NEWSORGANIZATIONS'NEWSLINKSHARINGSTRATEGIESONTWITTER: ECONOMICTHEORYANDCOMPUTATIONALTEXTANALYSIS By ChankyungPak ADISSERTATION Submittedto MichiganStateUniversity inpartialful˝llmentoftherequirements forthedegreeof InformationandMediaDoctorofPhilosophy 2018 ABSTRACT NEWSORGANIZATIONS'NEWSLINKSHARINGSTRATEGIESONTWITTER: ECONOMICTHEORYANDCOMPUTATIONALTEXTANALYSIS By ChankyungPak Thisdissertationexploresnewsorganizations'socialmediastrategiestodisseminatetheir newsstoriesasemergent quasi-editorialdecisions .Asintraditionaleditorialdecisions,how newsorganizationssharetheirnewslinksonsocialmediadeterminesthevisibilityofcertain storiesandmakescertainaspectsofnewsstandoutorbeunobtrusive.Byillustratinghow thesepracticesonTwitterresembleanddeviatefromtraditionalones,Iarguethatsocial mediaopenupanewpathbywhichnewsorganizationsmediatenewsinformation,rela- tivelyfreefromjournalisticnormsandroutinesembeddedintheoldermediaandtraditional editorialprocesses. Toanswerafundamentalquestion,thereareasonforanewsorganizationtobe strategiconsocialIpresentedaneconomicmodelbasedoncompetitionforlimited attentionofsocialmediausers.Inthismodel,thescarcityofusers'attentioncapacityrelative tothevolumeofinformationpropagatedviasocialmediacreatescompetitionbetweennews organizations.Themodelillustratesthatonenewsorganization'sattempttocaptureusers' attentionunderminesthechanceforotherorganizationstodoso.Thus,anewsorganization should strategically decidehowmanynewslinksitisgoingtoshareconsideringhowmany others wouldshare.Asimpleempiricaltestcon˝rmsthemodel'spredictionthatnews organizationswillreducetheproportionofnewslinkstheyshareonsocialmediaasmore newsispublishedbyallorganizations. Computationaltextanalysesshedlightonthemorequalitativeaspectsofnewsdis- seminationstrategiesonsocialmedia.First,usingarecentlydevelopedmachinelearning technique,StructuralTopicModel(STM),Iinvestigatenewsorganizations'selectivenews linksharingasanewlayerofgatekeeping.Theresultindicatesthatthecommonconcern thatcommercializedmediadrivesnewstowardhumaninterestratherthannewsworthiness iscrystallizedmorevisiblyonTwitterthanonnewswebsites.Further,acomparisonof theselectivelinksharingacrossdi˙erentmediatypesshowsthattopicselectiondi˙ersde- pendingonagiventopic'spopularityonTwitterandanewsorganization'sspecialtyinthe topic.Eventhoughanewsorganizationmayconsideracertaintopictobeimportantinits editorialdecision,sothattheorganizationhasbecome specialized inthattopicthroughout itshistory,itwouldnotsharemuchaboutthetopiconTwitterbecausepopularityinthe shorttermdominateslinksharingstrategies. Ifoundthatregionalmediaconveylessnegativesentimentthroughnewsstoriesthan othertypesofnewsorganizations.Thisseemstobeassociatedwiththelesscontroversial newstopicstheyfrequentlycovercomparedtonationalandonlinemedia.However,news paraphrasingforTwitterhomogenizesemotionalframingacrossdi˙erenttypesoforganiza- tions.Inparticular,regionalmediacatchuptoothertypesbyaddingevenmorenegativity onnewsparaphrasesforTwitter.This˝ndingprovidesanothersigni˝cantindicationthat socialmediastrategiesaregovernedbydi˙erentlogicthanthatwhichgovernstraditional editorialpractices. Majorempirical˝ndingsprovideevidencethatthesocialmediastrategiesofnewsorga- nizationsarealreadyfunctioningasaseparateinformation-mediatingprocess.Iarguethat thedistinctivenessofsocialmediastrategiesas quasi-editorialdecisions raisesapractical needtopubliclymonitornewsorganizations'behaviorsonsocialmediatolearnwhether theywillprovidenewsthatisinformativeanddiverseenoughfornewsreaders'informed decisions.Theautomateddatacollectionschemesandcomputationaltextanalysistech- niquesIadoptedinthisdissertationwillinformthedesignofinfrastructureforsuchpublic monitoring. Copyrightby CHANKYUNGPAK 2018 ToJ v ACKNOWLEDGEMENTS OneimportantlessonIlearnedfromyearsforthisdissertationisthatasingleperson's hardworkisnotevenclosetoenoughforadoctoraldegree.Iwouldliketothankallwho supportedmeinbothacademicandnonacademiccommunitiesduringmyPhDwork.HereI wouldliketoacknowledgeafewofthosewhomadespeci˝ccontributionstothisdissertation. Firstofall,IwouldliketothankmyfamilyIhaveinKoreaandhereinUS.Iamgrateful forhavinganamazinglife-parter,JinaYu.Notonlydidshesupportmesharingeveryday, butalsogavemanyinsightsthathelpedmeimproveresearchdesignsandwaystopresent myworktobroaderaudience.Chattingwithherataco˙eeshopandduringcommuting timewasmorevaluablethananythingIlearnedfromacademicsources. Iwasfortunatetohaveasupportivedissertationcommittee.HowRickWash,myadvisor andco-author,nurturedmethroughoutmyPhDworkisbeyondasimpleword,support. Hetaughtmehowtothink,readandwriteasaninterdisciplinaryscholar,byservingasa rolemodel.Healsoprovidedgenerous˝nancialsupportforalongtime.SteveWildman hasbeenagreatmentorgivingmeinvaluablewisdomaboutresearchandlife.Hislectures andcountlessconversationswithhiminspiredmymodelworkinChapter3,whichwasthe inceptionofthisdissertation.Hecarefullyreadmultipleversionsofdraftsandprovidedme withverydetailedcomments.AsmycommitteememberandachairpersonofDepartment ofMediaandInformation,JohannesBauerhasbeenagreatdefenderofinterdisciplinary researchinmedia˝eld,andalwaysenthusiasticinsuggestingwaystoexpandmywork. EstherThorsonprovidedmewithmanyinsightsabouthowIcansituatemyworkinthe realmofjournalismstudies,whichgreatlyhelpedmeforgemyacademicidentity. Inadditiontomycommittee,otherfacultymemberscontributedtomydissertationin variousways.EmileeRader,asaco-PIofBehavior,Information,andTechnologyLab (BITLab)atMSUandmyco-author,gavemegeneroussupportandagreatdealofadvice aboutresearchandbeingagoodscholar.KjerstinThorson,oneofmyco-authors,helpedme vi expandmyperspectivetopoliticalcommunication.WinsonPengcarefullyreadmyproposal draftandgavemevaluablecomments. IowethankstoBITLabmembers,YumiJung,WenjuanMa,ChrisFennell,JangheeCho, andHopeChidziwisanotonameafew,whosharedbrilliantinsightsinweeklymeetingsand everydayconversationsatthelabspaceforyears.Iwasfortunatetohavetwooldfriends, YoungwooJeungandJungyoungKim,whoworkontheirownPhDworkshereinUS.They havebeengenerousenoughtocatchmyrandomphonecallsanytimeandsharetheirthoughts aboutmyunorganizedideas. IwanttosayaspecialthankyoutoTHEfriendsfromInformationandMediadoctoral programatMSU;RobinBlom,KelleyCotter,JuliaDeCook,ShaheenKanthawala,Mel MedeirosandJanineSlaker.Theyhavebeenmygreatconversationpartners,co-authors, andaboveall,enthusiasticcheerleaders.TheyarethebestluckthathappenedtomyPhD life. Finally,CollegeofCommunicationArtsandSciences˝nanciallysupportedthisdisserta- tionintheformofaDissertationCompletionFellowship. vii TABLEOFCONTENTS LISTOFTABLES ................................... x LISTOFFIGURES ................................... xi KEYTOABBREVIATIONS .............................. xv CHAPTER1INTRODUCTION1 1.1NewsOrganizations'AdaptationtoSocialMedia................1 1.2SocialMediaStrategiesonTwitterasQuasi-EditorialDecisions.......4 CHAPTER2BACKGROUNDANDRELATEDWORK8 2.1SocialMediaandNewJournalism........................8 2.1.1IndividualandOrganizationalUseofTwitterforJournalism.....8 2.1.2TwitterasaNewsDistributionPlatform................11 2.1.3EmergingNewsOutletsandSocialMedia...............12 2.2NewsDisseminationStrategiesonSocialMedia................13 2.2.1WhatMakesSocialMediaDi˙erentasaNewsDistributionPlatform?13 2.2.2CompetitionforLimitedAttentiononSocialMedia..........15 2.2.3SelectiveNewsLinkSharingasGatekeeping..............17 2.2.4SentimentandNewsFraming......................19 2.3ContributionofDissertation...........................22 CHAPTER3ACONCEPTUALMODEL:ANECONOMICMODELOFNEWS LINKSHARING25 3.1Introduction....................................25 3.2CompetitionforLimitedAttentionModel...................26 3.3NewsLinkSharingunderUsers'LimitedAttention..............28 3.3.1AModelSetting.............................28 3.3.2SymmetricNashEquilibrium......................29 3.4Results:Over-sharingHypothesis........................30 3.4.1ComparativeStatics...........................32 3.5DiscussionoftheConceptualModelandAnEmpiricalHypothesis......33 CHAPTER4AUTOMATEDDATACOLLECTIONANDCOMPUTATIONAL ANALYSISOFNEWSDISTRIBUTIONONTWITTER35 4.1AutomatedDataCollection...........................36 4.1.1DataCollectingSoftware.........................36 4.1.2DescriptionofData............................37 4.2ComputationalTextAnalysis..........................40 4.2.1TextPreprocessing............................41 4.2.2SelectiveNewsLinkSharing:StructuralTopicModel.........42 4.2.3ParaphrasingforTwitter:MeasuringSentimentwithDictionaries..46 viii CHAPTER5ASIMPLETESTOFTHESTRATEGICLINKSHARINGMODEL49 5.1Introduction....................................49 5.2Analysis......................................50 5.2.1EmpiricalModels.............................50 5.2.2Results...................................51 5.3Discussion.....................................55 CHAPTER6SELECTIVENEWLINKSHARINGONTWITTER:ASTRUC- TURALTOPICMODEL57 6.1Introduction....................................57 6.2MotivationandRelatedWorks..........................58 6.2.1SocialMediaasANewsDistributionPlatform.............59 6.2.2NewsOrganizations'OnlineNewsChoice................60 6.3ComputationalIdenti˝cationofSelectiveNewsLinkSharing.........64 6.3.1NewsTopicCategorization........................64 6.3.2OverallPatternsofSelectiveLinkSharing...............67 6.3.3Di˙erenceAcrossOrganizations.....................72 6.4Discussion.....................................82 CHAPTER7NEWSPARAPHRASINGANDSENTIMENT87 7.1Introduction:NewsParaphrasingasaStrategy................87 7.2RelatedWorks...................................89 7.2.1ParaphrasingNewsonSocialMedia...................89 7.2.2NegativeFraminginOnlineNews....................91 7.3Results.......................................94 7.3.1NegativityofNewsbyOrganizationalTypes..............94 7.3.2StrengthofExtra-NegativeFraming...................96 7.4Discussion.....................................98 CHAPTER8CONCLUSION:SOCIALMEDIASTRATEGIESASQUASI-EDITORIAL DECISIONS103 8.1NewsLinkSharingasaStrategicChoice....................104 8.2SelectiveNewsLinkSharingasGatekeeping..................107 8.3Extra-negativeFramingforParaphrases....................110 8.4MovingForward..................................112 APPENDICES115 APPENDIXAPROOFSOFLEMMASANDPROPOSITIONSINCHAP- TER3116 APPENDIXBDETAILSOFDATACOLLECTION118 APPENDIXCADDITIONALSTMGRAPHSANDFULLOUTPUT124 APPENDIXDSENTIMENTANALYSISOUTPUTSBASEDONLSDAND AFINN134 BIBLIOGRAPHY135 ix LISTOFTABLES Table4.1:Thelistofnewsorganizationsincludedinthedataset............38 Table5.1:Regressionoftheproportionofnewsstoriessharedbynewsorganizations onTwitter....................................52 Table6.1:Wordslikelytoappeargiveneachtopic....................68 Table6.2:Summaryoftheselectivenewslinksharingstrategies.Numbersinparen- thesisarenumberoftopicscategorizedasagivenstrategy.........82 Table7.1:Linearregressionofsentimentgaponnewssentimentandnewsorgani- zationstype...................................98 TableB.1:ListofTwitteraccountsincludedinthedataset...............119 x LISTOFFIGURES Figure3.1:ConditionsforSNE..............................32 Figure4.1:AcomputationalmethodtoidentifysharedandunsharednewsonTwitter.37 Figure4.2:Theaverageproportionofnewssharingofeachnewsorganizationagainst theaveragenumberofnewsstoriesdailypublishedwithasmoothed trend.Eachdotrepresentsanewsorganization...............39 Figure4.3:DailyNumberofPublishedOnlineNewsStories.Thedataweren't collectedduringtheperiodbetweenthebluedottedlines..........40 Figure4.4:Textpre-processingprocedure.........................41 Figure4.5:Agraphicalrepresentationofsharing-dependenttopicmodel........45 Figure6.1:Abargraphforestimatedoverallproportionsof30topics.Theunitis fraction.....................................69 Figure6.2:Relativepropensitytobesharedonsocialmediaof30topics.The propensityismeasuredasafractionaldeviationoftheproportionofa newstopicamongsharedstoriesfromthatamongpublishedstories.The topicsareorderedbythefrequencyonnewswebsites.Thepropensity tobesharedandfrequencyonwebsitesarenotnotablyassociated....70 Figure6.3:Newspublicationandlinksharingpatternsbynewsorganizations:Pol- itics/Presidenttopic..............................74 Figure6.4:Newspublicationandlinksharingpatternsbynewsorganizations:Sport- s/Basketballtopic...............................75 Figure6.5:Newspublicationandlinksharingpatternsbynewsorganizations:Im- migrationtopic.................................75 Figure6.6:Newspublicationandlinksharingpatternsbynewsorganizations:In- ternational/MiddleEasttopic.........................76 Figure6.7:Newspublicationandlinksharingpatternsbynewsorganizations:In- ternational/FarEasttopic...........................76 Figure6.8:Newspublicationandlinksharingpatternsbynewsorganizations:Sport- s/Footballtopic................................77 xi Figure6.9:Newspublicationandlinksharingpatternsbynewsorganizations:Crime.78 Figure6.10:Newspublicationandlinksharingpatternsbynewsorganizations:Trans- portation....................................79 Figure6.11:Newspublicationandlinksharingpatternsbynewsorganizations:Pol- itics/Congresstopic..............................80 Figure6.12:Newspublicationandlinksharingpatternsbynewsorganizations:So- cial/Gendertopic................................81 Figure6.13:Thegatekeepingmomentumgeneratedbytheselectivenewslinksharing strategies.Topicsfallingontotheredareaswillbemorevisible,and topicsintheblueareawillbelessvisibleonsocialmediacomparedto traditionaloutlets...............................85 Figure7.1:Averagesentimentscoresofnewsstoriesandtweets:LIWCdictionary..95 Figure7.2:Sentimentlevelofnewsstoriesbynewsorganizationtype.........96 Figure7.3:Sentimentleveloftweetsbynewsorganizationtype............99 Figure7.4:Additionalnegativesentimenttotweetsbynewsorganizationtype....99 FigureB.1:AnexampleofshortenednewsURLinatweet...............123 FigureC.1:Newspublicationandlinksharingpatternsbynewsorganizations:Life/In- terviewtopic..................................124 FigureC.2:Newspublicationandlinksharingpatternsbynewsorganizations:Econ- omy/TechnologyEasttopic..........................124 FigureC.3:Newspublicationandlinksharingpatternsbynewsorganizations:So- cial/IdentityCon˛icttopic..........................125 FigureC.4:Newspublicationandlinksharingpatternsbynewsorganizations:Legal/Reg- ulationtopic..................................125 FigureC.5:Newspublicationandlinksharingpatternsbynewsorganizations:Econ- omy/Financetopic...............................126 FigureC.6:Newspublicationandlinksharingpatternsbynewsorganizations:UK topic......................................126 xii FigureC.7:Newspublicationandlinksharingpatternsbynewsorganizations:Up- datetopic....................................127 FigureC.8:Newspublicationandlinksharingpatternsbynewsorganizations:Sport- s/Otherstopic.................................127 FigureC.9:Newspublicationandlinksharingpatternsbynewsorganizations:En- tertainmenttopic...............................128 FigureC.10:Newspublicationandlinksharingpatternsbynewsorganizations:Re- gionalPoliticstopic..............................128 FigureC.11:Newspublicationandlinksharingpatternsbynewsorganizations:Weather topic......................................129 FigureC.12:Newspublicationandlinksharingpatternsbynewsorganizations:Health topic......................................129 FigureC.13:Newspublicationandlinksharingpatternsbynewsorganizations:Sub- scriptiontopic.................................130 FigureC.14:Newspublicationandlinksharingpatternsbynewsorganizations:ITax topic......................................130 FigureC.15:Newspublicationandlinksharingpatternsbynewsorganizations:So- cialMediatopic................................131 FigureC.16:Newspublicationandlinksharingpatternsbynewsorganizations:Ed- ucationtopic..................................131 FigureC.17:Newspublicationandlinksharingpatternsbynewsorganizations:Sport- s/Collegetopic.................................132 FigureC.18:Newspublicationandlinksharingpatternsbynewsorganizations:Me Tootopic....................................132 FigureC.19:Newspublicationandlinksharingpatternsbynewsorganizations:French topic......................................133 FigureC.20:Newspublicationandlinksharingpatternsbynewsorganizations:Ad- vertisingtopic.................................133 FigureD.1:Averagesentimentscoresofnewsstoriesandtweets:AFINNdictionary. Zeroscoreindicatesneutralsentiment....................134 xiii FigureD.2:Averagesentimentscoresofnewsstoriesandtweets:LSDdictionary. Zeroscoreindicatesneutralsentiment....................134 xiv KEYTOABBREVIATIONS LDA LatentDirichletAllocation NAS NetworkAgendaSetting SEO SearchEngineOptimization SMO SocialMediaOptimization STM StructuralTopicModel xv CHAPTER1 INTRODUCTION 1.1NewsOrganizations'AdaptationtoSocialMedia Technologicalchangesonaplatformcreateplatformusers'neweconomicincentives. AsarecentreportaboutFacebook'salgorithmchangein2018pointedout,Facebook's longstanding,thoughtfuldesigne˙orttomakeitsplatformanewpublicsphereunexpectedly discoveredanicheforrampantmisinformationaboutpoliticalissues(Ingram,2018,January 22).Asinmanycasesinvolvedwithtechnologicaluncertainty,thisunexpecteddiscovery evenappearstohavecausedtheplatformownerstogiveupontheirvisionfortheirbusiness, establishingtheirplatformasanonlinepublicsquare(Warzel,2018,January12).However, despitetheindispensableroleofmediaorganizationsforjournalism,theacademiccommunity andindustrystilllackanunderstandingoftheeconomybehindtheorganizations'adaptation andsurvivaltosocialmedia,comparedtohumanusers'reactions.Thisdissertationaimsto illustratehowneweconomicincentivesspurredbysocialmediaasthenewnewsdistribution platformfornewsorganizationsledtoanadditionallayerofinformationmediatingprocesses, inadditiontothetraditionaleditorialdecisions,potentiallywithdi˙erentlogic. Journalismisasocio-technicalnetworkinthathumanactorstypicallyjournalistsand theiraudienceandavarietyoftechnologiesrelatedtoinformationgenerationanddissem- inationcollaboratetode˝neandre-de˝neinformationmediatingpractices(Turner,2005). Themultiplicityofthehumanactorsandtechnologiesandtheirinterdependencemakesit hardtoevaluatetheimpactofatechnologicaladvanceonjournalism.Whenanewtechnol- ogyemerges,bothinformationgeneratorsandaudiencesadapttheircultures,routinesand norms,whichinturngivesfeedbacktothetechnologicaldesign. Tounderstandthiscomplexity,scholarshavefocusedonspeci˝caspectsofjournalism asanever-changing,socio-technicalnetwork.Somefocusonhowaudiences'culturesco- 1 evolvewithtechnologies,mutatingthesocialroleofjournalism(Singer,2005;Deuze,2008; Hermida,2010).Somescholarszoominonjournalists'relationshipswithinanorganization andfocusonchangesintheirexplicitandimplicitrulesandnews-makingprocesses(Cottle, 2007;Domingo&Paterson,2011).Otherscholarspayattentiontodeterminingthepowerof new,information-curatingtechnologies,suchasrecommendationalgorithms(Pariser,2011; Bakshyetal.,2015)andsocialnetworkingfeatures,ratherthanhumanactors(Hermida etal.,2012). Inthisdiscussion,whichtendstofocusonthenovelaspectsoftransformingthemedia environment,newsorganizationsasabusinessareoftenportrayedmerelyasvictimslosing theirjurisdiction.However,newsorganizationsarestillpowerfulactorsretainingscarce resources,comparedtootheremergentactors,suchasparticipatoryusersandplatform companies.Althoughtheirprivilegeisbeingundermined,newsorganizationshavereporting infrastructure,longstandinginstitutionalandpersonalrelationshipswithnewssourceswith power,andeditorialprocesses,whichgivethemauthorityforfactualinformation.Indeed, newsreadersstillputmoreweightonreportingfromnewsorganizationsthanon individualjournalistsorparticipatoryaudiences,especiallywheninformationthereaders seekismorepersonallyrelevant(Schmierbach&Oeldorf-Hirsch,2012;Thomsonetal.,2012). Thus,hownewsorganizationsreacttonewtechnologies,suchassocialmedia,stillhas signi˝cantimplicationsforusers'newsconsumptioninthetransformingmediaenvironment. Afocusonnewsorganizations'adaptationtonewonlinetechnologiespresnetaneed toconsidertheireconomicgoals.Thetraditionallyconceivedgoalwastoprovideobjective informationthatwasnewsworthyenoughtoinformcitizens'decisions(Hermida,2010).Hier- archicaleditorialdecisionsandafactcheckingprocessbytheprofessionaldeskhavebeenan outcomeofthesocialandorganizationalevolutionofobjectivismbasedoncheaperprinting anddeliverycostthatreachbacktothelate19thcentury(Gentzkowetal.,2006).However, therehasalwaysbeenatensionbetweenjournalisticnormsandthepursuitofpro˝tability injournalism(Gentzkow&Shapiro,2010).Byo˙eringanewnewsdisseminationplatform 2 thatisrelativelyfreefromthetraditionaleditorialdecisions,socialmediaarelikelytoadd weighttothepro˝tabilitymotivation. Asmorereaders˝ndnewsfromsocialmediaratherthantraditionalsources(Newman etal.,2016;Shearer&Gottfried,2017),newsorganizationshaverecognizedtheimportance ofsocialmediastrategies(Newman,2011).Becausethedistributionofnewsisunderthe in˛uenceoftheuniquesocio-technicaltraitsofsocialmedia,newsorganizationsstriveto understandwhathappenstotheirnewsasitisdisseminatedviasocialmedia(Diakopoulos, 2017).Thesetraitsincludeinformationoverloadbypushingtechnology(Weltevredeetal., 2014),informationcascade(Lerman&Ghosh,2010),selectiveexposurebasedonsocial networks(Bakshyetal.,2015),and˝lterbubblebyrecommendationalgorithms(Pariser, 2011;Bozdag&vandenHoven,2015).However,optimizingonlinenewsdissemination fortheiraudienceastheyunderstoodusingWebanalyticstechniquesincreasinglydetaches journalismpracticesfromthetraiditionalsocialgoalofjournalism(TandocJr,2016). Theshiftinggoalofnewsorganizationsislikelytoresultinnewdecisionrules.Thispoint hasapowerfulimplicationforthediscussionabouttheimprovementofsocialmediaalgo- rithms.Researchersoftensuggestdesignimprovementsbasedonusers'currentbehaviors, whichtheyhaveobservedbysurveysorobservationaldataanalysis.Yet,inanin˛uential paperineconomics,RobertLucasJr.arguedthatapolicyevaluationbasedonthecurrent behaviorofeconomicactorsisfalliblebecausethepolicyalsochangesactors'decisionrules (LucasJr,1976).Similarly,tounderstandhowactorswillreacttoalgorithmicadvances,one needstobasethediscussiononnewincentivesandmotivations,whichwillgovernthenew decisionrule.Inthisvein,Irelateempirical˝ndingsfromthebehaviorofnewsorganizations onTwitter,oneofthemostpopularsocialmediaplatformsusedasanewsdistributionplat- form,tonewlyfoundeconomicincentives,relyingontheeconomicmodelingandpredictions frommediastudiesinthisdissertation. 3 1.2SocialMediaStrategiesonTwitterasQuasi-EditorialDecisions Althoughdi˙erentactorsnotonlynewsorganizations,butalsoreadersandnewsaggre- gatorscontributetonewsdisseminationinavarietyofformsinthenewmediaenvironment, outcomesofsuchcontributionsoftenboildowntoafewcommonfunctionsthatselect,re- duce,and˝lterinformation.Whenanactorchoosestoshareanewsstoryonaninformation platform,thenewsstorybecomesmoresalienttoothers.Further,asanactorrephrases whatshereadsfromothersources,acertainaspectofthenewsbecomesmoresalient.These outcomesareanalogoustonewsorganizations'traditionalfunctions,suchasgatekeeping, agenda-setting,andframing,exceptthatthosefunctionsarenowdistributedacrossdi˙erent typesofactorsratherthanexclusivelyoccupiedbynewsorganizations. Recognizingthesimilarityinoutcomesbetweentraditionaljournalismpracticesanddis- tributedinformationmediatingprocesses,scholarshavetriedtoextendconceptsbuilton traditionaljournalismtoavarietyofwaysforactorstointerveneinnewsdissemination. Gatekeepingtheoryisextendedasmultiplestepsbyjournalists,individuals,strategiccom- municators,andalgorithms(Wallace,2018),oranetworkofmulti-directionalprocessesof thosesteps(Barzilai-Nahon,2008).Whethertoprovideahyperlinktocertaininformationis anewprimaryformofonlinegatekeepingaccordingtothesestudies(Dimitrovaetal.,2003). Asagatekeepingroleisdistributedacrossdi˙erentactors,thepowerofagendasettingisalso distributedasanoutcome(RussellNeumanetal.,2014;Vargoetal.,2014,2017).Scholars alsofoundthatnon-journalistshavethepowerofnewsframingastheyinterpretandre- interpretnewsasitispropagatedthroughnetworkedpaths(Meraz&Papacharissi,2013).I de˝ne quasi-editorialdecisions ,asanoverarchingtermthatencompassesthesevariousforms ofnewsinformationmediatingprocesses,whicharehappeningnontradiationallyoutsidethe customaryboundaryofjournalism. Inthisdissertation,Iconsiderthenewsdisseminationstrategiesonsocialmediaofnews organizationsasakindofquasi-editorialdecisions.Newsorganizationsdecidewhatinfor- mationtoshowreadersbychoosingnewslinkstoshareonsocialmedia,asingatekeeping. 4 Theyfocusonspeci˝caspectsofanewsstoryastheyparaphrasethestoryforasocial mediapost,e.g.asinmkaingheadlines.Thus,totheextentthatnewsreadersmigrateto socialmedia,andthatthenewsorganizations'choicesonsocialmediadeviatefromtheir traditionaleditorialdecisions,thesocialmediastrategiesarelikelytoplaeanadditional layerofinformationmediatingprocessontheaudience.Thisisasigni˝cantphenomenon, whichmayimpacteverydaynewsreading,ifoneconsidersthatsocialmediahavebecome anmajornewssourceforreaders.In2017,two-thirdsofU.S.adultsidenti˝edsocialmedia astheirnewssource,andtwentypercentofthemreportedthatthey often getnewsfrom socialmedia(Shearer&Gottfried,2017).SocialmediacomeaheadofTVasanewssource forthe18to24year-oldagegroup(Newmanetal.,2016). Irelatethisadditionallayeroftheinformationmediatingprocesstotheeconomicin- centivesofnewsorganizations,whichsocialmedianewlyestablishedasanewsdistribution platform.Althoughsocialmediaprovideanewopportunitytodistributenewsstories,they alsocreateanewcompetitiveenvironmentfornewsorganizations.Inadditionatonewssto- riesdisseminatedbythetraditionalmediaorganizations,readersalsoproducenews-related contentinvaryingdegreesonsocialmedia;theycreatetheirownnewsstories,writeblog posts,sharenewslinks,andcommentonthenews.Consequently,newsorganizationsmust competewiththisvarietyofinformationtoreachreaders.Fromtheperspectiveofreaders, allocatingtheir limitedattention todesirablenewsinformationisbecominganimportant task.Thus,capturingreaders'limitedattentionposesachallengetonewsorganizations thatareformingmediastrategies(Anderson&DePalma,2013).Newsorganizationsstrive tocaptureusers'attentionwithavarietyofmethods.Recentreportsshowthattheyin- creasinglyhiredatascientistsforviralmarketingandsocialmediaoptimization(SMO)and adoptsocialmediamonitoringplatforms(Rowan,2014,January2;Diakopoulos,2017).In manyorganizations,SMOspecialistsworkinthecontextofaudiencemanagementrather thanintheeditorialprocess,althoughthedegreetowhichthisisdonevaries(Roston,2015, January22;Elizabeth,2017,November14).Thisevidencestronglysuggeststhatthebe- 5 haviorofnewsorganizationsonsocialmediaislikelytobegovernedbyeconomicincentives stemmingfromcompetitionforsocialmediausers'limitedattention. Throughoutthisdissertation,IuseTwitterasatypicalcaseofsocialmediaasanews disseminationplatform.Thereareseveralreasonsforthischoice:First,theinterestinthe organizationlevelnewsdisseminationnaturallylimitsthechoicetoFacebookandTwitter, wheremostnewsorganizationsregularlysharehyperlinkstotheirpublishedstoriesviatheir o˚cialaccounts.In2017,thesetwoservicesareamongthethreesocialmediawebsites usedmostfrequentlyasapathwaytonewsstories(Shearer&Gottfried,2017).Although YoutubeandInstagramarealsosigni˝cantnewssources,thespeci˝cpurposeofthosesites videoandimagesharingpreventsthemfrombeingusedasageneralnewsdistribution platform.Second,TwitterretainstheprimaryfeaturesItrytocaptureasthemaindrivers ofthedisseminationstrategiesonsocialmediaofnewsorganizationsasdistinguishedfrom traditionaleditorialdecisions.Thosefeaturesareinformationoverloadintensi˝edbythe datastreamandinformationsuppliers'sensitivitytouserreactionsmainlydueprimarilyto theviralnatureoftheinformationdisseminationprocessonsocialmedia.Itiseasiertotie observationsfromTwitterwiththemainfeaturesofinterestcomparedtoFacebookbecause Twitterrelieslessonalgorithmiccurationandhasbeenrelativelystableinsophisticating theiralgorithms.Inaddition,TwitterAPIprovidesamoreaccessiblewaytocollectnews organizations'behaviors. 1 Apparently,observedpatternsfromTwittercannotbedirectly generalizedtoFacebook.However,mygeneralconjecturethatinformationoverloadandvi- ralityonsocialmediaasanewsdisseminationplatformgiverisetoanadditionalinformation mediatingprocessisstillvalidforFacebook,iftheconjectureisempiricallyveri˝ed. Thepurposeofthisdissertationistoilluminatethesigni˝canceoftheemergentinfor- mationmediatingprocessesgovernedbypro˝tabilityfromsocialmediaratherthanfrom traditionaljournalisticnorms.Todoso,I˝rstformalizeaneconomictheory;socialme- diausers'limitedattentioncreatesastrategicsituationwhereinnewsorganizationsshould 1 AlthoughTwitterhascuratingfunctionalitiessuchasaseparatesectionforrecommended tweetsandpromotedtweets,amajorpartofitsinterfaceremainreal-time. 6 considerotherorganizations'choicesinChapter3.UsingastatisticalanalysisofaWeb- scrapeddatasetdescribedinChapter4,Ishowthatnewsorganizations'behaviorsapprove themodel'spredictioninChapter5.SubsequentlyinChapter6,Ianalyzehownewsorga- nizationsselectivelysharenewslinksconditionedonthenewscontent.InChapter7,using computationaltextanalysistechniques,Ishowhowtheyaddanemotionalframeasthey paraphrasenewsstoriesforsocialmediaposts.Ifurtherdiscusshowthesebehaviorsdi˙er dependingontypesofnewsorganizations,ingaugingthepotentialimpactoftheemergent newsoutlets,whichactivelyusesocialmediaastheirmajornewsdistributionplatform. Majorempirical˝ndingsrevealthatnewsorganizationsselectivelysharepopularnews topicsonTwitter,deviatingfromtheireditorialdecisionsfortheirwebsite.Theyaddextra- negativeframingonnewsasparaphrasingfortweets,whichprovidesevidencethatnews organizations'socialmediastrategiesarealreadyfunctioningasaseparateinformationme- diatingprocesses.Theseresultsimplyapracticalneedtopubliclymonitorthesocialmedia strategiesofnewsorganizationsasquasi-editorialdecisionstolearnwhetherthepro˝tabil- itymotivationfromsocialmediawillleadtotheprovisionofnewsthatisinformativeand diverseenoughforcitizens'informeddecisions,aswehaveexpectedfromtraditionaljour- nalism.Theautomateddatacollectionschemesandcomputationaltextanalysistechniques IproposedtoapplyinChapter4willinformthedesignoftheinfrastructureforsuchpublic monitoring. 7 CHAPTER2 BACKGROUNDANDRELATEDWORK Thisdissertationfocuseson organizationallevel socialmediauseofnewsorganizationsto disseminatenewsstoriesthatpotentiallyimpactsnewsreading.Tomotivatetheresearch, Ibeginbyreviewingjournalismliteratureaboutthepotentialfornewjournalismthatis emergingfromsocialmediathatfocusesprimarilyon individualjournalists' Twitteruse thatbypassesthetraditionaleditorialdecisions.Althoughtheseworksraiseafundamental questionwhethersocialmediaweakenthegatekeeperroleofjournalismbyo˙eringameans tounveilthenews-makingprocesstheytendtopresumethatinformationfromindividual journalistshasasigni˝cantimpactonreaders.Builtonanassessmentthatinformationon socialmediasu˙ersfromacredibilityissue,Isuggestredirectingthefocustothebehaviors ofnewsorganizations,andtheirnewsdistributionstrategiesviasocialmedia.Imainlyuse Twitterasacaseof socialmediaasanewsdistributionplatform wherenewsorganizations regularlysharehyperlinkstotheirpublishednewsstories.Isubsequentlyreviewnewmedia theoriesandbehavioraleconomictheoriesthathintatsocio-technicaltraits,whichmayin˛u- encenewsorganizations'newsdisseminationstrategiesonsocialmedia.Finally,Imotivate myresearchbyarguingthattheuseofsocialmediabynewsorganizationscanemergeas animportantinformationmediatingprocessthatin˛uenceswhatnewsreadersthinkabout, andhowtheythinkaboutnews. 