Cognitive Artificial Intelligence The Invisible Invasion ...new media, therefore continuing to learn...

Cognitive Artificial Intelligence The Invisible Invasion of the Media Business

Cognitive Artificial Intelligence | Introduction

03

Cognitive Artificial Intelligence | Contents

02

Introduction 03

Why is this the right time for AI in media? 09

Use Cases 11

#1 Fightagainstfakenewsandoffensivecontent 11

#2 Customer service 15

#3 Narrative-editorial contribution and content creation 18

Relevance to the C-Suite 20

Benefitsandchallenges 21

Contact & authors 22

Contents

ArtificialIntelligence(AI)hasbeenaroundsincethe1950s, but AI and cognitive technologies have only recentlytakenoffinbusiness.

This statement is supported by current venture capital fundingforfirmswithcognitivetechnologies:In2016,over650AIstartupsraised$5bn,ninetimesasmuchasthe$0.59bnthatwereraisedin2012.Amongthemost active AI investors are Intel, Google, GE, and Sam-sung, who have invested in or acquired more than 50 AIstartupssince2011.

Forour2016GlobalCIOSurvey,Deloitteasked1,200IT executives to name the emerging technologies in whichtheyplantomakesignificantinvestmentsoverthenexttwoyears,and64percentincludedcognitivetechnologies.

Therapiddiffusionofnewtechnologiesinthedigitaland cognitive space is also changing the media industry andthewayweconsumenewsandentertainment.

The days when your only means of consuming content werethroughtheTVorprintednewspapersaregone.Today, the media landscape is more fragmented than ever - there are endless ways to watch your favorite showorreadthelatestnews.Socialmediachannelshave become the top destinations for content con-sumption and even enable consumers to generate theirowncontent.Inthemostrecentdigitalnewsre- port by the Reuters Institute, 51% state social media

areoneoftheirsourcesofnewseachweek,while36%say they are comfortable with automatically selected contentbasedonpastconsumption.Inparticular,younger generations are more comfortable with algo- rithmsthanwitheditors.Infact,only40%intheEUagree that they can trust news organizations and jour-nalistsmostofthetime.

Millennials and younger generations already spend more time streaming content than watching TV, which creates an enormous opportunity for both content cre-ators and advertisers to forge a personal relationship withendconsumers.Personalizedcontentisinhighdemand,sharedviasocialmediaplatformseveryday. It looks like power is shifting between traditional and new media, therefore continuing to learn about new technologies and their business solutions is crucial for understandingthenewmedialandscape.Thispaperaims to shed light on the next promising technology thathasthepotentialtodisruptmedia:Cognitiveartificialintelligence.Butwhyiscognitiveartificialintel-ligence ready to change the game for the whole media sector? In the following paragraphs, we will establish a common understanding of AI and cognitive technolo-gies before we examine uses cases in the media sector andtheirrelevancetotheC-Suite.

Introduction


05


04

What is cognitive AI? A stage in automation technology where cognitive tech-nology augments human decision-making capabilities, while gradually being able to perform certain judgment- based tasks independently in a manner similar to human beings.

What is AI? Whilesomesaythatmachinesarenot(yet)capableoftrueArtificialIntelligence,itisabroadlyacceptedideathat an AI refers to a system created by humans that is able to perform tasks, that would otherwise require a human being1.Moreprecisely:CognitiveAIisbasedon programs and/or computers with the following abilities2:

1. Cognition The ability to identify objects visually, understand and transcribe human speech, and understand texts

2. Memory The ability to hold knowledge and to store it somewhere

3. Learning The ability to create knowledge about the world that can be used for reasoning

4. Reasoning The ability to use knowledge about the world in order to deduct conclusions from available information

Interestingly, we now even see systems emerging that showthefirstsignsofimagination,whichofcourseisofspecial interest to the media industry where products areoftentheresultofcreativework.Weoutlinethevarious levels of sophistication of AI below to provide a frameworkwithsometangibleexamples.

RoboticsProcessAutomation(RPA)toolsandsolutionscurrently deployed are used to automate manual and repetitiveactivitiesthatrequirenojudgment.Insteadthey rely on a set of simple rules on which they base theirdecisions.Sincethosesolutionsrequirenoauto-mated reasoning, they are actually not an example of AI.Nevertheless,moresophisticatedroboticssolutionsalso provide automated reasoning and use machine learningtobecomebetterovertime.ThisisaclearsignthatthesolutionissomeformofAI.Iftheyareabletoprocess unstructured data like images, text and ges-tures and understand their actual meaning, one could callitcognitiveAI.

1DeloitteUniversityPress–DemystifyingArtificialIntelligence 2ARTIFICIALINTELLIGENCE–ArtificialIntelligence:Definition,Trends,TechniquesandCases-JoostN.Kok,EgbertJ.W.Boers,WalterA.Kosters,PetervanderPuttenandMannesPoel

Stages of AI

Source:“RoboticsProcessAutomation(RPA)andmoreadvancedautomation”,Deloittepresentation,October2016

• Used for rule-based processes, such as invoice processing exceptions• Addresses priority business problems driven by process breakages• Enables - Faster handling time - Reduced handling costs - Reduced error rates

• Processes requiring judgment such as commercial contract understanding, insights, and implications• Covers machine learning capability• Interprets human behavior

• Used for predictive decision making, such as Amazon Echo and Alexa• Dynamically self-managing and adaptable

• Systems that completely replicate human capabilites• Turing Test Definition: “A test for intelligence in a computer, requiring that a human being should be unable to distinguish the machine from another human being by using the replies to questions put to both”

Stage

RoboticsMimics HumanActions

Intelligent AutomationMimics/Augments Quantitative Human Judgment

Cognitive AutomationAugments Human Intelligence

Artificial General IntelligenceMimics Human Intelligence

Description


07


06

What are the technologies behind cognitive AI?WedistinguishbetweenthefieldofAIandtechnologiesthatemanatefromAI.Individualtechnologiesareget- tingbetteratperformingspecifictasksthatonlyhu- mansusedtobeabletodo.Wecallthesecognitivetechnologies, and it is these technologies that business and public sector leaders should focus their attention on3.

