[PDF] Finding Bias in Political News and Blog Websites

View - Download Finding Bias in Political News and Blog Websites

Previous PDF

Next PDF

FindingBiasin PoliticalNe wsandBlog Websites

SonalGupta

ABSTRACT

Newsandblo gwebsites oftenhavepol iticalbias(suchasRe- publican,Democratic)intheira rticles.Automaticdetection ofth ebiaswillimpro vepersonalized feedandcategorizat ion ofne wsandblogarticles.O urprojectaims topred ictRe- publicanvs.Democrat icbiasofnew swebsitesandpolitical blogsusingthephras es(a.k.a.meme s)they quoteintheir text.Wefo rmabiparti tegraphofwebsit es andmemes quotedbythewebsi tes.Th ealgorit hmstartswithlabelsof graphusinga simplelabelpr opagation algorithm.Oural- gorithmpredictslab elsbetterthansup ervisedclassification approachandotherbaselines.

1.INTRODUCTION

Politicsorientedwebdo cumentsandthelinksbetwe enthem provideusefulinform ationaboutthei rcontentandthetop- icsthey span.However ,findingpoliticalbia sinthearticles theypublish isahardproblembe causeo fthe hugetex tcor- pusfr omthearticle sandspar sehyperlinkinformation.An examplewouldbetover ifythecommonlyhel dbelieft hat FoxNews hasRepublic anbiasint heirreporting.Automatic detectionofpoliticalbiasi nwebsit escanhelpinimproving personalizedfeedofnewsarticlesan dcategorizat ionoft he websites. Hyperlinksbetweenthewebsitesi sanusefulinformati on forpre dictingbiasofwebsitesusingthe irlinkingpat tern. However,thehyperlinkin formation issparseandwebsites donot veryof tenlinktoother websites.Forexa mple,they mightwritethequotes tatingthes ource butnotexplic- itlylinkin gtoit.Inaddition, manyof theweb sitesmay linktom orepro minentnewssou rces,ratherthanthenon- prominentones.Thismaylead tohigherconcen trationof in-linksforfamouswebsites and verysmallnumbe roflinks toli ttleknownwebsites. Weca nalsoapp roachtheprob lembyconsideringfulltex tof thearti cles,commonlyknownassenti mentanalysis.Many ofth esentimentanaly sisliteraturebuildsuponvarious NLP techniqueslikesyntacticparsin g,semanticpar sing,negative andpositive wordsdictionarylook up.Howeve r,sentiment analysisisstillavery hardp roblemtosolv e,espe ciallyfor alargecorpus. Inthi sproject, weproposeanovelway ofpredi ctingpolitical biasofwebs itesusi ngthephrasestheyquoteintheirar ti- cles.Thesequ otephrases,or memes,areanuseful sourcefor predictingawebsite'sbia s.W ebelievethatthereareso me discriminatorymemesthatare quo tedbysimilarbia sedweb- sites,andwec anexploitth equoting patternof memesto identifylabelsforweb sites.Table1sho ws memeswithtop chi-squarescore(seeChapter1 3in[6])intheda taset.We canseet hatt hesememeshavep oliticalinclinationa ndare intuitivelypredictiveofthew ebsitesbiasthatcitethem moreoften. Webuildabipartite grapho fmemesandweb- sites,whereeac hmemeandwebsite formanodeand there isaned gebe tweenameme nodeandwebsiteno dei fthe websiteusesthatmem ephraseinits article.Wes tartwith inth egraph.In thisproject,weshow thatas impleiter- ativealgorithmth atcomputesbiasofweb sitesandmemes accordingtotheirneighbors inthe bipartitegraph,work s wellforthe predictio ntask.

2.RELATED WORK

Researchinsentimentana lysish aslargelybeenforfinding opinionorsenti mentinonline reviews.Recently,people havealsost artedlookingat miningopinionintextof news articlesandblogs.Ad amican dGlance[1]st udiedthetopics ofdiscu ssioninonlineblogsand the irlinkingpat terns.They concentratedonblogsthateithersupport ed Republicanor

Democraticparty.Theyfoundthatthegra phrepresent-

inglinksb etweentheblogs clearlyhastwobigconnected componentssuggestingthatblogswith aparticularpoliti- calbias refertosimilarb iasedblogs.The yalsoana lyzed the namesassocia tedwitheachoftheblogsandfoundthatblo gs withRep ublicanorDemocraticbiasrefermoreof Repu bli- canor Democraticpeop le,respectively.Thepa perreveals interestinginformationaboutbloggin gbiasontheInternet. However,itusesalotofpr iorkn owledge,like lis tsofthe blogsandtheirl abels( Republican/Democr atic)fromexter- nalwebsi tes,whichmightturnouttobev eryhardtodo fromte xtoftheblogs.Inad dition ,suchextern alinforma - tionisnota vaila bleformos toftheblogsandnewsarticles.

Also,theyshowe dthisphenomenon withonly40blogsi n

total.