2.1SocialMediaandNewJournalism 2.1.1IndividualandOrganizationalUseofTwitterforJournalism Journalismstudiesabouttheemergenceofnewjournalismonsocialmedia,particularlyon Twitter,tendtocon˝netheirattentiontoindividualjournalists'socialmediauseratherthan thatofnewsorganizations(Lasorsaetal.,2012).Byobservingjournalists'personalopinions, 8 thereportingprocessesofcoveringongoingevents,andhowtheygatherinformationfrom othersocialmediausers,newsreadersgettoobservehownewsisproduced.Ontheone hand,thistransparencymakesnewsmoreaccountable.Ontheotherhand,newsreaders mayreducetheirbeliefthatjournalistsandnewstheyproduceareobjective(Lawrenceetal., 2014;Mourão,2015). Theseworksrevealthefundamentaltensionbetweenmoreaccountabilityforthenews makingprocessandadherencetothejournalisticnorms,basedonmanualcontentanalyses ofarelativelysmallnumberofsampletweetsfromjournalists.Forexample,Lawrence etal.(2014)hand-coded1,946sampletweetsby430politicalreportersduringthe2012US presidentialelectionseason,andarguethatjournalismbecamesomewhattransparentinsofar asreportersoftenexpressedtheiropinionsandsharedanecdotesfromthereportingprocess withtweets.However,theauthorsconcludethatthenewsmakingprocessisstillone-way inthesensethatreportersdonotoftenseekinformationfromreadersorsharesubstantive informationaboutnewsmaking,suchasfact-checkingprocesses.Analyzing5,700tweets aboutthe˝rstpresidentialdebatein2012from430reporters,Mourão(2015)drawsasimilar conclusion;althoughjournalistsarewillingtodepartfromtheobjectivitynormbysharing theiropinions,humorandsarcasm,theytendtoreinforcetheirauthoritybybuildinga communityaroundthemselvesratherthancommunicatingwithanaudience.Lasorsaetal. (2012)summarizesthemaintensionbetweenthetraditionalnormofobjectivityandanewer tendencyofopinionsharingandhumorouscommentsasthenormalizationprocessthat wasSinger(2005)earliercharacterizedpoliticalblogsbyjournalists.AsLawrenceetal. (2014)putit,journalistsusenew a˙ordances ofTwittertobreakfromthenormwhilethe normalizingpowerofconventionsandroutinesisrecapturingthetweetsasintheone-way newsreportingofthetraditionalnewsoutlets.Indeed,Parmelee(2013)revealedjournalists feelthistensionbecausetheyareusingTwitterasanewsreportingtool. HownewsorganizationsuseTwitterontheorganizationallevelmaybedi˙erent,however, fromhowindividualjournalistsintentionallyutilizethea˙ordancesofTwitter.Empirical 9 researchshowsthattheorganizationaldecisionofAmericannewsmediatendstoconform toeconomicincentives(Gentzkow&Shapiro,2010)whereasithasbeenwidelystudied thatindividualjournalistscomplywithconventionsandroutinesestablishedinthelongrun (Lowrey,2009).Further,ingeneral,journalists'perceptionoftheroleofthemediadeviates fromthestructureofnewsorganizationswithwhichtheyarea˚liated(Zhuetal.,1997).This resultresonateswithmoregeneral˝ndingsfromorganizationstudies,inwhichnormswithin anorganizationfacilitatethedeviationofemployees'behaviorfromemployer'sexpectations, whichstemsfromchangesinincentivesinanewbusinessenvironment(Kaplan&Henderson, 2005).Althoughindividualemployeesinnewsorganizationsultimatelycomposesocialmedia posts,thereareindeeddi˙erencebetweenjournalists'postsandwhattheywritebehalf ofnewsorganzations.Forexample,Clearyetal.(2015)reportedthatnewsorganizations' o˚cialtweetsaremostlynewslinksharingwhereasjournalists'individualtweetsweremore frequentlyaboutpromotionandinteractionwithusers.Thedi˙erenceislikelytobecome clearerasnewsorganizationshiresocialmediaspecialistanddatascientiststocustomizetheir socialmediapostsratherthanrelyingonreporters(Roston,2015,January22;Elizabeth, 2017,November14). Thepotentialdisagreementbetweenindividualjournalists'Twitteruseandthatofnews organizationsraisesaquestionaboutwhethertheuseofsocialmediauseonanindividual level,howevernewitis,isthemostfundamentalchangethatpotentiallytransformsthe everydayreadingofthenews.Individualjournalists'tweetsarelikelytobesusceptibleto thecredibilityissuesfromwhichinformationonsocialmediagenerallysu˙ers(Schmierbach &Oeldorf-Hirsch,2012).Thomsonetal.(2012)reportthatpeoplebelieveo˚cialnews outletsmorethanindividualjournalists.Anetal.(2014)foundthatnewsreadersbelieve newsinformationsigni˝cantlylessonTwitterwhenthenewsistweetedbyindividualfriends indisastersituations.Thisassessmentbringsustoaneedtoexploreorganizationallevel strategiesfordisseminatingnewsstoriesviasocialmediatothoroughlyevaluatehowsocial mediaasanewinformationplatformcantransformjournalismandeverydaynewsreading. 10 2.1.2TwitterasaNewsDistributionPlatform Despitethepotentialofsocialmediatofacilitatenewtypesofjournalism,traditionalnews reportingstillplaysamajorroleinnewsdistributionandconsumptiononsocialmedia.Pew Research's2009-2010studyreportedthat,amongonlinenewsreaders,only 6% responded thatspecialinterestnewssitesaretheirfavoritenewssite,and 5% answeredbloggers'sites aretheirfavorite.Asubstantialportionstillpreferswebsitesrunbytraditionalnewsorga- nizations(Purcelletal.,2010).Collectingtrendingtopicsfromthe Tweetersphere ,Kwak etal.(2010)foundthatthemajorityoftweetswerereactionstoheadlinenews. Moreover,traditionalnewsoutletsdonotappeartotakeadvantageofthenewopportu- nitiesthatmanyobserversexpectTwittertoprovide.MediascholarsexpectedthattheWeb woulda˙ordnewsorganizationsachannelforthemutualinteractionbetweenthemedia anditsaudience(Chan-Olmsted&Park,2000),andane˙ectivepromotiontooltoattract youngeraudienceswhodonotregularlyaccesstraditionalmedia(Palser,2009;Chan-Olmsted etal.,2013).However,evidencethatnewsorganizationsareusingtheseopportunitiesis ratherweak.Forexample,Greer&Ferguson(2011)analyzedtweetsfrom488localTV stations,andfoundoutthatonly 23 : 3% of455commercialTVstationstweetforinteraction withnewsreaders.Similarly,Meyer&Tang(2015)recentlyhand-coded4,507tweetsfrom 60localnewsorganizations(TVandnewspaper),andonly 7 : 4% ofthetweetsfromlocal televisionstationsand 11 : 6% fromlocalnewspaperswereintendedforinteraction.Cleary etal.(2015)drewasimilar˝ndingintweetsfromCNNInternationalchannel.Thesestudies alsogenerallyconcludethattraditionalnewscompaniesarenotengagedinpromotionfor eithertheirwebsite,ortheorganization'sbrands(Greer&Ferguson,2011;Meyer&Tang, 2015). NewsorganizationsseemrathertouseTwitterasanothernewsdistributionplatform. Greer&Ferguson(2011)found 94 : 9% ofthecommercialTVstationstweettodisseminate theirnewsarticleswhereasonly 17 : 6% tweettopromotetheirprograms.Similarly,Meyer& Tang(2015)reportedthat 94 : 4% ofthetweetsfromTVstationsand 96 : 3% fromnewspapers 11 arefornewslinksharing.ThenewsdistributionthroughTwitterisapro˝tabletactic.Hong (2012)foundthattheTwitteruseandthenumberoftweetsbythenewsorganizations inducemoretra˚ctowardtheirnewswebsites.Therefore,newslinksharingwouldbea majorconcernwhennewscompaniescontemplatetheirsocialmediause. 2.1.3EmergingNewsOutletsandSocialMedia Asanewnewsdistributionplatform,socialmediabringfreshplayersintothemediaecosys- tembysustainingalternativenewsoutlets(Nichollsetal.,2016;Rauch,2015).Researchhas describedvariouskindsofemergingmediaenabledbytheInternet'scapabilityofmatching thenewserviceswithnichedemandinthelongtail(Anderson,2007).Earlierworksfocused on`nichenews'thattargetsnarrowedinterests,suchastechnologicalgadgetsandsubcul- tures.However,manyofthenichenewsventuresbecameapartofthetraditionalnews ecosystembyprovidingnarrowlyfocusednewsstoriestotraditionalmediaorganizations (Grueskinetal.,2011;Cook&Sirkkunen,2013). Othertypesofemergingnewsoutletsthatfollowedthenichenewsarenewscuration services,whichaggregatenewsinformationfromothernewsorganizations.Manynews curationservicesfollowedthenormalizationprocessbywhichtheybecamearegularnews organization.In2011,forexample,Weber&Monge(2011)foundthatHu˚ngtonPostwas playingaroleofhub,similartoGoogleNewsorYahoo!Newsbyapplyingamodi˝edversion ofthehub-authoritymodel(Hyperlink-InducedTopicSearch;HITS)(Kleinberg,1999)to acontentsharingnetworkbetweenonlinenewsoutlets.Sincethen,however,Hu˚ngton PosthasgrownasoneofthemostpopularnewssourcesintheUnitedStatesthatproduces itsownoriginalcontentevenaheadofCNN(Newmanetal.,2016).Fact-checkingwebsites emergedasanalternativeformofinformationsources,butaccordingtoLowrey(2017)'s assessment,nowtheirorganizationalbehaviorspartiallyconformtotraditionaljournalistic conventionsandroutines. Thesuccessoftheemergingnewswebsitesisdueprimarilytotheuseofsocialmedia. 12 Theseemergingnewsorganizationsarevigorouslyusingsocialmediaastheirmajordistribu- tionplatform(Newmanetal.,2016),andunlikemanylegacymedia,theyexplicitlyaimat theviralityofstoriestogainpopularityamongsocialmediausers.AccordingtoFarisetal. (2017)'sanalysis,tweetsfromonlineonlynewswebsitessuchasHu˚ngtonPost,Breitbart andTheHillweresharedmorefrequentlybyusersthanthosefrommostofthelegacymedia duringthe2016USpresidentialelectionseason.Theonlylegacymediawhosetweetswere sharedmorebysmallmarginsweretheNewYorkTimesandCNN.Thisisdisproportional tothesigni˝cantlysmallernumberoffollowerstheemergingnewsorganizationsretained; Hu˚ngtonPost,TheHillandBreitbarthad11.3M,2.93Mand896Kfollowersrespectively asofJanuary,2018whereastheNewYorkTimes,CNNandFoxhad40.8M,39.1Mand 16.9M,respectively.Thisobservationindicatesthattheemergingnewsorganizations'news distributionthroughsocialmediahasbeensuccessful. Further,recentworksalsoshowthattheirnewslinksdisseminatedviasocialmedia playanagenda-settingroleinpoliticaldiscourses.Benkleretal.(2017)foundthatthe vastmajorityofpoliticalagendaduringthe2016USpresidentialelectionemergedfromthe right-wingonlinemedia,suchasBreibartmuchaheadoflegacymediasuchasFoxNews. Additionally,Starbird(2017)foundthattheemergingpartisannewswebsitesareplaying aleadingroleonTwitterinformingconspiratorialalternativepoliticalnarrativesabout sensitivepoliticalissues,suchasmass-shootings.Becausesomeoftheemergingmediaare explicitlydenyingthetraditionaljournalisticconventionsandnorms,itismoreimportant toknowhowtheemergingmediaareutilizingsocialmediatogaugetheimpactofsocial mediaonjournalism. 2.2NewsDisseminationStrategiesonSocialMedia 2.2.1WhatMakesSocialMediaDi˙erentasaNewsDistributionPlatform? Ifsocialmediaareanewsdistributionplatformfornewsorganizations,doesconsumingnews onsocialmediaa˙ecttheexperienceofnewsreaders?Becausesocialmediahavemanysocio- 13 technicalcomponents,theirimpactonnewsconsumptioncanbedeliveredthroughmultiple paths.However,theroleofnewsorganizations'socialmediastrategieshasnotdrawnmuch attentionasoneofthepaths.Instead,asocialnetwork(Anetal.,2011;Bakshyetal.,2015; Hermidaetal.,2012)andrecommendationalgorithms(Bakshyetal.,2015;Pariser,2011) onsocialmediahavebeenthemainfociofliteraturethatinvestigatestheimpactofsocial mediaonnewsconsumption.The˝ndingsdi˙er,orevencontradicteachother,depending onthecontextinwhicheachstudywasconducted. Resultsabouttheimpactofasocialnetworkonnewsconsumptiondi˙erdependingon whetherthe weaktie orthe strongtie (Granovetter,1973)dominates.Anetal.(2011) andHermidaetal.(2012)commonlyreportthatsocialnetworksformedonanonlinesocial mediaplatformincreasenewsdiversity.Theirresultsconformtotheeaktieargument; usersareexposedtoinformationthattheywouldnotknowaboutwithoutsocialmediafrom theirfriends'recommendationwhodonotsharepreferences.However,aFacebookdata scienceteamrecentlyreportedcontradictoryevidenceanalyzingusers'newslinkselection onFacebook(Bakshyetal.,2015).Theyfoundthattheproportionofhardnewsstorieswith theoppositepoliticalviewavailabletousersdecreasesdramaticallyastheyfriendwithother users.Forexample,45%and40%ofhardnewsavailabletoconservativeandliberalusers arewiththeoppositepoliticalviewiftheusersarerandomlyexposedtonewsonFacebook. However,only24%and35%ofhardnewssharedbyusers'friendsarei.e. withtheoppositepoliticalviewforliberalusersandconservativeusersrespectively.This isbecauseFacebookuserstendtofrienduserswhosharesimilarpoliticalviews.Inother words,thestrongtiesdominateinthiscase. Afewrecenthuman-computerinteraction(HCI)studieshavefocusedonalgorithmsthat automaticallycurateinformationforsocialmediausers.Previousresearchhasdiscussed thepowerofthealgorithmstogovernaccesstoinformation(Kitchin,2016).Biasedaccess toinformationonlineduetothealgorithmiccurationmaycauseuserstoencounteronly viewpointsthatreinforcetheirexistingattitudes,whichpreventsminorityopinionsfrom 14 beingexpressedanddeliberated(Bozdag&vandenHoven,2015).Bakshyetal.(2015)also measuredhowmuchthealgorithmreducesconsumptionoftheoppositeviewonFacebook. Theyfoundthatconservativeuserssee5%lesscross-cuttingcontentintheirNewsFeeds comparedwithwhatfriendsshare,whereasliberalssee8%less. 2.2.2CompetitionforLimitedAttentiononSocialMedia Thisdissertationintroducesadditionalstructuralfactorinnewsconsumptionthroughsocial mediathatmaya˙ectnewsreaders'experiencetothediscussion.Thatfactorisnews organizations'newsdisseminationstrategiestoattractsocialusers'limitedattention.Asan economicentity,anewsorganizationislikelytoadaptitsbehaviortothesocio-technical characteristicsofsocialmediaasanewsdistributionplatformtoattractmorenewsreaders. Itssocialmediastrategies,asaresult,maya˙ectcontentthatnewsreadersendupconsuming throughsocialmediainturn. Newsorganizationshaveadoptedstrategiestocuttheclutteronotheronlinenewsplat- forms,suchasasearchengine.Searchenginesareapreviouslydominantonlinenewsplat- formforreadersoverloadedbythemixtureofindividualnewsarticlesfromdiversi˝edsources andallotherkindsofinformationavailableonline(Hermida,2010).AsManovich(2012) characterizesthem,searchengineusersqueryrelevantinformationwithkeywordstheyex- pecttobeassociatedwiththecontenttheylookfortoextractitfromdatabases.Knowing this,newsorganizationshavewidelyusednewsdistributiontacticstomaximizeachance thattheircontentisvisibleinasearchresult,whichisoftencalledsearchengineoptimization (SEO)(Dick,2011;Newman,2011;Giomelakis&Veglis,2015).Leadingnewscompanies suchastheNewYorkTimesandBBChiredSEOspecialistsandtrainedtheirjournaliststo publishnewscontent˝ttedtotheWeb(Giomelakis&Veglis,2015).TheseSEOtechniques includecustomizingnewstitlesforsearch,HTMLtagging,tailoringmetadataforimages, URLoptimizing,etc.Theyarepartlytechnicalchoicestoconformtoasearchalgorithm, butalsostrategicchoicesthatre˛ectanorganization'sonlineperformanceandcompetitors. 15 Usinganethnographicapproachandinterviews,Dick(2011)describesthatSEOspecialists innewsorganizationsengageinaclickstreamanalysisandacompetitoranalysisto˝ndthe bestpracticesofSEO.HealsofoundthatSEOinvolvessigni˝cantcostintrainingjournalists andfacilitatingcommunicationbetweenSEOspecialistsandthedesk. Asmoreaudiencesaremovingtowardsocialmediatoreadnews,newsorganizationsare likelytoadapttodi˙erentsocio-technicaltraitsofsocialmedia,inadditiontosearchengines. Manovich(2012)contrastsitwithsearchengines,assertingthatinformationonsocialmedia isorganizedasa datastream ratherthanadatabase.Onaninformationplatformthata datastreamrules,auserexperiencesacontinuous˛owofinformation.Weltevredeetal. (2014)relatesthestreamonsocialmediawiththetraditionaldistinctionbetween`push' technologyand`pull'technologyontheWeb.Asinformationiscontinuously`pushed'toa user,newinformationimmediatelyreplacesoldinformationinadatastream.Thus,auser islikelytomissrelevantinformationsweptbyirrelevantinformationpushedbyastream. Thiscancreatesevereinformationoverloadcomparedtoasearch-basedplatformwherea userpullsrelevantinformation. Newsorganizationsneedtoconsidertheinformationoverloadthattheirownandother newsorganizations'newssharingcangenerateonsocialmedia.Anderson&DePalma(2013) mathematicallymodelacompetitionamongmultiplecompaniestosequentiallypropagate informationtoconsumers.Havinganadvertisingmarketintheirmind,theauthorsassumed ascenariowhereaconsumerbecomesawareofthecompany'sproductonlywhenshechooses toseeinformationfromthecompany.AndersonandDePalmacallthissituation competition forlimitedattention becauseothercompaniesalsotrytoreachtheconsumerswhocannot processalltheinformation.Oneconclusionoftheirmodelisthatcompaniesarebetter o˙byrefrainingfromexcessiveinformationpropagationbecauseitwillmitigateconsumers' informationoverload. However,whethertopropagateinformationmaynotbetheonlystrategicconcernin sharingnewsonsocialmedia.AndersonandDePalma'smodelonlyre˛ectsalimitedaspect 16 ofinformationqualityinthatallconsumersinthemodelareassumedtoagreeonhow goodorbadtheinformationqualityis,andthissimpli˝cationallows˝rmstodecideonly howmanyinformationgoodstheyaregoingtopropagatetoconsumersdependingonthe informationqualitytheyproduceandcompetitionwithother˝rms.However,thequalityof newsinformation,inreality,maynotbesomethingeveryonecanagreeupon;anewsstory typicallycontainsdi˙erenttopics,hasaspeci˝cviewpointonacertainissueofinterest,and maybefactualoropinionloadedwhiletheremuststillbeadimensionofbeingbetteror worse.Therefore,newsorganizations'strategicconcernsarelikelytobericherthanonly theirlinksharingdecisions.Forexample,newscompaniesmaywanttosummarizeand paraphraseanewsstory(e.g.within280charactersonTwitter)sharedonsocialmediaso thatthecomposedsocialmediaposte˚cientlyrepresentstheoriginalcontent,butissuccinct enoughnottooverwhelmsocialmediausers'attention. Althoughnewsorganizations'newsdisseminationstrategiesonsocialmediahavenot beenwidelyanalyzed,recentmarketingliteraturerevealedthat,ingeneral,companiesare strategicaboutusingsocialmediaastheirbrandingplatform.Interviewing14marketing managersresponsibleforthesocialmediaactivityoftheircompany,Tsimonis&Dimitriadis (2014)foundthatthepresenceofcompetitorsonsocialmediaisoneofthemainmotiva- tionsofrunningtheirownsocialmediaaccount.Withasimilarapproach,Parveenetal. (2015)revealedthatsocialmediateamsrunbymanycompanieslearninformationabout competitorsviasocialmedia,andstrivetodevelopaninnovativesocialmediausedi˙erent fromtheircompetitors'. 2.2.3SelectiveNewsLinkSharingasGatekeeping Builtonthestrategicconcernsnewsorganizationsconsiderforsocialmedia,animmediate questionis,hnewsdotheychoosetoshareonsocialIfnewsorganizations selectacertainsubsetofnewsstoriestosharetheirlinksonsocialmedia,andiftheselective newslinksharingisastrategicchoice,itwillmanifestitselfascertainpatternsinsharednews 17 contentandunsharednewscontent.Then,thepatternwille˙ectivelyactasanadditional layerofgatekeepingimposedonsocialmediausers.Linksharingasgatekeepingisnota newphenomenoninjournalism.Dimitrovaetal.(2003)earlierfoundthatAmericanmajor newspapercompaniesusehyperlinksasagatekeepingtooltocontrolthenumberofexternal linkstoaninformationsource. Theselectivenewslinksharingonsocialmediacanbeunderstoodasaninter-media agenda-settingviagatekeepingbetweennewswebsitesandsocialmedia.Theselectivechoice ofinformationbynewsorganizationshasbeenexplainedundertheframeworkofagenda- settingbygatekeeping(Shaw&McCombs,1977;Shoemaker&Vos,2009).Inotherwords, newsorganizationsin˛uencepublicopinionnotonlybycarryingthroughtheiropinion, butalsobychoosingissuesforthepublictothinkabout(McCombs,2014).Thistheory expandedtowardinter-mediaagenda-settingtheory(Meraz,2011;Vargoetal.,2017)and networkagenda-settingtheory(NAS)(Vargoetal.,2014),inwhichsequentialchoicesof informationfrommediabyotheractorssuchasothernewsorganizationsorsocialmedia users,alsoexerttheagenda-settingpowerbyselectinginformation.Empiricalstudiesbased ontheseframeworksoftentrackchangesofissuesacrossdi˙erentmediaoutletsormedia platformstoknowwhichmediawerein˛uencers,orthosewhosetthenewsagendaofothers (Meraz,2011;Vargoetal.,2017). Whereastheseframeworksfocusedontheactors'intentionalchoices,suchasnewsorgani- zations'editorialdecisions,anotherdevelopmentoftheagenda-settingtheory,acknowledges thattechnologicalfactorscanimpactthesalientagendaaswell.Extendingtheagenda- settingtheorythatfocusesonalinearpropagationofinformationfromnewsorganizations toapublic,agenda-buildingtheoryconsidersreciprocalin˛uencesbetweenmultipleentities (Kiousis&Ragas,2015).Becausedi˙erententities,suchasnewsreaders,politicalparties, andindustrycomeintoplaytointeractinthisframework,howtechnologicalinnovations changehoweachkindofentitycommunicateislikelytoalsoin˛uencehowapublicagenda isbuiltinanunpredictedway.Thisassessmentresonateswiththerecentdiscussiontoview 18 atechnicalchoiceofpersonalizedinformationcurationalgorithmsasapublicmatterthat requiresmonitoring(Krolletal.,2016;Sandvigetal.,2016).Inasimilarmanner,this dissertationexploresthepossibilitythatusingsocialmediaasanewsdistributionplatform functionsasasocio-technologicalfactorthatimposesanotherlayerofinformationrestriction ontheformationofpublicopinion. 2.2.4SentimentandNewsFraming McCombsetal.(1997)expandedtheagenda-settingtheorytothesecondleveltheselection of attributes ofanewsobject.Theyarguethatnewsreportingin˛uences how tothink aboutareportedissuebyfocusingonorneglectingspeci˝caspectsofanewsobject,such asapoliticianorapolicy.Thesecond-levelagenda-settinghasacloserelationshipwith framing(McCombsetal.,1997)althoughtheformertendstofocusmoreonthechoiceof informationwhereasthelatterfocusesmoreontheattributionofresponsibilityforansocial issue(Scheufele,2000). Intheirseminalstudy,McCombsetal.(1997)acknowledgedtheimportanceofsentiment inshapinghowtothinkaboutnews.Byconnectingsocialissueswithpositive/negative/ neutralsentiment,newsreportinginformsreaders'judgmentaboutreal-worldissues(McCo- mas&Shanahan,1999).Researchhasparticularlyfocusedonnegativityinnewsframing. Trussler&Soroka(2014)hypothesizedthatnewsreaderspsychologicallyhaveapreference fornegativenewsframingbecausenegativityisfurtherfromhumans'innatelypositiveex- pectationandis,thus,consideredasignalformoreusefulinformation(Kahneman,1979). Indeed,Trussler&Soroka(2014)foundthatpoliticallymotivatednewsreaderspreferneg- ativenewsframingunderanexperimentcondition.Theysuggestthisresultasevidence forthedemandsideexplanationoftheprevalenceofnegativityinpoliticalreporting.Built onthisassessment,researchershaveappliedautomatedmeasurementsofsentimentinnews stories.Forexample,usinghumancodersandmultiplelexicons,Young&Soroka(2012) foundthatnewsframingisconsistentlybiasedtowardthenegative,particularlyforcrime 19 andforeignpolicytopics. Fortraditionalnewspaperreporting,journalistsmustcomposeheadlinesthatsummarize newsstoriestheyhavewritten.However,theheadlineisnotonlyameresummaryofnews content,butalsothemain`hook'ofanewsstorytoreaders(Molek-Kozakowska,2013). Inthesamevein,Bell(1991)listedthemainpurposesofheadlinesasa)summarizing,b) framingandc)attracting.Inotherwords,howthecontentissummarizedinheadlines in˛uencesnewsreaders'choicetoactuallyreadthestory.Thesimilaritybetweenheadlines andnewsparaphrasingforsocialmediaimpliesthatjournalists'choicesmadeforheadlines thathavebeenobservedinpreviousstudiescaninformusaboutpotentialchoicesmadefor socialmediaposts. Onlinenewsconsumptionislikelytomakethereaderattractingroleofnewsparaphrasing moreimportant,duetothecompetitionbetweenindividualnewsstoriesonline.Intraditional newsconsumptionwherenewsstoriesareconsumedasabundle(e.g.anewspaperorTV newsprogram),headlinessignalnewscontenttocompetewithotherarticleswithinthesame bundle.Thus,theheadlineshavetobetowrittentoshowastory'srelevancetoareader (Dor,2003).However,intheonlineenvironment,newsreaderstypicallynavigateacross di˙erentnewssources.Inthiscircumstance,newsorganizationshaveastrongincentiveto signaltheirstories'appealtoreaders.Moreoveronsocialmedia,thissignalingincentiveis likelytobeevenlargerbecauseusers'attentioniseasilydistracted.Indeed,researchershave detectedaproliferationofsensationalexpressionsinonlinenewsheadlinesfor click-bait .For example,Chakrabortyetal.(2016)andPotthastetal.(2016)showedsentimentpolarity (valence)canbeusedasapredictorofanautomatedclick-baitdetectionalgorithm. Negativenewssummaryhasbeenshowntohaveastrongimpactonnewsperception. Manipulatingheadlinesandleads,Priceetal.(1997)foundthatexperimentsubjects'opinion andemotionalvalenceaboutfundingtopublicuniversitieschangedependingonhowthe sameinformationisframed(con˛ict/humaninterest/consequence).Further,theyfound thatthechangedopinionimpactsthedecisionmakingaboutthepolicy.Withasimilar 20 experimentdesign,Zillmannetal.(2004)showedthatexperimentsubjectsspendmore timereadingnewsstorieswhentheyareframedascon˛ictorvictims'agonyratherthan misfortuneoreconomicloss.Thesestudiesimplythatnewsparaphrasing,suchasasocial mediapost,hasapowerof extra-framing ontopofframesinthemainbodyofnewsstories. Allinall,theliteratureIhavediscussedsofarimpliesthatnewsorganizationsare likelytobeengagedinstrategicbehaviorsonsocialmediathroughwhichtheyregularly disseminatetheirnewsstories,andthatsuchbehaviorscanactasanadditionallayerof informationmediatingprocess tonewsreaders.Thesestrategiescanin˛uencesocialmedia users'newsreadingasthoughtheyareeditorialdecisions.Justaseditorialdecisionshave in˛uencednewsreadersabouttoandwto(Bennett&Iyengar,2008) innewsproduction,attoandwtodistributethroughsocialmediamay alsoin˛uencenewsconsumptiontotheextentthattheyhavedistinguishedpatterns.The goalofthisresearchistoexploresuchstrategicbehaviors,particularlyonTwitter,and evaluatewhethernewsorganizations'newsdisseminationstrategiesonsocialmeidadeserve publicdiscussionasaninformationmediatingprocessthatpotentiallyhasasigni˝cantsocial impact. Althouhgafewpreviousworkshavecategorizedtypesofnewstweetsbynewsorgani- zations(Newman,2011),theirattentionhasbeenlimitedtoasocialmediapostassuch ratherthanitsrelationshipwithoriginalnewstext,duetothelargevolumeofnewstexts andthecostofthehand-codingapproach.However,thefocusontweetslimitsthescopeof aninquiryonlyto shared newsratherthan unshared news.Further,dataexclusivelyfrom tweetscontainsonlyinformationabout outcomes ofnewsorganizations'socialmediastrate- giesratherthan input .Thus,thepreviousapproachpreventsresearchersfromobservingthe decision-makingofnewsorganizationswithregardtotheirdistributionofpublishednews storiesonTwitter.I,instead,adoptalarge-scalenewsscrapingandacomputationaltext analysistoenablethecomparisonsbetweenshared and unsharednews,andbetweentext thatistailoredforTwitterandoriginalnews.Further,thecomputationalmethodsIadopt 21 provideaninfrastructureforpotentialpublicmonitoringofthesocialmediastrategiesfor futurediscussion. 2.3ContributionofDissertation Thisdissertationexploreshownewsorganizations'socialmediauseasanewsdistribu- tionplatformimposesanadditionallayeroftheinformationmediatingprocess,focusingon Twitter.Newsorganizationsstillplayacentralroleinaudience'saccesstoinformationon publicissues(Purcelletal.,2010),andtheyusemajorsocialmedia,suchasTwitter,mainly asanewsdistributionplatform(Greer&Ferguson,2011;Meyer&Tang,2015).Newsorga- nizations'behaviorsareunderthee˙ectofthesocio-technicaltraitsofsocialmediaasanews distributionplatform,andtheycanin˛uencewhatnewsreadersthinkandhowtheythink aboutpubicissues(Bennett&Iyengar,2008).Speci˝cally,newsorganizationsarelikelyto adoptstrategiestocopewithsocialmediausers'limitedattention,whichmaybiastheir newsreading.Thus,asmorenewsreadersusesocialmediaastheirsourceofinformation, whatnewsorganizationsshareandhowtheysharenewsonsocialmediamayalsoin˛uence theinformationtowhichnewsreadersareexposed. Newsorganizationspotentiallybothhelpandharmsocietybyselectivelychoosingnews- worthyinformation.Theyhelpnewsreadersunderstandtheworldbyreducingtheamount ofinformationtoprocess,buttheyalsomaybiasreaders'perceptionofpublicissuesdue toselectivelychoseninformationinfavoroftheirviewpoints.Thus,hownewsorganiza- tionschooseinformationhasbeencentraltotheevaluationofthesocialimpactofnews. Theinformationrestrictionbynewsorganizationsgenerallytakesplaceintwoways.First, newsorganizationsdecidetheselectionandsalienceofpublicissues(gatekeeping/agenda- setting).Second,theydecidetheselectionandsalienceofparticularaspectsofanissue (framing)(Scheufele,1999). Myresearchquestionaboutthequantityofnewsdistributedthroughsocialmedia(Chap- ter3and5)andahypothesisaboutselectivenewslinksharingconditionedonnewscontent 22 (Chapter6)aretolearnhowsocialmediaasanewsdistributionplatformin˛uencesthe selectionandsalienceofpublicissues.Asecondhypothesisaboutparaphrasingnewsfora tweet(Chapter7)aimsathoworganizations'newsdistributionviasocialmediain˛uences theselectionandsalienceofparticularaspectsofanissue.Further,socialmediaprovidesop- portunitiesforadiversi˝ednewssupply.Theysustainmanyonlineonlynewsorganizations astheirmajornewsdistributionplatform(Newmanetal.,2016).Thus,thisdissertation furtheraimstoillustratetheimpactofsocialmediaonhowblendsofnewsorganizations informreadersbyinclusivelyanalyzingdi˙erenttypesofnewsorganizations,e.g.magazines, nationalmedia,regionalmedia,andonlinemedia,acrossthedi˙erentchapters. Empirical˝ndingsfromTwittercannotbegeneralizeddirectlytootherplatforms.How- ever,thegeneralconclusionthatnewsorganizations'strategiesre˛ectingsocio-technical traitsofsocialmediaasanewsdistributionplatformcreatesanadditionalinformationme- diatingprocess,isexpectedtobevalidforFacebook,thesinglelargestsocialmediaplatform alsousedasaprominentnewsdistributionplatform.AsinTwitter,newsorganizationshave similarneedstoaddressusers'limitedattentiontoreachthem,andsensitivelyreacttousers' engagementinandsharingofnewsinformation.GiventhatFacebookrefersthelargestuser tra˚ctonewswebsites,organizationsareevenmorelikelytocustomizetheirnewsdissem- inationforFacebookaswell.However,thespeci˝cpatternsofnewsorganizations'choices maydi˙erbetweenTwitterandFacebookbecause,comparedtoTwitter,Facebookprovides functionsforamoreclosedsocialnetworkactivelycuratedbyalgorithms. Methodologically,Iusecomputationalapproachesfordataachievementandcontentanal- ysis,whichenablesansweringresearchquestionsthataredi˚culttoanswerwithtraditional approaches.Becauseitishardtoachieveanunbiasedsampleofconnectionsbetweennews storiesandsocialmediaposts,thehandcodingapproachthatneedsalimitednumberof sampledatacouldnotallowforquestionsaboutsuchconnections.Thecomputationaldata scrapingapproachandlarge-scaletextanalysistechniquesIsuggestinthisdissertationpro- videtoolstoanalyzewhatnewsorganizationsdidtotheirnewsstoriesforsocialmedia. 23 Thesetoolscanbefurtherappliedtoadditionalresearchquestionsaboutrelationshipsbe- tweenthenewsandsocialmedia.Becausemydatacollectionandanalysisprocessare fullyautomated,theyprovidetoolstoconstantlymonitornewsorganizations'socialmedia strategiesandsharetheinformationwiththepublic. 