• Computer vision Technology that allows computers to identify objects in images and to describe what is actually going on by using sequences of image-processing operations todecomposeimagedataintomanageablepieces.

Modern image recognition typically uses neural net-works.Technicallyaneuralnetworkisasetofnodes(neurons)thatarelinkedtogetherviaedges.Thenodesaredefinedbyanactivationfunctionandtheedgesgoinginandoutofthem.Eachedgecarriesaso-calledweight.Structurally,thenetworkisorganizedintheform of layers, starting with an input layer, followed by somehiddenlayersandendingwithanoutputlayer.For the task of plain image recognition, the structure used most widely is that of a Convolutional Neural Net-work(CNN),whichmimicsthestructureofthevisualcortexofthehumanbrain.Formorecomplextaskslike describing the actual content of a picture in natural language,onewoulduseadifferentstructure,e.g.aRe-currentNeuralNetwork(RNN).Theinputlayerisusedtocapturetheinputdata,i.e.forimagerecognitioneach node would receive the color information of a pix-eloftherespectiveimage.Thehiddenlayersareusedto store both the knowledge the network has gathered about the world and the logic it uses to reason about it.Theoutputlayerisusedtopresenttheresultsofthenetworksreasoning.Trainingsuchaneuralnetworkis

typically done in the form of supervised machine learn-ing.Thismeansthatthesystemlearnsfromasetofla-beled examples, in the case of image recognition these areimageslabeledbytheircontent,e.g.cat,dog,houseetc.Duringthetrainingprocess,thenetworkhastolookateachimageandmakeaguessaboutthelabel.Dependingonhowgoodtheguesswas,i.e.howfaroffit was from the true label, the weights of the network arechangedslightly.Thisprocessisrepeatedformanyiterations,typicallyuntilnosignificantimprovementcanbeachieved.Astheseneuralnetworkshavemanyhiddenlayers,i.e.theyaredeep,thisprocessisalsocalledDeepLearning.

• Speech recognition The ability to automatically and accurately transcribe humanspeech.

State-of-the-art speech recognition systems rely on neuralnetworks,too.Whilethetrainingprocessisbasically the same as with neural networks for image recognition, their structure is that of a Recurrent NeuralNetwork(RNN).Thesetypesofnetworksdifferfrom typical image recognition networks with regard tothedirectionoftheedgesconnectingthenodes.While in the former the edges always point in the directionoftheoutputlayer,i.e.forward,theedges ofthelattercanalsopointbackwards.

A RNN is able to handle time-dependent data such as speech, as well as text, through its ability during trai- ningtoremembersignificantevents.Thisisimportantforspeechdata(andothersequentialdata)wherewhatcamelasthelpsdeterminewhatcomesnext.Forin-stance,ifsomeonesays“theappleis”,thenextwordismorelikelytobe“red”than“Fred”,althoughbothmightsoundquitesimilar,especiallyoverthephone.

3http://www.theatlantic.com/sponsored/deloitte-shifts/demystifying-artificial-intelligence/257/

Source:“unsupervisedlearningofhierarchicalrepresentationwithconvolutionaldeepbeliefnetworks’,ICML2009,&Comm.ACM2011.HonglakLee,RogerGrosse,Rajesh Ranganath, and Andrew Ng

Deep neutral network example

• Natural language processing Theabilityofcomputerstoprocesstext(extractmeaningfromorgeneratetext),e.g.translateorsummarizeit.

For various reasons, natural language processing is regarded as a very hard problem that is still not com- pletelysolved.Whileaccurateandevensimpleapproa- ches exist to tackle problems like sentiment analysis or identifying the topics a text is about, systems that try to create a good natural language summary of a written text or that translate text from one language to another havestillfartogo.Someofthemainproblemsareambiguity, sarcasm, slang and novel words, inconsis-

tencies, typos and grammatical errors as well as com-plexorlongsentences.To give readers an idea of how NLPworks,onecanstartwithaverysimpleexample.Amuch-usedtechniquefortacklingveryspecificNLPproblems is to transform a text into a vector of word counts, thereby basically just answering the question, howoftendoeseachwordoccurinatext?Onecanthensimplycalculatedistancesbetweendifferenttextvectorstogaininsightintohowsimilarthosetextsare.Sentimentanalysiscanbedonesimilarly,e.g.byfirstassigningsentimentvaluestowords(highvaluestowords that indicate a positive sentiment and low values tothoseindicatingnegativesentiment)andthenmulti-plyingthosevalueswithtextvectors.

