RESEARCHARTICLE Why are some languages confused for others? Investigating data from the Great Language Game HedvigSkirgård1☯,Sea´nG.Roberts2☯*,LarsYencken3☯ 1TheWellspringsofLinguisticDiversityLaureateproject,DepartmentofLinguistics,SchoolofCulture, HistoryandLanguage,CollegeofAsiaandthePacific,AustralianNationalUniversity,Acton,ACT,Australia, 2Language&CognitionDepartment,MaxPlanckInstituteforPsycholinguistics,Nijmegen,Gelderland, Netherlands,3Independentresearcher,Melbourne,Victoria,Australia a1111111111 a1111111111 ☯Theseauthorscontributedequallytothiswork. *[email protected] a1111111111 a1111111111 a1111111111 Abstract Inthispaperweexploretheresultsofalarge-scaleonlinegamecalled‘theGreatLanguage Game’,inwhichpeoplelistentoanaudiospeechsampleandmakeaforced-choiceguess OPENACCESS abouttheidentityofthelanguagefrom2ormorealternatives.Thedatainclude15million Citation:SkirgårdH,RobertsSG,YenckenL guessesfrom400audiorecordingsof78languages.Weinvestigatewhichlanguagesare (2017)Whyaresomelanguagesconfusedfor confusedforwhichinthegame,andifthiscorrelateswiththesimilaritiesthatlinguistsiden- others?InvestigatingdatafromtheGreat tifybetweenlanguages.Thisincludessharedlexicalitems,similarsoundinventoriesand LanguageGame.PLoSONE12(4):e0165934. https://doi.org/10.1371/journal.pone.0165934 establishedhistoricalrelationships.Ourfindingsare,asexpected,thatplayersaremore likelytoconfusetwolanguagesthatareobjectivelymoresimilar.Wealsoinvestigatefactors Editor:NielsO.Schiller,LeidenUniversity, NETHERLANDS thatmayaffectplayers’abilitytoaccuratelyselectthetargetlanguage,suchashowmany peoplespeakthelanguage,howoftenthelanguageismentionedinwrittenmaterialsand Received:December22,2015 theeconomicpowerofthetargetlanguagecommunity.Weseethatnon-linguisticfactors Accepted:October20,2016 affectplayers’abilitytoaccuratelyidentifythetarget.Forexample,languageswithwider Published:April5,2017 ‘globalreach’aremoreoftenidentifiedcorrectly.Thissuggeststhatbothlinguisticandcul- Copyright:©2017Skirgårdetal.Thisisanopen turalknowledgeinfluencetheperceptionandrecognitionoflanguagesandtheirsimilarity. accessarticledistributedunderthetermsofthe CreativeCommonsAttributionLicense,which permitsunrestricteduse,distribution,and reproductioninanymedium,providedtheoriginal authorandsourcearecredited. DataAvailabilityStatement:Allrelevantdataare withinthepaperanditsSupportingInformation Introduction files. Objectivestudiesofsimilaritiesanddifferencesoflanguagesallowresearcherstoreconstruct Funding:SGRissupportedbytheMaxPlanck patternsofhistoricalchangeinlinguisticstructures[1],orsocialchangessuchasmigrations societyandanERCAdvancedGrantNo.269484 [2,3].However,knowledgeoflinguisticvariationisalsoimportantintheday-to-daylifeof INTERACTtoStephenLevinson.Thefundershad non-linguists.Beingconfrontedwith,andoftenconfoundedby,adifferentaccentoraforeign noroleinstudydesign,datacollectionand analysis,decisiontopublish,orpreparationofthe languageisperhapsauniversalhumanexperience;weareofteninsituationswherewehear manuscript. snippetsofunfamiliarconversationinapublicspaceorontheradio,andwetrytoidentify whatitiswehear.Peoplearesensitivetodifferencesinpronunciationandvocabularyandare Competinginterests:Theauthorshavedeclared thatnocompetinginterestsexist. quicktolinkthesedifferencestotheirknowledgeofculturaldifferences.Ourattitudesandthe PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 1/35 Whyaresomelanguagesconfusedforothers? waywechangeourbehaviourwhentryingtocommunicatepastthesedifferencesformpartof thebasisforlanguagechange,butisveryhardtostudy. Insomecases,identifyingnon-groupmembersbylinguisticmeansmaybeamatteroflife