24 CHAPTER3 ACONCEPTUALMODEL:ANECONOMICMODELOFNEWSLINK SHARING 3.1Introduction Inthischapter,Ipresentasimplemodelinwhichnewsorganizationschoosethenumber ofnewslinksthattheyshareonsocialmedia.Althoughhighlystylized,thepresentedmodel capturesafundamentalaspectofthecompetitionforthelimitedattentiononsocialmedia: to competitively supplyaninformationgoodtheamountofwhichexceedsauser'scognitive capacity. Severalobservationsindicatethatthisquantitativeaspectofthecompetitionisimpor- tant.Aboveall,thereissimplytoomuchinformationonsocialmediaforausertoprocess. Anaveragesocialmediauserisexposeddailyto54,000words,theequivalentofanaver- agenovel(Bennett,2013,July13).Asaresult,newsorganizationsarefailingtoattract socialmediausers'attentiontotheirbrands.AccordingtoarecentReutersInstitutereport (Newmanetal.,2016),only52%ofUSnewsreadersreportedthattheynoticedthenews brandwhentheyaccesstoanewsarticleviasocialmedia.Eventhoughauserchoosesto readanewsarticle,shewouldn'tstaywithinthenewsorganization'swebsite.Newsreaders whovisitnewsorganizationwebsitesviasocialmediatendtostayforoneminuteandforty secondsonlywhereasnewsreaderswhovisitnewswebsitesdirectlystayonaverageforfour minutesandthirtysixseconds(Mitchelletal.,2014).Inotherwords,atypicalnewsreader onsocialmediaclicksonanewslink,readsthenewsarticle,andleavesthenewsorganiza- tion'swebsite.Inthiscase,thegoalofthenewsorganizations'socialmediastrategyislikely tomaximizethenumberofclicksonnewslinkstheyshare. Afundamentalintuitionofthequantitativecompetitionforlimitedattentionisthatthe amountofotherorganizations'sharednewslinksonsocialmediacana˙ecttheprobability 25 thatmynewslinksattractsocialmediausers'attention.Thus,mydecisionwilldependon otherorganizations'behavior.Forexample,ifothernewsorganizationssharesomanynews linksthatuserswillnotpayattentiontoadditionalinformation,Iwillnotsharenewslinks ifthesharingincurscost.Thissituationinwhichmydecisiondependsonotheragents' decisionsisoftencalleda strategic situation,orsimplya game .Thus,thesimplemodelI suggestistoshowsocialmediausers'limitedattentioncanmakenewslinksharingonsocial mediaastrategicconcern. Themodelshowsthatthequantitativecompetitionforthelimitedattentionresultsin manynewslinkexcessivelysharingcomparedtonewsorganizations'collectivelyoptimal level,asanequilbrium.Sinceanindividualcompanycannotfalloutofthisequilibrium, itislikelytocarefullydevise qualitative meanstoattractusers'attentiontonewslinkit publishedasinSEOpractice.Inthefollowingchapters,Iwillintroduceotherqualitative socialmediastrategiesinwhichnewsorganizationmaybeengaged.Thisconjectureabout thequalitativestrategywillleadtoempiricalresearchquestionsIwillexplorewithalarge- scaletextanalysisofnewstextsandsocialmediapostsbynewsorganizations. 3.2CompetitionforLimitedAttentionModel InspiredbyAnderson&DePalma(2013)'sadvertisingmarketmodel,thissimplemodel focusesontheshort-termnewslinksharingdecisionafterthenewsproductionis˝nished. Themainfeatureofthismodelisthattherearemultiple˝rmsthatrepeatedlypropagate informationtouserswhohaveacognitivelimittotheamountofinformationtheycan process.Anderson&DePalma(2012)callthissituation competitionforlimitedattention . Users'limitedattentionorcognitivecapacitycreatesatrade-o˙forthe˝rmsbecause,onthe onehand,theywanttopropagatemoreinformationtoincreasetheprobabilityofreaching users'attention.Butontheotherhand,sharingtoomuchinformationwilloverwhelmusers' attention,whichinturnmakesusersunabletoprocessanymoreinformation. InAndersonandDePalma'smodel,˝rmsrepeatedlysendout one advertisingmessage, 26 onlyoneexposureofwhichhasanimpactonconsumer.Inotherwords,theirmodelcon- sidersasituationwhereconsumersbecomeawareoftheadvertisedproductupontheone exposure,whichgeneratesa˝xedamountofexpectedpro˝t;anyextraexposuredoesnot generatepro˝tfor˝rms.Thus,itresemblestheclassicaleadvliterature ineconomicswhereintheroleofadvertisingishavingconsumers˝ndoutaboutagood (Stigler,1961).Unlikeersuasiveadvmodels,AndersonanddePalmaassumethat repetitiveexposuretoanadvertisingmessagedoesnotincreasethelikelihoodofpurchasing theadvertisedproductasamodelsimpli˝cation.However,a˝xedamountofmarginalcost ofadvertisingoccurseverytimea˝rmsendsoutanadvertisingmessage.Inthissetting, ˝rmshaveanincentivetorepeatedlysendoutthesamemessagebecauseonlygreaterfre- quencyofthemessageincreasestheprobabilitythatthemessageisexposedtoaconsumer, attheexpenseofgreatercost. UnlikeAndersonandDePalma'sadvertisingmodel,Ipostulateasituationwherea ˝rmsendsouthyperlinksto multiple onlinenewsstoriescontainingadvertisements,and newsorganizationsgain˝xedpro˝tsfromtheonlineadvertisementwhenauserclicksthe hyperlink.However,anorganizationsendsouteachlinkonlyonce,whichpreventsalinkfrom incurringthepro˝tmorethanonce.ThissettingcontrastswithAndersonandDePalma's modelwhereonemessageisrepeatedlysentout.Thedi˙erentsettingsbetweenthetwo modelscomefromthedi˙erentrealitiesthattheytrytocapture,e.g.productadvertising vs.newsdissemination.Despitethedi˙erence,themainintuitionthatdrivestheoutcomes inthetwomodelsissimilar;one˝rm'spropagationofinformationgeneratesanegative externalitytocompetitorsinthatiteatsupusers'limitedattention.However,asel˝sh˝rm wouldnotcareaboutthisnegativeexternality,whichleadstoanoveruseofusers'limited attention.Inotherwords,asanoutcomeofthecompetition,newsorganizationsexcessively sharemanynewslinksonsocialmediacomparedtothecollectivelyoptimallevelofnews organizations. 27 3.3NewsLinkSharingunderUsers'LimitedAttention 3.3.1AModelSetting Themodelmakesafewsimplifyingassumptionsaboutidenticalnewsorganizationsand newsstoriesastheirproducts.Tobemorespeci˝c,(a)newsorganizationspublishthe samenumberofnewsstories,and(b)allthepublishednewsstorieshavethesameappeal tohomogeneousnewsreaders.(c)Also,themarginalcostofsharinganadditionalnews linkisthesameforallnewsstoriesregardlessofwhichnewsorganizationcreatedthestory. Thus,anarbitrarynewsorganization, i ,doesnothaveotherstrategicchoicesotherthan thenumberofsharedlinks, x i .Inotherwords,Iassumethattherevenuefromsocialmedia users'clicksonnewslinksisindependentofnewsorganizations'othersourcesofrevenuesuch aso˜inesubscriptionorvisitingawebpageviaasearchengine.Thisassumptionmeansthat thepresentmodelfocusesontheshorttermdecisionofsharingnewsstoriesthatarealready published.Thisassumptionmaynotbetoostrongbecause,formosttraditionalmedia ˝rms,onlinerevenuesarestilldwarfedbyo˜inerevenues.Notonlydoesthissimpli˝cation allowustoconcentrateonthecoreofthestrategicinteractionbetweennewsorganizations byeliminatingmanycomplicationsinreality,butalsoithelpsdeveloparesearchquestion aboutthedegreeofnewslinksharingonsocialmedia.Iassumethata˝xedcost, c ,incurs peranewslinkcreated.Thiscostspeci˝cationcanbeinterpretedasthee˙orttooptimize eachnewslinktodistributiononsocialmedia,asinSEOpractices(Dick,2011)suchasnews titleoptimizationorsummarizingthenewscontentforasocialmediapost. Eachof K newsorganizationspublishes M newsstories,whoselinksmaybesharedon socialmedia.Becausethevalueofnewsstoriesisassumedtobethesameandindependentof othernewsstories'value,thetotal K M newsstoriesareidenticaluptothenewsreader's choice.Thereare N newsreaderswhosubscribetotheonlysocialmediumthatexists.Each readercanreadonly T newsstories,whichrepresentstheir limitedattention ;ifthereare morenewsavailableonsocialmediathananewsreadercanread,i.e. KM>T ,shewill 28 readonlyasubsetoftheavailablenews.Newsreadersareexposedtonewsstoriessharedon socialmediainarandomorderandreadthenewsstoriesintheordertheyencounterthem withprobability 1 untiltheyreachthecognitivelimit T .Thus,eachreaderwillrandomly read T storiesamongthe KM availableinthecaseof KM>T . Inthissetting,arisk-neutralnewscompany i 'sexpectedpro˝tmaximizationproblemis asfollows: max 0 x i M E [ ˇ i ( x )]= Pr f astoryisread g Np c x i = min ˆ T P j x j ; 1 ˙ Np c x i ; (3.1) where x i denotesthenumberofnewsarticlessharedbynewsorganization i ,and p denotesan advertisingrevenuefromasingleclickonanewsarticle,whichIassumetobethesameforall clicks.Ifthereare more sharednewsarticlesthanasocialmediausercanread( KM>T ), theprobabilitythatanewsarticleischosenisthenumberofnewsarticleareadercanprocess ( T )overthenumberofnewsarticlessharedbyall˝rmsinthemarket( P j x j )duetothe equalprobabilityassumption.Ifthereare fewer newslinkssharedonsocialmediathan readerscanprocess( KM T ),allnewsarticleswillbereadwithprobabilityone.Thus, min f T= P j x j ; 1 g istheprobabilityforausertoreadanarticleinbothcases.Giventhis probability, min f T= P j x j ; 1 g Np istheexpectedrevenuefromasinglenewsstorythrough socialmedia.Thusthewholeexpression, E [ ˇ i ( x )] ,isnewsorganization i 'stotalexpected pro˝tfromsharing x i newsarticlesonsocialmediagivenother˝rmschoices. 3.3.2SymmetricNashEquilibrium IfocusonthesymmetricNashEquilibrium(SNE)asthemostplausibleequilibriumbecause parametersexceptthechoicevariable,i.e.thelinksharingdecision, x i arethesameacross all˝rms.LetusassumethereexistssuchachoiceinthesymmetricNashequilibrium, x . Then,therecanbethreepossiblecases. 29 1. Nolinksharing: x =0 2. Fulllinksharing: x = M 3. Partiallinksharing: 0 c . Inaddition,Iwillonlyconsideracasewhere Np c K K 1 forsimplicity.Thisassumption isnotveryrestrictive;rearrangingtherighthandsideoftheinequality,weget Np c (1+ 1 = ( K 1)) .For K =2 ,therightsideofthisinequalityis 2 c .Thus,theassumptionrequires onlythattherebeenoughrevenueinthemarkettosupporttwomessages,whichcouldbe one˝rmwithtwomessagesortwo˝rmswithonemessageeach. Letusde˝nethe newssupply-attentionratio , ˚ MK T .Thenthefollowingproposition showsthatpartialsharingissustainedasaSNEwhencompetitionforlimitedattentionis relativelystrong. Proposition1. Np c K 1 K isacutpointsuchthat,if ˚ isnogreaterthanthat,full-sharing isthesymmetricequilibrium,andif ˚ isgreaterthanthat,partialsharingcharacterizesthe equilibrium.Underthepartialsharingequilibrium,thenumberofsharednewsstoriesper ˝rmisdeterminedas x = Np c T ( K 1) K 2 . Proof. Seeappendix (SketchofProof)Iftherearefewernewsstoriesthananewsreadercanprocessinthe market,marginalpro˝tis˝xedat Np c .Inthiscase,aslongasthistermispositive, 30 fullsharingisSNE.Iftherearemorenewsstoriesthanreaders'cognitivecapacity,marginal pro˝tisdecreasinginthenumberoflinks.Then,partialsharingequilibriumpotentially existsatthelevelwheremarginalpro˝tiszero.Howeverinsomerange,marginalpro˝tmay staypositiveatthemaximumnumberofnewsstoriesnewsorganizationscanshare, M .In thatcase,fullsharingstillcharacterizesSNE.Otherwise,partialsharingissustainedasa SNE. Insum,boththepartialsharingandthefullsharingequilibriumarepossibledepending onthenewssupply-attentionratiorelativetopro˝tability.Therelativestrengthbetween howmanytotalnewsstoriesarepublishedcomparedtousers'limitedattentionandthe pro˝tabilityofsharingdeterminestheproportionofnewsstoriessharedonsocialmedia ( x M = Np c ( K 1) K T MK = Np c ( K 1) K 1 ˚ )among M publishednewsstories.InChapter5,this outcomewillbeusedtotestwhetheranewsorganizationindeedcaresaboutothers'quan- titativechoices. Notethat T MK ( K 1) K Np ismarginalnegativeexternalitywhenallnewsorganizations shareallstories.Whenallorganizationsshare M stories,oneorganization'smarginalpro˝t is T MK Np c T MK 2 Np .The˝rsttwotermsarethedirecte˙ectfromsharingthemarginal link,andthethirdtermistheindirectmarginallossfromloweredprobabilityforanews linktoclickedduetoonemorelinkonthesocialmedia.Theindirecte˙ectincurstoall othernewsorganizationsaswellasthegivenorganization.Becausethereare ( K 1) other organizations,thismarginalexternalityaddsupto T MK ( K 1) K Np .Then,theconditionfor fullsharingequilibriuminProposition1isequivalenttotheconditionwherethismarginal negativeexternalityatfullsharingislargerthanthemarginalcostofsharinganadditional link, c .Inotherwords,fulllinksharingisSNE. NowtheconditionforthreepossibleSNEcanbecharacterizedinatwo-dimentionalspace generatedbythemarginalpro˝t-costratio, Np=c ,andthenumberofstoryattentionratio, M=T asinFigure3.1. 31 Np c M T Slope: K 1 K 2 0 1 NoSharing Partial Full Figure3.1:ConditionsforSNE. 3.4.1ComparativeStatics Proposition2. (Excessivesharinghypothesis)Firmssharenofewerstoriesthanthe˝rst bestunderSNE. Proof. Toreplicatethe˝rstbest(FB)choiceofnewsorganizations,supposeallnewsorga- nizationscollude.Then,thecolludingentity'sproblemissimply: max 0 x KM min ˆ T x ; 1 ˙ Np c x; (3.2) where x denotesthetotalnumberofsharednewslinks.Itwillsharenostoriesifthemarginal costissohighthatanysharingisnotpro˝table( Np c ). Thus,nowsuppose Np>c .If KM T ,thecolluding˝rmwillshareall KM storiesas inthecompetitioncasebecausethepro˝tonthemarginallinksharedis Np c inthiscase, whichisassumedtobepositive. Suppose KM>T .Firstconsidermarginalpro˝tat x>T .Becausethepro˝tmaxi- mizationproblembecomes max T 0 .Thus,sharing T storiesisoptimal. Insum,each˝rmundercollusionsimplyshares M if MK T ,and T=K if MK>T aslongas Np>c .ThisisnogreaterthantheSNEchoicefor M if MK T ,and M or 32 Np c T K ( K 1) K if MK>T .Thelatterisequalorgreaterthantheformerundertheassumption, Np c ( K 1) K ,wemade. Letuscalleach˝rm'schoiceunderthecollusion x FB .Then,Proposition2issummarized as x x FB . 3.5DiscussionoftheConceptualModelandAnEmpiricalHypoth- esis Theexcessivesharingresultfollowsfromthefactthatthecompetitionforlimitedat- tentioncorrespondstothe thetragedyofthecommons (Hardin,1968;Ostrom,2015)where eacheconomicagent'schoiceexerts negativeexternality tootheragents'decision.If˝rm i sharesonemorenewslink,itputsmoreloadonnewsreaders'limitedattentionwhen therearesu˚cientlymanynewslinksonsocialmedia.Becauseoftheadditionallinkon socialmedia,newsreaders'distractiona˙ectsnotonlytheprobabilitythat˝rm i 'slinkis clicked,butalsotheprobabilitythatallother˝rms'linksareclicked.Ontheotherhand, ˝rm i getsbene˝tfromsharingmorelinksbecausethe˝rmincreasesitsshareofattention relativetoothernewsorganizations.Thenegativeexternalityfromsharingmorelinkcauses a game situationinwhicheachcompanyneedstobe strategic againsteachother.Because onecompany'spro˝tdependsonothercompanies'strategiesaswellasitsownstrategy,the companyneedstotakeaccountofdecisionsthatother˝rmswouldmadetooptimizetheir strategyforpro˝t.Theequilibriumnumberofsharedlinks, x istheresultofthisstrategic concern. Thestandardresultofthetragedyofthecommonsgamestatesthatusers'attentionasa publicresource (Anderson&DePalma,2009)willbeoverlyoccupiedrelativetotheoptimal levelbecauseanindividualcompanyinterestedonlyinitsownpro˝twouldnotcareabout thenegativeexternalitytoothercompanies.TheexcessivelinksharingIprovedcorresponds tothisgeneralconclusionofthetragedyofthecommonsgame.Further,thisresultimplies thatcompetitionreinforcesinformationoverloadonsocialmedia. 33 Anoutcomeofthemodelinwhichtheamountofsharednewslinksisastrategicdecision constitutesatestablehypothesis.Ifthelinksharingwerenotastrategicconcern,thepro- portionofnewsstoriessharedbyanewsorganizationwillnotbeassociatedwiththenumber ofotherorganizations'newsstories.Ifnewsorganizationsdonotperceivecompetitionas ameaningfulconcern,the perceived probabilitythatasharednewsstoryisclickedwillin- creaseinthenumberofsharednewslinksindependentofothernewsorganizations'decisions. Then,thepro˝tmaximizationproblemshouldbe max 0 x i M [min f T x i ; 1 g Np c ] x i .Bythe samelogicastheproofofProposition2,theoptimalchoiceshouldbeallnewsstoriesor themaximumnumberthattheirreaderscanprocess,i.e. M or T .Thus,theproportionof sharedstorieswillbe 1 or T=M ,whichdoesnotdependonothernewsorganizations'choices. Ontheotherhand,ifnewslinksharingisstrategicasinthepreviousmodel,theproportion is x =M = Np c ( K 1) K 1 ˚ = Np c ( K 1) K T MK ,whichdependsonthetotalnumberofstoriesall organizationspublished.Inotherwords,themoreothernewsorganizationspublish,the feweroftheirownnewsstoriesanewsorganizationshares.Thishypothesiswillbetested usingaWebscrapeddatasetdescribedinChapter4. Althoughthemodelinthischapterillustratesonestrategicaspectoftheamountofnews linksharing,oneshouldnotethatthemodelisbasedonstylizedassumptions.Aboveall, thismodelignoreshorizontalandverticalnewsqualities.Thus,thesimplyquantitative reactionstothenumberofotherorganizations'sharednewslinksisunlikelytobecloseto thefulldescriptionofcompetitionamongnewsorganizationsonsocialmedia.Forexample, howisa˝rmsselectionofnewstopicsin˛uencedbythetopicso˙eredbyitscompetitors? Arethereotherstrategictoolsthatnewsorganizationscanuseonsocialmediathansimply whethertoshareadecision?InChapter6and7,usingtextanalysistechniques,Iwilltake onsomeofthesequestionsaboutthequalitativeaspectsofnewslinksharingonTwitter. 34 CHAPTER4 AUTOMATEDDATACOLLECTIONANDCOMPUTATIONALANALYSIS OFNEWSDISTRIBUTIONONTWITTER Intheearlierchapters,Iproposedpotentialsocialmediastrategiesstemmingfromcompeti- tionamongnewsorganizationsonsocialmedia.However,theempiricalexaminationofthe existenceofsuchstrategiesposesmethodologicalchallenges,whicharemainlyinvolvedwith scaleofthedata.Forexample,totestwhetheranewsorganizationreactstothenumber ofnewsstoriespublishedbyotherorganizations,Ineedtohaveameasureforthetotal numberofnewsstories.Totesttheselectivenewslinksharing,Ineedtohaveasampleof publishednewsstorieswhichdonotmisrepresentthedistributionofnewscontentsbecause suchasamplingbiaswillchangetheestimatedproportionofsharednewscontent.However, itisimpossibletodistinguishnewscontentsbeforereadingthenews.Itraisesachallenge becausearesearcherhastoanalyzethedata˝rsttogetagoodsample. Eventhoughautomateddatacollectionallowsforbypassingsuchasamplingchallengeby collectingallthedataavailableonline,thesheeramountdataalsopreventsthetraditional hand-codingapproach.Forexample,thedatasetusedforthisdissertationincludesmore than400Knewsstoriesandmorethan150Ktweetswithembeddednewslinks.Further, myanalysesinvolvecomparisonsofthepairsofnewsstoriesandtweets,aswellasseparate storiesandtweetsontheirown.Thus,toanswerthequestionsposedinthisdissertation callsforcomputationaltextanalysismethodsaswellasautomateddatacollection. Inthischapter,IwillexplainthetechniquesIadoptedforthedatacollectionandthe analyses.Inadditiontotheutilityforthisstudy,theWebscrapingsoftwarebuiltonanopen- sourcedatabase, MediaCloud canbegenerallyusedforfurtherpublicmonitoringofnew organizations'behaviors.Moregenerally,Iapplytheanalyticaltoolsfrommachinelearning andsentimentanalysis˝eldsascomputationaltranslationsofmedia-relatedsocialscienti˝c researchquestions.Thesetechniquesarevaluablebecausetheyhelpexpandthefrontier 35 ofanswerablequestionsinsocialscience,andextendaone-shotobservationtolong-term monitoringofsocialphenomenaundertheevolutionaryprocess. 4.1AutomatedDataCollection 4.1.1DataCollectingSoftware ThedatawereautomaticallycollectedbyasystemofsoftwareIdevelopedforthisstudy withPythonlanguage.Thesystemhasthreemaincomponentstocollectnewsstories andtweetswithembeddedhyperlinkstothem,whichwererequiredforthesubsequent analyses.First,the newsscraper collectsnewsstoriespublishedoneachnewsorganization's website.Thissoftwareisbuilton MediaCloud ,anopensourcesdatabasedevelopedand maintainedbytheHarvardBerkmanKleinCenter 1 . MediaCloud monitorstheRSSfeedsof approximately60,000newsorganizationsaroundtheworldeverythirtyminutes,andstores meta-informationofthenewsstoriesfedviaRSS.Tomyknowledge,thisisthelargestset ofdataregardingnewsstoriespublishedonline.BecauseMediaClouddoesnotprovide contentofnewsstories,thenewsscraperopensalltheURLsinthelistfromMediaCloud anddownloadsthenewstexts. Second,the Twitterscraper collectstweetsfromeachnewsorganization'so˚cialTwitter accounts.Whennewsorganizationshavemorethanoneaccount,Twitterscraperdownloaded datafromalltheaccounts. 2 TheTwitterscrapercollectstextintweets,embeddedURLs,and thetweeteddateviaTwitterRESTAPI.Third,toidentifywhichnewsstorieswereshared onTwitterbynewsorganizations,the URLmatcher matchesURLsfromnewsdatasetwith URLsfromtheTwitterdataset.TheURLmatcherconductsmultiplestepsofURLpre- processingbecauseURLsareoftenshortenedtosavespacewithinatweet,ormodi˝edto havethemcontainusefulinformation.Thus,thepre-processingincludesURLunshortening andquerytermparsing.Additionally,itparsesoutnewstitlesornewsIDstomaximize 1 Availableathttps://mediacloud.org 2 SeeAppendixBtoseethelistoftheincludedaccounts. 36 Figure4.1:AcomputationalmethodtoidentifysharedandunsharednewsonTwitter. performanceofthematchingtask.ThisdatacollectionsystemisvisualizedinFigure4.1. Anewsstorywasrecordedassharedwhenitwasshared atleast once. 4.1.2DescriptionofData NewsorganizationsinthedatasetwerecollectedfromAlexaonlinetra˚cranking.Iretrieved thetop200websitesfromtheAlexanewsmediasection.Iaugmentedthislistusingalist ofmediathatwerein˛uentialduringthe2016USPresidentialElectionaccordingtothe recentHarvardBerkmanKleinCenterreport(Farisetal.,2017).Then,IchoseonlyUS newswebsitesexcepttheBBC,theGuardian,ReutersandAlJazeera,whichpotentially haveasubstantialimpactonnewsconsumptionintheUnitedStates.Among116news organizationsintheoriginallist,IremovednewsorganizationswhoseRSSfeedisnotwell monitoredbyMediaCloud(28organizations),andwhoseonlinenewsstorieswerenotparsed bytheNewsScraper(4organizations).Thisprocessresultedin84newsorganizations.These newsorganizationsarecategorizedforconvenienceasinTable4.1. Thedatasetincludesregionalmediaandonline-onlynewswebsites.Iincludedtheonline- 37 Table4.1:Thelistofnewsorganizationsincludedinthedataset. NationalRegionalMagazineOnline ABCNewsTheArizonaRepublicEconomistAlterNet AlJazeeraBaltimoreSunForbesBipartisanReport BBCBostonHeraldNewYorkMagazineBreitbart BloombergChicagoSuntimeTheAtlanticBuzzFeedNews CBSNewsDallasNewsTheChristianScienceMonitorCommonDreams CNBCDenverPostTheNewRepublicConservativeTribune FoxNewsDesMoinesRegisterTimeFactCheck MSNBCDetroitNewsFiveThirtyEight NBCNewsHoustonChronicleFreeBeacon NewYorkTimesMiamiHeraldsGatewayPundit NPRNewYorkPostHu˚ngtonPost PBSNewshourNewsdayIBTimes ReutersNJ.comInfowars TheGuardianOrangeCountyRegisterInquisitr USATodayOregonianMediaMatters WallStreetJournalOrlandoSentinelPolitico WashingtonPostSeattleTimesPoliticusUSA Sun-Sentinel TheIndianaStarSlate TheKansasCityStarTalkingPointsMemo TheMercuryNewsTheConversation ThePhiladelphiaInquirerTheDailyBeast ThePressDemocratTheDailyCaller TheSacramentoBeeTheHill TheStrangerTheIntercept TheTennesseanTheRoot WashingtonTimestownhall.com VanityFair Vice Vox WashingtonExaminer WesternJournalism Westword onlynewswebsitestoseeifthetraditionalnewsorganizationsandtheonline-onlyoutlets adoptdi˙erentsocialmediastrategies.ThedatawerecollectedbetweenNovember20, 2017/11/20andJanuary1,2018,andJanuary9,2018andJanuary28,2018forsixty-three days.Thedatacollectionincludesagapweekduringwhichdatawerenotcollectedbecause thesoftwarewasmistakenlyturnedo˙.Thisdatacollectionprocessresultedin435,355 newsstoriespublishedby84newsorganizations.Amongthem,152,555newstorieswere shared,resultingina 35 : 04% sharingproportion.Thenumberofpublishednewsstoriesby newsorganizationsvariedfrom 21 (InfoWars)to 30 ; 341 (WashingtonTimes).Thisvariation 38 Figure4.2:Theaverageproportionofnewssharingofeachnewsorganizationagainstthe averagenumberofnewsstoriesdailypublishedwithasmoothedtrend.Eachdotrepresents anewsorganization. stemsfromadi˙erenceoftheirmediatypes,fromwebsitesthatstemfromopinionblogs, suchasInfowarsortownhall.com,tomagazinessuchasEconomistorTime.Ontheother hand,dailynewsoutletswithblogsasapartoftheirwebsites,suchasWashingtonPost, orregionalpaperspartneringwithothernewsoutlets,suchastheSeattleTimespublish farmorenewsstoriesontheirwebsites.Theproportionofsharednewsstoriesalsovaried from 7 : 122% (theSeattleTimes)to 95 : 16% (Economist).Thisvariationdependsonhow manynewsstoriestheypublishontheirown.MagazinessuchasEconomistorTimetend toshareamajorityofwhattheypublishwhereasoutletswithpartneredcontentandblog postsbyjournalists,suchasWashingtonPostortheSeattleTimes,tendtoshareless.This relationshipisvisualizedinFigure4.2. Thenon-parametricallysmoothedtrendcalculatedusingtheLocalPolynomialRegression Fitting(Clevelandetal.,1992)appearstoshowanegativetrend(thebluelineinFigure 4.2).However,thereisapossibilitythattheoutliersthatpublishmorethan200newsstories aredrivingthenegativetrend.Theseoutliersturnoutnottoimpactthegeneral˝ndings.I 39 Figure4.3:DailyNumberofPublishedOnlineNewsStories.Thedataweren'tcollected duringtheperiodbetweenthebluedottedlines. willdiscusstherobustnessresultinChapter5. Figure4.3showsthedailypatternofthenumberofpublishedonlinenewsstoriesby the84newsorganizations.Itshowsaclear˛uctuationwherethenumberofpublished newsstoriesisloweronweekendsandholidays.Thatis,inadditiontothelowernumberof publishednewsitemsonSaturdayandSunday,the˝gurealsoshowsthatnewsorganizations publishmuchlessnewsonThanksgiving(November23,2017),Christmas,andNewYear's. Moreover,headingtowardtheendof2017,thenumberofnewsstoriestookadownward trend,whichseemsnaturalfortheholidayseason,andasharprecoveryafterthenewyear began.Thegraphicalobservationshowsthatthedailypatternisconsistentwithweeklyand seasonalworkpatterns,andprovidesausefulfacevalidityofthedata. 4.2ComputationalTextAnalysis Toanalyzenewsorganizations'potentialselectivenewslinksharingandtheextra-framing imposedintheprocessofnewsparaphrasing,Iadoptcomputationaltextanalysistechniques 40 Removecapitalization,punctuation Discardstopwords Stemming Otherreduction Discardwordorder(bagofwords) Figure4.4:Textpre-processingprocedure. frommachinelearningandsentimentanalysis.Toapplythesemethodstotextdata,news textsshouldbeprocessedintoamachine-readableform.Thissectionwilldiscussthepre- processing˝rstandthenexplainthetextanalysistechniques. 4.2.1TextPreprocessing Therichnessofwrittenandspokennaturallanguagemakeastatisticalanalysisdi˚cult. Tosimplifytextdatatoaformthatisstatisticallyanalyzableand,yetnottothepoint atwhichinterestingvariationinthedatavanishes,researchersoftenapplystandardpre- processingprocedurestogivencorpora(Manningetal.,2008).Figure4.4summarizesthe pre-processingproceduresIapplytonewsstoriesandsocialmediapostsforthisstudy. Theremovalofcapitalization,punctuationandstopwordsridstextsoflessinformative data,andreducesthedimensionalityofdata.Ina˝nalformoftextdata,eachwordorphrase standsforasinglevariable.Thiscancausedimensionalityproblemsforastatisticalanalysis 41 becausetheremustbemanywords(i.e.variablesinatextanalysis)relativetothenumber ofdocuments(i.e.observations).Removinguninformativewordsmitigatesthisproblemat theexpenseoflosingminimalvariationinthedata.Stemmingreferstoaprocedurethat achievesthebaseformofaword.Thisprocedurealsoreducesthedimensionalityofthe data,butremovesonlythefunctionalpartsofaword.Thisimpliesanassumptionthatthe functionsofwords,suchasclassorgrammaticalnumber,arenotimportantforthepurpose ofanalysis.IapplystandardPorterstemmingalgorithm(Porter,1980). Discardingwordordersisasimpli˝cationthatassumesawordorderisuninformative. Althoughthisisevidentlyafalseassumption,discardingwordordersisastandardpractice becauseincorporatingwordordersproduceslittleenhancementinmodelperformancerelative tothecomputationalcostthisprocesssaves(Manningetal.,2008).Thesimpli˝cation resultsinthe bag-of-words modelwherebyanauthorofatextpicksmultiplewordsfrom anidenticalandindependentworddistributiontocomposethetext.The˝naloutcomeof thepre-processingproceduresisa word-documentmatrix ;acolumnequalswords,andarow equalsdocuments.Entriesoftheword-documentmatrixstandforthecountofphrasesused inadocument.Aresearcherusesthismatrixjustlikeadatamatrixisusedinstandard statisticalmethods. 4.2.2SelectiveNewsLinkSharing:StructuralTopicModel ToseeifnewsorganizationstendtosharemorepopulartopicsonTwitter,Iuseamachine learningalgorithm,StructuralTopicModel(Robertsetal.,2016)thatappliestonatural languagei.e.newstextinthisdissertation.The˛ipsideoftheresearchquestion, typeofnewsislikelytobesharedonsocialistypeofnewsis not likelyto beThismeansthat,toknowaboutnewsorganizations'selectivelinksharingon newscontent,aresearcherneedstocomparesharednewsandunsharednews.Compared withmanypreviousworks,whichfocusonlyonsocialmediaposts,therequirementforthe comparisonbetweentwosetsofnewsstoriesdramaticallyincreasesthevolumeofdatathat 42 needstobeanalyzed.First,aresearcherneedstolookatnewsstoriesthatcontainmuch longertextthansocialmediaposts.Also,oneneedstolookatbothsharedandunshared newsstories. Manuallyreadingallthenewsarticlestodecidewhattypesofnewsaremorelikelyto besharedisanimplausibletask.Toillustratethis,supposethatanewsorganizationdaily publishesonehundrednewsstories,andaresearcherwantstoanalyzenewsstoriespublished bytwentynewsorganizationsforaweek(saysixdaysexcludingSunday).Theresearcher needsthentoanalyze12,000newsarticles,anddiscoverdi˙erentpatternsbetweenshared newsandunsharednews. StructuralTopicModel(STM)isanalgorithmthatpermitsautomatedanalysisforsuch aninquiry.Thisstatisticalalgorithmsimultaneouslydiscoverslatenttopicsinnewsstories andidenti˝estopicsandphrasesstatisticallyassociatedwiththelinksharingdecision.In otherwords,tocircumventtheimplausiblemanualapproach,byapplyingSTM,Ireplace theresearchquestionwithacomputationaltaskthatidenti˝estopicscorrelatedwiththe linksharingdecision. STMisarecentextensionoftopicmodels(Bleietal.,2003;Blei&La˙erty,2006;Blei, 2012)designedtoidentifylatenttopicsfromtheco-occurrenceofwords.STMextendstopic modelsbyembeddingregressionmodelswithinthem.Thatis,STMsimultaneously˝nd topicsandregressesthetopicstootherobservablevariables.Thisallowsfortestingthe associationbetweenlatenttopicsandthelinksharingprobability,whichwillbethemain interestofChapter6. Intheclassoftopicmodels,atopicisoperationalizedasthelatentdistributionofwords (orphrasesdependingontextmodels).To˝ndthelatentdistributions,atopicmodel assumesatextgeneratingmodel,inwhichanauthorchoosesatopicfromamultinomial distribution,andwords(orphrases)fromanothermultinomialdistributiongiventhetopic. Basedonthistextgeneratingmodel,themodelisestimatedbyBayesianmethods.Depend- ingonhowtospecifythepriorbeliefabouttheprobabilitythatatopicisdrawn,topicscan 43 haveonlynegativecovariancebetweenoneanother(Dirichletdistribution;LatentDirichlet Allocation),orcanhavebothpositiveandnegativecovariance(logisticnormaldistribution; CorrelatedTopicModel). STMisbuiltonthecorrelatedtopicmodelRobertsetal.