1 https://www.smithandcrown.com/what-is-an-ico/ https://www.smithandcrown.com/icos-crowdsale-history/ https://themerkle.com/5-of-the-most-successful-cryptocurrency-icos-to-date/ https://github.com/Scanate/UltimateICOCalendar https://icotracker.net/ http://icorating.com https://www.fastcompany.com/40446051/inside-the-ico-bubble-why-initial-coin-offerings-have-raised-more-than-1-billion-since-january

Input layer‘Image’

Hidden layers‘Filter recognizing low, medium and high level features’

Output layer

Cognitive Artificial Intelligence | Why is this the right time for AI in media?

09


08

Why is this the right time for AI in media?

1. Vast amount of and easy access to dataThanks to social media, mobile devices, and low-cost sensors, the volume of data in the world is increasing at a pace never seenbefore.Theamountofdatadoublesinsizeevery12monthsandisexpectedtoexceed40zettabytes(40billiontera-bytes, or in other words the capacity of 10 billion typical harddrives)by2020,comparedtoonly4zettabytesin20134.The unprecedented growth of data, especially in form of unstructured data, is critical to the advancement of machine learningasthemoredatathesesystemsconsume,the“smart-er”theybecome.

2. Big Data Analytics New techniques for managing and analyzing very large data sets,i.e.BigDataAnalytics,canbecreditedwithadvancesinAIbecause they use statistical models for probabilistic reasoning aboutimages,text,orspeech.ExposingAIsystemstolargesets of data helps to improve and train the decision-making process.

3. Processing powerMoore’sLaworiginatedaround1970andstatesthatthenumber of transistors in a computer chip will double in size approximatelyeverytwoyears,whichcontinuestoapplytoday.As data volumes have grown larger and analysis methods have become more sophisticated, the distributed networks that makedataandprocessingcapacity(clustercomputingviathecloud)accessibletoindividualusershavebecomeexponential-lymorepowerful.Today,wecanquicklyprocess,search,andmanipulate data in volumes that would have been impossible toprocessonlyafewyearsago.Thecurrentgenerationofmicroprocessorsdeliver4milliontimestheperformanceofthefirstsingle-chipmicroprocessorintroducedin1971.However,with regard to certain task this incredible advance is dwarfed by theimprovementthatcomesfrombetteralgorithms.

Wecanusedifferentcognitivetechnologiestodevelopusecasesthroughout the media value chain, starting with content produc-tion, moving to content distribution and ending with customer experience,asshowninthetableonthenextpage.

4https://www.emc.com/leadership/digital-universe/2014iview/executive-summary.htm

Rocket Engine Processingpower& algorithms

Rocket Fuel Large data sets

Source:AndrewNg

While these techniques clearly do not involve any sophisticated understanding of the content of texts, theydogivereasonableresults.However,muchmoresophisticated approaches are required for real under-standing.

Part-of-speechtagging,forinstance,isusedtode-termine which words in a sentence are nouns, verbs, articles,conjunctions,andsoon.Thisthenallowsustodeterminehowthedifferentpartsofasentenceinterrelateandtocorrectlyanswerquestionslike“whodidwhat?”.Part-of-speechtaggingisabasictoolforNLPandhelpstackleproblemslikedisambiguatinghomonyms.Itisusedtopreprocesstextforfurtheranalysis.Otherapproachesusingneuralnetworksforlanguage translation try to teach systems to under-standconcepts,e.g.theconceptoftime.Justafewyears ago, advanced translation systems were trained onbilingualcorpora(preciselytranslatedtext)andthen

translatedthedifferentpartsofasentencebylookingfor the closest representation of those parts in the oth-erlanguage.Newsystems(e.g.GoogleNeuralMachineTranslation)workdifferently.Thegoalhereistoteachthe system to really understand the message of a sen-tence.Whenitcomestotheactualtranslation,insteadof translating a sentence part by part, they generate a new sentence in the target language, that conveys the sameideatheoriginalsentencedid.Theadvantagesofthis technique should not be underestimated, because the earlier approaches required huge corpora for each pair of languages to be translated which often just do notexist.Thisnewapproachinsteadteachesthesys-tem to really understand languages, making translation basicallyabyproduct.Oncethesystemunderstandsa particular language, it can translate without specif-ically being taught to do so by progressing from this language to the idea being conveyed to expressing the ideainallotherlanguagesitknows.

“Languageisapartofourorganismandnolesscomplicatedthanit.”Ludwig Wittgenstein

Cognitive Artifi cial Intelligence | Use Cases – Fight against fake news and off ensive content

1110

Cognitive Artifi cial Intelligence | Why is this the right time for AI in media?

Cognitive technologies in the media value chain

Media Value Chain

Cognitive Technologies

Benefits of Cognitive

Technologies

Natural Language Processing

Speech Recognition

Computer Vision

Contentproduction

Media content is produced based on AI-supported writing, speeches and visualizations.

Narrative-editorial contribution and content creation

Speech-to-text trans-lation including creation of logical storyline based

on input

Narrative-editorial contribution and content creation

Speech-to-text trans-lation including crea-

tion of logical storyline based on input

Contentaggregation

Media content is curated and aggregated based on smart filtering by relevance and preferences.

Combatting fake news and violent content

Intelligent music suggestions based on

stress level of voice

Automated film scene curation and cutting of film material based on

language and audio (e.g. film music) analyses

Contentpackaging

Based on the media content, system can recognize in which format to pack the media in order to make it most appea-ling for target group.

B2B-transparency in media buys based on

visuals

B2B-transparency in media buys based

on voice

B2B-transparency in media buys based on

audio and language

Distribution

Automatic distri-bution of content through most effective channels based on media content.

