You are on page 1of 2

HenryKissingervs.

SentimentAnalysis
WALIDS.SABA,PhD CIO,Pragmatech walid.saba@pragmatech.com Thoseofusthatworkinnaturallanguageprocessing(NLP)knowverywellthatunderstandingnatural languagerequiresmassiveamountofcommonsenseknowledge,knowledgethatafiveyearoldhas e.g.,tablesdontlaugh,peoplesleep,elephantsdontfly,itmakessensetosayredcarbutnotred opinion,etc.Weimmediatelyandeffortlesslyunderstandwhatawaiterinarestaurantmeanswhenhe saysthecornertablewantsanotherbeerbecauseweknowtablesdonthavewants(andthey certainlydontdesirebeer),soitmustbesomepersonsittingatthecornertablewhowantsthebeer! Thisspecificphenomenon,whichiscalledmetonomyinthecomputationallinguisticsliterature,isbut oneofamultitudeofproblemsthatwestilldonothaveacomputationallyeffectivesolutionfor. Quantifierscoperesolutionisanotherphenomenonthatwestilldontquiteunderstand.InsayingJon boughtahouseoneverystreetinhisneighborhoodwedontmeanthereisasinglehousethatison everystreetinJonsneighborhood,ahousewhichJonbought,butinJonadvertisedahouseonevery streetinhisneighborhoodwecouldverywellmeanthattheresasinglehousethatJonadvertisedon everystreetinhisneighborhood. Withoutdelvingintothedetailsofanumberofphenomenainnaturallanguagethatwestilldonothave acomputationallyeffectivesolutionfor,letmejustsaythat,asofyet,theresnocomputerprogram thatcantrulyunderstandsimple,everydayspokenlanguage,notwithstandingalltheclaimsthatare beingmadeeitherforcommercialreasons,orsometimesbythosewhodonotquiteunderstandthesize oftheproblem(afterall,someasearlyasthe1950sthoughtthatwithinafewyearstheywouldhave programsthatcandoeffectivemachinetranslationwerestillwaiting,bytheway!) IamnotbeingnegativetowardsNLP.Imyselfworkinlanguageprocessing.Furthermore,Iamastrong believerthatweCANbuildsystemsthatunderstandordinaryspokenlanguage.However,Ibelievethe problemismuchmoredifficultthansomethink,andIbelievewearestillfarfromachievingthat monumentalchallenge.Whatwecandoatthemomentisunderstandwhatapieceoftextisabout thatis,whatthesubjectmatterofapieceoftextis,whatarethekeytopics,andwhat(named)entities arebeingmentionedandwhataretheirtypes(peoplevs.products,organizations,brands,companies, locations,etc.)Eventhissimpletask,hasnotbeenperfected,buttherearesystemsthatdoaverygood job(incidentally,weatPragmatechjustfinishedtheconstructionofonesuchsystemthatwebelieveis thebestinthisregard.) Iftherelativelysimpletaskofunderstandingwhatacertainpieceoftextisabouthasnotbeen perfected,itisbeyondmycomprehensiontohearsomespeakofsentimentanalysis.Sentimentanalysis isactuallymuchharderthanunderstandingsimpleordinaryspokenlanguage,which,asIarguedabove isaproblemthatwearefarfromsolving(recallthecornertablethatwantsabeer!) Tomakethepointthatnoserioussentimentanalysiscanatthispointbedone,Iwillhavetorefertoa famousdiplomat,knowntheworldover.IrecalloncehearingHenryKissingersaying(Ibelieveinan interviewwithCharlieRose):theUSistheworstplacetolivein,untilyoutrylivinganywhereelse.

ThesebrilliantonelinersareclassicHenryKissinger.InthisstatementDr.Kissingerwasclearlymaking anextremelypositivestatementabouttheUnitedStatesofAmerica,thatwithallitsimperfections,the USisstillthebestplaceintheworldtolivein.Realizingthepositivesentimentinthissentencetowards theUSclearlyrequiresdeepknowledgeknowledgeofculture,politicsandevenpsychology (intentions,etc.).Incidentally,removingtwowordsfromthissentenceturnsitintoanextremely negativesentenceabouttheUS: 1. TheUSistheworstplacetolivein,untilyoutrylivinganywhereelse(US+ve) 2. TheUSistheworstplacetolivein,trylivinganywhereelse(USve) ToinferthecorrectsentimentabouttheUSin(1)and(2)adeepanalysisandquiteabitofworld knowledgeisneeded,andsomachinelearningandstatisticalmethodsarehelplesshere(astheyare alsohelplesselsewhere,inmyopinion,butthatsanothersubject). Iftheaboveexampleisnotconvincing,heresanother(afterthisexample,youwillrealizeIcanmakean infinitenumberofexamples!): 3. Idontlikesmartphones,IhateiPhone,andIdontlike BlackBerryandIcertainlydislikeSamsung.Idont likethewholetechnology,whetheritsiOS,Android,or whatever.Allofthisstuffisjunk.Idontseeanyneedfor thistechnology. OK,OK,Imkidding.Idontreallymeananyoftheabove. Actually,myfeelingsarecompletelytheoppositeof everythingIsaidabove. Theoretically(andthusactually!)nostatisticalormachinelearningalgorithmcanlearnthattheabove patternindicatesapositivesentimenttowardssmartphones.Thisisnottheplacetomakeascientific proofofthisclaim,butIhopetheaboveexamplesareenoughtoconvincepeoplethatwhateverisnow calledsentimentanalysisisnothingmorethanguesswork. Hypeisaphenomenonthatisusedinmanydomains.Wehypemusicians,movies,politicians,products, andsoon.Butwhenitcomestohype,weintechnologysectoraremasters.RememberExpertSystems thesesystemsweresupposedtoencodefewrulesacquiredfromdomainexpertsandthenhelpus solveanyproblem.Thiswasover30yearsago.Hypeisnotbad,butwhenitisoverdone,itcanbevery damagingtoallofusinthefieldthisisactuallywhathappenedwithsemantictechnology! Whenclaimsaregrand,andwhenatechnologyfails,andmiserablyso,itshutsthedooronmany opportunitiesforprogress.Beforethetapofinvestinginlanguageprocessingisstopped,letsbehumble aboutourclaims.Beexcitedandthinkbig,butdontconvincethelaymanthatwealreadyhavesystems thatcaninferfromwhattheywritewhattheylikeanddontlike.Letsnotmakeanegativesentiment aboutsentimentanalysis.

You might also like