(2013,2014);Lucasetal. (2015).Thewordprobabilitygiventoatopicisassumedtobedeterminedbydocument covariates(beingsharedinmycase)inthepriorbelief.Thisspeci˝cationallowsforbotha identi˝cationoftopicsandtheregressionto˝ndtopicsrelatedtothesharingdecisionasa covariate. IuseSTMwithmodi˝cationstothemodelinregardtothenewslinksharingcontext. Inparticular,Ispecifyanewstextgeneratingprocessthatdependsontopicsandthelink sharingdecision.Formally,thegenerativeprocessforeachnewsarticlethatIassumedfora STMmodelis: 1. Formingattentiontotopics:Anewsorganizationformsanattention( ~ d )totopicsfor anarticle d ,whichisdeterminedbyanorganization,thenewssharingdecision,and theinteractionbetweenthetwo,whichareincludedin X d . ~ d j X d ; ˘ LogisticNormal ( = X d ; (4.1) 2. Formingawordprobabilitygivenatopic:Anewsorganizationformsaprobability tousewords( d;k )forthearticle d andgivenatopic k ,whichisinturndetermined bythebaselineweight( m ),thedeviationbyatopic k , k ,thedeviationbysharing decision g , g andtheinteractionbetweenthetwo i . d;k / exp ( m + k + g d + i = g d ) (4.2) 3. Foreachword( n )inanewsarticle( d ), n 2 1 ;:::;N d : Decisiononnewstopic:Anewsorganizationdrawsword'stopicbasedonthe attentiontotopicsformedin4.1. z d;n j ~ d ˘ Multinomial ( ~ ) 44 AttentiontoTopics NewsTopics 1 NewsTopics 2 NewsTopics N Words 1 Words 2 Words N WordProbability Shared? (2) (1) Figure4.5:Agraphicalrepresentationofsharing-dependenttopicmodel. Decisiononwords:Conditionalonthetopicchosen,anewsorganizationdraws anobservedwordfromthattopicaccordingtothewordprobabilityformedin 4.2. w w;n j z d;n ; d;k = z ˘ Multinomial ( d;k = z ) Thisformalspeci˝cation,alsovisualizedasa graphicalmodel inFigure4.5,allowsusto identifydiscriminatingwordsthatcontroltopics.Thevanillatopicmodelwithoutcovariates isconditionalonlyonthe˝rstandlaststrata(attentiontotopicsandworduseprobability) inFigure4.5to˝ndlatenttopics.Ontheotherhand,thetwostratainthestructuraltopic modelareinturnconditionalonthenewsarticlespeci˝clinksharingdecision( shared? ) intwoways.First,theattentiontoatopicdependsonthesharingdecision,which(1)in Figure4.5represents.Thiscorrespondstoequation4.1intheformalspeci˝cationabove.If inequation4.1ispositivelylargeforatopic k ,anewsarticlewiththetopic k islikelyto beshared.Second,wordfrequenciesdependonthesharingdecision,which(2)inFigure4.5 represents. Lastly,thestructuraltopicmodelsadoptthesameLaplacepriortothewordfrequency coe˚cients.Thus,thestructuraltopicmodeladoptsthesameautomaticvariableselection techniqueasintherecentdiscriminatingwordsalgorithmsMonroeetal.(2008);Taddy 45 (2013);Mitra&Gilbert(2014). Insum,thestructuraltopicmodelprovidesacomputationalwaytoidentifyseveral associationsbetweennewstextandthesharingdecision.First,itidenti˝esnewstopics andhowtheyareassociatedwiththesharingdecision.Second,itidenti˝esdiscriminating wordsaftercontrollingthetopics.Lastly,italsoidenti˝esdiscriminatingwordswithineach topic.Further,thisisawaytotranslatemyresearchquestionhowdonewsorganizations conditiontheirlinksharingonnewscontent?toacomputationallyanswerablequestion whattypesoflatenttopicsaremorelikelytobeshared? 4.2.3ParaphrasingforTwitter:MeasuringSentimentwithDictionaries Toanalyzehownewsorganizationsparaphrasetheirnewsstoriesforsocialmedia,Iadoptthe dictionarybasedword-countingmethodthathasbecomeaprimarywaytomeasureemotion orsentimentincomputationaltextanalysisworks.Inparticular,bycountingemotionwords innewsstoriesandtweets,Itestwhethernewsorganizationsaddemotionalframingasthey paraphrasetheirnewsstories.Forexample,ifnewsorganizationswanttomakenewsstories looksensationaltoattractsocialmediausers'attention,theparaphraseswillcontainwords withstrongsentiment(negativeorpositive)comparedtonewsstories. Identifyingwordslikelytoappearonaparaphrasecanbee˚cientlyconductedbycom- putationalalgorithmsthatcountwords.Althoughtheyhaven'tbeenappliedtonewsor- ganziations'paraphrasingforsocialmedia,dictionarybasedcountingwordshasbeenwidely adoptedtostudyotheronlinenewscontexts.Forexample,Horne&Adali(2017)counted wordsrelatedtosentimentsascategorizedinpre-existingdictionariestomeasuresentiment inRedditusers'newsparaphrase.Also,Berger&Milkman(2012)andKim(2015)applieda similarmethodtounderstandtherelationshipbetweensentimentinnewsstoriesandonline newspopularity.Iwilltakeasimilarwordcountingapproachtoidentifykindsofphrases likelytoappearinsocialmediapostsofnewsorganizations. Acaveatinthepreviousliteraturethatanalyzednewsparaphrasingisthatmostofthe 46 literatureanalyzedonlyparaphrasedtext.However,thisapproachmakesithardtoknow whethersentimentinanewstweetcamefromtheoriginalnewsstoryorfromthenewsor- ganization'sstrategy.Inanexceptionalwork,Andrew(2007)comparedheadlinesinmajor Canadiannewspaperswithactualnewscontentduringthe2004Canadianfederalelectionto focusonnewsorganizations'paraphrasingchoices.Withthisapproach,herevealedthatthe Canadiannewspaperschooseissuewordsratherthanhorseracewords(i.e.wordsthatde- scribeelectionasagame),andmorenon-neutral(biased)wordsinnewsheadlines,compared tooriginalnewsstories.Inasimilarway,mywordcountingapproachwillincludeacompar- isonbetweensocialmediapostsandoriginalnewstextstodistinguishnewsorganizations' paraphrasingchoicesfromnewscontent. Therearemanyavailabledictionariesdevelopedinthecontextofdi˙erent˝elds.These includeGeneralInquirerDatabase(GI)frompoliticalscienceStoneetal.(1966),DICTION fromcommunicationHart(2001),LinguisticInquiryandWordCount(LIWC)Pennebaker etal.,theRegressiveImageryDictionary(RID)Martindale(1975)andTAS/CMergen- thaler(1996)frompsychology,A˙ectiveNormsforEnglishWords(ANEW)frombehav- ioralscienceBradley&Lang(1999),DictionaryofA˙ectinLanguage(DAL)fromlitera- tureWhissell&Dewson(1986),WordNet-A˙ect(WNA)fromlinguisticsStrapparavaetal. (2004),andPointwiseMutualInformationwordlist(PMI)fromcomputationallinguistics Turney&Littman(2003).Mostrecently,bysynthesizingexistingdictionaries,Young& Soroka(2012)developedadictionarycalledLexicoderSentimentDictionary(LSD)tailored topoliticalnews. InChapter7,Iwillapplythreepopulardictionariestothedata;A˚n,LSDandLIWC. A˚nandLSDareusedbecausetheyarecustomizedfortwospeci˝cinformationformatsof interestinthissectiontweetsandnewsstories.Inotherwords,thischoiceisintendedto checktherobustnessofresultsacrossdi˙erentdictionariesdesignedfordi˙erentcontexts. LIWCischosenbecauseitcontainsthemostextensivewordsandwordcategories,andhas beenbyfarthemostoftenusedforacademicresearch.Ispeci˝callyusedthelatestversion 47 ofLIWC2015. 48 CHAPTER5 ASIMPLETESTOFTHESTRATEGICLINKSHARINGMODEL 5.1Introduction Itisnotfeasibletotestempiricallytheexcessivelinksharinghypothesisfromtheeco- nomicmodelinChapter3becausethe ˝rstbest choiceisunobservablefromthedata.How- ever,whethernewsorganizationstakecompetitionforlimitedattentionintoconsideration forthenumberofsharednewsstoriesistestable.AsillustratedattheendofChapter3,if theconceptualmodelismodi˝edsothatnewsorganizationsdonotconsiderthecompeti- tion,theproportionofsharednewsstoriesastheiroptimalchoiceshouldnotreacttothe numberofstoriesotherorganizationspublish.Ontheotherhand,accordingtotheorigi- nalmodelwithcompetition,theproportionofsharednewsstoriesintheequilibriumisa decreasingfunctionofthetotalnumberofnewsstoriespublishedbyallnewsorganizations. Speci˝cally,theproportionofsharednewsstoriesisdecidedas Np c K 1 K 1 ˚ = Np c K 1 K T MK .In otherwords,iftherearemorenewsstoriescomparedtosocialmediausers'attention,the proportiondecreases.Therefore,thedirectionofnewsorganizations'reactiontothetotal numberofnewsstoriesisindicativeofwhethernewsorganizationsconsidercompetitionas theysharenewslinks.Thiscomparisonprovidesatestablehypothesis:newsorganizations willsharealesserproportionofpublishednewsstorieswhenallnewsorganizationspublish morestories,iftheyconsiderthecompetitionforlimitedattention.Inthischapter,Iwilltest thishypothesisbyanalyzingnewsorganizations'dailynewslinksharingchoicesonTwitter. 49 5.2Analysis 5.2.1EmpiricalModels Whetheranewsorganizationconsidersthecompetitionornot,thetheoreticalmodelpre- dictsthatanorganizationreactstothenumberofitsownnewsstories.However,themodel doesnotsayhowdi˙erentlytheywillreacttotheirownandthetotalnewsbecauseallsto- riespublishedbyallorganizationswereassumedtohavethesamequalityinthemodel.In practice,thereisagoodreasonfortheconjecturethatnewsorganizationsmoresensitively reacttothenumberofitsownnews.UnlikethemodelinChapter3,eachnewsorganiza- tionislikelytohaveanumberof captured readerswhopreferthespeci˝corganizationto others.Inthiscase,thecapturedreaders'attentionwillbemoreresponsivetotheamount ofinformationfromthepreferrednewsorganizationthanthosefromothers.Forexample, whetherBreitbartpublishesalotofnewswouldnotmattertofar-leftreaders.Because thosecapturedreadersaremorelikelytoincurpro˝tforthepreferrednewsorganizations, organizationswillalsocarespeci˝callyaboutthenumberoftheirownnewslinkssharedon socialmediainordertonotoverloadattentionofthecapturedreadersaswellasthetotal amount. Figure4.2visualizestherelationshipbetweentheproportionofsharednewsstoriesand thenumberofpublishedstoriesonTwitter,brokendownbynewsorganizations.Thereare afewnewsorganizationsthatpublishasmanyas500dailystoriesbecausesomeregional paperspublishstoriesoriginallypublishedbyotheroutletswiththesameownerorbythose withwhomtheypartner.Thesenewsorganizationstendtosharefeweroftheirpartners' storiesthroughtheirTwitteraccount,presumablybecausetheyprioritizetheirownstories. Thispatternfrompartnershipmaymakethepotentialnegativeassociationbetweenthe numberofpublishednewsstoriesandthesharedproportionlookstrongerthanitreallyis. Thus,Irepeatedlyconductanalyseswithdatawithoutthoseoutliers. ThemaindatasetdescribedinChapter4isusedfortheregressionanalysis.Eachnews 50 storyinthedatasetistaggedwithwhetheritwassharedbyanewsorganizationonTwitter. Bycountingstoriessharedbyeachnewsorganizationeveryday,Icalculatedthedailypro- portionofsharednews.Thus,thedependentvariableistheproportion(percentage)ofnews storiessharedbyanewsorganizationoneachday,andtheindependentvariablesarethe numberofitsownnewsstoriesoneachday,andthetotalnumberofnewsstoriespublished byall84newsorganizationsincludedinthedatasetoneachday.Theimpactofthetotal newsimpliesanewsorganization'sstrategicsituationinthesensethatone'sactiondepends onothers.Ialsoincludedtypesofnewsorganizations,asinTable4.1,toseeifthedi˙erent typestendtoadoptdi˙erentnewslinksharingstrategies,andthedayoftheweektocontrol anapparentweeklycycle. Inadditiontothebasiclinearregressionmodel(Model1),Iwasabletotakeadvantageof thepanelstructureofthedataset.Inparticular,Model2isa˝xede˙ectmodelthatcontrols eachnewsorganization'sunobservabletendencyregardingthenewslinksharingproportion bylookingatthevariation within eachnewsorganization.Inthiscase,however,theimpact ofthetypesofnewsorganizationscannotbeexaminedbecausethetypesareaninvariant variable.Model3andModel4arereplicationsofthetwomodelswithouttheoutliers. 5.2.2Results Table5.1showsresultsoftheregression.Model1showsthatanewsorganizationreducesthe proportionofnewslinksitsharesonTwitterrespondingtoboththenumberofitsownnews storiesandthetotalnumberofnewsstoriespublishedbyallnewsorganizations.Ifanews organizationpublishesonemorenewsstory,thenitwillshare 0 : 0869% fewernewsstories (SE 0 : 0048 ).Ifthereisonemorenewsstorypublishedinthemarket,anewsorganization willshare 0 : 0011% fewer newsstories(SE 0 : 0003) .Thisresultcon˝rmsthehypothesisfrom themodelthatnewsorganizationsarestrategicaboutsocialmediausers'limitedattention. Theimpactofthenumberofthetotalnewsstoriesisnotcon˛atedwithareactiontothe numberofanewsorganization'sownnewsstoriesbecausethenegativeimpactofthetotal 51 Table5.1:Regressionoftheproportionofnewsstoriessharedbynewsorganizationson Twitter. VariableModel1Model2Model3Model4 Intercept79.9512***2.1521*81.0988***2.2692* (3.2813)(0.9231)(3.3776)(0.9568) OwnNews-0.0869***-0.0365***-0.0964***-0.0307* (0.0048)(0.0095)(0.0060)(0.0119) TotalNews-0.0011**-0.0017***-0.0011***-0.0019*** (0.0003)(0.0003)(0.0003)(0.0.0003) Tuesday0.84130.69110.84680.6868 (1.7496)(1.3602)(1.8092)(1.4092) Wednesday0.23310.18270.22280.1747 (1.7711)(1.3762)(1.8313)(1.4257) Thursday-0.3278-0.4177-0.3735-0.4245 (1.7608)(1.3687)(1.8210)(1.4182) Friday-1.3547-1.4576-1.4386-1.5114 (1.7165)(1.3353)(1.7751)(1.3836) Saturday-8.1999***-8.4452***-8.6746-8.9201*** (2.0238)(1.5682)(2.0948)(1.6261) Sunday-6.5905**-6.4503***-6.9454**-6.7895*** (2.0424)(1.5804)(2.1143)(1.6389) National-10.3072***-10.9206*** (1.9874)(2.0395) Online-12.4145***-12.5414*** (1.8272)(1.8545) Regional-15.8720***-15.1256*** (1.8649)(1.8998) R0.11400.02040.09460.0200 N4,5324,5324,3614,361 FixedE˙ectNoYesNoYes OutliersYesYesNoNo Signif.codes:0`***'0.001`**'0.01`*'0.05`.'0.1`'1 newsstoriesisobservedevenwhencontrollingoutthenumberoftheorganization'sown newsstories. Althoughthetotalnumberofnewsstoriesseemstohavearelativelysmallimpactas expectedsmallerthantheimpactoftheirownnews,itcanbetranslatedintoasigni˝cant impactinarealisticsituation.Forexample,possiblybecauseofasigni˝cantevent,suppose 100newsorganizationspublishedtenmorenewsstories.Theimpactoftheirownnewsis 52 onlytranslatedinto 0 : 869% fewernewslinks.Ifanewsorganizationwaspublishing100 newsstories,tenmorenewsstoriessimplycanceloutthisimpact,leavingthenumberof newslinksalmostthesame.However,tenmorenewsstoriesalsomeans1,000morenews storiesonthemarketinthisscenario.Itsimpactistranslatedinto 1% fewernewslink sharingaccordingtotheestimatedresults.Thisimpactwouldbecomeevenlargerifnews organizationsconsidercompetitionwithothernewssources,andnon-news-speci˝csources discussingapopularnewsissueareexcludedfromthedata. Thisnegativeimpactofthetotalnewsstoriesisrobustevenwhencontrollingtheid- iosyncraticcharacteristicsofnewsorganizationsusingthe˝xede˙ectestimation(Model2). Whiletheestimatedimpactofthenumberoftheirownnewsbecamesmaller( 0 : 0365 ;SE 0 : 0095 ),theestimatedimpactofthetotalnumberofnewsbecomesevenstronger( 0 : 0017 ; SE 0 : 0003 )thanitwasinModel1.Thus,the˝xede˙ectmodelprovidesastrongersupport forthestrategiclinksharinghypothesis. Toseeiftheorganizationsthatpublishmanystoriesontheirwebsitesdrivetheresults,I excludedthoseoutliersonthefarrightsideofFigure4.2.Inparticular,IexcludedReuters, TheSeattleTimes,USAToday,WashingtonPostandWashingtonTimesthosepublishing morethan200storiesdaily.Models3and4inTable5.1arethereiteratedversionsof Models1and2.However,theresultswithouttheoutliersarehardlydistinguishablefrom theearlierresults.Theonlysigni˝cantdi˙erenceisthattheimpactoftheamountofown newsestimatedby˝xede˙ectbecomeslessstatisticallysigni˝cantwhendatawithoutthe outliersareused(fromP-value<0.001toP-value<0.05).However,theestimatedimpact ofthetotalnewsonthemarketstaysasstatisticallysigni˝cantasthatinModel2,and itse˙ectsizebecameslightlylarger(from-0.0017to-0.0019).Thus,theestimatedresults withouttheoutliersarethesameorslightlyclosertothestrategicsituationbetweennews organizationsthattheconceptualmodeltriedtocapture. Thepatternfromeachdayoftheweekisasexpected;newsorganizationsshareasigni˝- cantlysmallerproportionofnewsstoriesonTwitterduringtheweekends(about 8% lesson 53 Saturdayandabout 6 : 5% lessonSunday).ThisresultisverysimilarinModel3. Finally,di˙erenttypesofnewsorganizationspredictsigni˝cantlydi˙erentproportionsof newsstoriessharedonTwitter,accordingtoModel1.Thebaselinecasehereismagazines, whichpublishasigni˝cantlysmallerproportionofnewsstoriescomparedtodailynews outlets.Thus,theytendtosharealargeportionofnewsstories(potentiallysu˙eringlessof theinformationoverloadproblem).Amongothertypesofnewsoutlets,nationalmediatend tosharealargerfractionoftheirnewsstoriesfollowedbyonlinenewsoutlets.However, thedi˙erencebetweenthesetwogroupswasnotlarge( 10 : 3072 ˇ 12 : 4145 ).Indeed,the di˙erencebetweenthetwocoe˚cientwasnotstatisticallysigni˝cantbyalinearhypothesis test( p =0 : 13 ).Ontheotherhand,regionalmediashareasigni˝cantlysmallerproportion oftheirnews( 15 : 8720 ;SE 1 : 8649 )thanbothnationalandonlinemediado.Thisisalso con˝rmedbylinearhypothesistesting( p< 0 : 01 and p< 0 : 001 respectively).Theestimated di˙erencesinproportionsofsharednewsstoriesbetweendi˙erenttypesoforganizationsin Model3wereverysimilartothoseofModel1. Thisresulthastwopotentialexplanations.First,becauseregionalnewsorganizations oftenpartnerwithnationalnewsorganizations,theymaynotsharemanyofstoriesfromthe partners.Second,because˝nancialsu˙eringprevailsamongregionalmedia,theymaybe lessengagedinsocialmediastrategies.Indeed,itisknownthatnationalnewsorganizations areengagedinsophisticatedsocialmediastrategiesandadjustthemselveswiththenew onlineenvironment(NYT,TheGuardian,BBC)asmuchasonlineonlyventures(Kew, 2016,August17;Rowan,2014,January2;Newman,2011)whereasregionalmediakeep downsizingtheirorganizations.However,theremaybeacounter-argumenttothesecond explanationwherethoselessengagedinsocialmediastrategiesleadstomoresharednews becausetheyresorttoanoldpracticetoautomaticallyfeedtheirnewsstoriesthroughsocial mediaaccounts. 54 5.3Discussion Asimplest,butintuitiveguessaboutnewsorganizations'socialmediastrategywouldbe sharingasmanystoriesastheybelievereaderscanprocessregardlessofotherorganizations' choices.However,ifnewsorganizationscareaboutthechancethattheirnewslinksare clicked,andcompetitionwithothermediaforusers'limitedattention,thereisanincentive toadjusthowmuchinformationtheywanttodisseminateasformalizedinthetheoretical modelinChapter3.Olderevidencerevealedthattweetsfromnewsorganizationswereau- tomaticallyfedasinRSS(Palser,2009;Armstrong&Gao,2010).However,morerecently, newsorganizationsareincreasinglyadoptingmanualcustomizationofnewstweets(New- man,2011),hiringdatascientiststodiscoveroptimalsocialmediastrategies(Rowan,2014, January2),anddevelopingsocialmediamonitoringplatforms(Diakopoulos,2017). Theempiricalresultsinthischaptersupportthehypothesisthatnewsorganizations indeedcareaboutusers'informationoverloadandwhatotherorganizationsdoasthey disseminatenewsstoriesonTwitter,hence strategic .Astheeconomicmodelpredicts,news organizationsshrinktheproportionofnewsstoriestheydisseminateviaTwitterwhenthe wholemarketpublishesmorenewsstories.Theyalsoreducetheproportionwhenthey publishmoreontheirown,whichimpliesthattheyalsocareaboutcapturedconsumers whoprefertheirstories.However,thisresultdoesnotcontradicttheideaofstrategiclink sharing.Newsorganizations'reactionstothetotalamountofnewspublishedinthemarket wasstatisticallysigni˝cantevenwhencontrollingthereactiontotheirowns.Theimpact ofthestrategicchoicewasevenstrongerthanthereactiontotheirownnewsinarealistic situation. Althoughthesimpleanalysisthatfocusesonlyonthequantitativeaspectoflinkshar- ingonTwitteromitsmanyqualitativeaspectsofpotentialstrategies,itclari˝esthatnews disseminationthroughTwitterissigni˝cantlygovernedbystrategicconcerns.Thisraises aconcernaboutwhetherthestrategiclinksharingonsocialmediawillprovidethesame su˚cient,fair,anddiversenewsinformationthatcaninformsocietyasweexpectedfrom 55 traditionaljournalisticnorms.Thismarketdrivennewsdisseminationcanstillsupportthe desirableprovisionofnewsiforganizations'strategiesarediverseenough.But,iftheyare highlypatterned,theimmigrationtowardTwitterandFacebookfornewsconsumptioncan restrictorbiasnewsperception.Thus,towhatextentnewsinformationdisseminatedvia socialmediadeviatesfromthetraditionaleditorialdecisionsisanimportantempiricalques- tiontogaugethesigni˝canceofthenewinformationmediatingprocessnewlyimposedby socialmedia. Neitherthegametheoreticmodelnortheempiricalresultsaddresses how newsorganiza- tionsre˛ectothernewsorganizations'behaviorsbecausebothexplainonlyoutcomesofthe interactionratherthanhowtheygetthere.Onerealisticscenarioisthatnewsorganizations preemptivelyshrinktheproportionofsharednewswhendominantlypopularnewsbreaks outtomitigatepotentialinformationoverload.Inthiscase,newsorganizationsmaywant tosharelessaboutlessimportantnewstopics.Thisscenariospeci˝callyevokesapossibil- itythatstrategicnewslinksharingcausesaconcentrationoncertainnewstopicsonsocial media.Thepotentialconcentrationissueisparticularlyimportantbecauseitcanreducethe diversityofnewsinformationonsocialmedia.Theempiricalassessmentofthedistribution ofsharednewstopicsisthesubjectofthenextchapter. 56 CHAPTER6 SELECTIVENEWLINKSHARINGONTWITTER:ASTRUCTURAL TOPICMODEL 6.1Introduction TheeconomicmodelsuggestedinChapter3impliesthatnewsorganizationscannot individuallyescapefromtheequilibriumwheretheyexcessivelysharemanystories.Although thelimitedattentionasapublicresourceprovidesatheoreticalbasisthatpredictsthe quantitativechoiceoforganizationsasanoutcomeofcompetition,theexcessivesharing equilibriumisbuiltonthesimplifyingassumptionsmadeforthemodel,suchasidentical newsinformationacrossdi˙erentorganizationsandstories.Inreality,factorsotherthan howmanynewsstoriesarepublishedmusta˙ectanewsorganization'sdecisiononwhether toshare.Asonesuchpossibility,inthischapter,Iinvestigate selectivenewslinksharing hownewsorganizationsconditionthelinksharingdecisiononnewscontent.Applying astructuraltopicmodel(STM)recentlydevelopedattheintersectionofpoliticalscience, statisticsandmachinelearning(Robertsetal.,2016)totheWebscrapednewsstoriesand tweets,Iwillshowwhichnewstopicsaremorelikelytobemorevisiblethanothersonnews organizations'Twitteraccounts. Thisstudyisparticularlymeaningfulforthepotentialimpactofsocialmediaasanews platformonnewsdiversitybecausetypesofnewsstoriessharedonTwitterwillin˛uencethe setofinformationtowhichusershaveaccess.Inotherwords,newsorganizations'selective linksharingispotentiallyanewformof gatekeeping (Dimitrovaetal.,2003)totheextent thatitsystematicallydeviatesfromtraditionaleditorialdecisions. 57 6.2MotivationandRelatedWorks Unlikethesimpli˝edmodelsuggestedinChapter3,newsqualityconsistsofavarietyof dimensions;thereisadistinctionbetweenfactualnewsandcommentary.Somenewsstories arein-depthwhereasothersareshortreports.Di˙erentnewsstoriesmaycontaindi˙erent viewpointsaboutthesamepublicissue.Asinglenewsreportmaybeaboutavarietyof publicissues.Moreover,newsreaders'preferenceforeachdimensionofnewsqualityis diverseaswell.Thus,eachnewsorganizationaimsatdi˙erentgroupsofnewsreaders;to useaneconomicterm,newsisa verticallyandhorizontallydi˙erentiatedgood (Gentzkow& Shapiro,2010).Anaturaleconomicbehaviorofa˝rmthatsellsdi˙erentiatedgoodsisto conditionitssupplyonthequalityofthegoods(Andersonetal.,1992).Thisisbecausethe qualitya˙ectsconsumers'valuationofthegood,whichinturndeterminesdemandthe˝rm faces.Thus,asfaras˝rm'sbehaviordependsondemand,italsodependsonthequalityof theproduct. Totranslatethisargumentintothecontextofthisstudy,newsorganizationswillcondi- tionnewslinksharingonnewscontent.However,itishardtoincorporatethemultidimen- sionalqualitydi˙erentiationintoamathematicalmodel;suchamodelwillhavetoomany parametersstructuredinacomplicatedwaytobesolvedtoproducepredictablehypotheses. Instead,onecanempiricallyexplorehowthenewslinksharingdecisiondi˙ersdepending onavarietynewsqualitiestoidentifytheselectivenewslinksharing.Althoughdrawing predictionsfromaformaleconomicmodelisnotfeasible,˝ndingsfromtheonlinejournalism literatureprovidesomeconjecturesaboutnewstypesmoreorlesslikelytobedisseminated viasocialmedia.Inthissection,Iwillreviewthisliteraturetocomplementthestandard economictheorywithwhatprevious˝ndingsaboutnewsorganizations'onlinenewschoices andtheirsocialmediapostingspredictabouttheirselectivenewslinksharingonsocial mediaasanewsdistributionplatform. 58 6.2.1SocialMediaasANewsDistributionPlatform Studiesaboutnewsorganizations'tweetshintatwhattypesofnewsstoriesnewsorganiza- tionsshareonTwitter.Initially,mediascholarsexpectedthattheWebwoulda˙ordnewsor- ganizationsachannelforthemutualinteractionbetweenmediaandaudience(Chan-Olmsted &Park,2000),andwouldbeane˙ectivepromotiontooltoattractyoungeraudienceswho donotregularlyaccesstraditionalmedia(Palser,2009;Chan-Olmstedetal.,2013).How- ever,evidencethatnewsorganizationsareusingtheseopportunitiesareratherweak.For example,Greer&Ferguson(2011)analyzedtweetsfrom488localTVstations,andfound outthatonly 23 : 3% of455commercialTVstationstweetsforinteractionwithnewsread- ers.Similarly,Meyer&Tang(2015)recentlyhand-coded4,507tweetsfrom60localnews organizations(TVandnewspaper),andonly 7 : 4% ofthetweetsfromlocaltelevisionsta- tionsand 11 : 6% fromlocalnewspaperswereintendedforinteraction.Clearyetal.(2015) drewasimilar˝ndingfromtweetsfromtheCNNInternationalchannel.Thesestudiesalso generallyconcludethattraditionalnewscompaniesarenotengagedinpromotionforeither theirwebsites,ortheorganizations'brands(Greer&Ferguson,2011;Meyer&Tang,2015). Armstrong&Gao(2010)concludedthattraditionalnewsorganizationsgenerallyextend theirconventionstoTwitterratherthanadoptingacustomizedstrategy. Traditionalnewsorganizationsratherusesocialmediaasanadditionalnewsdistribution platform.Greer&Ferguson(2011)foundthat 94 : 9% ofthecommercialTVstationstweets todisseminatetheirnewsarticleswhereasonly 17 : 6% tweettopromotetheirprograms. Similarly,Meyer&Tang(2015)reportedthat 94 : 4% ofthetweetsfromTVstationsand 96 : 3% fromnewspapersarefornewslinksharing.Thenewsdistributionthroughsocial mediaisapro˝tabletactic.Hong(2012)foundoutthatthesocialmediauseandthe numberoftweetsbythenewsorganizationsinducemoretra˚ctowardtheirnewswebsites. Therefore,newslinksharingwouldbethemainconcernwhennewscompaniescontemplate theirsocialmediause. Previousstudieshavefoundsomeevidencethatnewsorganizationsarestrategicindis- 59 seminatingtheirstoriesviasocialmedia.Armstrong&Gao(2010)speculatedthatthe abundanceoflinksto`sensationalistic'newswithtopics,suchascrimeandmanylife-style newslinksbylocalmediacomparedtonationalandregionalmediaarearesultofacon- siderationforthetargeteddemand.Althoughnewsnewslinksharingseemedautomaticfor manyorganizationsthatconsideredthesameparaphrasingforsocialmediaasheadlinesfrom originalstories(Palser,2009;Armstrong&Gao,2010),ArmstrongandGaoalsofoundthat someorganizationsaddnewsleadswhereasothernewsorganizationspostonlyheadlines. Further,morerecentstudiesfoundthatnewsorganizationsareadoptingmanualcuration optimizedforsocialmedia.AccordingtoNewman(2011),theBBCswitchedfromautomatic feedtomanualchoiceandeditingoptimizedforTwitterin2011,andthenumberoffollow- ershasdoubledsincethen.Recognizingthispotential,Facebookpublishedaguidelinefor strategicnewspostsbasedontheirownusernewsengagementstudies. 1 6.2.2NewsOrganizations'OnlineNewsChoice Althoughstudiesabouthownewsorganizations,as˝rms,chooseasubsetofpublished newsstoriestodisseminateviasocialmediaarerare,theirtraditionalgatekeepingpractice providessomeclue.Theselectivenewslinksharingdecisionforsocialmediaresemblesthe traditionalgatekeepingfunctionofjournalisminthatitdecidesthesalienceofasubsetof information.Onecaneven˝ndaprimitiveformofselectivenewslinksharingfromthe originalconceptionofeepintheliterature;Asnewsorganizationschoosenews likelytodrawmoreattentionortobemorenewsworthyonsocialmedia,inthepastawire- editorselectedstoriesfromamassofwirecopiesfromwireservices(White,1950).Thenews bulletinisanotheroldpracticeevenclosertonewslinkselection.Whereasthetraditionally understoodgatekeepingisaprocessthroughwhichjournalistsselectednewsinformationout of raw information,boththenewslinksharingandthenewsbulletingivemoresaliencetoa 1 https://www.facebook.com/notes/facebook-journalists/study-how-people-are-engaging- journalists-on-facebook-best-practices/245775148767840 60 subsetof published news(Bruns,2011).Indeed,intheonlinecontext,Dimitrovaetal.(2003) showedthatnewsorganizationsusehyperlinksasagatekeepingtooltocontrolout-tra˚cto informationsources. Recentdiscussionsaboutnewsaggregatingservicesshowthatwhathappensonthose platformsimposesinformationconstraintsonnewsreadersontopoftraditionaljournalism practice.Whennewsreadersrelyonseparateplatforms,suchasGoogleNewsorYahooNews, toreduceinformationtotheamountthattheycanintelligentlyprocess,theplatformcan haveastrongpowertodecidethesalienceofnewsinformation.Inthisvein,Carlson(2007) arguesthatGoogle'schoiceofnewshasadi˙erentmotivationfromtraditionalmedia's,and thatthemoveofthegatekeepingpowertotheaggregatorscanultimatelyunderminesthe traditionalvaluesofjournalism.Morerecently,thediscussionabouta˝lterbubble(Pariser, 2011)warnsthatthealgorithmicrecommendationofnewsstoriescandisturbthetraditional processthroughwhichlegacyjournalismprovidesqualityinformationaboutpublicissues becausenewsreadersfallintoafeedbackloopwherebynewsreadersbecometrappedbythe algorithmsinwhattheyareestimatedtoprefer. Althoughmediatheorieshaveexpandedthegatekeeperconcepttocapturemultiplegates withinanorganizationthathappensduringtheeditorialprocessandtheinter-mediagate- keeping(Shoemaker&Vos,2009).Theselectivelinksharingonsocialmediaisfullycaptured byneitherframeworkbecausegatekeepingonsocialmediahappens within anorganization, butalso between multi-platforms.Indeed,ifnewsorganizationsselectivelysharenewslinks onsocialmedia,thissuggeststhatthegatekeepingprocesswithinanorganizationcanbe governedbydi˙erentsocio-technicalconstraintsdependingonplatformsusedfornewsdis- semination. Althoughresearchaboutmulti-platformgatekeepingwithinanorganizationislimited, therearecomparisonsbetweenaudience'snewschoicesononlineplatformsandjournalists' hintsatsocialmediausers'newspreferences.Forexample,Zubiaga(2013)comparesthe NewYorkTimes'topnewsstoriesasaneditorialdecisionwithstoriesthatarethemost 61 popularonTwitterandFacebook.Analyzingmorethan56Knewsstories,hefoundthat professionaleditorstendtoselecthardnewssuchasUSpoliticsandinternationalrelations whereassocialmediauserslikerelativelysoftnewssuchasscience,fashionandtechnology. Ifnewsorganizationsoptimizetheirlinkselectionforthenumberofclicksorviralityon socialmedia,users'orientationtowardsoftnewswillmanifestitselfinnewsorganizations' choicesforsocialmediaaswell. Recentevidencefromthe˝eldsuggeststhatdecisionsinvolvedwithnewsdissemination viasocialmediaisrelativelyautonomousfromthetraditionaleditorialprocess.According tosocialmediaspecialists'testimoniestotheAmericanPressInstitute(Elizabeth,2017, November14),thedutiesandresponsibilitiesregardingsocialmediapostingarefractured ratherthanbeingtiedtothecentralizededitorialdeskdecisions.Althoughtheyvaryacross organizations,socialmediastrategiesareperceivedandorganizedasbusinessstrategies ratherthanjournalismpractices.In2015,MichaelRoston,aseniorsta˙editorsocialmedia atTheNewYorkTimes,wrotethatTheTimes'socialmediadesks'editorsjoinedthe AudienceDevelopmentdepartment,whichfocusesonsearchengineoptimizations,analytics andgrowth(Roston,2015,January22).Asaresult,manysocialmediaspecialistsfeelthat theirjobdoesnotresideatthecoreoftasksofnewsorganizations.workinsocialmedia alwaysfeltsomewhatremovedfromtherestofeditorial;likethewriterswereoperating independentlyandmyjobwastocatchandpromoteasmuchoftheircontentasIcould. Wedidn'tfeelin(Elizabeth,2017,November14). Thistendencyisescalatedbythewide-spreaduseofaudiencemetrictechniquesforsocial mediaoptimization.Whattopostandwhentopostonlineincreasinglydependonthereac- tionofaudiences.Formanysocialmediaspecialists,thebasicroutinethatdominatestheir jobisostandcoun(Elizabeth,2017,November14).Usinganethnographicalmethod, TandocJr(2014)foundthatuseofaudiencemetricandrecommendationalgorithmsindeed resultsinanewdrivetowardanadditionaldimensionofgatekeepingforonlinejournalism. Selectionanddeselectionofnewsstoriesonorganizations'newsdisseminationplatformsare 62 quicklydecidedaccordingtooutcomesofreal-timeaudiencemetricsorevenofautomated algorithms.Tomakethisprocessevenmoree˚cient,leadingnewsorganizationshavebeen developingsocialmediamonitoringanddecision-aidplatformsontheirown(Diakopoulos, 2017).Thedegreetowhichsocialmediarelateddecisionsareintegratedintoeditorialvaries. Althoughitispartiallyin˛uencedbyhowwellanewsorganizationdoesinthemarket,but it'snotde˝nite.Forexample,whereasTheNewYorkTimes'caseisleaningtowardabusi- nessorientationofsocialmediastrategies,Elizabeth(2017,November14)foundthatthe WashingtonPosttakesadi˙erentpathtoencouragesocialmediaspecialistsandnewsrooms toworktogether.Insomeregionalnewspapercompanies,averylimitednumberofsta˙, evenasingleperson,decidesthesocialmediaposting.Butsometimesitisentirelydispersed amongindividualreporterswithoutaidofsocialmediaspecialists,mainlybecauseofthe tight˝nancialconstraint.Thus,althoughitisalmostcertainthatselectivelinksharingwill deviatefromthetraditionalgatekeepingdecisions,itsdegreeforindividualorganizationsis hardtopredictbeforehand. Allinall,previousstudiesimplythatnewsorganizationsarelikelytobeengagedina gatekeepingtaskforsocialmediathatisseparatefromthetraditionaleditorialprocess,and thatthesocialmediagatekeepingstrategyislikelytobegovernedbyaperceivedmechanism bywhichnewsstorieswouldgoviralonsocialmedia.Thus,newsorganizationsarelikely tosharenewsstorieswithspeci˝ccomponents,suchastopicsthatareknowntocontribute tovirality.Previousstudiesaboutnewsviralityhaveidenti˝edsuchnewscomponentsthat helpvirality:Evidencegenerallyshowsthatemotionalarousal(intensity)isassociatedwith viralityonsocialmedia(Berger&Milkman,2012;Stieglitz&Dang-Xuan,2013).Focusingon newstopics,García-Perdomoetal.