Automated selection and distribution of

content including auto-mated scheduling

Automated selection and distribution of

content including auto-mated scheduling

Automated selection and distribution of content including

automated scheduling

Accessdevice

Intuitive and easy access to devices including remote controlling.

Cognitive TV program / media guide /

remote control

Cognitive music controling voice

(e.g. Amazon Alexa)

Cognitive news feed based on conversations

Consumer

Cognitive en-hancement of overall consumer experience.

Cognitive chatbots

Next levelcustomer Service

Next levelcustomer service

Time/cost efficiency

Functionality

Generate insights For the scope of this paper, we want to highlight three

use cases that will be relevant to the C-Suite in the media sectorfortheupcomingyears.

Fakenewsandoffensivecontent(e.g.violance,nudityetc.)continuetobetopicsofgreatrelevanceformostmediacompanies.Especiallyfakenewshavebecomesuch a concern for society that for example Facebook has launched a national print advertising campaign againstfakenewsintheUKandotherEuropeancoun-triestoeducatethepublicaheadofpoliticalelections.The problem arises when too much information is shared at a high frequency through the internet, therebynotleavingenoughtimeforamanualreview.As a result, and in absence of working and mature content monitoring solutions, companies such as Facebook hire more and more employees to review and fact-check content to prevent inappropriate infor-mationfromgoingviralandavoidpunitivefines.Thisis however not a long-term solution because, as the volume of data grows rapidly, the manual handling of misinformation and inappropriate content would atsomepointbecomeprohibitivelyexpensive.Manual fact-checking also introduces bias into the

screening process, as it is nearly impossible to prevent personal views, opinions, and beliefs of team mem-bersinfluencingtheirjudgment.Inadditiontofakenews,thereisalsothefightagainstoffensivematerial,i.e.contentwithviolenceand/ornudity,whichispos-tedonsocialmediaandmanuallyflaggedandrepor-tedbyusers.

Currently, it takes quite a long time time until these itemsareremoved.However,pressureisbuilding.TheGerman parliament for instance recently adopted the “Netzwerkdurchsetzungsgesetz”,or“Facebooklaw”,asitiscommonlycalled,whichintroducesfinesofupto50 million euros if problematic content is not removed fastenough.

Cognitiveartificialintelligencecanassistwithensuringdata veracity, identifying fake news and automating thecensorshipordeletionofsuchcontentintelligently.In the digital era, not only social media companies, butalsotraditional(news)mediacompaniescanusecognitive technologies to detect inappropriate content, helpingthemremaincredibleandtrustworthy.Thequestionsare:Howcantheydoitandwhereshouldthey start?

Use Cases

Cognitive AI has the potential to reduce themanualeffortofdataveracity,theidentificationoffakenewsandviolentcontenttoassurecredibilityandtrust.

#1 Fight against fake news and off ensive content

Cognitive Artificial Intelligence | Use Cases – Fight against fake news and offensive content

13


12

Current situationNot only Facebook, but also many news media compa-nies have editors or a designated team that fact-check theirstories.Theteamverifystheauthororcreatorofthe article, names stated, titles, geographic locations, numbers,sourcesused,andsoon.ThereareexternaltoolssuchasSpike,Hoaxy,orGoogleTrendsthatcanaidthefact-checkteamduringthisprocess.

However,notonlyisthistime-consumingandreliesonhuman judgment, it is also impeded by the pressure for clicks which has pushed many news organizations tonotlooktooclosely.Shouldastoryturnouttobefake, they will publish another story that reveals the “real”truthtopushthenewscycleanothertimeandgenerateanotherroundofclicks.Ofcoursetheseblacksheep thereby willingly undermine the trustworthiness ofnewsmediacompaniesasawhole.

Current challenges Therearethreemainchallengesregardingtheverifica-tionofcontent:

• Increasing amount of dataThe amount of content produced and shared is growing at a fast rate and presents a serious chal-lenge to fact-checking teams as they tackle their task manuallywithlimitedresources.

• Longer detection timeVerifyingsourcesandauthorstakesasignificantamount of time, which editors are often unwilling to invest, especially when they are under constant pres-suretoquicklyscoutthenextsensationalstory.

• BiasThe fact-checking team is made up of human beings with their own emotions and political views, which caninfluencetheirjudgmentofcontents.

AI-based solution With cognitive technology, media companies can more effectivelyidentifyfakenewsandcontentdisplayingviolenceornudity.Techniquesthathelpinclude(1)naturallanguageprocessing(NLP)and(2)ImageRec-ognition.

(1) OnecanstartbyusingNLPtechniquestoex-tract information from a text, be it an article or a postonasocialmediaplatform.Aclassificationmodelhelpsidentifythegeneraltopic(s)atextisaboutandwhetherornotatextisclickbait.Thelatter would work similar to how spam mail is identi-fied.Othertechniquescanbeusedforinformationextraction, which is basically about extracting structureddatafromatext.

Therearemanydifferenttechniquesforperformingtextclassification.Asimpleonewouldbeanunsu-pervised machine learning technique called k-means clustering applied to a collection of sample texts, which couldbeturnedintotextvectorsasdescribedabove.K-meanswouldthengroupthosetextsthataremostsimilarintodistinctgroups.Asthisisatechniqueforunsupervised learning, the engineer would have to look attheresultinggroupsandassignthemnames(e.g.appropriateandnotappropriate).Theresultingmodelcanthenbeusedtoclassifynewtext.Ofcoursethisisa very limited approach, much better yet more complex methodsexist.Anexamplewouldbeusingneuralnetworksandatechniquecalledword2vec.