ordeath,asintheBiblicalepisodeinvolvingthepronunciationofHebrews(h)ibboleth‘earof wheat’(Judges12:6).Morerecentexamplesincludethe20th-centurycivilwarinSriLanka whereallophonicdifferenceswereusedtodifferentiateTamilandSinhalesespeakers(see[4]). Indeed,linguisticdiversitycanbedrivenbytheneedtodifferentiatebetweenculturalgroups [5](seealsoThurston’sconceptof‘esoterogeny’[6]andLarsen’s‘nabo-opposisjon’[7]).In extremecases,abiasedattitudetowardsalanguagecanseverelyaffectperception,eventothe extentofreducingcomprehension[8,9]. Inspiteoftheimportanceofperceivinglinguisticdifferencesindailylife,therehavebeen fewstudiesexploringthefactorswhichinfluencepeople’sabilitytorecognisethelanguagesof theworldandtheirsimilarities.Thisgamepresentsuswithanewkindofdatathathasprevi- ouslynotbeenavailable,andoffersthechancetoinvestigatetheperceptionoflanguagesimi- larity.Inthispaper,weanalysetheresultsofanonlinegameplayedmillionsoftimesbya globalsampleofpeople,inwhichthegoalwastoaccuratelyidentifylanguagesbasedona 20-secondaudioclipofspeech.Basedontheplayers’behaviour,wecanextractinformation aboutwhichlanguagestheycaneasilyidentify,andwheretheygetconfused.Weusetheresults ofthisgametoinvestigatewhatfactorsmakealanguageeasytoidentifyaccurately,including non-linguisticfactorssuchaseconomicpowerandthequalityoftheaudiorecordings.We alsoassesswhetherlanguageswhichareoftenconfusedalsohavesomeobjectivelinguistic similarity,forexamplebeingcloselyrelated,geographicallyclose,orhavingsimilarsoundsys- temsorlexicon. Morespecifically,theresearchquestionsweaddressinthispaperare: • Q1:Whichlanguagesareconfusedforwhichothers? • Q2:Arethereanyasymmetriesofconfusion? • Q3:Whatfactorscanpredictwhetherplayersconfusetwolanguagesforeachother?Candi- datesinclude: – geographicalcloseness; – genealogy; – similarityofphonemeinventories; – lexicalsimilarity. • Q4:Whatfactorscanpredictplayers’accuracy?Candidatesinclude: – acousticqualityofthespeechsamples; – proportionofnon-nativespeakers(L2speakers); – totalnativespeaker(L1)speakerpopulation; – linguisticdiversityofthemaincountryinwhichthelanguageisspoken; – numberofcountriesthelanguageisspokenin; – “languagenametransparency”—howclearthelinkisbetweenthelanguagenameandthe nameofthemaincountry; – economicpowerofmaincountryasmeasuredbyGrossDomesticProduct(GDP); PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 2/35 Whyaresomelanguagesconfusedforothers? – Familiaritywiththelanguage,asmeasuredbythefrequencyofoccurrenceofthelan- guagenameinwrittenmaterialssuchastheEnglishandChinesecorporafromGoogle Books. • Q5:Isplayers’performancemoreaccuratelypredictedbylinguisticornon-linguistic factors? • Q6:Aresomephonologicalcuesmoreimportantthanothersinpredictingplayers’ accuracy? Identifyinglanguagesandrelationshipsbetweenthem Dividingacontinuumofrelatedlinguisticvarietiesintodiscretelanguagesandfamiliesisa difficulttask.AsEvans&Levinsonnote,“wearetheonlyspecieswithacommunicationsys- temwhichisfundamentallyvariableatalllevels[oflinguisticorganization]”[10].Themost commonmethodsofseparatingvarietiesintolanguagesarebasedontestsofmutualintelligi- bility,sharedliterarytraditionand/orsharedlexicons[11,12].However,speaker/signercom- munitiesmayhavetheirownperceptionsandlabellingofgroupidentity(cf[13]).Theeditors oftheEthnologuerecognisethat“alanguageismoreoftencomprisedofcontinuaoffeatures thatextendacrosstime,geography,andsocialspace.Thereisgrowingattentionbeinggivento therolesorfunctionsthatlanguagevarietiesplaywithinthelinguisticecologyofaregionora speechcommunity”[14].Languageusers’ownperceptionofwhataresalientcontrastsand importantcategoriescanoftendiffersignificantlyfromwhatlinguistsfocuson.Often,catego- rizationsbylanguageusersarepoliticallymotivated.Forexample,NorwegianandSwedishare considereddifferentlanguagesbytheirrespectivelanguagecommunities,despitetherebeing largegroupsofspeakersfromthetwocommunitieswhoareabletounderstandeachotherand alargesharedlexicon(cf.