(2017)foundthatcon˛ict/controversy,humaninterest andoddnewscategoriesaremorelikelytobesharedbyTwitterusers.Thus,assuming thatnewsorganizationsareadaptingtosuchnewsdemandonTwitter,onecanhypothesize thatnewsorganizationswillsharenewstopicslikelytocontainhighemotionalarousal, controversy,humaninterest,oroddity.Further,astheemergingonlinenewsorganizations 63 havebiggerstakesinsocialmediatodisseminatetheirstories,thispatternwillbemore evidentforthem.Iwilltestthesehypothesesbyanalyzinginthenextsectionwhichnews linkseachnewsorganizationsharedonitsTwitteraccount. 6.3ComputationalIdenti˝cationofSelectiveNewsLinkSharing IdentifyingtheselectivenewslinksharingonTwitterrequiresatwo-stepanalysisto˝nd out(a)whatanewsstoryisaboutand(b)howthenewscontentisassociatedwiththelink sharingdecision.Becauseofthedi˚cultytoachievearepresentativesampleofpublished newsandthelonglistofnewsorganizations,theamountofdatainvolvedwiththisanalysis becameenormous.Inthischapter,IadoptamachinelearningtechniquecalledStructural TopicModel(STM)(Robertsetal.,2016)toanalyzemorethan40Kweb-scrapedonline newsstoriesfrom84outlets.STMwillsimultaneouslyidentifynews topic acomputational operationalizationofnewscontentand˝ndtherelationshipbetweenthenewstopicsand observablevariablesofinterest,suchaswhichorganizationpublishedthenewsstoryand whethertheorganizationsharedthenewsonTwitterinthiscase. 2 6.3.1NewsTopicCategorization Categorizingnewsbasedonitscontentisalongstandingmethodologicalissueofmedia studies.Dependingonnewssamplesandthescopeofresearchquestions,sometraditional newscategoriesmaybeirrelevantorcollapsible,butotherunusualcategoriesmaybecome important(Sjøvaag&Stavelin,2012).Aclassoftopicmodels,includingSTM,overcomesthis challengebyusingabottom-upapproachwherebytopicsare discovered fromco-occurrence ofwordsacrossrelateddocumentsratherthanpre-supposingnewscategories. However,thatthetopicmodelsrequireresearcherstopre-specifythenumberoftopics (i.e.newscategories)posesachallengetothisstudy.Thechallengeis,slightlydi˙erent, butconceptuallysimilartothetraditionalquestionformanualcontentanalysis:arethe 2 SeeChapter4foradetaileddescriptionofthemethod. 64 discoveredtopicsbasedonthenumberoftopicsassumptionappropriatefortheresearch question?Thisisaharderquestiontoanswerwhentopicmodelsareappliedtoahighly heterogeneouscorpusoftextsuchasabodyofnewsstoriespublishedbymultiplenews organizations.Ifaresearchersetsthenumberoftopicstoolarge,thenthemodelislikelyto identifyonlytypesofnewsorganizations,suchasonline,national,entertainmentoriented, Texasbased,etc.,usingorganization-speci˝ctextualtraitsratherthancommontopics.In thiscase,aresearcherwillbeunabletocomparedi˙erentbehaviorsforthesamenews topicsbecausetopicoverlapacrossdi˙erentorganizationswouldbelimited.Ifthenumber oftopicsistoolow,ontheotherhand,aresearcherislikelytomissinterestingvariations acrossdi˙erentnewscontents. Researchershavesuggestedmultipleindicestocompareperformanceofdi˙erenttopic numberassumptionsfromtheperspectiveofthestatisticalmodelcomparison.Forexample, Bleietal.(2003)suggestedtheheld-outlog-likelihoodapproach.Thisapproachcompares likelihoodsfromheld-outdataasinthecommoncross-validationusingestimatedpa- rametersbasedondi˙erentchoicesoftopicnumber.Otherscholarssuggestedanapproach closertohumanintuitionoftopics;ifatopiciswell-de˝ned,wordsthatcharacterizeatopic shouldappearfrequentlyinthesamedocument(coherence)(Mimnoetal.,2011;Newman etal.,2010),andwordsthatcharacterizedi˙erenttopicsshouldnotbeinthesamedocument frequently(exclusivity)(Bischof&Airoldi,2012).However,theseautomaticproceduresto choosetheoptimalnumberoftopicsarenotoftenadoptedforappliedworkbecausethey tendtodeviatefromthechoicesofhumancoders(Changetal.,2009).Moreover,thedif- ferentproposedindicesareofteninconsistentwitheachother,asinthisstudy,andtheory doesnotdictatewhichindexshouldbeprioritized.Afterall,thenumberoftopicstendsto beupforresearchers'carefulchoices,basedontheirresearchquestions.Inthiscase,the interpretationoftopicmodelsisthebestcategorizationofnewstopics giventhe numberoftopics Forthisstudy,Ichosethirtytopics.Themainreasonforthischoiceistonotbreakdown 65 topicstothelevelatwhichnewstopicsbelongsolelytospeci˝cregionalmarkets.Thisisnot toignoreregionaltopics;infact,themotivationistheopposite.Howdi˙erenttypesofnews organizationstreatregionaltopicsdi˙erentlyisofgreatinteresttothisstudy.However,if themodeltoo˝nelycategorizesnewstopicssothatNewYorknewsisdistinguishedfrom Texasnews,oneendsupwithatrivial˝ndingthat,forexample,NewYork-basednews organizationsarelikelytoshareNewYorknewstopics.Indeed,IfoundthatSTMbeginsto identifynewsaboutspeci˝cregionswiththirty-twotopicsfromthisstudy'sdataset.Thus,I decidedtolimitthenumberoftopicstothirtysothatthemodelcanidentifymorebroadly labeledregionaltopics. Table6.1summarizestheidenti˝edtopicsandwordsmostlikelytoappearinagiven topic.Ilabeledeachtopicbyinspectingthelistoftheassociatedwordsandrandomlycho- sennewsstoriesamongthoseestimatedtobemostlikelytobeaboutthespeci˝ctopic. Formostcases,theassociatedwordsprovideacoherentinterpretationabouteachtopic. Forexample,theoutputispreciseenoughtodistinguishdi˙erentsports:theSports/Bas- ketballtopicisassociatedwithwordssuchas`3-pointer',`3-point'and`rebound'whereas `nfc',`touchdown',`n˛'representtheSports/Footballtopic.Thesports-relatedtopicswill beparticularlyimportantfortheregionalmedia'sstrategyinwhatfollows.Furthermore, theoutputdistinguishestopicsrelatedtonationalpoliticse.g.Politics/Presidentand Politics/CongressfromtheRegionalPoliticstopic,whichisassociatedwithwords,such as`mayor',`counti',`neighborhood',andnamesofregionalcommunities.Thereareafew casesinwhichthenewsformatorfunctionalitydominatenewscontent,suchasSubscription, AdvertisingandFrench(language).Althoughthesetopicsareoflessinterestinthisstudy, identifyingthemwasunavoidablebecausenewswebsitesincludeafewnativeadvertising oradvertisingforitsown,andMediaClouddoesnotdistinguishthem.However,STM's successfuldetectionofnon-newsmessageshelpednarrowmyfocusonactualnewsstoriesof myinterestforthisstudy.Moreover,thesecategoriesarenotfrequentinthedataset,and arelesslikelytocontainsubstantivenewscontentoverlappedwithothernewstopicsbecause 66 mosthavenon-newsrelatedfunctions(e.g.subscriptionandadverstising)orcharacteristics (e.g.Frenchlanguage). Figure6.1visualizesthefrequencyofeachtopicamongallthenewsinthedataset.The distributionshowsasensiblemixofnewstopics:Topicsoftenclassi˝edas softnews suchas LifeStyle,Crime,andsportsrelatedtopicsoccupytopranksofthelist.However,thereisa goodamountof hardnews withtopicsrelatedtonationalpolitics,economyandinternational relationsaswell.Althoughthereisonlytopicdirectlyrelatedtoregionalissues,Regional, thesports-relatedtopicsandtopics,suchasEducation,TransportationandCrimearelikely tobemorerelatedtoregionalissues. 6.3.2OverallPatternsofSelectiveLinkSharing TheregressionwithintheSTMmodelassociatesportionsofthelatenttopicswithobserved covariates(a)whetheranewsstorywassharedonTwitter,(b)indicatorsofnewsorga- nizations,and(c)interactionsbetweenthetwo.Testimatedcoe˚cientsforthecovariates mean(a)howmuchmoreoftenagiventopicis sharedoverall comparedtothetopic'spro- portionamongthetotalnewsstories,(b)howmuchmoreoftenagiventopicis published by anewsorganizationcomparedtootherorganizations,and(c)howmuchmoreoftenagiven topicis sharedbyanewsorganization comparedtootherorganizations.Inotherwords,this regressionallowsdetectingbothtopicsmoreorlesslikelytobesharedoverall,andtopics thataremoreorlesslikelytobesharedbyspeci˝cnewsorganizations. Figure6.2presentstheoverallpropensity.Thezerointhemiddleisthetopic-wisebaseline proportionamongallthepublishednewsstories,whichinfactvariesacrossdi˙erenttopics. Thepointsrepresenthowtheproportiondeviatesfromthebaselinewhenonlylookingatthe sharedtopics.Thus,itcanbeinterpretedasarelativepropensityoftopicstobeshared.In whatfollows,ABCwaschosenasabaselinetovisualizehowdi˙erentorganizations'selective linksharingdeviatesfromnational,traditional,and(relatively)neutralmediaoutlet.The topicsaresortedfromthemostcommontotherarest. 67 Table6.1:Wordslikelytoappeargiveneachtopic. LabelWordswithHighProbability LifeStyle/Interviewyeah,youv,im,ive,your,weird,oh,mayb,stu˙,shes,funni,guy,crazi,guess,laugh,youll, ok,hey,id,somebodi Crimepolic,custodi,arrest,sheri˙,prison,feloni,jail,detain,sentenc,murder,prosecutor,juri, patrol,indict,prosecut,convict,testi˝,plead,shoot,˛ed Sports/Basketball3-pointer,3-point,rebound,turnov,basket,streak,nba,halftim,foul,overtim,score,basketbal, teammat,bench,lineup,shoot,quarter,ankl,de˝cit,game Economy/Techsoftwar,retail,compani,inc,googl,batteri,appl,buyer,user,maker,innov,manufactur,sale, custom,industri,regulatori,ceo,internet,electr,technolog Social/Con˛ictdemocraci,planet,reader,earth,radic,mainstream,cultur,phrase,context,truth,racist, wealth,rhetor,principl,narrat,enemi,concept,societi,notion,religi International/MiddleEastjerusalem,syria,saudi,islam,isra,embassi,arab,israel,refuge,iraq,muslim,iran,minist, parliament,terrorist,civilian,diplomat,ministri,u.n,democraci Legal/Regulationlawsuit,regulatori,court,suprem,constitut,regul,commiss,legislatur,lawyer,ban,legal, environment,discrimin,justic,permit,attorney,su,religi,amend,enact Economy/Financebitcoin,currenc,in˛at,economist,index,investor,stock,billion,growth,pro˝t,equiti,market, strategist,forecast,hike,buyer,sector,asset,price,invest UKprinc,queen,marriag,marri,royal,bbc,caption,divorc,harri,copyright,kate,pregnant,wed, girlfriend,birthday,uk,shes,actress,london,babi Politics/Presidentmueller,˛ynn,fbi,hillari,clinton,russia,trump,russian,presidenti,judiciari,counsel,obama, barack,probe,testi˝,sander,investig,strategist,politico,indict Updatesp.m,a.m,vs,dec,pic.twitter.com,t.co,jan,8,9,7,aug,feb,nov,6,pm,2017,nashvill, decemb,4,oct Sports/Footballnfc,touchdown,vike,n˛,lineback,postseason,quarterback,cornerback,playo˙,yard,o˙sea- son,eagl,rooki,matchup,panther,bradi,bu˙alo,teammat,stadium,bowl Sports/Otherbasebal,tiger,leagu,cup,soccer,championship,tournament,player,stadium,roster,sport, o˙season,premier,squad,golf,olymp,captain,athlet,australian,club Entertainmentalbum,disney,music,song,theater,comedi,movi,studio,net˛ix,singer,˝lm,sing,episod, artist,band,festiv,funni,actor,premier,viewer Social/Genderfranken,inappropri,harass,lawsuit,resign,alleg,complaint,misconduct,settlement,sexual, accus,kiss,investig,sta˙er,apolog,ethic,inquiri,employe,uncomfort,workplac Regional/Politics˝re˝ght,evacu,homeless,santa,shelter,mayor,counti,neighborhood,francisco,˛ame,diego, orang,downtown,smoke,oakland,los,hurrican,˛ee,rent,˝re Weathertemperatur,snow,outdoor,˝sh,tree,bird,salt,winter,weather,inch,ice,hike,rain,hotel, ocean,mountain,denver,tourist,sea,forest Transportationpasseng,airlin,airport,˛ight,navi,plane,port,crash,vehicl,pilot,transport,crew,highway, driver,patrol,truck,tra˚c,rescu,accid,ship Healthdiseas,patient,clinic,medicin,cancer,addict,drug,medic,health,doctor,brain,alcohol,dr, hospit,nurs,surgeri,depress,studi,suicid,treatment Subscriptionsubscrib,newslett,editor,inbox,advertis,reader,click,robot,pleas,jersey,york,veri˝,cabl, privaci,occasion,commentari,columnist,publish,editori,manhattan Taxdeduct,tax,trillion,taxpay,debt,˝scal,insur,economist,revenu,corpor,premium,incom, hike,provis,enrol,de˝cit,wage,treasuri,reform,repeal SocialMediayoutub,user,instagram,facebook,googl,app,platform,internet,onlin,blog,fake,web,video, social,mail,camera,audio,footag,content,compu Immigrationdaca,bipartisan,shutdown,mcconnel,schumer,immigr,deport,legisl,congression,gop,senat, republican,sen,democrat,congress,capitol,lawmak,short-term,mitch,rep Educationstudent,teacher,educ,campus,graduat,school,nonpro˝t,teach,enrol,donat,academ,balti- mor,district,donor,colleg,grade,fundrais,chariti,volunt,taught Sports/Collegencaa,freshman,coach,footbal,iowa,championship,espn,sophomor,bowl,athlet,tennesse, mississippi,michigan,ohio,miami,basketbal,alabama,teammat,georgia,oklahoma Politics/Congressvoter,ballot,moor,roy,alabama,elect,gop,nomine,republican,poll,strategist,democrat, re-elect,nomin,senat,presidenti,legislatur,vote,jone,doug MeTooweinstein,women,sex,hollywood,femal,harvey,male,rape,gender,actress,woman,nbc,gay, workplac,men,sexual,rose,charli,uncomfort,discrimin International/FarEastkorea,missil,korean,nuclear,china,chines,sanction,russia,japan,moscow,russian,diplomat, kim,u.n,ministri,regim,un,olymp,weapon,deleg Frenchn,la,de,e,dalla,h,el,del,franc,b,m,v,p,co,l,f,s,r,fort,k Advertisingemb,0,copi,cup,code,paus,bowl,video,email,link,courtesi,kid,super,sale,camera,recipi, tip,mom,tast,facebook 68 Figure6.1:Abargraphforestimatedoverallproportionsof30topics.Theunitisfraction. 69 Figure6.2:Relativepropensitytobesharedonsocialmediaof30topics.Thepropensity ismeasuredasafractionaldeviationoftheproportionofanewstopicamongsharedstories fromthatamongpublishedstories.Thetopicsareorderedbythefrequencyonnewswebsites. Thepropensitytobesharedandfrequencyonwebsitesarenotnotablyassociated. 70 Theredoesnotseemtobeaclearcorrelationbetweenthefrequencyofanewstopicand thetopic-wisepropensitytobeshared.Ifthereweresuchacorrelation,thepointsinFigure 6.2shouldhavebeenroughlylaidonaline,whichdoesnotseemtobethecase.Forexample, whereasSocial/Con˛ictandInternational/MiddleEasttopicsarerelativelycommon,they arelesscommonamongthesharedtopics.Ontheotherhand,socialmediauserswillsee Sports/Collegetopicwithahigherchancethanwhenthey˝ndnewsfromwebsites.This doesnotnecessarilymeanthatsocialmediauserswillseemoreSports/Collegetopicthan Economy/Techtopicbecausetheresultsshowonlydeviationsfromthepublishedproportion ratherthanactualproportionsonTwitter.Forexample,becausetherearemanymorenews storiesaboutEconomy/Techpublishedonwebsites,evenifnewsorganizationsareslightly lesslikelytosharethemonTwitter,therearelikelytobemoreEconomy/Technewsstories thanSports/Collegeones.Whattheresultsdomeanisthatanewsreaderislesslikelyto seeEconomy/TechnewsfromTwitterthanfromnewswebsites. TheobservationthatthechanceofatopicbeingsharedonTwitterdoesnotmatchwith itsfrequencyonnewswebsitesimpliesthatthepublicationdecisionandthelinksharing decisiondonotnecessarilyfollowthesamelogic.Acommonknowledgeaboutapublication decisionisthataformaleditorialprocesswithinanewsorganizationdetermineswhatto communicatetoreadersaccordingtonewsworthiness.Thisdecisionisgovernedbytradi- tionaljournalismnorms(Hermida,2010).Ontheotherhand,recentobservationsabout newsorganizations'SocialMediaOptimization(SMO)seemstobemoregovernedbyeco- nomicincentives(Newman,2011;Diakopoulos,2017).Thus,thedi˙erencebetweenthe patternsonwebsitesandontheirsocialmediaaccountsarelikelytorepresentdi˙erentsets ofinformationgiventonewsreaders,whichstemfromdi˙erentgoverningprinciples.Ifnews readers'migrationtosocialmediaremainsconsistentinthefuture,thenewschoicesmade byeconomicincentivesmaysigni˝cantlyimpactwhatnewsreadersseeasnews. Typesofnewsmorelikelytobesharedseemtocon˝rmtheconjecturethattheshar- ingpropensityisgovernedbyeconomicincentivesratherthannewsworthiness.Thetopics 71 withhigherpropensitytobesharedareonestraditionallycategorizedas`soft'topics,such asCrime,Sports/BasketballandSports/Football.Indeed,amongeleventopicsthathave alargerproportiononTwitterthanonnewswebsites(topicsontherightofthedotted lineinFigure6.2),theonlynewstopicthatisdi˚culttocategorizeassoftnewswasRe- gionalPolitics.Ontheotherhand,allothertopicstraditionallycategorizedas`hard'topics, suchasInternational/MiddleEast,Economy/FinanceandPolitics/President,havealower proportiononsocialmediathanonwebsites(topicsontheleftofthedottedline).Ifthe commonwisdomthatreaderstendtoprefersoftnewstohardnewsstillholdsforTwitter, newsorganizations'incentivestosharenewslinkslikelytobeclickedmayhavecreatedthis sharingpatternEconomicallydrivengatekeepingforsocialmediapushesconsumablenews towardsofttopics. 6.3.3Di˙erenceAcrossOrganizations Althoughthepreviousanalysisofoverallnewslinksharingrevealsageneraltendencyfor newslinkssharedonTwittertobebiasedtowardsoftnews,thepatternmaydi˙erdepending onthestrategiesofindividualnewsorganizations.Previousliteraturehasshownthatnews publishingpatterns,whichistraditionallyconceivedasgovernedbyvaluesandconventions ofjournalism,hingeonnewsorganizations'di˙erentincentives.Forexample,Lischka(2014) illustrateshowdi˙erentmediaownershipstructuresandmediamodesleadtodi˙erentvol- umesofnewstopics,di˙erenttones,anddi˙erentnewsinterpretations.Ifsocialmedia strategiesaremoresubjecttoeconomicincentivesthanthetraditionaleditorialdecisions, thedi˙erencesofnewschoicesacrossnewsorganizationswillbeevenmoreconspicuousthan thenewspublishingpattern.Althoughlittleisknownaboutnewstopicssharedonsocial media,thereisevidencethatdi˙erenttypesofnewsorganizationse.g.publicvs.com- mercialTVstationsandTVstationsvs.newspaperstweetindistinctpatterns(Greer& Ferguson,2011;Meyer&Tang,2015).Further,asillustratedintheoverallpatterninthe previoussection,thereisnoguaranteethatthenewspublishingpatternandtheconditional 72 newslinksharingpatternareanalogous. Thus,therearetwoquestionsregardingthepotentiallyvaryingselectivenewslinksharing strategiesacrossdi˙erentnewsorganizations:(a)Dothelikelihoodsthatacertaintopicis shareddi˙eracrossdi˙erentorganizations?(b)Doesthegapbetweennewspublishing andnewslinksharingpatternsdi˙eracrossdi˙erentorganizations?Intheregressionin theSTMmodel,coe˚cientsfornewsorganizationindicatorsrepresenthowmuchmoreof anewstopicanewsorganizationpublishesthanthebaseline(ABC).Andcoe˚cientsfor theinteractiontermsbetweenthelinksharingdecisiondummyandthenewsorganization indicatorsrepresenthowmuchtheproportionofatopicamongsharednewsstoriesdeviates fromtheproportionamongallthepublishedstories.Thus,the˝rstsetofcoe˚cientsshows thepublishingpatternwhereasthesecondsetofcoe˚cientsshowshowthelinksharing deviatesfromthepublishingpattern.Toillustratehowdi˙erenttypesofnewsorganizations adoptdi˙erentlinksharingstrategies,Iwillvisualizethesecoe˚cientsbrokendownbynews topicswithclearpatterns. Newspublishingpatternsaccordwithorganizationtype. Estimatedproportionsof publishednewstopicsaccordwithcommonsensepatterns;nationalmediaandonlinemedia publishmoreaboutnationalpoliticse.g.Politics/PresidentandPolitics/Congresswhereas regionalmediapublishmoreabouttopicsthatarelikelytobepreferabletoregionalaudience e.g.RegionalPolitics,Sports/Basketball,Sports/College,Sports/Football.Figure6.3 and6.4presentatypicalexample.Thetransparentbarsmeanhowmuchproportionofa publishedtopicforeachorganizationdeviatesfromthebaseline(ABC)whereasthesolidbars meanhowmuchproportionofatweetedtopicforeachorganizationdeviatesfromthegiven organization'spublishingpattern.Thus,thetransparentandsolidbarscanbeinterpreted aseachneworganization'spropensityto publish agivennewstopiconitswebsiteandthe propensityto share linksaboutagiventopiconitsTwitteraccountrespectively.Thegreen barmeansthatanewswebsitehasapaywall,andtheredbarmeansnopaywall.InFigure 73 Figure6.3:Newspublicationandlinksharingpatternsbynewsorganizations:Politics/Pres- identtopic. 6.3,thenationalandonlinemediaaremorelikelytopublishaboutPoltics/Presidenttopics thantheregionalmedia.Ontheotherhand,theregionalmediaaremuchmorelikelyto publishaboutSports/Basketballtopics(Figure6.4). Althoughonlinemedia'sfavoredtopicsresemblethoseofthenationalmedia,onlinemedia tendtopublishmoreaboutsocialissues,suchasImmigration(Figure6.5),SocialMedia, andSocial/Gender,whereasnationalmediatendtopublishmoreaboutinternationaltopics e.g.International/MiddleEast(Figure6.6)andInternational/FarEast(Figure6.7). Areversionarystrategyiscommonforfavoredtopics. Whenitcomestoindivid- ualorganizations'selectivenewslinksharingpatterns,themostcommonstrategyacross di˙erentmediatypesisa reversionarystrategy ,wherebynewsorganizationssharelessof aparticularnewstopicwhentheypublishalotaboutthetopic.Forexample,Figure6.5 showsthatnationalandonlinemediafrequentlywriteaboutImmigration,butsharelesson Twitter.Ontheotherhand,Figure6.8showsthatmanyregionalmediapublishalotabout Sports/Footballtopic,buttheytendtosharelessoftheirlinksonTwitter.Anaturalout- 74 Figure6.4:Newspublicationandlinksharingpatternsbynewsorganizations:Sports/Bas- ketballtopic. Figure6.5:Newspublicationandlinksharingpatternsbynewsorganizations:Immigration topic. 75 Figure6.6:Newspublicationandlinksharingpatternsbynewsorganizations:Internation- al/MiddleEasttopic. Figure6.7:Newspublicationandlinksharingpatternsbynewsorganizations:Internation- al/FarEasttopic. 76 Figure6.8:Newspublicationandlinksharingpatternsbynewsorganizations:Sports/Foot- balltopic. comeofthisstrategyistomakedi˙erenttypesofnewsorganizationslookalikeonTwitter, ersingbacktotheav Thisstrategyisreasonablewhennewsorganizationsdonotwanttobetoospecialized inagiventopicandreachabroaderaudiencebase.Forexample,althoughtheWallStreet Journalpublishesmoreabout˝nancethanothersdo,tooverwhelmgeneralsocialmedia userswiththeprofessionalanalysesof˝nancemaynotbeapro˝tableapproach.Giventhe users'limitedattention,theWallStreetJournalislikelytomissachancetoattracttheir attentionwithmorebroadlypopularnewstopics.Indeed,theWallStreetJournalobviously publishes 25% pointsmoreaboutEconomy/FinancetopicthanthebaselineABCinthedata, butitshares 16% pointsfewernewslinksaboutthesametopiconTwittercomparedtowhat itpublished.Asimilarexplanationcanbeappliedtoregionalmedia'sreversionarystrategy fortheSports/Footballtopic(Figure6.8).Althoughtheregionalmedia'sspecialtyisin regionalnews,andtheyarelikelytoexpectregionalfootballfanswhovisittheirwebsites, theymayhavedi˙erentincentivesonTwittertoreachouttobroadersocialmediausers.If 77 Figure6.9:Newspublicationandlinksharingpatternsbynewsorganizations:Crime. thisisthecase,regionalmediaareunlikelytosharetoomuchaboutregionalfootballgames. Inotherwords,aspecializedtopicwithoutbroadpopularappealislikelytogravitatetoward theaverageproportionbytheforceofthereversionarystrategyonTwitter. Theoppositedirectionofthereversionarystrategyisalsopossible.Inotherwords,when anewsorganizationisnotspecializedinagiventopic,itmaywanttosharemoretoattracta broadrangeofsocialmediausersiftheorganizationbelievesthatthegiventopicispopular. NationalandonlinemediatendtousesuchareversionarystrategyforCrimeTopic(Figure 6.9)whereasregionalmediatendtodosowithTransportationtopic(Figure6.10). Astrategicretreatisanothercommonstrategyfornon-specializedtopicsfor regionalmedia. Unlikespecializednewstopicsinwhichnewsorganizationsoftencounter- balancewithlesslinksharing,newsorganizationstendtoshareevenfewernon-specialized topics.Forexample,regionalmediapublishlessaboutInternational/MiddleEasttopics, andshareevenlessthanthatonTwitter(Figure6.6).Theyalsobehaveinasimilarmanner withthePolitics/Presidenttopic(Figure6.3).Thisisunderstandablebecauseifanews 78 Figure6.10:Newspublicationandlinksharingpatternsbynewsorganizations:Transporta- tion. organizationdoesnothaveenoughresourcestocoveragiventopic,andbroadsocialmedia usersarenotparticularlyinterestedinagiventopic,newslinksaboutthetopicarenotlikely toattracteitherthespeci˝ctargetgrouporthebroadnewsdemandonsocialmedia. Ontheotherhand,therewasnoevidencethatothermediatypesalsoadoptstrategic retreat.Thisisamorepredictableresultfornationalmediabecausetheytendtohave asymmetricallylargerreportinginfrastructurecomparedtoregionalmedia.Inotherwords, theremaynotbemanynewstopicsinwhichnationalmediaconceivealackofexpertiseto thelevelwheretheywanttoadoptstrategicretreat.Thereasononlinemedia,whichoften haveaslimitedareportinginfrastructureasregionalmediado,donotadoptthestrategic retreatismurkier.Onepossibleexplanationisthatonlinenewsorganizationssupplement newstheycannotdirectlycoverwiththereportingofothernewsorganizations,withor evenwithoutlegitimatepartnerships(Weber&Monge,2011).Inthiscase,theymaynot particularlyconceivetheirlackofexpertiseinacertainnewstopic. 79 Figure6.11:Newspublicationandlinksharingpatternsbynewsorganizations:Politics/- Congresstopic. Aconcentrationstrategyisusedforspecialized and populartopics. Thereare somehottopicsaboutwhichnewsorganizationspublishmore and sharemorelinks.For bothnationalandonlinemedia,Politics/Congressissuchahottopic(Figure6.11).Onthe otherhand,Social/Genderisahottopicforonlinemedia,butnotfornationalmedia(Figure 6.12).Thisdi˙erenceseemsrelatedtotheearlierpointthatonlinemedia'sspecialtyismore insocialissuesthannationalmedia's. Inthemeantime,regionalmedia'shottopicsarequitedisconnectedfromnational/online media.Figure6.9and6.4showthatregionalmediatendtobothpublishandsharemore ofCrimeandBasketballtopics.Again,thesearetopicsonwhichregionalmediahave traditionallyfocused,andwhichtheirregionalreadersarelikelytoexpectfromthem. Insum,theresultsofthestructuraltopicmodelspottedthreenotablenewslinksharing strategiesonTwitter:(a)reversiontotheaverage,(b)strategicretreat,and(c)concentra- tion.Choicesfromthesestrategiesdi˙erdependingonthelinkbetweentopicsandtypesof newsorganizations.Thereversionstrategywascommonwhenacertaintopicwasanews 80 Figure6.12:Newspublicationandlinksharingpatternsbynewsorganizations:Social/Gen- dertopic. organization'srelativespecialty(comparativeadvantage)e.g.Immigrationfornationalme- diaandSports/Footballforregionalmediabutthetopic'spopularitywasmodest.When newsorganizationsdonothaveaspecialtyinacertaintopic,andthetopicisnotparticularly popularfortheexpectedaudience,newsorganizationsseemtoretreatfromsuchnewstopics e.g.International/MiddleEastforregionalmedia.However,whennewsorganizationsare specializedinacertaintopic,anditispopular,theywillconcentrateonthosehottopics. Thecategorizationofthetopicsbythedi˙erentstrategiesissummarizedinTable6.2. Thenumbersinparenthesesindicatethenumberoftopicsthatfallintoeachcategory.As mentionedabove,thereversionarystrategyisthemostcommoncategory.Thereisgreat similaritybetweenonlineandnationalmediainregardtothereversionarystrategywhereas theregionalmediaaredisconnectedfromthetwo.Therearefewerhottopicsonwhicheach typeoforganizationsconcentrates.Again,regionalmediaaredisconnectedfromnational andonlinemediaregardinghottopics.Regionalmediaretreatfromafewtopicsthatare nottheirtraditionalfocusofcoverage.Itwashardertodetectdiscerniblepatternsfrom 81 Table6.2:Summaryoftheselectivenewslinksharingstrategies.Numbersinparenthesis arenumberoftopicscategorizedasagivenstrategy. ReversionRetreatConcentration Magazine International/FarEast Health Weather President(4) National CrimePolitics/Congress FinanceLegal/Regulation(2) Politics/President Weather Immigration Social/Gender Tax International/MiddleEast Sports/Others(9) Online CrimePolitics/Congress Life/InterviewSocial/Gender Politics/PresidentLegal/Regulation(3) Weather Immigration MeToo(6) Regional Sports/FootballInternational/MiddleEastSports/Basketball EducationPolitics/PresidentEconomy/Technology Life/InterviewSocial/IdentityCon˛ict(3)Update(3) Sports/Others RegionalPolitics Legal/Regulation Transportation(7) magazinesbecausefewermediaareincludedinthedatasetcomparedtoothertypes. 6.4Discussion The˝ndingsfromthepooleddatashowthelinksharingpropensityofanewstopicdoes notaccordwiththefrequenciesamongallthepublishedstories,con˝rmingthatnewsorga- nizations'linksharingisgovernedbyadi˙erentlogicfromtheonethatgovernstraditional gatekeepingdecisions.Traditionaleditorialdecisionsarethoughttobegovernedbyjour- nalisticnorms(Hermida,2010).Instead,therehasbeenaconcernthateconomicincentives drivetraditionalreportingtowardpopularity(Lischka,2014).However,theevendistribu- tionbetweenhardnewstopics,suchasSocial/IdentityCon˛ict,International/MiddleEast andLegal/Regulation,andsoftnewstopics,suchasLife/Interview,Crime,andSports/Bas- ketball,illustratesthatthemarketdriveisnotin˛uentialenoughtobanishthehardnews onnewswebsites. 82 However,thedeviationoftheselectivelinksharingforTwittere.g.highsharingpropen- sityofsportstopicsvs.lowsharingpropensityofpolitics/internationalrelationrelated topicsismoreconsistentwiththewidelyexpectedoutcomesofmarketdriveinjournal- ism'sgatekeepingfunction.TheconjecturethatgatekeepingonTwitterismoremarket driventhanthetraditionaloneisalsosupportedbyrecentobservationsfromwithinnews organizations.Thoughtheoutcomesofthenewslinkselectionhavesimilarmodalitytothose ofthetraditionalgatekeeping,itisnotalwaysano˚cialpartoftheeditorialdecisions,and anindependentsocialmediateamdecidesthelinkselectioninmanyorganizations(Rowan, 2014,January2;TandocJr,2014;Elizabeth,2017,November14).Thelogicbehindthis decisionresemblesthatoftheearlierSEOpracticesthatmainlyaimatmaximizingthepop- ularityofanewslink.However,theselectivenewslinksharingisamoreexplicitformof gatekeepingthanSEOinthatnewsorganizationshavefullcontroloverwhichnewslinks shouldbe invisible ontheirsocialmediaaccounts(TandocJr,2014).Thus,theselectivelink sharingislikelytoprovideapowerfulmomentumtowardpopularity-biasedgatekeepingon thetopofthetraditionalgatekeeping. Althoughthepooledanalysisofthetopic-wiselinksharingpropensityindicatesthatthe selectivelinksharingcarriesoutthelongstandingconcernsaboutthebiastowardsoftnews onTwitter,suchaconclusionmaystillbehasty,becauseofthedominanceofspeci˝ctypes oforganizationsinthedata.Forinstance,becausetherearemanyregionalmediainthe dataset(Table4.1),theirstrategiesmightdominatetheoverallpatterns. Theorganization-levelanalysesrevealmoredetailsofselectivelinksharing.The˝nding thatnewsorganizations'linkchoicedi˙ersacrossorganizationsshowsthattheircriteriafor selectivelinksharingcannotbereducedtotheaone-dimensionaldegreeofpopularity.The observationthatthesametypesofnewsorganizationsoccasionallyadoptacoherentselective linksharingstrategyimpliesthatinherentcharacteristicsofanorganization,aswellas popularityasanexternalfactor,arealsoasigni˝cantdeterminantoftheselectivelinksharing strategy.Itentativelycharacterizetheotherdimensionasa specialty ofanewsorganization. 