(2) In case a text comes with a picture or video, which technically is just a sequence of pictures, one can useamodelforimageclassificationtoidentifyitscontent.Facebookforinstanceusesneuralnetworks to add tags about the content of a photo toeveryphotothatisuploaded.Ifatagisregardedas problematic, the photo/video can simply be removed.

Tobuildsuchamodel,onewouldfirsthavetocreatea database containing many labeled pictures which actuallycontaintheunwantedcontent.Usingmachinelearning,thesecanthenbeusedtotrainamodel(e.g.aneuralnetwork)whichscreenseverypicturepostedonawebsite.Thesameprincipleappliestovideosbutrequiresmuchmoredataprocessingpower.

(3) In a next step, the information extracted from the content in previous processing steps can be used for fact checking, which could be simply done by comparing the now structured information with information from fact databases or comparing the information with information extracted from reliablesources.

Impact on media companies TheuseofcognitiveAIallowsmediacompaniestofightfakenewsandviolentcontentmoreeffectivelyandwithhigherreliability.Itcansupporteditorsandjournalistsin making informed decisions about publishing a speci- ficpieceofcontent,asithasthepowertocopewiththevast amount of data and search the whole internet for proof to verify a piece of content, while also eliminating political and emotional judgment human beings have oncertaintopics.ThemainbenefitsofcognitiveAIinthefightagainstfakenewscomemainlyintwoforms:(1)costsavingfromeliminatingtime-consumingfactchecks, so editors and journalists can focus on their mainwork,and(2)highertrustworthiness.

“Ifwearenotseriousaboutfactsandwhat’strueandwhat’snot,ifwecan’tdiscriminatebetween serious arguments and propaganda, thenwehaveproblems.”Barack Obama (Guardian)

Cognitive Artificial Intelligence | Use Cases – Customer Service

15


14

Example Imagine CNN tweets about a certain topic, how can cognitive AI tech-nology support other media news companies to verify the piece of information?

The cognitive technology could consist of a model trained to identifyclickbaitsimilartoaspamfilterthatlearnstorecognizespam.ThroughNaturalLanguageProcessing,italsounderstandsthe semantic meaning by assessing the subject, headline, text, and locationofspecificcontent.

Through the analysis of text, pictures, metadata, history, and comparison to past content, the accuracy of context presented receives a score in an attempt to understand the context without relyingonexternalsources.

The cognitive technology measures and scores the reputation basedontrustworthyrankingwebsites(e.g.Alexawebrank)byconsideringfeaturessuchasdomainnameortraffichistory.

The cognitive technology can analyze other websites and check for the same or similar facts to weigh it against credible media sources andassignascore.

Language

Reputation Facts

Context

LanguageNo buzz words identified,correctuseof grammarScore: 5/5

ReputationVerifiedaccount,realfollower base, posts regularlyScore: 5/5

ContextContent posted on Twitter, scans history of past postsScore: 5/5

Facts Similar facts found on other websites, verifiedsourcelinkScore: 4/5

Total Score19/20:Contentistrustworthy and most likely true

CNN

Netflixplanstospendupto$8billion on programming next year cnn.it/2ifoVSX.

The digital world has introduced new opportunities for media companies, especially in building a personal rela-tionship with customers as they switch more and more to on-demand content and subscription-based pay-mentmodels,likethevideostreamingservicesofferedbyNetflixorAmazonPrimevideo.However,customershave more power than they used to and are becoming moredemandingthanever.Today’scustomersexpectcustomer service on their own terms, which is why media companies need to keep customer service satis-factionhighinordertostayaheadofthegame.

According to an IBM study, 91% of customers have called a customer service more than once to solve the sameissue.Furthermore,90%complainedthattheyhad been put on hold for too long, while 80% had to repeatthesameinformationtoadifferentpersondur-ingthesamecall.5Poorcustomerserviceisthemainreason why half of all customer service calls remain unresolved,whichleadstosignificantlossesofcusto- mersformanycompanies.

As a result, more money is invested to improve custom-erserviceandtheoverallcustomerexperience.Netflix,for instance, has opened a new European customer service hub in Amsterdam just to keep customers hap-pywiththeirservices.CognitiveAIcanaugmenthumanintelligence to increase customer satisfaction and make theirlivesbetteraswellassavesignificantresourcesforthemediacompaniesandincreaseretentionrate.In fact, IBM predicts that by 2020, more than 85% of customer interactions will be handled without human involvement.6

5IBM(https://www.ibm.com/think/marketing/customer-service-of-the-future-is-powered-by-artificial-intelligence/) 6IBM(https://www.ibm.com/think/marketing/customer-service-of-the-future-is-powered-by-artificial-intelligence/)

Cognitive AI can augment hu-man intelligence to increase customer satisfaction and maketheirlivesbetter.

1.

Identify customer

2.

Verify problemand assign the right customer service represen-tative

3.

Identify solution

4.

Provide solution or rework case (if solution fails)

5.

Escalate and reassign case to appropriate customer service representative

6.

Close the case & store data collected to improve custo-mer service

#2 Customer Service


17


16

Current situation Current customer services are usually provided by phone, email or live chat with people who have been trained to assist customers with every issue they could potentiallyhave.Afterreceivinganissue,whichwecalla customer service case, it goes through the following simplifiedprocess:

Bots have been implemented to assist or even perform certainstepsintheprocessoverthephoneorviachat.For example, the bot can name potential causes for a problem based on the information provided or it can decidewhichquestiontoasknext.WiththehelpofAI, these bots are becoming smarter and can take on moreandmoretasks.

