPh{&*Zfk9R%TLia(kJF2QESF`qA|6-NV6rH(jVNU24wH&p&#zip`tlq-%#u z;k>hx)2%yq@&cO)UPL!h0=sBT25Q?_GtX2%6ykw9Y{dL`?}S0)r+kZxT}@2*`mSe_ zFNc(EZge*_%>vq@QzT06%W7_Hd}sL4D>yi~qvJ9R<+NrxbOWE6nVB{Zfm=H1)Z0Ec zH9eK$IDA;@`X|bNw+?uJ%g8976RMj=@!a?l7`>onjqe6dsj0p{bn?v-wrihReQ#q&gxlB^A2a z3T9ex)sCXk9R|q@NuNd7#ZD`8T&-7p)u8(LE~O+Sl;^E;hPS4HYBVOEg1S$%&D^F? zrB@gm7w0rZwWU$5cQ!Ek(W6IztF2cEDz|R^gpI-5X{WEx(kEUivkO0Gl`}p(+)`Jh zvn5gF_;I7G%P+^klab;?K5bK#-j~cwNPgZU3u#`Bo*|u)cPbHMtGwV&T@Gg+S>Ne z|EEv85Ft1W3$vQcF;5ji`M5l{|&#Zsil?rjq;BpNlOh2^Zv_0 z#z`1lo*lx$xctHRg0AjCg%pTQ3*)WQdMmWdl1OoAU%aTGoerjxdkvBpodSdvq(;PV zSXihND=f;^MNL~$P-lEM)af6BxjGuZ>iWloB2a@%RLy>y$B~tll`-V1OMymIwq5KCG;SH|VM6N_GH>F+4n(y*pwmV~3F32Y2-m5H)F$g!{tG1;6j*O%x5YzcIIJyTP%4cY z0vQ)w*5zA?%hI;(+d++%%I|NX*rbm~S%+MT)gokm-XHUA19*STb`_dY{$FE!eOL5Q zg)S2Z_`|G{aeWUov%R=7wP%Ae;~z6!-uYIW&pT7Qa=88NjDYS3!pBzA!!7z5?g$H^ zq5SxcDKSQQCkSq;u)g-J9NLl$4a1nd8>h)|Qu-@xMacvjGEjZ3n)l+OfkK z2#PU42kD$Id9y0rA)%pk3cGgg8u<4IB#il_Yd(wtkmIDJTS|FV?cm^a3Xrnzy+P&x z#25PMH#i(9Q?U7Lu?VPwxQv6H9jwF&X!uY8k*`6``T$|$Uy>+UUB=P9iskPkaG9)p z`vph`%D+|qN;$OVtdo<|*Rmg_ZvMnCU%vtn zY$Cj18Q0a*<2u=Of`K6%br%Bcz{VLNgVH}sOVz4x;4!{>^QPz6cneIAIp#8r%dPyj z))X(uxpvu7ley@=v5EaStg0wj%tWpMCR^Zwmz9I*8ZSB46d$80*05KgfS~@&j9(Hzeb;loDD7S6Hxl6RW z#@lMDso4Vk5y1c^A~6_b7#j>(EMW9O&YhkKa>^5dxGIt>Dc|IEbn@VFBYi|j#QH^wp3=qIez`NSNq8ghVi0M9!4XoPlWo|I*Xb6S+9oyUgs6?zq_5Go9_tG_ zrp-`OtLI(Ovv1;>Ab!JFmk)k;{rchbc8Kl0Jv}fpNdCxx<(@rz_zxdOfi<{x4Ki@2 z^ZNL>sdbFk30R5fLOZ9G4wViL9}pb@=OwVu(bxbl_*WKM4p8(77CX+~fLkfqj`L)i`AugtBshmUUjykGPY z()-2z9zGi1;-qv`RRv*B=H~8YuNbhjpX>?`4ZV5&y8K3joxQ!9x_X)Yq%Q7VS2sE@ z559HL?A9aZvlV(7P<@oookJ0~f(DKH!NcGY%H{@L)xyRKoNw~?@So4LvR|=H+umFO zbivn(LQg}MMmtmg-mO~8|(H%7`0e%`t83rBs0HoZ0n;fRokCXSrl zz_wWQRS3hM8dmOneRiL?yWclj_)CcR=pe7kq<{%!tIsaQD}xiz7^Sz2T>RA6*H=*i zmoj3yDF0?sWTei%r8&6le0(Go62&u)E#>HaDlc9>J}0-^N5ow^G@ z4ZeeXB!Y~?QriQ01+V369G5AOU{bqvi%Xt}6+xPVU2rV#KB!WH*0n+!