83 Thepublishingpatternsoforganizationsseemtomakethenamingappropriate;national mediasigni˝cantlypublishmoreaboutinternationaltopicsandnationalpolitics,andregional mediapublishmoreaboutsportsandregionalissues.Iusethetermecialttomean anewsorganizationisgoodItisrelatedtonewsorganizations'capabilities,such asreporters'traininglevelsatdi˙erentexpertise,thenumberofreporters,theexpertiseof editorialsta˙s,connectionstocertainnewssources,etc.Althoughthesecapabilitiesare changeable,theywillcharacterizenewsorganizationsatleastinashorttimespanbecause theytendtoco-evolveslowlywithreaders'long-termexpectations. Overall,speci˝cselectivelinksharingstrategiesseemtobedeterminedbytheintersection ofthepopularityofanewstopicandanewsorganization'sspecialty.Themostconspicuous strategyfromthiscrosswasthe reversionarystrategy forspecializedand non -populartopics andthe concentrationstrategy forspecializedandpopulartopics.Whereasthereversionary strategyactsasa homogenizingmomentum fornon-populartopicsinthesensethatdi˙erent newsorganizationswillsharewithsimilarproportionsofagiventopicunderthisstrategy,the concentrationstrategycanbediversifyingbecausedi˙erenttypesofnewsorganizationswill sharemoreabouttheirspecializedhottopics.Additionally,non-specializedandunpopular topicswillbelessvisiblebythe strategicretreat .ThesepredictionsaresummarizedinFigure 6.13.Theoverallresultimpliesthat newscanbediverse(specialized)onsocialmediaonly forpopulartopics. Totakeanotablescenariobasedonthe˝ndings,nationalmediashare moreaboutpopular,nationalpolitics,andregionalmediasharemoreaboutregionalsports. Onthecontrary,newsorganizationsarelesslikelytosharespecialized,butlesspop- ulartopicsbythereversionarystrategy.ThosetopicsareImmigration,International/Far Easttopicsfornational/onlinemedia,andSports/FootballandEducationtopicsforregional news.Lesspopular,specializedtopicsmayincludetopicsthatarecomplicatedenoughto requireprofessionaltreatmenttocoverbutdonotstimulateinterestfrombroadaudience. Somearguethatcoveringthesetopicshelpsreaderseasilyunderstandthosecomplexissues andmakeinformeddecisions(Yankelovich,1991).Inthatsense,thereversionarystrategy 84 Figure6.13:Thegatekeepingmomentumgeneratedbytheselectivenewslinksharingstrate- gies.Topicsfallingontotheredareaswillbemorevisible,andtopicsintheblueareawill belessvisibleonsocialmediacomparedtotraditionaloutlets. maymeanasocio-technicalmomentumwhichweakenstheroleofjournalismthatinforms publicdecisions.Acomparisonbetweencivictopicsfallingintotheconcentrationandre- versionarystrategiessomewhatsupportsthisconjecture.TopicssuchasPolitics/Congress orSocial/Genderareindeedimportant,butvulnerabletosensationalcon˛ictiveframing whereasImmigrationandEducationlacktwoopposingsides˝ghtingagainstoneanother. Inotherwords,althoughnewssharedonsocialmediaincludesafairamountofcivicissues, thesharednewsislikelytohighlightthesensationalaspectofthoseissues. Themomentum,whichsocialmediaasanewsdistributionplatformgeneratesdoesnot seemfavorabletoregionalmedia.Whattheyconcentrateonarelikelytoberegionaltopics, butmostlycon˝nedtolocalsports.Manyauthorsrecentlydiscussthepotentialofthe regionalmedia'srevivalfromahyperlocalapproach(Metzgaretal.,2011;Lowrey,2012). Yet,regionalmedia'sselectivelinksharingstrategyshowsthattheyfocusonlocalsports, whichbroaderlocalaudiencesexpect,refrainfromlocalcivicissues,andgiveuponnational issuesforsocialmedia.Forsocialmediatohelphyperlocalnews,localmediashouldbeable 85 toreasonablybelievethatthereisabroadaudiencewhowouldratherpayattentiontothose issuesonsocialmediathanotherlocalissues.Thisdoesnotseemaneasytask. Selectivenewslinksharingasareactiontotheirenvironmenthasanimplicationforthe designofinformationcuratingalgorithmsonsocialmedia.Theconclusionthatnewsorga- nizations'selectivelinksharingisanoutcomeoftheiradaptationtoinherentandexternal situationsimpliesthattheywillalsoadapttopotentialchangesinnewsdemandonsocial mediacausedbychangesinanalgorithm.Thisadaptationprocessislikelytobecomequicker andsmoother,determinedbythedatascientiststheyhire(Rowan,2014,January2)andthe socialmediamonitoringplatformstheyadopt(Diakopoulos,2017).However,quickadapta- tionislikelytomakenewsorganizationsvulnerablebecausetheirdemanddependsonhow theplatforms'decisionchangesusers'newsreadingbehavior.Basedonthetwo-sidedmarket theory,economistspointoutthatthedependenceonmonopolizedplatformsmaysuppress independentcontentproviders,asnotablyintheMicrosoftcase(Evans,2003).Indeed,there isevidencethatFacebook'salgorithmchangegreatlyimpactednewsorganizations'online revenue(Brown,2018,April18).Therefore,thediscussionofadesirablealgorithmdesign shouldincludeconsiderationaboutthepotentialstrategicadaptationofnewsorganizations aswell. Finally,theresultsfoundinthischaptermaybesubjecttothetopicofpopularity cyclealthoughthedatasetincludesafairamountoftimesixty-threedays.Forexample, STMidenti˝esMeTootopicasseparatefromSocial/Gendertopic.Thus,verifyingthe conclusionsIdrawrequiresananalysisofalongertimeperiodbeyondtime-sensitiveissues. ThedatacollectionschemeandtheautomatedtextanalysisIapplyherecanbeusedas aninfrastructurefortheconstantmonitoringofnewsorganizations'selectivelinksharing strategy. 86 CHAPTER7 NEWSPARAPHRASINGANDSENTIMENT 7.1Introduction:NewsParaphrasingasaStrategy Thusfar,Ihaveconsideredonlythe`whethertoshare'decisionasanewsorganization's choiceinitssocialmediastrategy.However,thatisnottheonlypossiblechoiceanews organizationcanmakewhenitdisseminatesnewslinksonsocialmedia.Inthischapter,I focusonhowtheorganizationparaphrasesanewsstoryforasocialmediapost. Toincorporatethemultiplicityofnewsorganizations'choices,mysimpli˝edconceptual modelinChapter3canbeextendedtoamorerealisticproblemasfollows: max x i ˝ i ( x i ; x i ) Np i c i ( x i ) (7.1) Now x i isavectorthatcontainsnewsorganization'smultiplestrategicchoicesonsocial media.Andthefunction ˝ i ( ) isafunctionthatmapsthechoicestotheprobabilitythata newsorganization, i 'snewsstoryisreadbysocialmediausers.Asin3.1,theprobability isalsodeterminedbyotherneworganizations'choices, x i .Costofsocialmediastrategies c i ( ) isallowedtodi˙erforeach˝rminthismodel,whichimpliesasymmetricequilibria. However,unlessonehasastrongsimplifyingbeliefaboutwhattheplausiblechoiceset isandwhatthefunctions ˝ i ( ) and c i ( ) looklike,itishardtosetupasolvableformal modelbasedontheequation7.1.Instead,inthischapter,Iagainempiricallyexplorenews organizations'additionalchoicesnewsparaphrasingbyanalyzingtweetsembeddedinnews links. Thereisevidencethatnewsorganizationsarestrategicallyparaphrasingtheirnewsstories forsocialmediaposts.Althoughnewslinksharingisdeemedautomaticforsomeorgani- zations,consideringthesametextinsocialmediapostswasusedasheadlinesfororiginal storiesuntilafewyearsago(Palser,2009;Armstrong&Gao,2010),ArmstrongandGao 87 observedsomedi˙erencesacrossorganizations.Forexample,someorganizationsaddapart ofaleadtoasocialmediapostwhereasothernewsorganizationsuseonlyheadlines.More recentreportsshowthatnewsorganizationstakeonestepfurthertoadoptamanualcura- tionprocessoptimizedforsocialmedia.AccordingtoNewman(2011),in2011,theBBC switchedfromtheautomaticfeedtothemanualchoiceandeditingoptimizedforTwitter, andasaresult,thenumberoffollowersdoubled.Recognizingthepotentialofsocialmedia optimizationfornews,Facebookpublishedaguidelineforstrategicnewspostsbasedontheir ownuser-newsengagementstudies. 1 Therecentsurgeofonline-onlyemergingmediamakesthemanualsocialmediaopti- mizationmorecommonbecausetheydependheavilyontheviraldisseminationoftheir newsstoriesasamajormeansofnewsdistribution.AccordingtoarecentWiredarticle, BuzzfeedreliesonmachinelearningtechniquesandA/Btestingtooptimizeitsrsocialme- diastrategiesforvirality(Rowan,2014,January2).Similarly,Breitbart'sco-founderLarry Solovrecentlysaidthatthecompanyhasaseparatesocialmediateamandthattheyare verycarefulabouthowtheyportraytheirstoriesonsocialmedia(Kew,2016,August17). Indeed,Reisetal.(2015)foundthat51%oftweetswerephraseddi˙erentlyfromonline headlinesin5,182tweetstheycollectedfromfourmajornewsorganizations'accountsin 2014(TheNewYorkTimes,BBC,ReutersandDailyMail). Inthischapter,Ifocusonwhetherandhownewsorganizationsaddsentimentasthey paraphrasetheirnewsstoriesforTwitter.Althoughtheliteratureonnewspopularityhas generallyreachedaconsensusthatnegativeframingfornewscontentsandheadlinesincreases newspopularityorviralityonsocialmedia,researchersdonotoftendistinguishbetween negativeframingfromasocialmediastrategyandnegativeframingintheoriginalstory. However,thedistinctionisnecessarytolearnwhetherasocialmediastrategyindeedadds negativitytonewsframingseparatefromthetraditionalreporting.Thus,inthischapter, Icomparethesentimentofnewstweetswiththatoftheoriginalnewsstories.Iconjecture 1 https://www.facebook.com/notes/facebook-journalists/study-how-people-are-engaging- journalists-on-facebook-best-practices/245775148767840 88 thatnewsorganizationsconsistentlyadd extra-negative framingastheyparaphraseoriginal storiesbecausesocialmediastrategiesaremoreeconomicallydriventhanthetraditional editorialdecisions.Asmultipleexperimentstudieshaveexihibited,negativenewsframing increasesreaders'engagementandchangesopinions(Priceetal.,1997;Zillmannetal.,2004; Trussler&Soroka,2014).Thus,theextra-negativityasanoutcomeofsocialmediastrategies mayhaveasigni˝cantimpactonnewsreadersonsocialmedia. 7.2RelatedWorks 7.2.1ParaphrasingNewsonSocialMedia Textisoneofthemostsalientcomponentsofasocialmediapostthatcontainsamajor portionoftheinformation(Hu&Liu,2012).Andhownewscontentisparaphrasedlargely a˙ectsthepopularityofanewsstoryonsocialmedia(Horne&Adali,2017).Whereasmany socialmediapostslackvisualcomponents,virtuallyallofthemcontainsomeelementsoftext. Associalmediaisatextmediaaboveallthings,newstextsinsocialmediahavebeenoften analyzed(Kwaketal.,2010;Newman,2011).Textinsocialmediapostscanhavedi˙erent topics(Zhaoetal.,2011)orsentiments(Hansenetal.,2011)fromothertextsbecausea lengthlimitisimposedbysocialmediausers'cognitivecapacity 2 orbyatechnologye.g. 280characterlimitsonTwitter. Observationsthattextonasocialmediaposta˙ectsthepopularityofanewsarticle, andthatthetextisnotameresummaryofnewscontent,implyanincentiveforanews linksharertocomposeaparaphrasingtextinawaythatincreasesthepopularityofthe socialmediapostwiththenewslink.Previousstudiesdiscoveredafewtextfeatureson socialmediapostswithanewslinkwerepositivelyassociatedwiththeirpopularity.Hansen etal.(2011)foundthatnegativesentimentinasocialmediapostwithanewslinkwasmore 2 Theonlineadvertisingliteratureconsistentlyreportsthenegativecorrelationbetween lengthofadvertisingmessageandtheclick-throughrate(Baltas,2003;Robinsonetal.,2007). Inthecontextofsocialmedia,DeVriesetal.(2012)foundbrandpostswithlengthytexts onFacebookhasasmallernumberoflikesfromusers. 89 likelytoberetweetedwhereaspositivesentimentincreasesapersonaltweet'spopularity. Further,newssharersonsocialmediaoftenaimatpopularitywhentheyparaphrasenews stories.Horne&Adali(2017)foundthatmanynewssharersonRedditparaphrasenews storiesindi˙erentwaysfromoriginalnewsheadlines.LikeHansenetal.(2011),Horneand Adalialsofoundthatnegativesentimentandmoreemotionalwordsinaparaphrasepredict greaterpopularityofanewsstoryonReddit.Moreover,intheiranalysis,aparaphraseof newscontentpredictedthenewslink'spopularitywithmoreaccuracythanthenewscontent itself.Inotherwords,theassociationbetweenanewsparaphraseandnewspopularityon Redditwasnotnecessarilydeterminedbytheoriginalnewscontent.Thisdistinguishable andsigni˝cantpowerofnewsparaphrasingonsocialmediaislikelytoprovideanincentive forjournaliststomakestrategicchoiceswithittoattractmoreattention. Beingstrategicforparaphrasingisnotanewtaskfornewsorganizations.Foratraditional newspaperreport,journalistsmustcomposeheadlinesthatsummarizenewsstoriesthey publish.However,theheadlineisnotonlyameresummarizationofanewscontent,but alsothemain`hook'ofanewsstorytoreaders(Molek-Kozakowska,2013)similartonews paraphrasingforsocialmediaposts.Inthesamevein,Bell(1991)listedthemainpurposesof headlinesasa)summarizing,b)framingandc)attracting.Inotherwords,howthecontent issummarizedinheadlinesin˛uencesnewsreaders'choicetoactuallyreadthestory.This similaritybetweenheadlinesandnewsparaphrasingforsocialmediaimpliesthatjournalists' choicesmadeforheadlinesthathavebeenobservedinpreviousstudiescaninformusabout potentialchoicesmadeforsocialmediaposts. Althoughcomponentsinmainnewstextsthatincreasenewsdemandarestillnotablein thestudiesofheadlines,manyresearchersfocusonsensationalisminheadlines.Forexample, Tenenboim&Cohen(2015)conductedacontentanalysisonheadlinesof15,431onlinenews storiespublishedonanIslaeliwebsite,andfoundoutsensationaltopicsandcuriosityarous- ingelementsinheadlinesleadtomoreclicks.Alsoapplyingqualitativelinguisticanalysis, Molek-Kozakowska(2013)showedthatsensationalisminheadlineshasmultipledimensions: 90 illocutions(aimsofjournalists),themes,narrative,evaluation(valuejudgment)andproxim- ity.Theprevalenceofsensationalisminnewsheadlinealsopredictsstrongsentimentinnews paraphrasing.AccordingtoUribe&Gunter(2007)'sempiricalworkbasedonahand-coding approach,sensationalnewsaboutpoliticsandcrimeissuestendtocontainstrongsentiment. 7.2.2NegativeFraminginOnlineNews Intheirseminalstudy,McCombsetal.(1997)acknowledgedtheimportanceofsentimentin shapinghowtothinkaboutnews.Byconnectingsocialissueswithpositive/negative/neutral sentiment,newsreportinginformsreaders'judgmentaboutreal-worldissues(McComas& Shanahan,1999).Researchhasparticularlyfocusedonnegativityinnewsframing.Trussler &Soroka(2014)hypothesizedthatnewsreadershaveapsychologicalpreferencefornegative newsframingbecausenegativityisfurtherfromhumans'innatelypositiveexpectations,and isthusconsideredasignalformoreusefulinformation(Kahneman,1979).Indeed,Trussler &Soroka(2014)foundthatpoliticallynewsreadersprefernegativenewsframingunderan experimentcondition.Theysuggestthisresultasevidenceforthedemandsideexplanation oftheprevalenceofnegativityinpoliticalreporting.Builtonthisassessment,researchers haveappliedautomatedmeasurementsofsentimentinnewsstories.Forexample,using humancodersandmultiplelexicons,Young&Soroka(2012)foundthatnewsframingis consistentlybiasedtowardnegativity,particularlyforcrimeandforeignpolicytopics. Negativesummarizationhasbeenshowntohaveastrongimpactonnewsperception. Bymanipulatingheadlinesandleads,Priceetal.(1997)foundthatexperimentsubjects' opinionandemotionalvalenceonfundingtopublicuniversitieschangedependingonhow thesameinformationisframed(con˛ict/humaninterest/consequence).Furthertheyfound thatthechangedopinionhadanimpactondecisionmakingaboutthepolicy.Witha similarexperimentdesign,Zillmannetal.(2004)showedthatexperimentsubjectsspend moretimereadingnewsstorieswhentheyareframedascon˛ictorvictims'agonyrather thanmisfortuneoreconomicloss.Analyzinganaggregatetimeseriesofpublicopinion 91 andsentimentinnewsoneconomicissues,Soroka(2006)arguesthatnegativesentiment containedinnewsstorieslargelya˙ectsopinionchangenotonlybecausemediacoverage tendstobenegativeratherthanpositive,butalsobecausenegativeinformationtendsto haveagreaterimpactontheimpressionsofnewsreaders.Thesestudiesimplythatnews paraphrasing,suchassocialmediafeedshasthepowerofextra-framing,ontopofframing inamainbodyofnewsstories. Althoughnewsorganizations'paraphrasingonsocialmediahasnotbeendirectlyana- lyzed,˝ndingsfrompreviousworksonnewsdemandshowthatnewssentimentandpopular- ityarecloselyrelated.First,mediastudiesgenerallyagreethatthenegativesentimentina newsstoryispositivelyassociatedwithanewsdemandinatraditionalnewsreadingsetting. Newhagen(1998)conductedanexperimentinwhichsubjectswatchnewsimagesthatevoke negativeemotion(anger,fearanddisgust).Hefoundthatimagesthatevokeangermost increasesubjects'approachtotheimages,followedbyimagesthatinducefearandimages thatinducedisgust.Healsofoundthattheimagesthatinducewerethemostmemorable. Heexplainsthisresultwithabiologicalde˝nitionofanger,whichisthephysiologicalresult ofterritorialviolation.Thisresultimpliesthatacertainemotioninanewsstorycaninduce newsreaderstomore,whichmayinturna˙ectdemandforthenews.Indeed, Trussler&Soroka(2014)foundthatapoliticallyinterestedpersonislikelytoreadapolitical newsstorywithanegativetoneandthatframespoliticsasagameratherthaninformative newsthatusesaneye-trackingexperiment. Yet,someresearcherspointoutthatthegeneralityofthenegativeframingshouldbe understoodwithcare.Inparticular,thepopularityofsuchastrategyanditse˙ectiveness candi˙erdependingonnewstopics.Forexample,focusingonlyonhealthrelatednews, Kim(2015)foundthatonlypositivesentimentinhealthnewsispositivelyassociatedwith itslikelihoodtobereadandsharedonsocialmedia.Similarresultswerefoundfornews paraphrasing.Bymeasuringthesentimentscores(valence)of69,907topnewsheadlines fromfourmajornewswebsites(BBCNews,DailyMail,theNewYorkTimesandReuters), 92 Reisetal.(2015)foundthatthemajorityoftheheadlineshasnegativesentimentoverall, butthelevelofthedominancevaries,dependingonnewstopics.Forexample,newsstories fallingunderatopiccategory,mostfrequentlyhavenegativeheadlines.However, storiesaboutsportsorscience&technologytendtohaveneutralheadlinescomparedto othertopics.Inthecontextofmystudy,thisimpliesthatthenegativeframingfornewsand socialmediawilldi˙eracrossnewsorganizationsdependingonthenewstopicstheymostly cover. Further,thepotentialnegativityfromnewsparaphrasingislikelytobecleareronsocial media,whereeconomicincentivestoelicitmoreclicksdominates,comparedtotraditional headlinemaking,whichisapartoftraditionaleditorialdecisions.Intraditionalnewscon- sumptionwherenewsstoriesareconsumedasabundle(e.g.anewspaperorTVnews program),headlinessignalnewscontenttocompetewithotherarticleswithinthesame bundle.Thus,theheadlineshadtobewrittentoshowastory'srelevancetoareader(Dor, 2003).Howeverintheonlineenvironment,newsreaderstypicallynavigateacrossdi˙erent newssources.Asaresult,attractingnewsreaders'attentionwithaheadlinebecameamore challengingtaskbecauseitalsoneedstotakeaccountofcompetitionwithothersources beyondthegivenbundle.Inthiscircumstance,newsorganizationshaveastrongincentive tosignaltheirstories'appealtoreaders.Moreoveronsocialmedia,thissignalingincentive islikelytobeevenlargerbecauseusers'attentioniseasilydistractedbyamixtureofvar- ioustypesofinformation.Indeed,researchershavedetectedtheproliferationoftheuseof sensationalandparticularlynegativeexpressionsinonlinenewsheadlinesfor click-bait .For example,Chakrabortyetal.(2016)andPotthastetal.(2016)showedsentimentpolarity (valence)canbeusedasapredictorofanautomatedclick-baitdetectionalgorithm. Allinall,newsorganizationsarelikelytoframetheirnewsnegativelytoattractreaders. Newsparaphrasingforsocialmediawillbemorenegativebecauseitismoregovernedby theeconomicmotivationcomparedtotraditionaleditorialdecisions.Thus,Ihypothesize thatnewsorganizationsapply extra-negativeframing tonewsastheyparaphraseforsocial 93 media.Testingthishypothesisrequiresthecomparisonofthesentimentofnewstweetswith thesentimentoftheoriginalnewsstories.Inthischapter,Iwillconductthecomparison usingthreedi˙erentdictionariesA˚n,LSDandLIWC.Further,theliteratureimpliesthat di˙erenttypesofnewsorganizationsthatcoverdi˙erenttopicsarelikelytoshowdi˙erent levelsofnegativeframing.Iwilltestthisconjecturebycomparingtheextra-negativeframing acrossdi˙erenttypesoforganizations. 7.3Results Inthissection,Ipresentresultsfromthedictionary-basedsentimentanalysis.Measured sentimentsofnewscontentthateachnewsorganizationpublicizesonitsnewswebsiteand Twitteraccountwillrevealthattheextra-negativityfromparaphrasingfortweetsisconsis- tentacrossdi˙erentorganizations.Subsequently,usinganinferentialanalysis,Iwilldiscuss howtheextra-negativitydi˙ersdependingontypesofnewsorganizations.Alltheanalyses inthissectionwereiteratedusingthethreedi˙erentdictionariesA˚n,LSDandLIWC tovalidatetheresults.ResultsomittedtoavoidrepetitionareincludedinAppendixD. 7.3.1NegativityofNewsbyOrganizationalTypes Thepreviousliteraturehasfoundgenerallythatnegativityinanewsstorytendstoincrease itsdemand,andnewsorganizationstakeadvantageofthistendencytoattractmorereaders. Thisimpliesthattweetsaboutnewsstoriesarelikelytohaveasystematicbiastoward negativeframing.Althoughresultsinthecontextofonlinenewsconsumptionareyetlimited andmoredebatable,thereismuchevidencethatnegativityinsocialmediapostsattracts moreattention.Further,Iconjecturethatnewsorganizations'socialmediastrategiesare largelygovernedbyeconomicincentivesstemmingfromcompetitionforlimitedattention. Ifthisconjectureholds,paraphrasesofnewsstoriesforTwitterarelikelytocontainmore negativesentimentcomparedtooriginalstories,whicharepresumablygovernedmoreby journalismnorms. 94 Figure7.1:Averagesentimentscoresofnewsstoriesandtweets:LIWCdictionary. Figure7.1presentsresultsofthesentimentanalysisbasedontheLIWCdictionarybroken downbynewsorganizations.Theredbarmeanstheaveragesentimentscoresofnewsstories andthebluebarrepresentstheaveragesentimentscoresoftweetswithembeddednewslinks totheoriginalnewsstories.Thescore50indicatesneutralsentimentforLIWC.Theresults showthatthetoneofnewsstoriesisindeedconsistentlynegativeacrossdi˙erentnews organizations.Inparticular,whereasregionalnewsmediaoccasionallyhavepositivetones, onaverage,mostofothertypesofnewsmediahavenegativetones.Moreover,thisresult isrobustacrossallthreedictionaries. 3 Excludingregionalmedia,themediawithpositive tonesinclude;Forbes,CNBC,USAToday,InquisitrandVanityFair.Theseseemrelatedto aspeci˝cfocusofmedia,suchas˝nanceorentertainment,andthenonpartisanorientation itpursues. Tocheckifthispatternholdsconsistentlyacrossdi˙erentdictionaries,Figure7.2visual- izestheaveragenewssentimentofeachtypeofnewsorganizationsusingthreedictionaries, 3 ResultsbasedonA˚nandLSDdictionariesareFigureD.1andFigureD.2inAppendix D. 95 Figure7.2:Sentimentlevelofnewsstoriesbynewsorganizationtype. A˚n,LSDandLICW.Indeed,nationalandonlinenewsmediahavethemostnegativetones whereasregionalmediahavetheleastnegativetonesregardlessofdictionaries.Theranks betweennationalandonlinemediadi˙erdependingonthedictionary,butthedi˙erence betweenthetwoismuchsmallerthanthedi˙erencewiththeothertwotypes.Theonly signi˝cantdi˙erenceacrossdi˙erentdictionarieswasthattheanalysisbasedonA˚nresults inpositiveaveragetonesofmagazinesandregionalpapersratherthannegative. 7.3.2StrengthofExtra-NegativeFraming Thenewsdisseminationthroughsocialmediamayraiseyetanotherlayerofnewsnegativity issuetheextra-negativityofnewsparaphrasingforsocialmedia.Tomanysocialmedia userswholookonlyatthenewsparaphraseswithoutclickingthelink,thepotentialextra- negativitycoupledwithinsu˚cientinformationaboutthenewstopicmaygreatlya˙ectusers' newsperception.Ifthesocialmediastrategyislargelygovernedbyeconomicincentives comparedtonewspublishing,theextra-negativityonTwitterislikelytooccurevenfor regionalmedia,whichhadmoderatesentimentontheirwebsites. Again,Figure7.1showsthenegativityoftweets(bluebars)measuredbytheLIWC 96 dictionary. 4 Formostnewsorganizations,bluebarsarelowerthanredbars,whichmeans thattweetshavemorenegativesentimentthantheoriginalnewsstories.However,thereare someexceptions,particularlyamongonlinemediasuchasBreitbart,ConservativeTribune, andWesternJournalism,whichareallknowntobeconservativepartisannewsoutlets.A notableaspectoftheseexceptionsisthatthesenewsorganizationstendtohaveevenmore negativetonesontheiroriginalnewsstoriesthanothers.Thus,thelackofextra-negativity maybeonlybecausetheiroriginalnewsinformationisalreadyverynegative,andnotbecause theydonotframetweetsnegatively. Asimpleregressionanalysiscanextendthisvisualanalysisbasedonaveragestoanews storylevelanalysis.Inaddition,theregressionallowsforexplicitlyrevealingwhetherdif- ferenttypesofnewsorganizationsaddtheextra-framingtoadi˙erentextent,whichisnot easytodemonstratewithavisualanalysis.TheregressionmodelIranhasthesentiment gapbetweenatweetandanoriginalnewsarticleasadependentvariable,whichrepresents theextra-negativityanewsorganizationaddsasitparaphrasesitsownstoriesforatweet. Thegapisexpectedtobeositivwiththeextra-negativity.Havingthesentimentofthe originalarticleasacontrolvariableallowsfordistinguishingtheextra-negativeframingfrom negativityinanoriginalstory.Theunitofanalysisiseacharticle-tweetpair. Table7.1presentstheregressionresults.Theresultsappeartobeconsistentwiththe expectations;morenegativenewsstorieshavelessextra-negativityfortweetparaphrasing becausetheyarealreadyframednegatively.AlthoughA˚nmeasureresultsintheopposite sign( 0 : 0446 ),itwasnotstatisticallysigni˝cant.Also,itshouldbenotedthatthe R 2 is extremelylowfortheA˚n-basedandLSD-basedregressionsalthoughLSDresultedina signi˝cant˝nding.Themainreasonforthisisthattheypickuplessvariationfromthegap betweenanewsstoryandatweetbecausetheyincluderelativelimitednumberofwordsin thedictionariescomparedtoLIWC.Althoughthesizeofthedictionariesislargeenoughto pickupthevariationfromaggregatedtext,theytendtomissinformationwhentheyare 4 Again,resultsconsistentwiththosebasedonotherdictionariesarereportedinFigure D.1andD.2inAppendixC. 97 Table7.1:Linearregressionofsentimentgaponnewssentimentandnewsorganizations type. VariableA˚nLSDLIWC Intercept0.4771***0.2575***-15.3137*** (0.0293)(0.0124)(0.3809) NewsSentiment-0.04460.0751***0.5853*** (0.0082)(0.0078)(0.0028) National-0.0928***-0.0462***-2.5787*** (0.0310)(0.0132)(0.3825) Online-0.0514**-0.0287*-2.6715*** (0.0327)(0.0138)(0.4007) Regional-0.1601***-0.0739***-2.0643*** (0.0311)(0.0132)(0.3815) R20.00150.00170.2110 N74,93190,593166,159 todetectsmallinformationfromagapbetweenasinglenewsstory-tweetpair.Thus,the regressionbasedonLIWCseemsmorereliableforthisanalysis. Evenaftercontrollingthesentimentoftheoriginalstories,regionalnewsmediatend toaddmorenegativityastheyparaphrasetheirnewsstoriesfortweetsthannationaland onlinemedia( 2 : 0643 > 2 : 5787 > 2 : 6715 ). 5 However,thenationalmediaandonline mediashowasimilarlevelofextra-negativeframing.Thismeansthatalthoughmagazines andregionalmediaarerelativelypositiveontheirwebsites,theycatchupwithonlineand nationalmediaonTwitterbyaddingmorenegativityonnewsparaphrasing.Inotherwords, newsparaphrasingforTwitternegativelyhomogenizesdi˙erenttypesofmedia.Thisis visualizedinFigure7.3wherethesentimentsarenottoodi˙erentcomparedtoFigure7.2. 7.4Discussion Newsparaphrasinghasalwaysbeenanessentialpartofthenewsmakingprocess,andits framingimpacthasbeenproventobesigni˝cant.Althoughheadlinemakingtraditional paraphrasinghaslongbeenaccusedofbeingsensational,itisapartofthetraditional 5 Notethatthisresultmeansthebaselinecategory,magazine,hasthelargestgap,which isexpectedtobepositivewiththeextra-negativity. 98 Figure7.3:Sentimentleveloftweetsbynewsorganizationtype. Figure7.4:Additionalnegativesentimenttotweetsbynewsorganizationtype. editorialdecision,likelytobegovernedbyjournalisticnorms.Ontheotherhand,recent testimonieshaveshownthatasimilartaskofparaphrasingforsocialmediaisoftenout- sideofnewsroomElizabeth(2017,November14);Roston(2015,January22),andlargely reliesonnumbersfromaudiencemetricandautomatedalgorithms(TandocJr,2014,2016). Consideringtheobservationthatsocialmediausersoftenreadonlytheparaphrasednews withoutclickingembeddedlinks(Horne&Adali,2017),itisreasonabletoworryaboutthe impactoftheadditionalmechanismwherebynewsorganizationcanimposenewsframing onsocialmedia,andmonitoritasaquasi-editorialdecision.Bycomparingsentimentsof 99 tweetsbydi˙erenttypesofnewsorganizationsandoriginalnewsstories,Ifoundevidence thattheincentivesfromthedemandofsocialmediausersledorganizationstoplacemore negativeframingonnewscomparedtotheoriginalstoriesinthischapter. ThereportedresultsaregenerallyconsistentwithconjecturesImadebaseduponthe literature.(a)Onlinenewsstoriesarenegativelyframedoverall,and(b)regionalmedia likelytocoverlesscontroversialissuesandframenewsinasigni˝cantlylessnegativeway comparedtonationalandonlinemedia.(c)Socialmediapostsarelikelytobemoredriven byaudience'sreceptionthanjournalisticnormsandareconsistentlymorenegativelyframed thanoriginalstoriesacrossdi˙erentoutlets.(d)Inaddition,regionalmediasigni˝cantly catchuptoothertypesofmediawithextra-negativeframingforTwitter. Thenegativenewsframingofnational/onlinemediaintheresultshasthepotentialto a˙ectthenewsperceptionofonlinenewsreaders,asscholarshavepointedout(Newhagen, 1998).Althoughregionalmediasupplieslessnegativeframing,itmightbelimitedtocertain newstopics,andmanyreportevidenceoftherapiddeclineofregional/localjournalismin theriseofonlinejournalism(Curran,2010;Nielsen,2015).Althoughsomeseepotential fromtheInternettorebuildastronglinkbetweennewsreadersandthelocalcommunity (Lowrey,2012),recentevidenceshowsthatsocialmediatendstofunctionasamomentum towardthecentralityofnationallegacymediaornation-wideemergentformsofonline-only media(Hodson&Lindgren,2017).Ifregionalmediaisindeedlosingthecompetitionfor limitedattentiononsocialmedia,users'newsperceptionmaybecomenegativelybiasedas anincidentalimpactofthenegativityofnational/onlinemediaobserved. Giventhatsocialmediausersreadonlyparaphrasednewsmostoftimewithoutclicking embeddednewslinks,theextra-negativeframingforsocialmediaexacerbatessuchimpacton newsperception.Furthermore,newsorganizations'sensitivitytomarketdemandonsocial media,enhancedbydata-drivensocialmediamonitoringtechnologies(Diakopoulos,2017), createsaconditionvulnerabletoafeedbackloopspiralingtowardnegativenewsframingon socialmedia.Thenegativityinnewsparaphrasinginducesnegativenewsperception,which 100 inturnpotentiallymakessocialmediausersdemandmorenegativeframing. Theobservationthatregionalpapersarecatchingupisconsistentwiththeconjecture thatnewsparaphrasingforsocialmediaismoredrivenbytheaudience'sreceptioncompared totraditionaleditorialdecisions.Ifjournalists'framingfornewsstoriesismoregovernedby theirtraditionaldecisionaboutnewsworthiness,regionalnewshasareasontohavemoderate sentimentfromtheirless-controversialregionaltopicsthannationalandonlinemedia.On theotherhand,ifparaphrasingisgovernedmorebypro˝tabilityfromsocialmediausers' clicksandengagement,thereisnoreasonthatsocialmediapostsfromregionalmediaare lessnegativelyframedthanthosefromothertypesofmedia.Inthiscase,thegapbetween thetwosentiments(i.e.theextra-negativeframing)shouldbemuchlargerforregionalmedia asobservedinthepresentedanalysis. Althoughnewsreadersareincreasinglysusceptibletosuchmarketdrivennegativefram- ingonsocialmedia,theyhavebeenrelativelyfreefromnormativediscussions,suchasan evaluationofitsimpactandwhatweshoulddoaboutitasasociety.Althoughsensation- alismoftraditionalheadlineshasfrequentlyraisedresearchers'concern(Chakrabortyetal., 2016;Molek-Kozakowska,2013;Blom&Hansen,2015),thetraditionaldecisioncouldstill resorttothepossibilityofself-regulationbasedonjournalists'valuesfromtheirlong-term trainingandexperienceasjournalists.However,itishardertoexpectthatnewsparaphras- ingforsocialmediawillbecontrolledunderasimilarself-regulationbecausesocialmedia strategiesaremorelikelytobesubjecttoeconomicprinciples.Ifthedominanceofaudi- ence'sreceptionfornewsparaphrasingisthecase,thentheresultingparaphrasesarelikely tobehomogenizedtore˛ectsocialmediausers'newsdemand.Surely,itisstillpossiblethat diversityinsocialmediausers'newsdemandwillinduceadiversityofparaphrases.However, theempiricalresultsinthischaptershowthatframingfromnewsparaphrasingonTwitter iscurrentlyundergoingaratherhomogenizingmomentum,disconnectedfromoriginalnews stories.Thisraisesaneedforthepublicmonitoringofparaphrasingonsocialmediaasnews organizations'quasi-editorialdecisions. 