Current challengesThere are four main challenges current customer servicesystemswithinamediacompanyface:

• Inconsistency of serviceProvidingaconsistentlyhighqualitycustomerser-viceexperienceisessential.However,inrealityitcanvarymassivelybetweenagents.Customersdemandconsistent and accurate information, no matter whichagenttheytalkto.

• Inefficient information sharingFinding the solution requires the agent to master a complex process and know where to look for information.Thisisoftenatime-consumingprocessandleadstoinefficiency,asmostcontentisstoredasunstructuredtext.

• Availability and time to solutionMost customers want a solution now, therefore the solution-findingprocessnotonlyneedstobecomefasterandsmoother,butalso24/7.

Drawbacksofcurrentbots:Eventhoughbotsarebecoming smarter, they still lack human-like features tomaketheconversationmorenatural(e.g.Alexa,Siri)and pull the majority of answers from a predetermined directory of responses because they fail to capture the meaningbehindcertainquestions.

Cognitive AI-based solution WithcognitiveA.I.,companiescanoffersuperiorcus-tomer service by combining the power of bots with that ofhumansbymakinguseoftheAI-basedtechnologies:

ThesystemcanuseNaturalLanguageProcessing(NLP)toextractstructureddatafromnaturallanguage,and use this information to create a more capable chatbot.Thiswayitcananalyzeadditionalsourcesofunstructured data, such as publicly available customer information,findsimilarcasesinthecasehistoryandalsolookforcluesinotherdata(errorcodes,supportdocuments,technicaldata)tofindcustomer-specificsolutions.Behindthescenes,advancedAIsolutions(incontrasttoe.g.asimpledecisiontreemodel)presenttheirsolutionstogetherwithaconfidencelevel,whichcan be used to decide whether or not the answer is presented directly to the customer or if the case shouldbeforwardedtoahuman.TheagentwouldtheninstantlyhaveaccesstotheresultsoftheAI’sanalysis, which in turn helps them resolve the case faster.Oncethecasehasbeensolvedsuccessfully,de-pending on how it was solved, machine learning could beusedtoimprovethebots’abilities.

Technically, creating a continuously learning system can be done by using an optimization algorithm called ‘stochasticgradientdescent’.Oneofthemostbasicoptimization algorithms behind machine learning is the ‘gradientdescent’algorithm.Itisanalgorithmwhichiterativelyupdatesamodel(the‘brain’ofthecognitiveAI)bylookingatallofthetrainingdataandimprovingitinrelativelybigsteps.Bycontrast,with‘stochasticgra-dientdescent’themodelisupdatedjustalittlebiteachtimeatrainingexampleislookedat.Besidesallowingcontinuous learning, this method is also computa-tionally much cheaper, which is important when the trainingdatasetsarereallybig.However,itsdownsideisthatthelearningpathismuch‘bumpier’,meaningthat single updates can easily make the system a tinybit‘dumber’foracoupleofiterations.However,deciding on whether or not one really wants to have a continuously learning system is actually a balancing act,asthiscanalsoleadtoreallyproblematiceffects.Agood example for things having gone quite wrong is a

chat bot from Microsoft which quickly became racist by incorporating the knowledge it gained from the conver-sationsithad.ThishighlightshowimportantgoodAIconvergencewillbecomeinthefuture.

Impact on media companies The use of cognitive AI to enhance customer service canbenefitmediacompaniesinthefollowingways:• 24/7customerserviceandreducedtimetosolution- finding• Cost saving from decreased labor-intensive support services • Highercustomersatisfactionandretentionrate• Personalizationofcustomerservicebasedon customer data records

The‘AndChill’NetflixchatbothelpscustomerscopewithNetflix’over-whelmingfilmlibraryandrecom-mends movies based on a few questionsvialivechat.Recommen-dations themselves are generated viaNetflix’recommendersystem,anothercognitiveengine.*

Source:ScreengrabANDCHILL

Messages DetailsAnd Chill

Sure! What’re you in the mood for?

I’m looking for a Netflix movie to watch

Something like The Interview where they get into crazy situations!

Check out Talladega Nights! Will Ferrell always brings the laughs. Here’s the trailer: andchill.io/tal-ladega

Awesome, thanks! :-* :-*

:-D

*75%ofNetflixviewershipisdrivenbyrecommendations”Source:BusinessInsider8http://www.businessinsider.com/netflixs-recommendation-engine-drives-75-of- viewership-2012-4?IR=T

Cognitive Artificial Intelligence | Use Cases – Narrative-editorial contribution and content creation

1918

Current situationCurrently, editors screen relevant channels to identify upcoming(hot)topicsor,insomecases,areinformedbycontactsaboutimportanteventsandnews.Withthe increasing amount of data, especially unstructured data,thatisavailableonline,itbecomesdifficultforeditors to analyze and identify relevant content in real time.

Current challengesThis leads to time delays between a current event and the moment it is reported or published, as well as editors spending their valuable time on research instead ofcreatinghigh-qualitycontent.