=*j1O-q^T-;Lqu^PkN*vlGJUq8p7 z8)t%BO_QZ0ND}AWyF+SfkQ@&6VF3{-QmAL7b}m>9J5K4d>x0+dXZB$b2=nqB9-aKu zs~y3Qb_|uBVf+YIQoqbS4nTB=gd_Xju>Uk*QdRgX#uS{7(ZS3xDj+&(6N6 zs`~BAmo!vnHiBr(@U0l|rBa zv0+8}2NHgd>-}%HFU7rStgWe;nJzr1{Z6k;m&?zFhbLc`VY^15}I)^RA%`vfkR_i~ZbpCjcc6aIc_XxJPI zlEp=U@GL>ftv`8F!KcJo8jA0EaiZsRS%Arc^AnIh)n~$qdaY@w_O$Uj{?nH`Ja_x~ z`MqHBC{9a5<^)XvPxSHY?ky|fl=)z$@gEIGM@Kc{2>obfq`+o*c{%ilm}Q5+%K(MQ z2;zQWO61<*TsF-S(@Zz`l&o7hFG$8_hDQoI`l3&$kE+bG=J$`1t4zPHZtVdJ^m_0B zS*^=36F8ylV%J?Ov`t0ld?f0VsqOMvwO?q(A_v-g?Xj^0f-78pL=Xsx@DoS|g>}0@ ztScGeIl@JVP!Q}ZZokKWq6E!q2%s1{K0fT-2w?!J*xCU$?`jn9jv}%I8|Z%u3JSoX zkM7ti5t_zqw{PE$@QH)N(qa&QCDgzs9mW$>3@kki>|vpyr|xklj8M_ivLdN?#1OxW z_HAxr@-L(~&FM`t9)rivyX+gyJ68ZJ6e>-8U^3h`3~-Qw0{g@Qx646_E0(*+0JwY0 zwP=4i1LNp@WCpOCaci5{T#|890|T=Nr$c-OtHz7~RvL0Eh?9<{CNigOU%v2jb3Ysp zBPaR*HX%2M!3OZO{=j6^HiV_1Zo!F!83){>*K=mNfa=YQ7oW-&B1Z2znnO^dM0cr2 z(7ZTD>PCexh9OGtqM1H@pD>iB<9RH@OZ2p&V#u+6iYCNw2s?Hl5iqv(C5+v7=E{{72v4rteft;T6U-?hj1P~HEdD!_CyVE|+i1K! z-Q6)oyU5KI;YYK-iW{WiAtxtS=qg30XTtZN6crYR%<^4&9c%?=CUVl| z*Ljh8nah9R>D7{AX*J8=lT>Q+F|BjJpvYeE0IRSe*?}b-^kIc65&@ZVgy#}SQU;_D zat|Cr{B$62-`>50j)sP84r>5_;_$n-PB{YO1mY6{q}-=}$N-3A+$cDQFF-LgEDRKo z$DPTShlg9n`47Q@2azo)_rHtxnADNL3w-45-R}sQ#zC5wm*5Zww7$<_4fciOIv~Z= zGT4#T8r-R#Sp)Z#t<{VN59SYH>#))0Ap3e?6S=ZlasNt=A#^D2A(U?;SMYnIHI*|> z_4R>I8lD|}@I*>NVt_3`5g7+A-9)-gd$kv*&z*Z56?H&?mxo8mz~K3(^KxQZIv7dF zESu`i%S=opXE(ZaYXLJM4mX&dgAITJ0hM!j1Cx>(FzUHIO#`JQJ2&^X=OW(0k00;Y zV38wKR#k;;Sp_86g`A`iP!T$9|1Hp~?DlG;xS%%9X1{qO(8(gm$2W$r)M?@5Yg_v06lb{2~wu$XS+G!|0{~KJ6ix+pZV~oXfM6yCL;Gj?U&%OKhVe%p3 z3F20RPP=6x+Mwyde*)#ebKw;gjc_eYO->&9{ae~<5XiR@OOFu*4sU#B!QYF~QBec0 z6#>9d1dV!HT1bv3JJdHQYllT(pgplzL`k%E2t9WE_=l{l+Hc=vY)6SpwijXgRW!RH z%vItz>qQ!#=ua*HOn>_HsnIocb=p3sK}&ID(Kv@BB_##+?i5lweL8pn4v*?Ax7jSNDQfx>h3tWHfG% z025?~1^NF_ykHT-Tu$Ttj^cSbU#>|53 zOT}LXR*eiVb3RVa&y|&1_pxw?79cH!M||I`077DQAoH**35vq4py(7KqtF5QAri~o z^v=R!WM+Qg@aMF8(79cSdu{-sEWLz2_scZd9jkQ@QqX&^B9#&=(4**cJV*;YS|Z4V!i z^?rsS9Tyjud_F29G7@|^3ttWuQ(awsU{|#I%O(_EkmVxC{@WhRPW4K`oy5q`)HXgc zUqz^!$TLGT+D6N=Hs2sS^7Ch+hg3Ws6(*Wmo0={hr$6heHGP0)PzCoq_!c?R_I8KS z#zf>r6bNoeDMQU#1P4?-riYdH@YsY1-xKx+3B&h!^U2oh=M7V|?=N!@IK;&vaNb9h-UcAtJeuUlX5T8VxqSKZ zX(grYTel*N)za1m!vPMF^%6aq+&g3^mnOT#Ec@lqN6@1Y@80%cV06?GI~m#;?h2&& zVjQQ3MutMpjtWgoKE%cKqF|Edf~Hq5hSUBVHN^D9?95EUzr-0zE-(Z##jp6e8t0FT zis~2HBeSgvLk8*e!-sdWA4fq)RfKHN_ARV}*hBfyl);xYtbvY(J5xM!hTXd=h!q1C zK+r)&C6`oAo#GT1UxxmLqDCnXW94H_O-)~4nb@Z2c8aIb(c?oy%@Av%j@<#u$xBK) zBQHx83kgp3Mjh4D6t(K`g^|dFFcd>%0#iya$HldC(m( z6gcc45~s#`?Dl!