[15]). Thereareatleastthreetypesofrelationshipbetweenlanguages:temporal,spatialandsocial. Forexample,thegradualdifferencesleadingfromMiddleEnglishtoModernEnglishforma temporalcontinuumofvariationwithinthesamecommunity.Linguisticborrowingdueto contactcanresultinaspatialcontinuum,asfoundinthearealpatternsofEastAsia(e.g.Can- tonese,Mandarin,JapaneseandKoreaninoursample)orMainlandSoutheastAsia(Khmer, Burmese,LaoandThai).Often,languagesarerelatedbothintime(i.e.genealogy)andinspace (i.e.geographicalproximity),asinthecaseoftheclosely-relatedGermanicdialectsofBelgium, theNetherlands,Germany,Switzerland,LuxembourgandAustria.Languagescanalsovaryon asocialscale,thisisforexamplereflectedindifferentlanguagevarietieswithinthesamelan- guagedependingonsocio-economicstatus(cf.thebasilect-acrolectcontinuumincreole languages). Despitevariationalongthetemporal,spatial,andsocialdimensions,communication betweenspeakersofquitedifferentvarietiesisbothpossibleandproductive,suggestingthat wemustpossessaneffectivecapacityforperceptuallyhandlingvariation.Inthispaper,we takeararechancetoinvestigatenon-linguists’perceptionoflinguisticvariationacrossthe world.Weexcludedguessesmadebyplayersfromcountrieswherethetargetlanguageisan officialordefactoofficiallanguage,becauseitislessinterestingthattheyareabletoaccurately identifythetarget.Thismeansthatwearemostlikelygettingdataonhowoutsiders(e.g.non- speakersofthetarget)perceivelanguagesimilarity. Predictions Thereisnoresearch,asfarasweareaware,whichexplorestheperceptionofawiderangeof languagesbyoutsiders(individualswhoarenotfamiliarwiththetargetlanguages)andhow PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 3/35 Whyaresomelanguagesconfusedforothers? thiscomparestolinguists’objectivemeasurementsofsimilarity.Thereishoweverrelevant workinperceptualdialectology(i.e.dealingwithverycloselyrelatedvarieties[16]),suchas workonlisteneridentificationofregionalvarietiesofEnglishfromtheUnitedStates[17,18]. Listenersareabletoidentifyregionalvarietieswellabovechance,andtheperceptionofsimi- laritydiffersbetweenlistenersfromdifferentdialectbackgroundsandbetweenlistenerswith varyingknowledgeofrelevantsociolinguisticcues. Listenersusedifferencesinspeechvarietiesascuestosocialidentity(e.g.[19]),leadingto theformationofstereotypesaboutcertainspeechsoundsorwaysofspeaking(e.g.[20]).These attitudesmay,inturn,affecttheperceptionofspeechsounds(e.g.[8,21]). AnotherstudyshowedthatAmericanchildrencouldidentifyAmericanEnglishandIndian Englishvarietiesinaforcedchoicetask,butwerelessgoodatdistinguishingAmericanEnglish andBritishEnglish[22].However,thechildrendidlinkpicturesoffamiliarculturalitems(e.g. housesandclothing)tomorefamiliarvarieties.Thissuggeststhatperceptualandcultural knowledgeofvariationemergeatayoungage.Indeed,youngadultswithhigh-functioning autism(whohavegoodperceptualprocessing,butpoorsocialprocessing)canidentifydialect varieties,butarelessgoodatlinkingthemwithsocialstereotypes[23],suggestingapossible dissociationbetweenthetwoskills. Anexperimentalstudyoftheabilityoflistenerstodistinguishbetweentwoforeignlan- guages(Germanvs.Russian)hasbeencarriedout[24].Participantswereexposedtorecord- ingsofthesamespeakersproducingdifferentlanguages.Theyperformedabovechance,and