101 Regardingpublicmonitoringofsocialmedia,thereportedresultssuggestthatLIWC showsabetterperformancethantheothertwodictionaries.Themainreasonseemsthat agenerallypurposedLIWCdictionarycontainsalargersetofwordsand˝ner-groundsen- timentscoresassociatedwiththewordswhereasLSDandA˚naredesignedforaspeci˝c domainLSDfornewsandA˚nfortweetsandcontainamuchshorterlistofwords. LSDandA˚ncontainenoughinformationtopickupaveragesentimentfromlargesample oftextasIcon˝rmedfromtheconsistencyofnewssentimentacrossdi˙erentdictionaries. However,theydonotseemtohavesu˚cientinformationtodetectlimitedinformationfrom thesentimentgapbetweenatweetandanewsstory.Interestingly,thisconclusionsome- whatcontradictsthepreviousdictionary-basedsentimentanalysisthatcalledforaneedfor publicmonitoring.Forexample,Young&Soroka(2012)arguedthatmonitoringnewsneeds anews-customizeddictionary.However,thepresentsettingthatcomparesnewstotweets resultsintheoppositeconclusion,i.e.thatweneedageneraldictionary.Itseemsthat thereisatrade-o˙betweengeneralizabilityandcustomization.Toincreasetheprecisionof sentimentinformationfromone(each)typeoftext,itisbettertocustomize.However,to comparetwodi˙erenttypesoftext,suchasanewsstoryandatweet,itisbettertohavea generaldictionary. 102 CHAPTER8 CONCLUSION:SOCIALMEDIASTRATEGIESASQUASI-EDITORIAL DECISIONS AlthoughInternettechnologiesprovidestrongmomentumtowardcommunication inasensethatusers,whousedtobethemereaudienceofinformation,canbringtheirvoices intojournalists'workviathenewvehicles(Singer,2009),thenewsmaking-distribution processisstill structured bypowerfulactors.Informationplatformowners,suchasFacebook andGoogle,nowhavesigni˝cantcontroloverthepriorityandrelevanceofinformationfor individualusers(Pariser,2011).Newsorganizations,albeitweakened,retainaprivileged positionovernonprofessionaljournaliststoaccesspubicinformationbecauseoftheirorga- nizedreportingandeditorialinfrastructure,˝nancialcapital,andhistoricallyaccumulated connectionstoinformationsources.Journalismisanetworkcomposedofmultiplehuman actorsandtechnologies(Turner,2005).Yet,itisanetworkstructuredbypowerfulactors whopossessthemeansofproductionanddistribution. Asananalysisofonekindofsuchpowerfulactors,Iinvestigatednewsorganizations' socialmediastrategiesasanemergent quasi-editorialdecision ,focusingonnewsdissemi- nationviaTwitter.Iconsiderthenewsdisseminationstrategiesonsocialmediaasaform ofthenewsorganizations'adaptationtosocio-technicaltraitsofthenewnewsdistribution platform.Drawingonrecentreportsthatnewsorganizations'socialmediastrategiesisdis- connectedfromtraditionaleditorialdecisionsittermsofpersonnel,organizationalstructure anddecisionprocessesRoston(2015,January22);Elizabeth(2017,November14);TandocJr (2014,2016),Itestedwhetherthisdisconnectionindeedproducesdi˙erenceinvisiblenews topicsandframingbetweennewswebsitesandorganizations'Twitteraccounts.Basedon thiscomparison,Iarguedthatthedi˙erentmechanismswherebythe quasi-editorialdecisions forsocialmediaaremadecanimposeanadditionallayerofinformation-mediatingprocess, relativelyfreefromjournalisticnormsandroutines,asreaders˝ndnewsinformationfrom 103 socialmediamoreoften. Monitoringsuchaquasi-editorialdecisionisnotasimpletask.Thesocialmediastrate- gies,asanadaptationprocess,arestillin˛uxpartiallybecauseoftheinertiaintheolder regimeofjournalism,andpartiallybecauseoftheceaselesslychangingtechnologies.Further- more,myresearchquestions,whichinvolvetwodi˙erentformsoftext,newsandtweetsfrom arapidlygrowingnumberofnewsorganizations,makethemanualanalysishardlyfeasible. Fortunately,thecomplicatedandeverydaymutatingonlineinformationmediatingprocess isrecordedontheInternetinanalmostcompleteform.Thus,Iproposedacomputational systemofdatacollectionandautomatedtextanalysisfortheconstantmonitoringofnews organizations'socialmediastrategies.Thisapproachwillexpandourfrontierof (Schudson,1998)overthepublicinformation˛owtowardcommunicationthat ishappeningviasocialmedia,anemergentnewsdistributionplatform. 8.1NewsLinkSharingasaStrategicChoice Afundamentalquestionbeforegaugingnewsorganizations'socialmediastrategiesin- cludes, Isthereareasonforanewsorganizationtobestrategiconsocialmedia? Iusethe term,basedontheeconomicde˝nition,torefertoasituationwhereeveryactor's decisiondependsononeanother's.InChapter3,Ifoundareasonforapositiveanswerto thisquestionfromsocialmediausers'limitedattention.IntheeconomicmodelIsuggested, thescarcityofusers'attentioncapacityrelativetothevolumeofinformationpropagated onsocialmediacreatescompetitionbetweennewsorganizations.Thismodelimpliesthat users'attentioncanbeconceivedasa publicresource ,whichisasourceofeconomicbene˝t thatanyonecanaccesstoexploit,butwhoseamountislimited(Ostrom,2015).Asmost publicresources,theoutcomeofmymodelrevealsthatonenewsorganization'sattempt tocaptureusers'attentionunderminesachanceforotherorganizationstodoso.Inother words,sharingmorenewslinkstoattractsocialmediausers'clickeatsupusers'attention, resultinginlessopportunityforthelinksofothernewsorganizationstobeinusers'aware- 104 ness.Thiscompetitionresemblesasituationinwhicheveryoneisshoutingtobeheard. BecauseIshout,apersonnexttomealsohastoshouttobeheard(Anderson&DePalma, 2013).Becausepeopleshout,Ihavetoshoutevenlouderinturn.Bysimilarlogic,anews organizationshould strategically decidehowmanynewslinksitisgoingtoshareconsidering howmany others wouldshare.Althoughrecentmediastudieshavefocusedontheimpact ofinformationoverload,asakindofbounded-rationality,onindividualinformationchoice behavior(Liang&Fu,2016),competitionforalimitedattentionmodelshowsthatnewsor- ganizations'adaptationtosuchhumanboundednesscanalsogiverisetonewcommunication patternsinanemergentmediaenvironment. Aswithmostpublicresources,thepredictedequilibriumofthemodelisthatthelimited attentionofsocialmediausersisover-exploited.Roughly,newsorganizationswillshare toomanynewslinkstobeatotherorganizations'links.Becauseeveryonetriestobeatone another,therewillbesomanynewslinksonsocialmediathatuserswillbeabletoprocess onlyasmallportionofthesharednewslinks.However,ifnewsorganizationscooperated jointlytomaximizetheirpro˝tsratherthancompete,theywoulddisseminatefewerlinksto saveusers'attention. Thisresulthasanimplicationonthediscussionaboutpublicaccountabilityofsocial mediaplatforms,whichwasdramaticallyescalatedbytheCambridgeAnalyticacase,in whichtheprivateinformationofusersthatwasretainedbyFacebookwaspoliticallyabused. ThefollowingcontroversyastoresponsibilityofFacebookandotherInternetplatformshas reachedanagreementthattheplatformowners'self-prescriptionwillnotfundamentally resolvetheriskofsimilartroubles(Vasuetal.,2018).However,giventhatsocialmedia ownershavebeenconsideredtechnologycompaniesrelativelyfreefromsocialaccountability, itisdi˚culttoestablishagroundingprinciplethatallowsforpublicengagementwiththe informationplatforms. Theconclusionofmymodelthatusers'attentionisapublicresourcecano˙ersuch aprinciple.AsOstrom(2015)argues,governanceofapublicresourcesusceptibletoover- 105 exploitationisapartofthekeyinfrastructureforsuccessandsustainabilityofdi˙erentlevels ofcommunities.Becausedemocracydesperatelyneedspublicattentiontocivicissues,as Berger(2011)pointsout,andifattentioncanbeperceivedasapublicresource,companies thatpro˝tfrom(re)distributingusers'attentionwouldhavesocialresponsibilityforthe commercialuseoftheresource.Thisargumentmaybeusedtowarrantpubliclygoverning onlineplatformstopreventoveruseorabuseofattentionthescarceandvaluableresource ofsociety.However,althoughthenaturaltimeislimited,humanattentionmaybeextended bytechnologies,suchasnewdevicesandinformationcurationalgorithms,partiallyprovided bytheplatformcompanies.Thus,towhatextentusers'attentionisalimitedresourceisan empiricalquestionthatcallsforfurtherresearch. Whetherattentionisover-exploitedishardtotestbecausethe˝rms'collectivelyoptimal useofattentionthetheoreticalcomparisontargetisnotobservablefromthedata.Yet, themodelstillproducesatestableprediction:organizationswillreducetheproportion ofnewslinkstheyshareonsocialmediaasmorestoriesarepublishedbyallorganiz Thispredictionresonateswiththepropositionthatnewslinksharingis`strategic'inthe sensethatoneorganization'sdecisionreliesonothers.TheregressionresultsinChapter5 basedonnewslinkssharedonTwitterareconsistentwiththisprediction.Theyarealsoin accordwithrecentreportsthatnewsorganizationshavehiredsocialmediaspecialistsand adoptedsocialmediamonitoringplatformstooptimizetheirsocialmediastrategiestothe reactionsofbothusersandcompetitors(Kew,2016,August17;Rowan,2014,January2; Diakopoulos,2017).Astechnologiestheyadoptforsuchmonitoringdevelop,sensitivityof theirreactionisexpectedtoincreaseinthefuture. Realistically,newsorganizations'strategicchoicemaynotbecon˝nedtodecisionsabout thenumberofnewslinktheywillshare.Forexample,newsorganizationsmaywantto concentrateonsomepopularnewstopics,whentheyshrinktheproportionofsharednewsas areactiontothelargevolumeofpublishedstories,asthemodelpredicts.Insteadofselecting newslinkstoshare,theymaywanttoadoptdi˙erentcontentandsignalingthisdecision 106 bymanipulatingtextonsocialmediaposts.However,neitherthemodelnorthesimple regressionabouttheaggregatedproportionofsharednewslinksallowsforinvestigating thosequalitativedimensionsofsocialmediastrategies.Thischallengeraisesaneedtotake adeeperlookattherelationshipbetweennewscontentandnewsorganizations'choiceson socialmedia. 8.2SelectiveNewsLinkSharingasGatekeeping Anewsstory,asaninformationgood,hasmultiplequalitativecharacteristics,suchas length,topic,factuality,opinion,etc.Basiceconomictheorydictatesthatanewsorganiza- tion,likeaneconomicentity,naturallyconditionsitsstrategyaboutthesecharacteristicsto maximizepro˝ts.Inthecontextofnewsdistributiononsocialmedia,thiseconomicpre- dictionshouldmanifestitselfas selectivenewslinksharing conditioningthelinksharing decisiononnewscontent.Whetheranewsorganizationsharesaspeci˝cnewslinkonsocial mediahasapotentiallysigni˝cantimpactonthevisibilityofanewsstory.Userswillbemore likelytobeawareofthenewsstory,andhaveeasieraccesstoitwhenthereisahyperlinkto it.Inthissense,hyperlinkscanfunctionasanewformofinformation`gates.'(Dimitrova etal.,2003). Theoutcomeofthelinksharingdecisiondi˙erentiatedaccessibilitytonewsisclose tothetraditionalgatekeepingbyjournalists.Bothtasksdecideaccessibilitytocertainin- formationbycontrollingitsvisibilityonamedium.However,principlesgoverningthetwo decisionsmaybedi˙erent.Traditionalgatekeepinghasbeenarguablyconceivedasgov- ernedbyjournalisticnorms.Relyingonjournalismnormsandvalues,journalistscarefully decidewhatinformationisworthcitizen'sknowing.Yearsoftrainingforjobethicsbased onprofessionalismandaccumulatedexperienceforobjectivereportinginform`newsworthi- ness.'Ontheotherhand,newsorganizations'competitionforusers'attentionimpliesthat newslinksharingismorelikelytobegovernedbyeconomicincentives;themorelikely usersaregeneratingadvertisingpro˝t,thebetter.Accordingtorecentreports,thedecision 107 makingforsocialmediapostingisoftenseparatefromtraditionaleditorialdecisionsorga- nizationallyandpersonnelwise,andisoftenmoreattachedtoamanagementsidewithinan organization(Roston,2015,January22;Elizabeth,2017,November14).Thistendencyis acceleratedbecausesocialmediastrategiesaredeterminedbyaudiencemetricsandauto- maticalgorithms(TandocJr,2014,2016;Diakopoulos,2017).Giventhatgatekeepingisa criticalroleofjournalismforcitizens'informeddecisions,thedominanceofeconomicvaluein decidingaccessibilitytonewscanposeachallengetotherighttobeinformed.Withacom- putationaltextanalysistechnique,StructuralTopicModel,Itestedwhethernewscontent oftendistributedviaTwitterisindeedsigni˝cantlydi˙erentfromthetraditionalgatekeeping discussedinChapter6. Theoverallcomparisonbetweenproportionsofpublishedtopicsandsharedtopicspro- videsevidencethattraditionalgatekeepingandtheselectivelinksharingdoindeedfollow di˙erentlogic.WhereasCrimeandSports/Basketballtopicswerefrequentlypublishedand evenmorelikelytobedistributedviaTwitter,Social/IdentityCon˛ictandMiddleEasttop- icsarealsooftenpublishedbutlesslikelytobesharedbyorganizations.Ingeneral,softnews topics,suchassportsandcrimearemorelikelytobesharedbynewsorganizationsthan hardnewstopicsrelatedtopoliticsandinternationalrelationsalthoughmanyhardnews topicsareamongthemostfrequenttopicsonthenewswebsites.Thisresultindicatesthat thepopularconjecturethatcommercializedmediadrivenewstowardhumaninterestrather thannewsworthinessiscrystallizedmorevisiblyonsocialmediathanonnewswebsites.If decisionmakersforlinksharingconsiderexpectedpro˝ttobemoreimportantthannews- worthinessaseditorspresumablydo,thisdiscrepancymakessense.Butitisnotyetde˝nite whetherthenewgatekeepingonsocialmediaimpliesadi˙erentregimeoftheinformation mediatingprocessfornewsreaders.Itraisesaquestion,weinformedenoughwiththe newwayofdecidingthevisibilityofneThisisnotonlyanormativequestion,butalso anempiricalonebecause,evenifanindividualorganization'slinksharinghasastrongbias inanundesirableway,thecollectionofnewsorganizationsmaystillprovideagoodmixof 108 newsonsocialmedia. Totakeadeeperlookathownewsorganizationsasacollectionselectivelysharenews links,Ialsobrokedowntheirselectivelinksharingbyindividualorganizationsandbroader categoriesoforganizationtypes.Theoutcomerevealeddi˙erentstrategiesofselectivelink sharing.Newsorganizationssharemorenewsoncertainhottopics(concentrationstrategy), lessonotherlesspopulartopics,buttheypublishedmuch(reversionarystrategy).The latterstrategycharacterizesthediscrepancybetweentraditionalgatekeepingandselective linksharing.Eventhoughanewsorganizationconsidersacertaintopictobeimportantin theireditorialdecision,thustheybecame`specialized'throughouttheirhistory,theywould notsharemuchaboutitonsocialmediabecausepopularityintheshorttermdominates socialmediastrategies.Thus,Isuggestedthatspecialtyandpopularityastwomajordeter- minantsofselectivelinksharing.Specialtyisconnectedtowhatnewsorganizationsconsider themselvestobeasagroupofjournalistswhereaspopularityismorerelatedtoshort-term pro˝tability. Topicsonwhichnewsorganizationsconcentrateandfromwhichtheyrefraindi˙eracross di˙erenttypesofnewsorganizations.Nationalandonlinemediaconcentrateonparticular politicaltopicsthatcanbereadilydepictedasahorse-˝ghtingwhereasonlinemediaalso concentrateonSocial/Gendertopic.However,theyrefrainedfromotherhardtopics,such asMiddleEast,ImmigrationandFinance.The˝ndingthatthesetypesoforganizationsare highlylikelytoshareafewhardtopicssomewhatmitigatestheearlierconcernthatthelink sharingisbiasedtowardsoftnews.However,thesharedhardnewstopicsarestillfocusedon speci˝ctopicsthatappealtothepopularpreferenceofapolarizedperceptionaboutpublic issues.Inthatsense,themarket-drivennewsdistributionviasocialmediaseemstobias informationavailableforinformedpoliticalthinking. Regionalmedia'shottopicsweresports,mostlyBasketball,presumablyduetothespe- ci˝cperiodofdatacollection.Ontheotherhand,theyrefrainfromregionalissuessuchas education,regionalpoliticsandtransportationeventhoughtheyare`specialized'inthese 109 topics.Thismayresultinanunfortunatesituationwhereregionalpapersaremostlycon- ceivedasinformationsourcesforregionalsportsasopposedtothevisionofhyper-locality asregionalmedia'sviablesurvivingstrategy(Metzgaretal.,2011).Moreover,regionalme- diaretreatfrommanytopicstheyarenotspecializedin,suchasinternationalrelations.If shrinkingtowardaregionalsportmediadoesnotprovideenoughdemandtorestoretheir business,thisoutcomedoesnotseemfavorableforthefutureviabilityoftheirbusiness. Theconcentrationonhottopicsdoesnotseemfavorabletothepluralistidealofdiversity- basedinformeddecisionsfordemocracy(Dahl,2005;Walker,1991).Becauseoverallpopular- ityistheonlypro˝tsourceonsocialmedia,thisoutcomeseemsalmostinevitable.Further, itisnotanoutcomeasingleorganizationcan`decide'toovercome.The˝ndingsimply thattheimprovementofsocialmediadesignshouldbebasedonnewsorganizations'eco- nomicreactionstochangestoenhancenewsorganizations'incentivestowardmorediverse linksharing.Thisapproachwillrequirefurtherin-depthmonitoringofthebehaviorsofnews organizationsandtheinteractionsbetweennewsorganizationsandplatforms. 8.3Extra-negativeFramingforParaphrases Newsorganizationshavemorepotentialchoicesthanwhethertosharewhentheydis- tributenewsviasocialmedia.InChapter7,Ifocusedontextinatweetoccuringwithnews links.Textisoneofthemostsalientformofanorganizations'choice,notonlybecause itcomesalmostcertainlywithnewslinksonsocialmedia,butalsobecauseourexisting knowledgeabouttraditionalheadlinesandonlinepoliticalcommentsdictatesthatnews paraphrasinghassigni˝cantimpactsonreaders'newsperception.Inparticular,Ifocusedon thesentimentthattheparaphrasingforatweetaddstotheframesofanoriginalnewsstory. Theliteraturegenerallyagreesthatnegativesentimentbothincreasesreaders'attentionto news(McCombsetal.,1997;Trussler&Soroka,2014)andchangesreactionstonews,such asreadingtime,attitude,opinion,etc.(Priceetal.,1997;Zillmannetal.,2004;Soroka, 2006;Newhagen,1998).Asaresult,newsstoriesareconsistentlynegativelyframed(Young 110 &Soroka,2012). However,thedi˙erenceinthedegreeofnegativeframingacrossdi˙erenttopics(Young &Soroka,2012)predictsthatdi˙erenttypesofnewsorganizationswithspeci˝cfocuses `specialty'tousetheterminChapter6willhavedi˙erentlevelsofthenegativeframing.For example,regionalmediathatcoversregionalissues,suchasregionalpolitics,education,and transportationarelikelytoframenewslessnegativelythannational/onlinemedia,which coversmorepartisanissues.Yet,socialmediastrategieswheretheaudience'sreceptionis expectedtogovernmayputadi˙erentlayerofsentimentonnewsbecausetheoriginalnews storyisparaphrasedforasocialmediapost.Inotherwords,theattentiondrawingfunction ofnegativeframingisexpectedtoprevailonsocialmedia. Thedictionary-basedsentimentanalysisofnewsandTwitterdatacon˝rmsthatnews storiesareconsistentlynegativelyframedinnewsorganizations.Inaddition,regionalmedia conveyedlessnegativesentimentthroughnewsstoriesthanothertypesofnewsorganizations. Further,paraphrasingonTwittershowedevenmorenegativitythanoriginalnewsstories overallregardlessofthetypeoforganization.Althoughtheseresultshaveasigni˝cant implicationonreaders'newsperceptionintheirownright,giventhewaningindustryof regionalmediaintermsofonlinepresence,aseparateanalysisofasingletypeoftextcannot distinguishbetweennegativityinnewsstoriesandnegativityfromparaphrasingstrategy. Toseesystematicallythesourceofthenegativityobservedonnewsparaphrases,Ialso conductedaninferentialanalysisthatregressesthenegativitygapbetweenastoryanda newsparaphraseonoriginalnewssentimentandtypesoforganizations.Theresultsindi- catethatregionalmediacatchuptoothertypesbyaddingevenmorenegativityonnews paraphrasesfortweets.Contrastingwiththesigni˝cantlylowlevelofnegativityinnews storiesfromregionalmedia,this˝ndingprovidesadditionalsigni˝cantindicationthatsocial mediastrategiesaregovernedbyadi˙erentlogicthanthatgoverningtraditionaljournalism practices.Inotherwords,newsorganizationsseemtoparaphrasenewsmostlytoattract usersratherthantofairlyrepresentthenewscontent. 111 Inparallelwithselectivelinksharing,theoverallresultsinChapter7implythatsocial mediastrategiesthatarelargelydrivenbyaudience'sreceptionoutsidenewsroomwillgreatly a˙ectwreadersthinkaboutaswellastothinkabAgain,ascalable monitoringsystemfornewsstrategiesiscalledfortodecidewhetherwewillbesatis˝edwith theemergingquasi-editorialprocesstoguaranteewell-informeddecisionmakingforsociety. 8.4MovingForward Inthisdissertation,Iinvestigatednewsorganizations'socialmediastrategiestodissem- inatenewsstoriesaspotentialquasi-editorialdecisions.InChapters3and5,Ishowedthat newsorganizations'behavioronsocialmediaisgovernedbyeconomicmotivation.Iusedan theoreticalmodelingandanempiricalanalysisofnewsorganizations'quantitativelinkshar- ingdecisionsagainstthedecisionsofothersonTwitter.Theseresults,combinedwiththe observationthatnewsorganizations'tasksonsocialmediahaveformsandfunctionssimilar tothoseoftheirtraditionaleditorialdecisions,suchasgatekeepingandframing,posethema- jortheoreticalandpracticalconcernofthisdissertationaboutwhethereconomically-driven quasi-editorialdecisionsaresigni˝cantlydi˙erentfromtraditionalones.Usingcomputa- tionalmethods,IillustratedthatthegatekeepingandframingforTwitterindeedgenerate signi˝cantlydi˙erentpatterns.Theywerebiasedtowardpopularityonsocialmedia,result- inginconcentrationonspeci˝ctopicsandadditionallynegativeframing. Althoughthisdissertationillustratedtheemergenceofanewinformation˝lteronTwit- ter,fortworeasons,itistooearlytosaythatthenew˝lterilluminatesadi˙erentregime ofjournalism.First,althoughTwitterisoneoftwosocialmediathroughwhichmostnews organizationsregularlysharenewslinks,Twitteraccountsforasmallportionofnewsread- ingthroughsocialmedia.Ibelievethatthegeneralconclusionthatsocialmediaasanews distributionplatformcreateanadditionalinformationmediatingprocess,canbegeneralized toFacebook.FacebookalsoretainsthecharacteristicsthatIargueareamaindrivingforce ofdistinctivenewsdisseminationstrategiesusers'limitedattentionandvirality.However, 112 thespeci˝cpatternsofthestrategiesmightbedi˙erentbecauseFacebookalsohasdi˙erent traitsthanTwitter,suchasamoreactivealgorithmiccuration,relativelyclosednetworks, andalargertextlimit.Thus,applyingsimilaranalysestotheFacebookcasewillhelpdraw amorecompletepictureoftheemergentinformationmediatingmechanismthatexistson multipleplatforms.Second,wedonotknowcompletelysigni˝canceoftheimpactthenew ˝lterhasonusersbecauseuserbehaviorsaremostlyomittedfromtheanalysesinthisdis- sertation.Withmultipleplatformsdesignedforinformationsharing,newsorganizationsare nolongeranexclusivesourceofnewsformostaudiences.Thus,weneedafurtherstudy thatcantestwhethernewsorganizations'strategiesindeedimpactnewsperceptions.This typeoffuturestudycannotbebasedonsmallsampleseither.Togaugetheimpactofnews organizations'socialmediastrategies,oneneedstomeasurehowoftenaudiencesareexposed tosuchstrategies,whichcanhardlybemeasuredrelyingonthememoryoftheaudience.A possiblepathistoconnectnewsreaders'logdataandasurveyontheirperceptions. Fromtheperspectiveofpolicy,newsorganizations'socialmediastrategyisamovingtar- get.Becausenewsorganizationsarelikelytoreacttoaudience'sreception,theirstrategies mustbechangedtoenvironmentalvariablesthata˙ectusers'newsdemands,justasplat- forms'algorithmsandnewsorganizations'businessmodels,change.Thus,todecidewhether theemerging˝lterindeedcallsforintervention,weneedaconstantmonitoringofit.The methodstolinkdi˙erentsourcesofdataandmachinelearningtechniquesthatIsuggested canextendexistingopensourceplatformswithasimilargoal,suchasMediaCloud,tomake themonitoringsystemsmorespeci˝callygearedtowardmeaningfulquestions.Further,the empirical˝ndingsinthisdissertationonlydescribeoutcomesoforganizations'socialmedia strategies,butmechanismswherebyuseofaudiencemetricsandorganizationaldisconnect- ednessofsocialmediaspecialistsleadtospeci˝cstrategiesstillremaintobeexplained. Understandingwtheygetwillhelpuspredictnewsorganizations'adaptation tofutureenvironmentalchanges,whichwillinformplatforms'betteralgorithmdesignand potentialpolicyintervention. 113 Althoughtheframingstudyinthisdissertationfocusesonlyonsentiment,studiesofnews paraphrasinginthetraditionalmediaenvironmentencompassamuchwiderconceptionof framing.Ingeneral,aframingconceptincludesboththeselectionofideas(Entman,1993) andanarrationoftheselectedideas(Gamson&Modigliani,1994).Althoughprevious attemptstocomputationallyoperationalizeframingconceptshaveappliedtopicmodelsto newstext(Boydstunetal.,2013),theyfailtodi˙erentiatebetweennewstopicsandframing becausethechoiceofalgorithmsisnotbasedonthetheories.Ibelieveexplicitmodelingof thewordselectionprocessandsequentialwordchoice,complyingwiththeoriginalideaof theconceptratherthanapplyingaready-madealgorithm,isamoreplausiblewaytothe framingconcept.Newlydevelopedalgorithmscanalsobeavaluableextensionofthepublic monitoringsystemfornewsorganizations'socialmediastrategies. 114 APPENDICES 115 APPENDIXA PROOFSOFLEMMASANDPROPOSITIONSINCHAPTER3 A.1ProofofProposition1 If KM T ,weknowthatfullsharingcharacterizestheequilibriumaslongas Np>c becausethepro˝tonthemarginallinksharedis Np c inthiscase,whichassumedtobe positive. If KM>T ,themarginalpro˝tcanbelowerthan Np c .Letusconsideran uncon- strained versionofequation3.1withouttheupperboundfortheprobabilityforalinktobe clickedatone: max 0 x i M E [ ˇ i ( x )]= T P j x j Np c x i (A.1) Supposethatallothernewsorganizationsexceptthe˝rm, i ,arechoosingtheSNEchoice, x .Thenthemaximizationproblembecomes: max 0 x i M T ( K 1) x + x i Np c x i The˝rstorderconditionofthisproblemis 1 : TNp ( K 1) x [( K 1) x + x i ] 2 c =0 InanSNE, x i = x .Thus, 1 Thesecondorderconditionholdsbecausethederivativeofthemarginalpro˝tis 2 MT ( K 1) Np= [ M ( K 1)+ x i ] 3 ,whichisnegativeasfaras x i isnonnegative. 116 TNp ( K 1) x K 2 x 2 = c () x = Np c T ( K 1) K 2 (A.2) Notethat,when x i = x forall i ,each˝rm'spro˝tis c K K 1 ,whichisstrictlypositive. Thus, x isnotconstrainedbytherequirementthatpro˝tbenon-negative. Nowbycomparingunconstrainedoptimalchoice x withthenumberofnewsstoriesan individualorganizationpublished, M ,wecan˝ndtherangeforfullsharingandnosharing equilibrium.Thatis,if x M ,newsorganizationscannotincreasethenumberofnews linksharedonsocialmediamorethan M eventhoughthemarginalpro˝tispositive.Hence, fullsharingisanSNEinthiscase.Rearrangingtheinequality,weget ˚ = MK T Np c K 1 K forfullsharingcondition.PartialsharingwillbesustainedasanSNEotherwise. 117 APPENDIXB DETAILSOFDATACOLLECTION B.1O˚cialTwitterAccounts NewsorganizationsoftenmaintainmorethanoneTwitteraccounts.Forex- ample, Time onlyhasoneo˚cialaccountwhereas WashingtonPost hasWashingtonPost, WashingtonPostWorld,WashingtonPostSports,andsoon.Totakeaccountofthis,the Twitterscrapercollectstweetsfromallthe sub-accounts .However,therearecaseswhere thedistinctionbetweeno˚cialaccountsandthosemaintainedbyindividualjournalistsis unclear.Tomakeadistinction,IcodedtheTwitteraccountsaccordingtofollowingsteps: (a)Searchaccountswithanewsorganization'sname(notthedescription),(b)Chooseac- countswiththeorganization'snameintheaccountname(acronymsareaccounted,e.g.WP =WashingtonPost)and(c)Ifanaccountdescriptionstatesthattheaccountismaintained byindividualreporters,theaccountwasexcluded. Thesestepsaimtocountonlysocialmediastrategiescomeoutoftheorganizationlevel ratherthanajournalistlevelorsomebureau(e.g.EconomyorInternational). B.2HowScraperWorks ThenewsscraperopensalltheURLfromMediaCloudandextractsnewstext.Thetext extractionisconductedbyapopularopen-sourcedetectorGoose,whichhasbeenanorigi- nallyJavabasedopen-sourceprogram,andnowitisimplementedwithScalaandPython. GooseusesrulestosearchthroughHTMLtagstodetecttext.Thenewsscraperalsouses threestrategiestoextractpublisheddate.The˝rstpriorityistoparsethetextandmatch datewithregularexpressions.Ifitdoesn'twork,thewebcrawlerextractsdatesfromURL. Ifthatdoesn'tworkeither,thenewsscraperlooksformetadatathatcontainsdate.To seethesealgorithmsworkwellenough,Isampled10newsstoriesfromeachnewsorganiza- 118 TableB.1:ListofTwitteraccountsincludedinthedataset. OrganizationTwitterOrganizationTwitter CNNCNNWashingtonPostwashingtonpost cnnbrkpostpolitics cnniPostWorldNews CNNPoliticspostlive NewYorkTimenytimesWashPostDC nytimesbookswapodesign TheGuardianguardianWashPostPR guardiannewsPostVideo GdnPoliticswashposthelp GuardianUSPostGraphics guardian˝lmPostSports guardianopinionPostStyle guardianmusicwpmagazine GuardianTravelWaPoFood guardiancultureWaPoTravel guardianstylepost_lead guardianlifeBBCBBCWorld guardian_sportBBC guardiantechBBCBreaking SocietyGuardianBBCTechnology guardianscienceBBCLondonNews GuardianDataBBCSport mediaguardianBBCPolitics GuardianEdubbchealth GuardianMTNBBCRD guardianweekendBBCScienceNews GuardianFashionBBCNewsEnts guardianstageForbesForbes guardiang2ForbesTech guardianecoForbesLife guardianworldFoxNewsFoxNews GdnVoluntaryFoxSports GuardianGDPFOXTV GuardianTeachHu˚ngtonPostHu˙Post g2˝lmandmusicHu˙PostPol GuardianBooksBloombergbusiness guardianreviewtechnology guardiancitiesmarkets GdnSocialCarebpolitics GuardianJobsluxury USATodayusatodayeconomics USATODAYmoneyWallStreetJournalWSJ usatodaytechWSJPolitics usatodaysportsWSJSports usatodaydcWSJMarkets USATODAYhealthWSJAsia USATOpinionWSJopinion usatodayhssWSJbusiness usatodayvideoWSJeurope usatodaystyleWSJIndia usatodaylifeWSJMoneyBeat usatodaytravelWSJNY usatodaymagsWSJbreakingnews usatodaymlbWSJecon usatodaynbaWSJtech USATODAYeatsWSJafrica usatodayweatherReutersReuters usatodayn˛ReutersWorld usatsimgReutersBiz USATODAYBooksReutersOpinion 119 TableB.1(cont'd) OrganizationTwitterOrganizationTwitter CNBCCNBCReutersChina CNBCnowreuterspictures CNBCTopStoriesReutersUK cnbcafricaReutersLive CNBCPoliticsReutersUS CNBCiReutersIndia NewYorkPostnypostReutersTech nypostsportsReutersAfrica nypmetroReutersSports nypostbizReutersPolls NYPfashionReutersPolitics USNewsusnewsReutersShowBiz USNewsEducationspecialreports USNewsHealthNBCNewsNBCNews TheAtlanticTheAtlanticNBC TheAtlNewsTheHillthehill TheAtlEducationTheHillOpinion TheAtlPolilticshilltransport TheAtlPhotoTheHillEvents TheAtlVideoTimeTIME TheAtlTechCBSNewsCBSNews TheAtlGlobalCBS TheAtlCultureCBSTweet TheAtlHealthABCNewsABC EconomistTheEconomistABCNetwork EconomistEventsabcnews EconAmericasABCPolitics EconWhichMBATheDailyBeastthedailybeast ECONdailychartsChicagoTribunechicagotribune EconEuropeChiTribEnt EconUSchicago_homes EconAsiaChiTribBiz EconCultureChiTribBooks EconEconomicsChiTribuneAuto EconBizFinChiTribFood EconSciTechChiTribPhoto TheMercuryNewsmercnewsAlJazeeraAlJazeera Nj.comnjdotcomAJEnglish HSSportsNJAlJazeera_World NJ_SportsAJENews NJcomsomersetAJEVideos NJentertainmentAJEWeather Monmouth_NJTheDailyCallerDailyCaller NJ_PoliticsTheDC_Opinion Bergen_NJTheDCSports njerseypoliticsTheDCPolitics NJ_Morristhedctechnews NYJetsNewsTheRootTheRoot butler_njMiamiHeraldMiamiHerald SeattleTimesseattletimesHeraldSports SeaTimesJobsMiamiHeraldFood seatimesprepsHeraldOpEd SeaTimesPhotoHeraldBusiness seatimesbizMiamiHeraldLive SeaTimesSportsThePhiladelphiaInquirerphillyinquirer SeaTimesOpinionInquisitrtheinquisitr 120 TableB.1(cont'd) OrganizationTwitterOrganizationTwitter WashingtonExaminerdcexaminerInquisitrSports TheConversationConversationUSIQShowbiz WashingtonTimesWashTimesInquisitrLife wtimespoliticsInquisitrGaming WashTimesLocalInquisitrFunny DetroitFreePressfreepInquisitrHealth freepsportsiqcontest OregonianOregonianInquisitrWorld OregonianBizInquisitrTech OregonianSportsIQScience hssports_oliveTheSacramentoBeesacbee_news oregonianstumpSacBeeEditBoard OregonianPolBaltimoreSunbaltimoresun DallasNewsdallasnewsbaltsunarts DallasISD_NewsBaltSunVid dmn_collegesBaltSunHealth DenverPost denverpostBaltSunSports DenverPostPicksPBSNewshourNewsHour denverpoliticsPBS denveropinionNewsHourWorld DPostSportsSunSentinelSunSentinel DenverPostBrkPhotoSSentinel dpcommunitySFLEventsForYou DPRockiesSSCourts denverbusiness denverpostliteThePressDemocratNorthBayNews PostBroncosNewsdaynewsday OrangeCountyRegisterocregisterNewsdayHSsports AlterNetAlterNetNewsdaySports TheKansasCityStarKCStarNewsdayBiz KCStarHSNewsdayHealth OrlandoSentinelorlandosentinelNewsdayOpinion OSCrimeNewsdayEnt OSPhotoBostonHeraldbostonherald OSLakeCountynpr.org/programsNPR orlandosportsnprpolitics OSentinelBiznprscience ChicagoSunTimessuntimesnprbusiness suntimes_hoopsnprclassical suntimes_prepsSlateSlate CSTbreakingSlateMoneybox CSTeditorialsSlateVideo SunTimesCHIBiparisanReportBipartisanism suntimes_sportsWesternJournalismWestJournalism chicagosmusicTalkingPointsMemoTPM DetroitNewsdetroitnewsInforwarsinfowars DetNewsOpinionFiveThirtyEightFiveThirtyEight TheStrangerTheStrangeNewYorkMegazineNYMag TheNewRepublicNewRepublicMediaMattersmmfa TheIndianaStarindystarFreeBeaconFreeBeacon TheChristianScienceMonitorcsmonitorTheIntercepttheintercept csmonitorphotoVanityFairVanityFair CommonDreamscommondreamsViceVICE TheTennesseanTennesseanvicedocs TNMusicNewsVICESports tnsportsVICESportsAU 121 TableB.1(cont'd) OrganizationTwitterOrganizationTwitter BuzzFeedNewsBuzzFeedVICESportsNZ BuzzFeedNewsVICERECORDS BuzzFeedFoodvicetech BuzzFeedBooksVICE_NZ buzzfeedpartnervice_money BuzzFeedPolFactCheckfactcheckdotorg BuzzFeedFashionIBTimesIBTimes BuzzFeedQuiztownhall.comtownhallcom BuzzFeedAnimalsDesMoinesRegisterDMRegister BuzzFeedCelebRegisterVisuals fwdWestwordDenverWestword BuzzFeedSportsPoliticopolitico BuzzFeedWorldBreitbartBreitbartNews BuzzFeedGeekyPoliticusUSApoliticususa BuzzFeedEntMSNBCMSNBC ConservativeTribuneconserv_tribuneVoxvoxdotcom GatewayPunditgatewaypundit tions,andmanuallyseeifthetextanddateextractionworkedwell.Ionlyincludednews organizationsforwhichtheextractionalgorithmworkedwellwith9out10newsstories. Whenthere'spaywallonanewsorganization'swebsite,IuseSeleniumtoscrapenews text.SinceSeleniumallowsautomatesearchingthroughaWebbrowser,itallowsforlogging inasifausermanuallytypeinanaccountnameandapassword,andextracttextfrom thedisplayedtext.Althoughmanywebsiteshavepaywalls,mostofthemhadapartial paywallinasensethat,whenauserusesURLfromRSSfeed,theyshowthefulltext.Thus, theSelenium-dependentscraperwasonlyusedtotheNewYorkTime,WallStreetJournal, WashingtonPostandEconomist. B.3TwitterScraper Toscrapenewscompanies'tweets,IuseTwitterAPI's GETstatusesuser_timeline methodthatreturnsupto 3 ; 200 themostrecenttweetsfromauser'saccount.Scrap- ingnewsorganization'salltweetsisstraightforwardbecausethetimelinereturnedbythe methodisequivalenttowhatauserseesonscreen. 