|CM{<5BIA~n3xveYwGmkWJd)C#e2V-_TC1-w0je>I&9xA2)o9Yp z+}uwU7$jjTXV0=Jg?!dKQ#-3!sVy&mAKVPpy+qVA&6Ugt`2cUJtp1S3$!(&8ET`y#MAhsR zc zhTkQ=^V@voV2{KPl9GWt_X@Ci0InC|tX+tUGqe5L3dH3{BDW7CuEVRD4@}8)GE`}lB{Qn3< zXzd~O5~%UdpBpGEQ?*9zz-N!tlu4dE89#2DB5$>wagJ^Yae&(_~{g{Z@<7PZ`>Cq{{ zcai~v18m4pJMNWhfSD(3H{Jp|I$Hs0itpD@rEL?U8ZagbIH3+`$@N36z>df{D2|N7@w3QUsph6i$`WkKZ{^KW}~={DW!dN)FVO;Pg}Iqt=l^34#l1t2B(Hc6tp*r+Mx z9bH`D3`IYIEO-@xxvnlWNOZAt%E}j>#Siuv`V2-dEo~5H zUPL%}cu2lJeht;`#R~+CdSlYH+}zgj;R{`o41S+mP2Jr~fpn0m%`qXn@9*d5v^+Ba zp060hvMYRhP)ke8%#0;iFpAs7xSgY(Ak@g|09m1$BM~^{QLvqjl7LXzj{ioS*~o;C zGW{EUCQY$ng8!6TC>I%7YF5@!X=!OvPFfifwx5d`Iay)ck1O#57(26q-7jI*@eXZ8 zGzN0lZ@|s8V-yMV&i9;N|1w>uX7zBL9o)JSqWBV;8qI!TLpZ%C7h$g`DJkj&3??N6 zFu7P?U$67_+Q$+r+~QJCz|sH70zjuvQQo(<^UE;?Lu=opr*o1bYr$o=a0%Yp z9k$H>{X9fkWTp0{9oxFO<=<7h6~Dbl3|q6<5J4|N!2sfmCgb5Z*HI3_jJ6ii;Fb+! z1fb=hq+=I_B@veI&wbq&Hj;bc5-ChM9e*9lwtB zB1CZXKLrBJ+atOkaFmq7UhT-2clx#Si%oY$lCB#>M{odzn$MqKiqR%__I`}|nJ)9_ zC-^1ijDdj}uA!3-uU_i& z2%2trV7`O2%F+T5Iw72u*w5WLO8GNNh_~V1OT+X1K*_*E@d>axpL=XbIR7!G)h5yNikmZ)%%zt zGG8RjD1fj9BGFKYjY~MiyjKV27M2vX79h?qprY ztN{`YrEog~cY3I(>})OYwC^l-I_6)gn`?HU)z4Q*G$uf00hU)vy^qLwQ-{8m7EIR! z5WOrSSV+_q1pseG7(ZQsR44P1wMMgdML4x}_4OMw16xSH@5N@BTD5NJ%=7Qv4P*pr z(hCvDJj?%$uK)k>PYvx1YwK)0KVr_qCzVC-Jkg5z#oHF$?THOd9n6U^Ak& z*FOiyBW-@S5av}76WzJ*NQKZWor=18vgqN)Z9qkP{TU$D5V6Z{W?e=SAz6}d76`xo zMi%w1U9)w7DA||aK)lh2Vc>eEIY}zzkWrJ8ww8UZt3xgW&-?}H()b(GMIb|;g%GaC zA2`0L-_G?Qb|5Jw72_He8u}7`f|1eb>*o^gfe?~3e*L!@8Q<#b;~^2`@cE&MzJ$7h z>b{nZ8Cv$^I2q?cUu;9R1)m4Wt>-E1Kg5DG!fAN@lZ~X#?KZGE`2Y zfaL@2yA5#A5Z-YtXx~V|cs~=9^R}8c}oqy-&B|6|_{W(7Vn6Yo+SElXbbgi{@ zyj4O1{M7UqIYot$2#)N9>5a?-xVv}nUQ78tz5Y4sq{3a8*}=bog2fCPFSS`L zY&tX;;i=cv*2+(Nu~fCTp5Wl1a2L#xzjB3?M>%-djuQaqVO6*RE{W`j0q9v8ed1S- z^j28TpFSxx-4TYhjg~$2WA?>$L&HXu@i~keg?={oee$2Kj{SDJ*Z_@I?THUd zdlgUKd!L(&yumU+7I)~BT#F(NGb2TAIFR!-3>Ymj&&#;=`D!+i26vvaFW0;ccM=^u^4_0i-xA_)tKLg?lT zG13qg?W<8>&TRTeDWuX-afY9Z>r8*;8ymwEH&PKZT$aA-hBOpyso=)%&STcT^ZK{o zeBsq1%t$VgCP)DKb#lv%L`aHQifct9xil zO7}7$6tPEFoL0nus!LpU;O-jXF^mNEPVm~L82S%b>g>s*oJs-Rj*N_KZ*PZ{ji`9& zxUHvQxyilB^y)(*BHxjrf>Q@A7R6LeS$V}2(LU-O_IFDX4Rxj)7 z8+VKakzV|>8C8X1!