theirsuccesswaspredictedbytheirlevelofexposuretothevariety,withevenoccasionalexpo- sureviamediaaidingperformance.Thissuggeststhatlistenersaresensitivetodifferences betweenlanguagesaswellasbetweendialects. InthispaperweanalysetheresultsoftheGreatLanguageGame,apopularonlinegamecre- atedbyoneoftheauthorsofthispaper,LarsYencken.Playerslistentoasampleofspeechand mustguesswhichlanguageitisfromasetofalternatives.Weexplorewhichfactorspredict whichlanguagesareguessedaccuratelyandwhatfactorspredictwhichlanguagesareconfused foreachother.Wetestedfourfactorsforconfusion(similarphonemeinventories,geographic proximity,genealogy,andsharedlexicon)andninefactorsforaccuracy(acousticqualityof audioclip,proportionofL2speakers,totalL1speakerpopulation,linguisticdiversityofmain country,numberofcountriesthelanguageisspokenin,languagenametransparency,eco- nomicpowerofmaincountryandfrequencyofoccurrenceofthelanguagenameinwritten materials).Thereareotherpossiblefactorsonecouldconsider,butthesearetheonesthatwe predictedwouldbeinteresting,andwereabletomeasure.Anexampleofafactorthatwould beinterestingbuthardtomeasureismutualintelligibility:linguistscommonlydividespeech varietiesintolanguagesbasedonmutualintelligibility(andsharedlexicon),butnomeasure- mentofmutualintelligibilityofallthelanguagesinoursampleexists. Themostobviouspredictionwewouldmakeisthatplayerswilldifferentiatelanguages basedonphonologicalproperties.Inotherwords,languageswhichsoundmoredifferentmay beeasiertodistinguishfromeachother.Particularphonologicalfeaturesmaybemoresalient thanothers(toneorretroflexconsonants,forexample).Lexicalitemsarealsoapotentialcue towhichspeakersmightpayattention:wepredictthatthemorelexicalitemsareshared betweentwolanguages,themoreoftenthelanguageswillbeconfusedforeachother.Since phonologicalfeaturesandlexicalitemsdiffusethroughborrowingandhistoricaldescent,play- ersmayalsofinditeasiertodistinguishbetweenlanguagesthatarefurtherapartinspace(spo- kenindifferentcountries)ortime(distantlyrelated,ornotatallrelated). Anotherobviousfactorthattheplayermightrelyonissimplyknowledgeofthetargetlan- guage.Thiseffectcanbelimitedtosomeextentbyexcludingresponsesfromcountrieswhere thelanguageisofficialordefacto-official(aswedobelow);butitcannotbeentirelyeliminated. PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 4/35 Whyaresomelanguagesconfusedforothers? Theprobabilityofexposuretothetargetlanguagemayberelatedtothenumberofpeoplewho speakit,weconsiderbothL1andL2speakerpopulations. Similarly,theplayermayrelyontheirculturalknowledge:aplayermightnotknowalan- guageverywell,butmighthavehearditbeingspoken(e.g.infilmsorothermedia)orknowa fewfactsaboutit.Aplayer’sexposureorculturalknowledgeofalanguagemaydependonthe ‘globalreach’ofthelanguage,whichcanbereflectedinthepopulationsize,thelevelofindus- trialisation(asmeasuredby,forexample,grossdomesticproduct,mentionsinwrittenmate- rial,howmanycountriesitisspokeninandproportionofL2-speakers).Forexample, MandarinhasmorespeakersthanSpanish,butSpanishisanofficiallanguageinmanymore countries,whichmayincreaseitsglobalreachdespitehavingfewerspeakers. Wealsoconsideredthelinguisticdiversityofthetargetlanguage’smaincountry,asaproxy forculturaldiversity.Countrieswithaunifiedcultural‘brand’mightbemoreeasilyrecognisa- bleandsalientinpeople’sculturalknowledgethanacountrywithdiverseculturesandlan- guages(thisshouldbynomeansbetakenasanendorsementofmonolingualismbythe authors,webelievethatdiversityandmultilingualismaremostimportantandshouldbe encouraged). Languagesdiscussedmorewidelymightbebetterknown.Weestimatethisbylookingat