MemeChi-Squarescore

Joethep lumber95.18

youcanp utlipstic konapig90.63 yesweca nyeswec an78.50 thecha ntisdrillbabydril l70.00

IBarackHusseinObamadosolemnlyswear52.79

ouroppon entissomeonewhoseesAmericaits eemsasbeing soimperfect imperfectenough thathe'spallingaroundwithte rroristswhowouldtarget theirow ncountry51.81 notgodbl essAmericag oddamnAmerica 51.81 thefunda mentalsofoureconomyarestrong 49.87 abusesofearmarkspen din gbycongressItoldt hecongressthanksbutnothanks forthatb ridgetonowhere49.87 heis notspreadi ngtheweal tharound46.02 Table1:Memes withto pchi-Squaredis tanceintheda tasetwhen 90%websitesarelabeledwiththe irbias. Godboleetal.[2]anal ysedse ntiments ofnewsand blogson alargescale.Theyrelyonmanuallycre atedpolaritylists andused WordNetforfind ingsimilarwords.They thenas - signp olarityofeachentityint ext andaggregatesth epolar- ityusing statisticaltec hniques.Thisworkdoesnotexploit anylinkinf ormationamon gthearticles,andwhetherthe y appearedonsamewebsit eor not.Mat tThomasetal.[7] investigatedtranscriptsofU.S.C ongressionalfloordebates andproposed analgorithmthatusessen timenta nalysisof discussionstopredictwhetherornot thespee chsupportsor opposestheproposedlegislat ion.Thesyste misusefulwhen weha vewellformatt eddatawi thwholediscussionsinorder tous ethecontext todeterm inethespeaker'sinc lin ation. This,howev er,isnotthecasewithnews andbl ogarticles whereisno 'discussion' amongvariou speople,andhenceit isha rdtomakeuseo ftheco ntext.

3.APPRO ACH

Wege nerateabipartitegraphof mem esandwebsitesfrom ourdatase t.Amemenoderepres ent samemeinthed ataset andaweb siten odecorrespondstoapolitically biasedweb- site.Th ereisanedgebetween ameme nod eanda website nodeifthe blogorne wswebsi teusedthem emeint heir text.Wecaneitherh avedirected edgesf rommemeno des tow ebsitenodesorvice-ver sa.Inthisproject, webuildthe graphwith directededgesfrom memestowebsites. Given aninit ialsetoflabeledweb siten ode s,ourgoalistofin d biasofall theother websitenodes inthegr aph.Letthe thesetofed ges.LetVw?Vbeth esetofweb sitenodes andVm?Vbeth esetofmem enodes.Le tV l w ?Vwbe theiniti allabelednodes.We assignweights(w dem ,wrep) toea chnodetomeasureits biastowa rdsRe publicansand

Democrats.Then,foreachnodeno tinV

l w ,weiteratively changetheweights ateachstep dependinguponthebias of itsneighbo rs.ThealgorithmisdescribedinT able2. Thealgo rithmissimilartoHubsandAut horiti esapproach suggestedin[4]when wea ssumet hatmemesarehubsand websitesareauthorities ,andrunt healgorithmforbothla- belsindepen dentlybutnormalizethescoresateachno de. Notethata swearenormal izings coresa tnodelev el,damp- 1:t=1

2:whileNumno desinVwthatchange bias>?do

3:iftisod dthen

4:Vt=Vm

5:endif

6:iftiseven then

7:Vt=Vw-V

l w

8:endif

9:fora llvi?Vtdo

10:w dem (vi)= P v j ?Ngh(v i w dem (vj)

11:wrep(vi)=

P v j ?Ngh(v i wrep(vj) 12:w dem (vi)= w dem (v i w dem (v i )+wrep(v i

13:wrep(vi)=

wrep(v i w dem (v i )+wrep(v i

14:endfor

15:t=t+1

16:endwhile

Table2:Psuedo -codeo fouralgorithm.Roughly,at

eachiteration ,democratic(orrepublican)s coreof scoresofits nei ghbors.Wenormal izethescoresat eachnode.

PrecisionRecall

Republican8015.4

Democratic88.889.9

Table3:Precis ionand recallforbothclasseswhen

weus epredict ionofbiasfromthewebsitename eningfactoran dsplittingbiasa ccordington odedegree(see [3])doe snotmakeanyd i ff erence. Weal sotriedPage Rankalgorithmtr immedtowardsbipar- titegra ph.WerunthePageRank algori thmsepara telyfor bothlabels ,andtriedtwoapproa chesforno rmalizingscore s. Oneisto norma lizeat eachnodesothateachlabelwei ght iseq uivalenttoprobability,andth eotheris tonormalize scoresaccordingt othenodetypes(memenodesand web- sitenodes). Theybothgivesimila rresults. Asea chmemeactsas afeature,we triedselecti ngmeme squotesdbs_dbs27.pdfusesText_33

[PDF] Biathlon - Reiseservice Strobl

[PDF] Biathlon Weltcup am Holmenkollen

[PDF] Biathlon Weltcup Antholz Weltmeister schaften Biathlon Hochfilzen

[PDF] BIATSS 15-16 n°2 - UFR Lettres et sciences humaines

[PDF] Bib achats Sept 2012

[PDF] BIB DE Telekom Antrag 23.7.09

[PDF] BIB Nouveautés adultes Mai 2015.pub - Anciens Et Réunions

[PDF] Bib n°20 - Bibliomer

[PDF] Bib n°30 - Bibliomer

[PDF] Bib-12022 - United Nations Economic Commission for Africa

[PDF] Bib-12883 - France

[PDF] Bib-44942 - Anciens Et Réunions

[PDF] Bib-46666

[PDF] Bib-47992

[PDF] Bib-50997 - Fabrication