Cognitive AI-based solution Cognitive AI is likely to help editors identify key pieces of information to report, as they emerge, by applying its ability to process and analyze large amounts of data inamuchfastermanner.Asystemthatreallyunder-stands the information, compared to one that would just do something simple like a key word search, can linkdifferentdatasourcesinrealtime.Forinstance,asystem powered by cognitive AI can connect data from social media with search trends and weather data and therebyderivefact-basedcontent. A real-life example of such a system is a bot that used Natural Language Generation to write a Los Angeles Times report once it discovered an earthquake had happened,whichwasactuallyalmostinstantaneously.ThiswayitmadetheLATimesthefirstmajoroutlettoreport the event, which is quite important in a business

wherenothingisolderthanyesterday’snews.Asalreadymentioned,NaturalLanguageProcessing,andespeciallyNaturalLanguageGeneration(NLG),arestillchallengingtasks.Understandingandcreatinghumorousandsarcastictextisexceptionallydifficultfor cognitive systems to master - with the exception of unintended humor!

However,ifonewantstoprocessorcreatefact-basedtext,today’ssystemsareadefiniteoption.NLGworksbestifthegoalisconveyinginformation.Articlesaboutrecurring topics often follow quite a standardized scheme, which is why technically rather simple systems willusuallysuffice.Automatingtheproductionofthesekinds of articles can be done by reusing text snippets andcustomizingthematdefinedpositionswiththemostcurrentdata.Onecoulddescribethisasatem-plate-basedapproach.Moreadvancedsystemsexist,but the technical concepts behind even some of the mostadvanced,likethe‘TodaiRobot’,acognitivesys-temdesignedtopasstheadmissiontesttoTokio’seliteuniversity,resemblemorethatofaclever‘copy+paste’strategythanthatofahumancreativewritingprocess.However,whiletheresultsofsuchsystemscanbequiteimpressive–theessayswrittenbythelatestversion of the Todai robot, for instance, were better thanthosebymoststudents–awritingstrategybasedon copying and pasting is problematic, and not just fromacopyrightperspective.Whiletheresultsofsucha system might not be ready to print, they could help acceleratethejournalisticresearchprocessinfuture.

#3 Narrative-editorial contribution and content creation

Cognitive Artificial Intelligence | Use Cases – Narrative-editorial contribution and content creation

Example of Narrative Science’s investment report7:

Based on the informative nature of AI-produced texts, amajorpartofAI“news”programscanbefoundinthefinancialindustry.Almost60%oftheclientsofNarrative Science, a company that provides AI narrative programs,areinfinancialservices.Severalothercom-panies are also active in the AI market with applications that can derive expert-level reports, analyze current social media streams, and help brands and others to understand the impact of branded content on their au-diences.Anotherchallengeisthatoncetheunderlyingalgorithms have been understood, bots could be used tofloodtheinternetwithfalseinformationonseveralchannels.AnarrativeAIwouldlinkthisinformationandcomeupwithnewsthatprovidesfalsefacts(“fakenews”).Informationthereforeneedstobeverifiedbeforepublication.

7https://www.wired.com/brandlab/2015/04/news-flash-ai-startups-reinventing-media/ 8http://futurecontent.co/automated-content-can-algorithms-write-your-content/

Impact on media companies Media companies can save on writing and editing costs byusingAItocreatebasiceditorialtexts.EditorsandjournalistscanalsobenefitfromAIbyoutsourcingworkorpre-workonarticles.Thisfreesuptimeforthem to focus on deeper investigations and more so-phisticated articles leading to potentially higher quality injournalism.

Example of a sports article from The Guardian written by Wordsmith8:

“Theenergysectorwasthemain contributor to relative performance, led by stock selection in energy equipment andservicecompanies.Interms of individual contrib-utors, a position in energy equipment and services com-panyOceaneeringInternatio- nal was the largest contributor toreturns.”

“Itwasaseasonfortheagesfor Leicester City as they lifted thePremierLeagueTrophyand were crowned champions ofEngland.LeicesterCityfeaturedoneoftheleague’smost skillful attacks, netting 68goals.JamieVardyledthewaywithanincredible24goals.Inadditiontotheirof-fensive prowess, Leicester City possessed one of the strongest defenses in Eng-land.”

Cognitive Artificial Intelligence | Benefits and challenges

21

Cognitive Artificial Intelligence | Relevance to the C-Suite

20

The progress of cognitive technologies within AI has the potential to revolutionize businesses and the respective operating and business models in themediaindustry.

Relevance to the C-Suite

Embracing cognitive AI and the adoption and imple-mentation of the respective technologies and solutions isrelevanttoalmostallCxOareas:

• Enable new revenue sourcesIncorporatingArtificialIntelligencetechnologiesintobusiness models opens new revenue sources, such as selling AI applications and managed services to othermediaplayers(e.g.fullyautomatedvideocap-turingandarchivingorAI-basedcontentcreation).

• Establish new partnerships / M&A strategyDependingonthelevelofAIintegrationintotheoverall operating model, the new technologies will alsorequireCEOstoforgenewstrategicpartner-shipsand/orintegrateAIintotheM&Astrategy.

• Realize cost savingsAI solutions can increase business automation as well as support processes and related decisions usu-ally performed by humans and thereby decrease the number of employees required as well as respective costs.Especiallyinsupportprocesses(customerservice,finance,HRetc.)thereisgreatpotentialforcostsavingsthroughtheemploymentofAI.Inthecontent creation process for fact-based news, AI also has the potential to relieve the editorial teams of less creativetasks.

• Introduce error-proof and auditable processesDuetoprocessautomationinalmostallareasandprocesses, all transactions and activities will be digi-tallyrecordedandthereforeeasilyauditable–resul- ting in less preparation as well as internal compliance andexternalauditingefforts.