thefrequencyofthenamesofthelanguagesintheGoogleBooksN-gramcorpusofEnglish andChinesetextsfrom1800–2000[25]. Anotherfactoristhequalityoftherecordings.Wepredictthatclearer,higher-quality recordingswillfacilitatelanguagerecognitionandthereforeleadtomorecorrectguesses. Finally,playersmightusevariousnon-linguisticheuristics.Forexample,participantsmight ruleoutcandidatelanguagesthattheyknow,andthereforededucethecorrectanswer,though thisishardtotestgivenourdata.Anotherheuristicistochooselanguagesbasedongeo- graphicorculturalproximity.Forexample,aplayermightthinkthatalanguagesoundslike Russian,butthenseethatRussianisnotpresentedasacandidateanswer.However,itmaybe thatSlovakisanalternative,theplayerreasonsthatRussiaandSlovakiaareculturallycon- nected,areinthesamepartoftheworld,orformpartofthesamelanguagefamily,andsothey maychooseSlovakdespitenotknowingwhatSlovaksoundslike.Thiswouldpredictthatcon- fusionbetweenlanguagesisrelatedtothegeographicdistancebetweenthem,butalsothat playersmightconfuselanguageswhicharegeographicallyclose,butquitedifferentintermsof theirgenealogy,forexampleLatvian(Baltic,Indo-European)andEstonian(Finnic,Uralic). Thisheuristiccanbeappliedtosomelanguagesmoreeasilythanothers.Forexample,the geographiclocationof“ScottishGaelic”istransparentfromthename,whilethisisnotthecase for“Shona”,whichisspokeninZimbabwe.Thename“Kannada”maybeentirelymisleading forWesternplayerswhomightlinkitwith“Canada”,whenitisactuallyaDravidianlanguage spokeninIndia.Therefore,weconsiderthetransparencybetweenthelanguagename,andthe languagemaincountry.Thisisreferredtoas‘languagenametransparency’andcanalsobe linkedtothepreviouslydiscussedideaofa‘cultural/nationalbrand’. Thus,inrelationtoQ3above,wepredictthattwolanguagesaremorelikelytobeconfused iftheyare: • geographicallyclosetoeachother; • closelyrelatedhistorically; • similarintheirphonemeinventories; • similarintheirlexicon. InrelationtoQ4,wepredictthatalanguageismorelikelytobeguessedcorrectlyifitis: PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 5/35 Whyaresomelanguagesconfusedforothers? • representedinthegamewithspeechsamplesofhighquality; • spokenasasecondlanguagebyalargeproportionofthetotalnumberofspeakers; • spokenbyalargepopulationasanativelanguage(L1); • mainlyspokeninacountrywheremostpeoplespeakthesamelanguage(lowlinguistic diversity); • spokeninmanycountries; • transparently-named; • spokenmainlyinacountrywithgreateconomicpower; • oftenmentionedinChineseGoogleBooks; • oftenmentionedinEnglishGoogleBooks. Finally,sinceperceptionmightbelinkedtoculturalknowledgeorshapedbylinguistic experience,wepredictthatpatternsofconfusionwilldifferbetweenplayersfromdifferent countries. Caveats Thedatainthecurrentstudyprovidesuswithauniqueopportunity,butalsocarrieswithit certainproblems.Ontheonehand,thereisalmosttoomuchdata:therearemorequestionsto beaskedthanwecanpossiblycoverhere.Wethusrestrictourselvestoanalysingdatapertain- ingtolanguagesandcountries,andleavemorefine-grainedanalyses,suchasattheregionalor statelevel,forfutureresearch.Ontheotherhand,thereisnotenoughdata.Thedataonplayer identityislimited,weonlyknowthelocationoftheirIP-addresses.Furthermore,theselection ofparticipantsisbiased(thegamerequiresacomputerandtheInternettoplay,andwasdevel- opedanddistributedincertainsocialcircles,andrequiressomeknowledgeofEnglish).Lastly, thesampleoflanguagesinthegameisnotrepresentativeoftheworld’slanguages.Wetryto controlfortheseimbalances,butcanonlydrawtentativeconclusions.Wewilldiscusslaterin