=x;ZF(}b>txWRgkOpn!|x*zrY%xCMV+Mbn;Wz-kwB*a``E|`nHEAK4mn2B zFgM7^7y~#U!3!v}#;$R0Ou4j}{pR_N?3`ERTfwf(`i zU8FP+bf&U0S%tHp2;XiytiBbKzIrd2^e_ixuQL1qfpkBQgnS1|3HCk=TfD94ZJYx$md0 zLL_ar4r&XW!L*bVOi-iHAxVvJ`m9++dwbq7k5Rk|=w6VygDw@jPWmS)HcENI^s^aH z(;S(RSrwkA2`>zH7U@oYAvY<~q<_KRapQ7^eNLY#rZN|3Qju^nMxOje2(og&Y1sd-&fItfKno9?TN0H#SGQSy5u6CGq zfSKs6AtuA!d%>lCC{m4MJY#E-P13=IK`+5)rmX;R<>d(htzckKt#$Ka-RFk=kb%(D z3_?rCCO_GWw|Es!9vutpvSBH+ z-F?AAup{npYf3ju=-Ua?oX1z61~5rvPqt#D;wWfTTiFBD5LVca@ew@0Ns@S{u%vLW_vg2LnZ|H)1 z{gYn2_+DiSl7)A>*>Zqd5dlRS;2B7J@MQ3c0#0v)k&8k^jT!&f z(Q{!5ol21z$qK*~Zhf4x#TDkC5l%;@wL(cl@5x;JQ5lrF62Bi=UBu-PKSMEjm6YV< z;(`e$#?3;!YXS4+q|BPOHUbmx@s1>eD_3f7zsC$L-a&;JR7?0*dB*+^bHs6Sd3?C7 z(xgPtz1Y(&hj-d1tDG2*uRbk|g3El5iiwpq$f6o1DyyVzO2CPd;$otL5#wER(t8e0 z@xvzDvjcSe`OVg29Xyue^~)Qn2N(>+}ehC1Yq_QLqPh(%+#L`oPh zM1!1Um-15ou13b=C%o#1+1K!cOG#T42$RodM!&VDk@;mC`X}_6?#sFG-${R?U3sHa z3nfuPqM*%c+k=m854>uP@G1;nVWCDYOI;_|G0y~j*Yky(yd6DVysasDEcE10NUvri zBYV=&ihQhfUQ9sAIySYE-HsMsUf%MXmmi4!`F#o|g`C|E3hfIQn%=7;lxf#477`ww zl#tN4FzK5NAe(lnzYd)c@)K$e%3*neEj@qgB(A`a2OC} z0H7WLGx3}+`{oz7&uJn}EiuP*R;*?B^U;6r%WvW9xbkqsHH2i1or(I6> z^8DautxX^x}rvA^viL4MD3sOXXNU^t14dfPp`%ctD1k@V8J z|BwHn_)f@;6obJ_QqV$Grb%52iCI_>wA9pl_z#U1gYqmdJ1W$m(o?Ymxt5|6Lr~g# z-o<*d z7_VZoGXjf)rb*BLf3up@k^E6uUs0P5t~EQUmPtSkgAc3=9nugC>%Yn9~A21NIIUK zt*ov1=+wOc)D1W%qGVTH4s|ZVh2uP+vZ_#v8zYPPbxe$szUn~;t--KoXdZgUX<$i9 ze1-iBMBpnD-E{NN-q`&S-)6^rk&}^et-gMVzUmJpoRp(k%PhcpD{9I-R+1aa#VirT zZIY_2Q2X?DmpvIMrx75eRCCNg5VF>veeo^k{?y4;0bVW^bFdru$amYLWTa}e1jsAD zt~uyt*g=^^iyC;miUttyCQ^K67#Efg-z&9=Y{WD;eUJx{#DK~tD$j+#%g^j<>aem ztb}_1UJcz09O))!40Mm+Ymk`;^;p`(A0ZK9)O1~*_9gGah0SR0N-oNC(dmKju<36j za%E_d4^}W#z&M$dX1Xscg1}XtOEkAg{Lj|3`PyLcLm^9Lwa+=r!#9wgQKh4|iEP|0 zOrNv8fAwg`ab7%4z0gTBiF*Te2Ppga?!L3d41G|ar!e(tT85$XXxdN7-Ek|a+k%2d zKz;zE`02U4o`Igdzy4ZdV>F9CD3#Hts;Ic_+*xwaixPaCvR$yg(#IA_^7jKR4x&>J(L;hjP zI*I7al>}@dyK@Z;K?xm$wl)kve1fq&_o(sIu*N?tSRbHv<(+fRA*%dNcP8CgW>~@Q;hQT!kY%FovYyF0EodBol3+dGNPTb+tB_tWhvu$AbV1)XmYJupEkjpG__jfvs1+mGyVukqxDf?*_D zw*hWiTVwmt{}!D)hvM$j?8W;_m47KWREiq84{F=XLEho~_=@zg-TBx3jYlm?7qFS< zbo}Xp8`sK=jdE3ZgUlG%W`XQb(Lk7CXlPuS?$z8F+YxXLCD)WXNIM1ckz+kNv9Cb{ z!bAczB!Am_!