• Eliminate legal and regulatory risks Potentiallegalandregulatoryrisksforthedistribu-tionofexplicitorillegal(usergenerated)contentscan be avoided, once AI monitoring tools for audio-visualcontentbecomemoreefficient.

• Streamline operations / customer serviceCognitiveAIsolutionsprovideCEOswithgreatercontrol over operations and related costs due to the potential replacement of humans with AI, which enables“bots”todomorethanjustexecutesimpleandrepetitivetasks.AdditionallyclientsatisfactionandbrandperceptioncanbesignificantlyenhancedbyincreasingtheleveloffirsttimeresolutionwithAI-basedandself-learningcustomerservicesolutions.

• Streamlining of program monitoring Especially the automated detection of fake news as wellasprohibitedcontentscouldrelieveCCOsofalot oftheirpainpoints.Itisgettingmoreandmorecom- plex as well as time and resource-consuming to scan the vast amount of information pieces and detect fake news in the process of creating compelling and well- researchedprograms/audiovisualcontentpieces.

• Automated content creationAdditionally, the assembly and aggregation of de-scriptive news feeds / news tickers without adding opinion will be largely automated, thus the amount ofstaffworkingintheeditorial/programofficescanbe reduced and / or redirected to create higher-value contentwithajournalisticambition.

• Improved securityAI can erase the biggest source of failures, as well as securitybreaches,i.e.thehumaninterface.AIena-bles processes and operations to be fail-proof and moreresistanttosecuritythreats.

We have seen that AI has the potential to open new areasofactivitiestomediaplayers.Inaddition,AIhasthe potential to solve some of the operational issues throughprocessautomation.

Benefitsandchallenges

Ultimately, the impact depends on how actively media playersdrivetheadoptionofusecases.Goingforward,therewillbeaneedtoexpandonAIbenefitsandover-cometheremainingchallenges.

BenefitsOrganizationsgainmajorcompetitiveadvantagesbyusingcognitivecomputing.CognitiveAIisakeylever for enabling future business growth by enhancing cus-tomerengagement,productivityandefficiency.Com- panies from various industries have already achieved arangeofbeneficialoutcomesviatheircognitiveinitiatives.

ChallengesTop adoption challenges include the cost of technology and security concerns, as well as gaps in the maturity of toolsanddeveloperskillsets.Inconclusion,thebene-fitsofadoptingandintegratingAIincoreandauxiliaryoperationsareplenty,ashighlightedabove.Mediaplay-ers should take a long-term view on AI and its potential to add value to the enterprise in both their current andnewbusinessmodels.Therewillbechallengesto adopting AI, as with any new technology that holds

thepromiseofsignificantdisruption.However,mediaplayerscouldbenefitfromjointlyrealizingcollaborationand alliances around AI to establish industry standards and avoid wrong investment decisions in non-standard AItechnologies.Therearealreadymanycompaniesinthe(digital)mediasectorwhichareinvestingheavilyinthedevelopmentofmatureAIsolutions.

“Forgetdigital–cognitivebusinessisthefuture.“Ginni Rometty (IBM)

Contact & Authors

Contact

Alexander MoggPartner | ConsultingTel.: +49 (0)151 5800 [email protected]

Cognitive Artificial Intelligence | Contact & Authors

Authors

Mirko GramatkeDirector | ConsultingTel.: +49 (0)151 5800 [email protected]

Johannes GegnerManager | ConsultingTel.: +49 (0)151 5800 [email protected]

Sascha BauerSenior Manager | ConsultingTel.: +49 (0)151 5807 [email protected]

Marcus BoenischSenior Consultant | ConsultingTel.: +49 (0)151 5800 [email protected]

Viet Dung Bruce LeConsultant | ConsultingTel.: +49 (0)151 5807 [email protected]

22

This communication contains general information only not suitable for addressing the particular circumstances of any individual case and is not intended to be used as a basis for commercial decisions or decisions of any otherkind.NoneofDeloitteConsultingGmbHorDeloitteToucheTohmatsuLimited,itsmemberfirms,ortheirrelatedentities(collectively,the“Deloittenetwork”)is,bymeansofthiscommunication,renderingprofessionaladviceorservices.NoentityintheDeloittenetworkshallberesponsibleforanylosswhatsoeversustainedbyanypersonwhoreliesonthiscommunication.

DeloittereferstooneormoreofDeloitteToucheTohmatsuLimited,aUKprivatecompanylimitedbyguarantee(“DTTL”),itsnetworkofmemberfirms,andtheirrelatedentities.DTTLandeachofitsmemberfirmsarelegallyseparateandindependententities.DTTL(alsoreferredtoas“DeloitteGlobal”)doesnotprovideservicestoclients.Pleaseseewww.deloitte.com/de/UeberUnsforamoredetaileddescriptionofDTTLanditsmemberfirms.

Deloitteprovidesaudit,riskadvisory,tax,financialadvisoryandconsultingservices to public and private clients spanning multiple industries; legal advisoryservicesinGermanyareprovidedbyDeloitteLegal.Withagloballyconnectednetworkofmemberfirmsinmorethan150countries,Deloittebrings world-class capabilities and high-quality service to clients, delivering theinsightstheyneedtoaddresstheirmostcomplexbusinesschallenges.Deloitte’sapproximately263,900professionalsarecommittedtomakinganimpactthatmatters.

Cognitive Artificial Intelligence The Invisible Invasion ...new media, therefore continuing to learn...

Documents

Transcript of Cognitive Artificial Intelligence The Invisible Invasion ...new media, therefore continuing to learn...