thepaperhowthiskindofgamecouldbechangedtobetterservelinguisticresearch,andpres- entanalternative—LingQuest. TheworkofCharlotteGooskensandcolleaguesonmutualintelligibilityofEuropeanlan- guagesdemonstratesthatspeakers’linguisticknowledgeismorecomplicatedthanitmight seem.Forexample,speakers’attitudestowardsthenativespeakersofalanguagemightinflu- encetheirperceptionofthatlanguage[8].Thisissomethingthatwewillbeawareofaswe interpretourresults,butthatweunfortunatelycannottestforinarigorousmanner. Materialsandmethods Thegame “TheGreatLanguageGame”(henceforthGLG;seehttps://greatlanguagegame.com/)wascre- atedbyLarsYenckenin2013.Theoriginalintentwastoincreasepublicawarenessofthelin- guisticdiversityofurbanareasofAustralia,theUnitedKingdomandtheUnitedStates.Inthis sectionwepresentthegame,thelanguagesinthegameandthefactorsthatwehaveinvesti- gatedinrelationtoplayers’abilitytoaccuratelyidentifylanguagesandtheirconfusion. TheGLGworksasfollows:theplayerlistenstoa20-secondaudioclipofnaturalspeech fromoneof78languages.Theymustthenidentifythecorrectlanguagefromasetofalterna- tives.ThealternativesarerepresentedwiththeircommonEnglishnames.Theplayercan PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 6/35 Whyaresomelanguagesconfusedforothers? Fig1.AscreenshotfromtheGreatLanguageGame. https://doi.org/10.1371/journal.pone.0165934.g001 makeachoiceatanytime,thereisnotimelimitorrewardforbeingspeedy.Theycanalso playtheclipmorethanonce.Wedonotknowhowplayersbehavedhere.Aftertheymakea choice,theyaretoldwhethertheywerecorrectorincorrect,andareshownthecorrectanswer. Inthefirstround,theplayerhastwoalternatives,thecorrectanswerandonedistractor.After everythreecorrectanswers,thenumberofdistractorsincreasesby1,uptoamaximumof10 optionsintotal.Whentheplayerhasmadethreeincorrectchoices,thegameends.Theyare thengivensomeinformationaboutthethreelanguageswhichtheyguessedincorrectly,asan encouragementtolearnmoreaboutthem.TheinstructionallanguageofthegameisEnglish, butthereisverylittletextthatneedstobereadinordertoplaythegame.Fig1isascreenshot ofthegameasitappearsinawebbrowser. Inthispaper,wemaketheassumptionthattheselectionoflanguagesastargetsinthegame israndom(seeS1Appendix),althoughwetakeintoaccountthenumberoftimesthelanguage wasofferedasatargetlanguage.Givenanaudioclip,theincorrectalternatives(distractors) providedareselectedrandomlywithuniformlikelihood.Thedistractorsandthecorrectalter- nativearedisplayedasoptionsinalphabeticalorder.Inpractice,thealternativesaplayer receivesmayvastlyalterthedifficultyofaquestion.Forexample,theplayermightbegivena setoflanguagesthatbelongtoverydifferentlanguagefamilies(e.g.Polish,UrduandTongan), mostlikelymakingthetaskeasier.Alternatively,theymightbegivenasetofveryclosely relatedlanguages(e.g.Swedish,IcelandicandNorwegian):amuchhardertask.Theyarenot awardedanypointsforselectingadistractorthatisinsomewaymoresimilartothetarget thantheotherdistractorsare.Forexample:ifthetargetisSamoanandthealternativesare Samoan,TonganandNorwegian,youwillbeequallywrongpickingTonganasyouwould havebeenpickingNorwegian,despiteTonganbeingcloselyrelatedtoSamoan. Fromhereonwards,wedescribeinmoredetailthegame’sinventoryoflanguagesandits playerbase.OurdiscussionisbasedontheGLG’spubliclyavailable2014-03-02confusion dataset(seehttp://lars.yencken.org/datasets/languagegame/andsupportinginformationS8 Data),andassociatedaudiofilesfromthattime(asmartphonecloneofGLGwascreatedin 2015byIsaacDrachman,called“Linguini”whichusesmanyofthesameaudioclipsasGLG, butcurrentlynodatahasbeenreleasedforLinguini,sononehasbeenconsideredinthis paper). PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 7/35 