`tXe_VtuIhoJJkZG^iPt$UW4nb~&+SQhS$u77%Z&O1m(=FgsdX`pwn zUk79_f17HRl{$g=J%A#?Mvs6?H ztTUBEo_)jwhIJAQ&z_sWQ_*mPtyUdXZ@3kM@&^RxM9=~?-&wgiSp0~Fz>o@yh`5#X zcjV#uRdsCQC5c;zi4YS9Ts1s~^Ry}!d-N<-_rH@!yDNS3a={5rnqOFF4_f*d*L9Tl zts~#^%kemXcpOVhgL8(x_9+Y>;sy@|1Fpxq4H(65eZkl9bk%+$L)y{P^B0H@WpQAJ z4So1z?pz3;s~jz1$Vz?>PUY z#yT(g?c*xkAbcszHf&I!!urj6$Aio88PZ~1@#JQ7T?jf~TWnxi<*Ql&Cm(Ey*>Bo0 zL3Y6zg-*eCUf~9leH2+|fsOmxYG{OmU5vF>g%`OzNWWhD`Rj|f5M`3cAa#K@dWiz4 zaRn>9zy7#iA;wVD4y5j`2@)|*>zlz+Tj$h$T^2ia+{Zr;xJpoPa9EPOrvpL!oB!tl zZB){TCeEz7AnO54Rfow9CIL?8-Mw=hc3`t|4YYo-G++M!(-`l-jypN@S4;{K`h;-t*Ex>z#l2q z>@q{*d{zEGYfT6LpYx^vY_CD{5+u~~G46$RT2(T8wIl28Q|*I$W>T7eFwoYf)yPLh zMhdggttQ0BL%!8}6I^+<(cVdkgjwlHmU!TVow#JSL5zVlZQkHuTCi&(ZH3IhKKm#9 zcRly2b%#3?X$~I4%llhL23#>&C)`9q604;ksw1zVv?`ASogT0Xn1p{!EC)N9brNHy z$($Ofbx<&UiUmQTX=3_aO|LqZ%D8FC8d>Dw`G_@J-Rb!c83(%0l*^T+N9*Cdp*g~& z@8^y7HsAHX&O3IgS@2Ht1uDb59hL<3{lrpZVLDA~De?qte0h@+h-QeHl<)U)uW=w9pkky9~a zH=--5j~8BRv-+L=V3Ol#pKpM6k_5{8!i8BUPKbaDhKA}V76pdF(1vvqrS#O<&7fhK znEWz#2Sp|@t*l$t<3;|sn!36=|E+i}Yc=s+f+aCq&Ctm~(k?-JPfTwCQ`;DjjotHD8err#v2ezb$ zYi8RFjGp@nW)u9=6h@UeG;x_fhx4T6_I4;rMh#Ht0>%n#16yo^#@a(%BKB7?%;Vng zE;5G6YPs_m@J4fG$I$|^rKL~((>= zWLh^|(kUIekw)T@wz{md@{3Dny_`4*z0)&9ki}F88~Y;KUrqfZ=m(3>zTA245ooM9 z94}aU*KkN=;F18K&}$TI(@tRSLj=1bSC*2Ri!0ppg}6dNwXR;c#Y8Ake*KZQ%*}(C zO7GuCe3VG#fr;#E^dI`*FXHt>PYYOL4Ma(;`&zb`R(OVh7G25V@XZ!}mZ%Up`Xwy?giVj(WtVq_6Z>4v*CfZ^^uNoKtJboO&5)^OTdD#S%1yrl=9qAFvlg(Ja%2T*}(4s-Zcm%(}(IhtYB~-7F z`^1JYx@YTaOsW{wQ7a#-r)$WS=(edcKf@Ra@G0T7rCtLK8X#xr{=jul zMDsWg>0%wjj8_$O~YR?p-0+^lJuQX>;qlOmjleq)9oim?VF2j0x%I{@A> z6yFg66!NTm>C`pD0HL_dv7FuqVGw#+VscpY?}P;hd)x(J zro-*!%(}3ARJl1X@d)gkoR)z>yipQQ^!bMop?10RJ);xo?fSn$ec?!IJTw^uTzWP- zDlF`BzXo*4;D8+S(^j-|yDL@UN0eIM6*Rvn~#9VVNjTVLE}~IFy@rvsvi# zj~%OKW}Z@}6*Ra2Pb;kogyXJ(_YHgzQ&#n;aA;LnKxLZ#HuR{W^#A{$X} zlJeVyv~B)=y-|tJ3cR%RLXJlNZReYi;x_(j7iRjH$W?PV&+CG7z_xjbDedfqhU(4uI44$z z1IY>w=)m|^$NU%{08|ow6X5S%rm)^y`ImLl(gdvT5DP0foR!-9#ZfkTp z-3YZ|tlV}_Bb)e!2>=pOWDxFlY3{xjAVrkLa0SI>-)w9|^#f&ZUQsOu<5)7z&U;xf zlj?R4S+s0ZdRErydjTj!Xa4G34LU5U&Zo&<|7?Q&VPRfi-Xu7xM|>5BMloXTM(`=6IUx0C~tggSYXa6R?