1 1 https://dev.twitter.com/rest/reference/get/statuses/user t imeline 122 FigureB.1:AnexampleofshortenednewsURLinatweet. B.4URLMatcher Torelatenewslinkembeddedintweetstooriginalnewsarticles,IalsodevelopaURL matcher.Thissoftwareworksinthreesteps.TheURLmatcherextractsaURLfromtweets scrapedbytheTwitterscraper.ThistaskisstraightforwardbecauseanembeddedURL istaggedinaJSONformasTwitterAPIprovides.Thenit˝ndsanoriginalformofthe URLifitisshortened.MostofURLsembeddedinsocialmediapostsareshortenedasin FigureB.1sothatitwouldnottakeupmuchspaceorthecharacterlimit.Theoriginal formswereidenti˝edbysendingHTTPrequestusingeachshortenedURLs.SinceURLs oftencontainquerytermsattachedtousethemforavarietyoffunctionality,suchasthe tra˚csourceidenti˝cation,theURLmatcherparsesout.Asa˝nalnormalizationprocess, theURLmatcheronlyextracts`path'partthatrepresentsuniquelocationwithinawebsite tominimizethefalsenegativecasesfromthematching.Duetoapopularsearchengine optimization(SEO)strategy,majorityofthepathscontainsomeformofnewsheadlines. However,sometimestheyarenumericnewsIDs.Afterthenormalizationprocess,anews storyislabeledas`shared'whenthenormalizedURLfromanewsorganization'swebsiteis alsofoundfromthelistofnormalizedURLsfromthenewsorganization'sTwitteraccount. 123 APPENDIXC ADDITIONALSTMGRAPHSANDFULLOUTPUT C.1AdditionalGraphs FigureC.1:Newspublicationandlinksharingpatternsbynewsorganizations:Life/Inter- viewtopic. FigureC.2:Newspublicationandlinksharingpatternsbynewsorganizations:Econo- my/TechnologyEasttopic. 124 FigureC.3:Newspublicationandlinksharingpatternsbynewsorganizations:Social/Iden- tityCon˛icttopic. FigureC.4:Newspublicationandlinksharingpatternsbynewsorganizations:Legal/Reg- ulationtopic. 125 FigureC.5:Newspublicationandlinksharingpatternsbynewsorganizations:Economy/- Financetopic. FigureC.6:Newspublicationandlinksharingpatternsbynewsorganizations:UKtopic. 126 FigureC.7:Newspublicationandlinksharingpatternsbynewsorganizations:Updatetopic. FigureC.8:Newspublicationandlinksharingpatternsbynewsorganizations:Sport- s/Otherstopic. 127 FigureC.9:Newspublicationandlinksharingpatternsbynewsorganizations:Entertain- menttopic FigureC.10:Newspublicationandlinksharingpatternsbynewsorganizations:Regional Politicstopic. 128 FigureC.11:Newspublicationandlinksharingpatternsbynewsorganizations:Weather topic. FigureC.12:Newspublicationandlinksharingpatternsbynewsorganizations:Health topic. 129 FigureC.13:Newspublicationandlinksharingpatternsbynewsorganizations:Subscription topic. FigureC.14:Newspublicationandlinksharingpatternsbynewsorganizations:ITaxtopic. 130 FigureC.15:Newspublicationandlinksharingpatternsbynewsorganizations:SocialMedia topic. FigureC.16:Newspublicationandlinksharingpatternsbynewsorganizations:Education topic. 131 FigureC.17:Newspublicationandlinksharingpatternsbynewsorganizations:Sports/- Collegetopic. FigureC.18:Newspublicationandlinksharingpatternsbynewsorganizations:MeToo topic. 132 FigureC.19:Newspublicationandlinksharingpatternsbynewsorganizations:French topic. FigureC.20:Newspublicationandlinksharingpatternsbynewsorganizations:Advertising topic. 133 APPENDIXD SENTIMENTANALYSISOUTPUTSBASEDONLSDANDAFINN FigureD.1:Averagesentimentscoresofnewsstoriesandtweets:AFINNdictionary.Zero scoreindicatesneutralsentiment. FigureD.2:Averagesentimentscoresofnewsstoriesandtweets:LSDdictionary.Zeroscore indicatesneutralsentiment. 134 BIBLIOGRAPHY 135 BIBLIOGRAPHY An,Jisun,MeeyoungCha,KrishnaGummadi&JonCrowcroft.2011.Medialandscapein twitter:Aworldofnewconventionsandpoliticaldiversity.In Proceedingsoficwsm2011 , AAAI. An,Jisun,DanieleQuercia,MeeyoungCha,KrishnaGummadi&JonCrowcroft.2014.Shar- ingpoliticalnews:thebalancingactofintimacyandsocializationinselectiveexposure. EPJDataScience 3(1).12. Anderson,Chris.2007. Thelongtail:howendlesschoiceiscreatingunlimiteddemand . RandomHouse. Anderson,SimonP&AndréDePalma.2009.Informationcongestion. TheRANDJournal ofEconomics 40(4). Anderson,SimonP&AndréDePalma.2012.Competitionforattentionintheinformation (overload)age. TheRANDJournalofEconomics 43(1). Anderson,SimonP&AndréDePalma.2013.Shoutingtobeheardinadvertising. Man- agementScience 59(7). Anderson,SimonP,AndreDePalma&JacquesFrançoisThisse.1992. Discretechoice theoryofproductdi˙erentiation .MITpress. Andrew,BlakeC.2007.Media-generatedshortcuts:Donewspaperheadlinespresentanother roadblockforlow-informationrationality? HarvardInternationalJournalofPress/Politics 12(2). Armstrong,CoryL&FangfangGao.2010.Nowtweetthishownewsorganizationsuse twitter. ElectronicNews 4(4). Bakshy,Eytan,SolomonMessing&LadaAAdamic.2015.Exposuretoideologicallydiverse newsandopiniononfacebook. Science 348(6239).132. Baltas,George.2003.Determinantsofinternetadvertisinge˙ectiveness:anempiricalstudy. InternationalJournalofMarketResearch 45(4). Barzilai-Nahon,Karine.2008.Towardatheoryofnetworkgatekeeping:Aframeworkfor exploringinformationcontrol. JournaloftheAssociationforInformationScienceand Technology 59(9). Bell,Allan.1991. Thelanguageofnewsmedia .BlackwellOxford. Benkler,Yochai,RobertFaris,RobertsmHal&EthanZuckerman.2017.Study:Breitbart- ledright-wingmediaecosystemalteredbroadermediaagenda.Tech.rep.ColumbiaJour- nalismReview.https://www.cjr.org/analysis/breitbart-media-trump-harvard-study.php. 136 Bennett,Shea.2013,July13.Socialmediaoverload-howmuchinformationdoweprocess eachday? SocialTimes. http://www.adweek.com/socialtimes/social-media-overload/ 488800. Bennett,WLance&ShantoIyengar.2008.Aneweraofminimale˙ects?thechanging foundationsofpoliticalcommunication. JournalofCommunication 58(4). Berger,Ben.2011. Attentionde˝citdemocracy:Theparadoxofcivicengagement .Princeton UniversityPress. Berger,Jonah&KatherineLMilkman.2012.Whatmakesonlinecontentviral? Journalof marketingresearch 49(2). Bischof,Jonathan&EdoardoMAiroldi.2012.Summarizingtopicalcontentwithword frequencyandexclusivity.In Proceedingsofthe29thinternationalconferenceonmachine learning(icml-12) , Blei,David&JohnLa˙erty.2006.Correlatedtopicmodels. Advancesinneuralinformation processingsystems 18. Blei,DavidM.2012.Probabilistictopicmodels. CommunicationsoftheACM 55(4). Blei,DavidM,AndrewYNg&MichaelIJordan.2003.Latentdirichletallocation. the JournalofmachineLearningresearch 3. Blom,JonasNygaard&KennethReineckeHansen.2015.Clickbait:Forward-referenceas lureinonlinenewsheadlines. JournalofPragmatics 76. Boydstun,AmberE,JustinHGross,PhilipResnik&NoahASmith.2013.Identifying mediaframesandframedynamicswithinandacrosspolicyissues.In Newdirectionsin analyzingtextasdataworkshop,london ,. Bozdag,Engin&JeroenvandenHoven.2015.Breakingthe˝lterbubble:democracyand design. EthicsandInformationTechnology 17(4). Bradley,MargaretM&PeterJLang.1999.A˙ectivenormsforenglishwords(anew): Instructionmanualanda˙ectiveratings.Tech.rep.TechnicalreportC-1,thecenterfor researchinpsychophysiology,UniversityofFlorida. Brown,Pete.2018,April18.Facebookstrugglestopromote`meaningfulinteractions'for localpublishers,datashows. ColumbiaJournalismReview. https://www.cjr.org/tow_ center/facebook-local-news.php. Bruns,Axel.2011.Gatekeeping,gatewatching,real-timefeedback:newchallengesforjour- nalism. BrazilianJournalismResearch 7(2). Carlson,Matt.2007.Orderversusaccess:Newssearchenginesandthechallengetotradi- tionaljournalisticroles. Media,Culture&Society 29(6). 137 Chakraborty,Abhijnan,BhargaviParanjape,SouryaKakarla&NiloyGanguly.2016.Stop clickbait:Detectingandpreventingclickbaitsinonlinenewsmedia.In 2016ieee/acm internationalconferenceonadvancesinsocialnetworksanalysisandmining(asonam) , IEEE. Chan-Olmsted,Sylvia,HyejoonRim&AmyZerba.2013.Mobilenewsadoptionamong youngadultsexaminingtherolesofperceptions,newsconsumption,andmediausage. Journalism&MassCommunicationQuarterly 90(1). Chan-Olmsted,SylviaM&JungSukPark.2000.Fromon-airtoonlineworld:Examin- ingthecontentandstructuresofbroadcasttvstations'websites. Journalism&Mass CommunicationQuarterly 77(2). Chang,Jonathan,SeanGerrish,ChongWang,JordanLBoyd-Graber&DavidMBlei. 2009.Readingtealeaves:Howhumansinterprettopicmodels.In Advancesinneural informationprocessingsystems , Cleary,Johanna,EisaalNashmi,TerryBloom&MichaelNorth.2015.Valuingtwitter: Organizationalandindividualrepresentationsatcnninternational. DigitalJournalism 3(6). Cleveland,WilliamS,EGrosse&WMShyu.1992.Localregressionmodels. Statistical modelsinS,editedbyJohnM.ChambersandTrevorJ.Hastie Cook,Clare&EsaSirkkunen.2013.What'sinaniche?exploringthebusinessmodelof onlinejournalism. JournalofMediaBusinessStudies 10(4). Cottle,Simon.2007.Ethnographyandnewsproduction:New(s)developmentsinthe˝eld. SociologyCompass 1(1). Curran,James.2010.Thefutureofjournalism. Journalismstudies 11(4). Dahl,RobertA.2005. Whogoverns?:Democracyandpowerinanamericancity .Yale UniversityPress. DeVries,Lisette,SonjaGensler&PeterSHLee˛ang.2012.Popularityofbrandpostson brandfanpages:Aninvestigationofthee˙ectsofsocialmediamarketing. Journalof InteractiveMarketing 26(2). Deuze,Mark.2008.Understandingjournalismasnewswork:Howitchanges,andhowit remainsthesame. WestminsterPapersinCommunicationandCulture 5(2). Diakopoulos,Nicholas.2017. Theroutledgecompaniontodigitaljournalismstudies chap. Computationaljournalismandtheemergenceofnewsplatforms,NewYork,NY: Routledge. Dick,Murray.2011.Searchengineoptimisationinuknewsproduction. JournalismPractice 5(4). 138 Dimitrova,DanielaV,ColleenConnolly-Ahern,AndrewPaulWilliams,LyndaLeeKaid &AmandaReid.2003.Hyperlinkingasgatekeeping:Onlinenewspapercoverageofthe executionofanamericanterrorist. JournalismStudies 4(3). Domingo,David&ChrisPaterson.2011.Makingonlinenews(volume2). NewsroomEthnog- raphyintheSecondDecadeofInternetJournalism,NewYork:PeterLang . Dor,Daniel.2003.Onnewspaperheadlinesasrelevanceoptimizers. JournalofPragmatics 35(5). Elizabeth,Jane.2017,November14.Socialmediateamstoday:Asummaryofwhatwe learned. AmericanPressInstitute. https://www.americanpressinstitute.org/publications/ reports/strategy-studies/social-media-teams-today/. Entman,RobertM.1993.Framing:Towardclari˝cationofafracturedparadigm. Journal ofcommunication 43(4). Evans,DavidS.2003.Theantitrusteconomicsofmulti-sidedplatformmarkets. YaleJournal onRegulation 20. Faris,Robert,HalRoberts,BruceEtling,NikkiBourassa,EthanZuckerman&YochaiBen- kler.2017.Partisanship,propaganda,anddisinformation:Onlinemediaandthe2016us presidentialelection.Tech.rep.BerkmanKleinCenter. Gamson,WilliamA&AndreModigliani.1994.Thechangingcultureofa˚rmativeaction. Equalemploymentopportunity:labormarketdiscriminationandpublicpolicy García-Perdomo,Víctor,RamónSalaverría,DanielleKKilgo&SummerHarlow.2017.To shareornottoshare:Thein˛uenceofnewsvaluesandtopicsonpopularsocialmedia contentintheunitedstates,brazil,andargentina. JournalismStudies Gentzkow,Matthew,EdwardLGlaeser&ClaudiaGoldin.2006.Theriseofthefourth estate.hownewspapersbecameinformativeandwhyitmattered.In Corruptionand reform:Lessonsfromamerica'seconomichistory ,UniversityofChicagoPress. Gentzkow,Matthew&JesseMShapiro.2010.Whatdrivesmediaslant?evidencefromus dailynewspapers. Econometrica 78(1). Giomelakis,Dimitrios&AndreasVeglis.2015.Employingsearchengineoptimizationtech- niquesinonlinenewsarticles. StudiesinMediaandCommunication 3(1). Granovetter,MarkS.1973.Thestrengthofweakties. Americanjournalofsociology 78(6). Greer,ClarkF&DouglasAFerguson.2011.Usingtwitterforpromotionandbranding: Acontentanalysisoflocaltelevisiontwittersites. JournalofBroadcasting&Electronic Media 55(2). Grueskin,Bill,AvaSeave&LucasGraves.2011. Thestorysofar:Whatweknowaboutthe businessofdigitaljournalism .ColumbiaUniversityPress. 139 Hansen,LarsKai,AdamArvidsson,FinnÅrupNielsen,ElanorColleoni&MichaelEt- ter.2011.Goodfriends,badnews-a˙ectandviralityintwitter.In Futureinformation technology ,Springer. Hardin,Garrett.1968.Thetragedyofthecommons. Science 162(3859). Hart,RoderickP.2001.Redevelopingdiction:theoreticalconsiderations. Theory,Method, andPracticeinComputerContentAnalysis 16. Hermida,Alfred.2010.Twitteringthenews:Theemergenceofambientjournalism. Jour- nalismPractice 4(3). Hermida,Alfred,FredFletcher,DarrylKorell&DonnaLogan.2012.Share,like,recommend: Decodingthesocialmedianewsconsumer. JournalismStudies 13(5-6). Hodson,Jaigris&AprilLindgren.2017.Isnoelectionnewsgoodnews?:The2015canadian electionandlocallyrelevantnewsontwitter.In Proceedingsofthe8thinternational conferenceonsocialmedia&society ,39.ACM. Hong,Sounman.2012.Onlinenewsontwitter:Newspapers'socialmediaadoptionandtheir onlinereadership. InformationEconomicsandPolicy 24(1). Horne,BenjaminD&SibelAdali.2017.Theimpactofcrowdsonnewsengagement:A redditcasestudy. arXivpreprintarXiv:1703.10570 . Hu,Xia&HuanLiu.2012.Textanalyticsinsocialmedia.In Miningtextdata , Springer. Ingram,Mathew.2018,January22.Facebook'slatestchangeswillprobablymakemisinfor- mationworse. ColumbiaJournalismReview. https://www.cjr.org/innovations/facebook- newsfeed-changes.php. Kahneman,Daniel.1979.Prospecttheory:Ananalysisofdecisionsunderrisk. Econometrica 47(2). Kaplan,Sarah&RebeccaHenderson.2005.Inertiaandincentives:Bridgingorganizational economicsandorganizationaltheory. OrganizationScience 16(5). Kew,Ben.2016,August17.Digiday:Breitbart`dominating'socialmediaand`punchesabove itsweight'. Breitbart. http://www.breitbart.com/big-journalism/2016/08/17/digiday- breitbart-dominating-social-media-and-punches-above-its-weight/. Kim,HyunSuk.2015.Attractingviewsandgoingviral:Howmessagefeaturesandnews- sharingchannelsa˙ecthealthnewsdi˙usion. JournalofCommunication 65(3). Kiousis,Spiro&MattRagas.2015.Implicationsofthird-levelagendabuildingforpublicre- lationsandstrategiccommunication. ThePowerofInformationNetworks:NewDirections forAgendaSetting 8. 140 Kitchin,Rob.2016.Thinkingcriticallyaboutandresearchingalgorithms. Information, Communication&Society 20(1). Kleinberg,JonM.1999.Authoritativesourcesinahyperlinkedenvironment. Journalofthe ACM(JACM) 46(5). Kroll,JoshuaA,JoannaHuey,SolonBarocas,EdwardWFelten,JoelRReidenberg,DavidG Robinson&HarlanYu.2016.AccountableAlgorithms. UniversityofPennsylvaniaLaw Review 165. Kwak,Haewoon,ChanghyunLee,HosungPark&SueMoon.2010.Whatistwitter,asocial networkoranewsmedia?In Proceedingsofthe19thinternationalconferenceonworld wideweb ,ACM. Lasorsa,DominicL,SethCLewis&AveryEHolton.2012.Normalizingtwitter:Journalism practiceinanemergingcommunicationspace. Journalismstudies 13(1). Lawrence,ReginaG,LoganMolyneux,MarkCoddington&AveryHolton.2014.Tweeting conventions:Politicaljournalists'useoftwittertocoverthe2012presidentialcampaign. JournalismStudies 15(6). Lerman,Kristina&RumiGhosh.2010.Informationcontagion:Anempiricalstudyofthe spreadofnewsondiggandtwittersocialnetworks. Icwsm 10. Liang,Hai&King-waFu.2016.Informationoverload,similarity,andredundancy:unsub- scribinginformationsourcesontwitter. JournalofComputer-MediatedCommunication 22(1). Lischka,JulianeA.2014.Di˙erentrevenueincentives,di˙erentcontent?comparingeconomic newsbeforeandduringthe˝nancialcrisisingermanpublicandcommercialnewsoutlets overtime. EuropeanJournalofCommunication 29(5). Lowrey,Wilson.2009.Institutionalroadblocks:Assessingjournalism'sresponsetochanging audiences. Journalismandcitizenshipnewagendasincommunication Lowrey,Wilson.2012.Journalisminnovationandtheecologyofnewsproduction:Institu- tionaltendencies. Journalism&CommunicationMonographs 14(4). Lowrey,Wilson.2017.Theemergenceanddevelopmentofnewsfact-checkingsites:Institu- tionallogicsandpopulationecology. JournalismStudies 18(3). Lucas,Christopher,RichardNielsen,MargaretRoberts,BrandonStewart,AlexStorer& DustinTingley.2015.Computerassistedtextanalysisforcomparativepolitics. Political Analysis 23(2). LucasJr,RobertE.1976.Econometricpolicyevaluation:Acritique.In Carnegie-rochester conferenceseriesonpublicpolicy ,vol.1,Elsevier. Manning,ChristopherD,PrabhakarRaghavan,HinrichSchützeetal.2008. Introductionto informationretrieval ,vol.11.CambridgeuniversitypressCambridge. 141 Manovich,Lev.2012.Datastream,database,timeline:Theformsofsocialmedia. SoftwareStudiesInitiative 27.http://lab.softwarestudies.com/2012/10/data-stream- database-timeline-new.html. Martindale,Colin.1975. Romanticprogression:Thepsychologyofliteraryhistory .Hemi- spherepublishingcorporation. McComas,Katherine&JamesShanahan.1999.Tellingstoriesaboutglobalclimatechange: Measuringtheimpactofnarrativesonissuecycles. CommunicationResearch 26(1). McCombs,Maxwell.2014. Settingtheagenda:Massmediaandpublicopinion .JohnWiley &Sons. McCombs,Maxwell,JuanPabloLlamas,EstebanLopez-Escobar&FedericoRey.1997. Candidateimagesinspanishelections:Second-levelagenda-settinge˙ects. Journalism& MassCommunicationQuarterly 74(4). Meraz,Sharon.2011.Usingtimeseriesanalysistomeasureintermediaagenda-settingin˛u- enceintraditionalmediaandpoliticalblognetworks. Journalism&MassCommunication Quarterly 88(1). Meraz,Sharon&ZiziPapacharissi.2013.Networkedgatekeepingandnetworkedframing on#egypt. Theinternationaljournalofpress/politics 18(2). Mergenthaler,Erhard.1996.Emotion-abstractionpatternsinverbatimprotocols:Anewway ofdescribingpsychotherapeuticprocesses. Journalofconsultingandclinicalpsychology 64(6). Metzgar,EmilyT,DavidDKurpius&KarenMRowley.2011.De˝ninghyperlocalmedia: Proposingaframeworkfordiscussion. NewMedia&Society 13(5). Meyer,KellyM&TangTang.2015.#socialjournalism:Localnewsmediaontwitter. InternationalJournalonMediaManagement 17(4). Mimno,David,HannaMWallach,EdmundTalley,MiriamLeenders&AndrewMcCallum. 2011.Optimizingsemanticcoherenceintopicmodels.In Proceedingsoftheconferenceon empiricalmethodsinnaturallanguageprocessing ,AssociationforComputational Linguistics. Mitchell,Amy,MarkJurkowitz&KennethOlmstead.2014.Social,searchanddirect- pathwaystodigitalnews.Tech.rep.PewResearchCenter.http://www.journalism.org/ 2014/03/13/social-search-direct/. Mitra,Tanushree&EricGilbert.2014.Thelanguagethatgetspeopletogive:Phrasesthat predictsuccessonkickstarter.In Proceedingsofthe17thacmconferenceoncomputer supportedcooperativework&socialcomputing ,ACM. Molek-Kozakowska,Katarzyna.2013.Towardsapragma-linguisticframeworkforthestudy ofsensationalisminnewsheadlines. Discourse&Communication 7(2). 142 Monroe,BurtL,MichaelPColaresi&KevinMQuinn.2008.Fightin'words:Lexicalfeature selectionandevaluationforidentifyingthecontentofpoliticalcon˛ict. PoliticalAnalysis 16(4). Mourão,RachelReis.2015.Theboysonthetimeline:Politicaljournalists'useoftwitter forbuildinginterpretivecommunities. Journalism 16(8). Newhagen,JohnE.1998.Tvnewsimagesthatinduceanger,fear,anddisgust:E˙ects onapproach-avoidanceandmemory. JournalofBroadcasting&ElectronicMedia 42(2). Newman,David,JeyHanLau,KarlGrieser&TimothyBaldwin.2010.Automaticevalu- ationoftopiccoherence.In Humanlanguagetechnologies:The2010annualconference ofthenorthamericanchapteroftheassociationforcomputationallinguistics , AssociationforComputationalLinguistics. Newman,Nic.2011.Mainstreammediaandthedistributionofnewsintheageofsocial discovery.Tech.rep.ReutersInstitutefortheStudyofJournalism,UniversityofOxford. Newman,Nic,RichardFletcher,DavidA.L.Levy&RasmusKleisNielsen.2016.Reuter institutedigitalnewsreport2016.Tech.rep.ReutersInstitutefortheStudyofJournalism, UniversityofOxford.http://www.digitalnewsreport.org/. Nicholls,T,NabeelahShabbir&RasmusKleisNielsen.2016.Digital-bornnewsmediain europe.Tech.rep.ReutersInstitutefortheStudyofJournalism,UniversityofOxford. Nielsen,RasmusKleis.2015. Localjournalism:Thedeclineofnewspapersandtheriseof digitalmedia .IBTauris. Ostrom,Elinor.2015. Governingthecommons .Cambridgeuniversitypress. Palser,Barb.2009.Hittingthetweetspot:newsoutletsshouldusetwittertoreachelusive andvaluableaudiences. AmericanJournalismReview 31(2). Pariser,Eli.2011. The˝lterbubble:Howthenewpersonalizedwebischangingwhatweread andhowwethink .Penguin. Parmelee,JohnH.2013.Politicaljournalistsandtwitter:In˛uencesonnormsandpractices. JournalofMediaPractice 14(4). Parveen,Farzana,NoorIsmawatiJaafar&SulaimanAinin.2015.Socialmediausageand organizationalperformance:Re˛ectionsofmalaysiansocialmediamanagers. Telematics andInformatics 32(1). Pennebaker,JamesW,MarthaEFrancis&RogerJBooth.???? Linguisticinquiryand wordcount:Liwc2001 .Mahway:LawrenceErlbaumAssociates. Porter,MartinF.1980.Analgorithmforsu˚xstripping. Program 14(3). 143 Potthast,Martin,SebastianKöpsel,BennoStein&MatthiasHagen.2016.Clickbaitdetec- tion.In Europeanconferenceoninformationretrieval ,Springer. Price,Vincent,DavidTewksbury&ElizabethPowers.1997.Switchingtrainsofthought: Theimpactofnewsframesonreaders'cognitiveresponses. Communicationresearch 24(5). Purcell,Kristen,LeeRainie,TomRosenstiel&KennethOlmstead.2010.Understandingthe participatorynewsconsumer.Tech.rep.PewResearchCenter.http://www.pewinternet. org/2010/03/01/part-5-news-gets-personal-social-and-participatory/. Rauch,Jennifer.2015.Exploringthealternativdialectic:Whatternative meanstoahybridaudience. Communication,Culture&Critique 8(1). Reis,Julio,FabricioBenevenuto,PVazdeMelo,RaquelPrates,HaewoonKwak&Jisun An.2015.Breakingthenews:Firstimpressionsmatterononlinenews.In Proceedingsof theinternationalconferenceonweblogsandsocialmedia2015 , Roberts,MargaretE,BrandonMStewart&EdoardoMAiroldi.2016.Amodeloftextfor experimentationinthesocialsciences. JournaloftheAmericanStatisticalAssociation 111(515). Roberts,MargaretE,BrandonMStewart,DustinTingley,EdoardoMAiroldietal.2013. Thestructuraltopicmodelandappliedsocialscience.In Advancesinneuralinformation processingsystemsworkshopontopicmodels:Computation,application,andevaluation ,. Roberts,MargaretE,BrandonMStewart,DustinTingley,ChristopherLucas,JetsonLeder- Luis,ShanaKushnerGadarian,BethanyAlbertson&DavidGRand.2014.Structural topicmodelsforopen-endedsurveyresponses. AmericanJournalofPoliticalScience 58(4). Robinson,Helen,AnnaWysocka&ChrisHand.2007.Internetadvertisinge˙ectiveness:the e˙ectofdesignonclick-throughratesforbannerads. InternationalJournalofAdvertising 26(4). Roston,Michael.2015,January22.Don'ttrytoohardtopleasetwitterandotherlessons fromthenewyorktimes'socialmediadesk.http://www.niemanlab.org/2015/01/dont- try-too-hard-to-please-twitter-and-other-lessons-from-the-new-york-times-social-media- desk/. Rowan,David.2014,January2.Howbuzzfeedmasteredsocialsharingtobecomeamedia giantforanewera. Wired. http://www.wired.co.uk/article/buzzfeed. RussellNeuman,W,LaurenGuggenheim,SMoJang&SooYoungBae.2014.Thedynamics ofpublicattention:Agenda-settingtheorymeetsbigdata. JournalofCommunication 64(2). 144 Sandvig,Christian,KevinHamilton,KarrieKarahalios&CedricLangbort.2016.Automa- tion,algorithms,andpolitics|whenthealgorithmitselfisaracist:Diagnosingethical harminthebasiccomponentsofsoftware. InternationalJournalofCommunication 10. Scheufele,DietramA.1999.Framingasatheoryofmediae˙ects. Journalofcommunication 49(1). Scheufele,DietramA.2000.Agenda-setting,priming,andframingrevisited:Anotherlook atcognitivee˙ectsofpoliticalcommunication. MassCommunication&Society 3(2-3). Schmierbach,Mike&AnneOeldorf-Hirsch.2012.Alittlebirdtoldme,soididn'tbelieve it:Twitter,credibility,andissueperceptions. CommunicationQuarterly 60(3). Schudson,Michael.1998. Thegoodcitizen:Ahistoryofamericanciviclife .FreePress. Shaw,DonaldLewis&MaxwellEMcCombs.1977. Theemergenceofamericanpolitical issues:Theagenda-settingfunctionofthepress .WestGroup. Shearer,Elisa&Je˙reyGottfried.2017.Newsuseacrosssocialmediaplatforms2017.Tech. rep.PewResearchCenter.http://www.journalism.org/2017/09/07/news-use-across- social-media-platforms-2017/. Shoemaker,PamelaJ&TimothyVos.2009. Gatekeepingtheory .Routledge. Singer,Jane.2009. Journalisminthenetwork .Routledge. Singer,JaneB.2005.Thepoliticalj-blogger:`normalizing'anewmediaformto˝toldnorms andpractices. Journalism 6(2). Sjøvaag,Helle&EirikStavelin.2012.Webmediaandthequantitativecontentanalysis: Methodologicalchallengesinmeasuringonlinenewscontent. Convergence 18(2).9. Soroka,StuartN.2006.Goodnewsandbadnews:Asymmetricresponsestoeconomic information. JournalofPolitics 68(2). Starbird,Kate.2017.Examiningthealternativemediaecosystemthroughtheproduction ofalternativenarrativesofmassshootingeventsontwitter.In Icwsm , Stieglitz,Stefan&LinhDang-Xuan.2013.Emotionsandinformationdi˙usioninsocialme- timentofmicroblogsandsharingbehavior. JournalofManagementInformation Systems 29(4). Stigler,GeorgeJ.1961.Theeconomicsofinformation. Journalofpoliticaleconomy 69(3). Stone,PhilipJ,DexterCDunphy&MarshallSSmith.1966. Thegeneralinquirer:A computerapproachtocontentanalysis. MITpress. 145 Strapparava,Carlo,AlessandroValituttietal.2004.Wordneta˙ect:ana˙ectiveextension ofwordnet.In Lrec ,vol.4,Citeseer. Taddy,Matt.2013.Multinomialinverseregressionfortextanalysis. JournaloftheAmerican StatisticalAssociation 108(503). TandocJr,EdsonC.2014.Journalismistwerking?howwebanalyticsischangingthe processofgatekeeping. NewMedia&Society 16(4). TandocJr,EdsonC.2016.Journalisticautonomyandwebanalytics. TheRoutledgeCom- paniontoDigitalJournalismStudies Tenenboim,Ori&AkibaACohen.2015.Whatpromptsuserstoclickandcomment:A longitudinalstudyofonlinenews. Journalism 16(2). Thomson,Robert,NaoyaIto,HinakoSuda,FangyuLin,YafeiLiu,RyoHayasaka,Ryuzo Isochi&ZianWang.2012.Trustingtweets:Thefukushimadisasterandinformation sourcecredibilityontwitter.In Proceedingsofthe9thinternationaliscramconference , Trussler,Marc&StuartSoroka.2014.Consumerdemandforcynicalandnegativenews frames. TheInternationalJournalofPress/Politics 19(3). Tsimonis,Georgios&SergiosDimitriadis.2014.Brandstrategiesinsocialmedia. Marketing Intelligence&Planning 32(3). Turner,Fred.2005.Actor-networkingthenews. SocialEpistemology 19(4). Turney,PeterD&MichaelLLittman.2003.Measuringpraiseandcriticism:Inferenceof semanticorientationfromassociation. ACMTransactionsonInformationSystems(TOIS) 21(4). Uribe,Rodrigo&BarrieGunter.2007.Aresensational'newsstoriesmorelikelytotrigger viewers'emotionsthannon-sensationalnewsstories?acontentanalysisofbritishtvnews. EuropeanJournalofCommunication 22(2). Vargo,ChrisJ,LeiGuo&MichelleAAmazeen.2017.Theagenda-settingpoweroffake news:Abigdataanalysisoftheonlinemedialandscapefrom2014to2016. NewMedia &Society Vargo,ChrisJ,LeiGuo,MaxwellMcCombs&DonaldLShaw.2014.Networkissueagendas ontwitterduringthe2012uspresidentialelection. JournalofCommunication 64(2). Vasu,Norman,BenjaminAng,ShashiJayakumarTerri-Anne-Teo,MuhammadFaizal& JuhiAhuja.2018.Fakenews:Nationalsecurityinthepost-truthera. PolicyReport, Singapore:CENS-S.RajaratnamSchoolofInternationalStudies 25. Walker,JackL.1991. Mobilizinginterestgroupsinamerica:Patrons,professions,andsocial movements .UniversityofMichiganPress. 146 Wallace,Julian.2018.Modellingcontemporarygatekeeping:Theriseofindividuals,algo- rithmsandplatformsindigitalnewsdissemination. DigitalJournalism 6(3). Warzel,Charlie.2018,January12.Facebookcouldn'thandlenews.maybeitnever wantedto. BuzzFeedNews. https://www.buzzfeed.com/charliewarzel/facebooks-news- feed-change-is-a-retreat-from-the-public. Weber,MatthewS&PeterMonge.2011.The˛owofdigitalnewsinanetworkofsources, authorities,andhubs. JournalofCommunication 61(6). Weltevrede,Esther,AnneHelmond&CarolinGerlitz.2014.Thepoliticsofreal-time:A deviceperspectiveonsocialmediaplatformsandsearchengines. Theory,Culture&Society 31(6). Whissell,CynthiaM&MichaelRJDewson.1986.Adictionaryofa˙ectinlanguage:Iii. analysisoftwobiblicalandtwosecularpassages. PerceptualandMotorSkills 62(1). White,DavidManning.1950.ThekeepAcasestudyintheselectionofnews. JournalismBulletin 27(4). Yankelovich,Daniel.1991. Comingtopublicjudgment:Makingdemocracyworkinacomplex world .SyracuseUniversityPress. Young,Lori&StuartSoroka.2012.A˙ectivenews:Theautomatedcodingofsentimentin politicaltexts. PoliticalCommunication 29(2). Zhao,WayneXin,JingJiang,JianshuWeng,JingHe,Ee-PengLim,HongfeiYan&Xiaoming Li.2011.Comparingtwitterandtraditionalmediausingtopicmodels.In European conferenceoninformationretrieval ,Springer. Zhu,Jian-Hua,DavidWeaver,Ven-HweiLo,ChongshanChen&WeiWu.1997.Individual, organizational,andsocietalin˛uencesonmediaroleperceptions:Acomparativestudyof journalistsinchina,taiwan,andtheunitedstates. Journalism&MassCommunication Quarterly 74(1). Zillmann,Dolf,LeiChen,SilviaKnobloch&CoyCallison.2004.E˙ectsofleadframingon selectiveexposuretointernetnewsreports. CommunicationResearch 31(1). Zubiaga,Arkaitz.2013.Newspapereditorsvsthecrowd:ontheappropriatenessoffront pagenewsselection.In Proceedingsofthe22ndinternationalconferenceonworldwide web ,ACM. 147