_nO4OZyngWZ zlP9t8u~Mt4#=c(D>nebztr-DeWzFxU)4%yjV-+I40Afdxgu)d5_h-9ika(y+8&GO&ShHpV?Vld0yTIqPP08Sv z8N@p>;@x&{_*#PD0&|I%U+*{LIXgOL<>Y*soHX|I?SAmF%0&9JgsxZ`J}} +In 2016, [Lei, Barzilay and Jaakkola](https://people.csail.mit.edu/taolei/papers/emnlp16_rationale.pdf) wrote about a new architecture for rationale extraction from NLP models. The aim was to generate a new model that could extract a "short and coherent" justification for why the model made a particular prediction. +***NB: I've simplified the mathematical notation from the original paper here in order to communicate the concepts more easily.*** -The idea is actually quite simple. Firstly, let's assume we're starting with a classification problem where we want to take document **X** and train a classifier function **F(x)** to predict label **y** based on the text in the document (e.g. X is a movie review and y is positive or negative sentiment). - +The idea is actually quite simple. Firstly, let's assume we're starting with a classification problem where we want to take document **X** and train a classifier function **F(x)** to predict label **y** based on the text in the document (e.g. imagine that X is a movie review and y is positive or negative sentiment). {{
}} -What Lei, Barzilay and Jaakola propose is that we add a new step to this process. We're going to introduce **G(X)** - a generator- which aims to generate a rationale **R** for the document. Then we're going to train our classifier **F(X)** to predict **y** not from the document representation **X** but from the rationale **R**. Our new process looks something like this: +In order to extract rationales, Lei, Barzilay and Jaakkola propose introducing **G(X)** - a generator component - to our model. The generator aims to generate - you probably guessed it - a "short and coherent" rationale **R** for the input documents. The brilliant part is that if we plug our generator into our existing architecture in the right way, we don't need to explicitely train it. + +This can be achieved by inserting the generator **G** between **X** and our original classifier **F(X)** (referred to in the paper as the encoder or **enc**) and having the classifier learn to predict **y** from the generated rationale **R**. Then we can randomly sample rationales from **G** and encourage it to generate rationales that help **F** to correctly predict **y** during back-propagation. + +**G** is constrained to encourage it to make short rationales (otherwise, it could propose **R=X** which wouldn't help us interpret anything). It is also discouraged from selecting distant groupings of words (this speaks to the 'coherent' part - we don't want a rationale that's just a random selection of words). +{{
}} + +In order for this model to work, it must be possible for the generator **G** to be able to disentangle key information from the input document **X** in such a way that it's possible for **F(R)** to predict the true output label **y** from it (e.g. "is there anything better" as captured in the figure above might be a concise way to summarise a very positive movie review). + +This architecture provides an excellent first pass at our rationale extraction task and Lei, Barzilay and Jaakkola show that it works well on the multi-aspect sentiment problem set out in their paper. However, there are a number of improvements that can be made that [we explore below](#solution-2-re-thinking-cooperative-rationalization-yu-chang-zhang-and-jaakola-2019) + +Since we are now going to focus on the improved solution, I have not provided any code for this one. However, you can find the author's [original code here](https://github.com/yala/text_nn/) and a more modern [PyTorch implementation of their work here](https://github.com/yala/text_nn/) - both are fairly well written and clear. The latter is probably more understandable these days given the former is implemented in the now-obsolete [theano](https://groups.google.com/g/theano-users/c/7Poq8BZutbY/m/rNCIfvAEAwAJ%3E%60_?pli=1) framework. + + +## Solution 2: Re-thinking Cooperative Rationalization (Yu, Chang, Zhang and Jaakola, 2019) + +In a follow up paper in 2019, [Yu, Chang, Zhang and Jaakola](http://people.csail.mit.edu/tommi/papers/YCZJ_EMNLP2019.pdf) - Jaakola being one of the authors from the previous paper - propose an improved rationale extraction approach that offers significant performance improvements over the earlier Lei et al paper. + +The new framework attempts to control for a number of shortcomings: + + 1. The earlier framework does not allow the generator **G(X)** any direct visibility onto the desired outcome **ỹ** (predicted **y**) which may lead to a poor solution. + 2. The encoder (**F(R) → ỹ**) does not have any direct visibility of the text excluded from the rationale and thus may not be performing optimally. + 3. Whilst most classification tasks have clues that are typically short phrases, some tasks require much more complicated context-dependent combinations of words and features that cannot be adequately captured by a rationale leading again leading back to point 2. + 4. Due to prior focus only on **R** (point 2.), there is no control over information *not* included in the rationale which means that the generator may condense summaries down to collections of pertinent feature words (e.g. positive or negative adjectives in our sentiment analysis example) and omit contextually relevant information (we might want to know what was good or bad). + +The proposed solution is a re-jig of the architecture so that the encoder once again predicts **y** directly from the full input document **X**, the rationale generator **G** has access to this predicted value **ỹ** (again where **ỹ** denotes the *predicted* value for **y** according to our model ). This new coupling of the encoder and generator is referred to as a single component known as the *introspective generator* in the paper. + + +{{
}} + +We also add a couple of new downstream predictor components: + * **P(y|R)** which is basically a clone of **F(R) → ỹ** that's reconfigured to accept a rationale from our new *introspective* generator and predict the probability of **y** given **R**. + * **P(y|Rc)** which is known as the *complement* predictor and tries to predict the likelihood of **y** given **R**'s complement, **Rc**, that is, all of the words that are *not* included in the rationale. + +The authors refer to this configuration as a three player co-operative game if you count the introspective generator and the two predictors as the players.