Whyaresomelanguagesconfusedforothers? Fig2.GeographicallocationsofthelanguagesincludedintheGreatLanguageGame.ThemapwascreatedinRusingthepackage maps[26].CoordinatescomefromGlottolog[12]. https://doi.org/10.1371/journal.pone.0165934.g002 Thelanguagesinthegame Atthetimeofwriting,theGLGfeatures78languages.Forfulldetailsonthelanguagesinthe game,theirgenealogyandtherateatwhichtheyareguessedaccuratelyinthegame,seesup- portinginformationS1Data.Fig2showsthegeographicallocationofthelanguages,withloca- tionstakenfromGlottolog[12]. Thesampleoflanguagesinthegameisnotbalancedaccordingtogeographyorlinguistic relatedness.Thisisevidentfromthemap:EuropeisoverrepresentedandtheAmericas completelyabsent.39ofthe78languagesarefromtheIndo-Europeanlanguagefamily.Thisis becausethesamplewasoriginallydesignedtoreflectthelinguisticdiversityofurbanareasof Australia,theUnitedKingdomandtheUnitedStates,forthepurposeofspreadingawareness aboutlinguisticdiversityinthesenationstothegeneralpopulation. Thesampleoflanguagesisalsodependentonaccesstospeechsamples.Thespeechsamples usedinthegamearemostlydrawnfromlargeradiobroadcasterswhichbroadcastshowsand newsinnon-majoritylanguagesoftheirrespectivecountries.Themajorityofthespeechsam- plescamefromAustralia’sSpecialBroadcastingService(SBS),VoiceOfAmerica(VOA),and BritishBroadcastingCorporation(BBC).SBS,VOAandBBCbroadcastradioandTVinlan- guagesspokenbysignificantpopulationsinAustralia,theUnitedStatesandtheUnitedKing- domrespectively.OthersmallersourcesofspeechsamplesforthegameareTheHindu, Phonemica(China),DeutscheWelle,TeluguOne,NHKJapan,ORFEins(Austria),SR(Swe- den),CBC(Canada),RFI(France),andthePacificandRegionalArchiveforDigitalSourcesin EndangeredCultures(PARADISEC).PARADISECisaarchivefordigitalconservationof endangeredlanguagesandcultures,withafocusonthePacificregion.Atthistime,theGLG doesnotrepresentthefullsetoflanguagesavailablefromthesebroadcasters(Englishand KreyòlareforexamplenotpresentdespitebeingcoveredbyVOA),norallthelanguagesof PARADISEC. This,incombinationwiththefactthattheaimofthegameistospreadawarenessabout linguisticdiversityincertainspecificareasinWesterncountries,resultsinunder- PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 8/35 Whyaresomelanguagesconfusedforothers? representationofindigenouslanguagesofotherareas,suchastheAmericas(0languages, “Americas”herestandsforbothSouthAmericaandNorthAmerica.Asasidenoteitisinter- estingtoobservethatdespiteVOAbeinglocatedintheUnitedStatestheydonotbroadcast newsinanyindigenouslanguageoftheAmericas,withtheexceptionofKreyòl.Kreyòlisa French-lexifiedcreole)andAfrica(8languages)andover-representationofothers,suchas Europe(33)andAsia(31). Audioclipscomefromrandomlychosen20secondsegmentsofbroadcasts,identifiedas beingfreefromexcessivenoise,musicorcode-switching(mixingofmultiplelanguages).Most clipsareofaninformalnature,andmayincludevariablenumbersofspeakers,laughter,back- channellingandoverlappingdialogue.Duetotheirdifferentsources,clipsalsovaryinaudio quality.Inthecaseoftelephoneinterviews,audioqualityalsovariesbetweenspeakerswithin thesameclip.Elicitedaudiorecordingscreatedinacontrolledlabenvironmentmighthave lessvariationintermsofacousticquality,typesofvoices,speedetc.Thismakethemmore standardisedandeasytocomparetoeachother,moresuitableforascientificcontrolledexper- iment.However,theclipsusedinthisgamehavetheadvantageofbeingmoresimilartoactual naturaldiscourse,hencemorefaithfultotheactualexperiencesofpeopleperceivinglanguages outintheworld. Forthepurposesofcomparingtheresultsofthedatafromthegametootherdatabaseswe havelinkedeachlanguageinthegametoitsappropriatecodeintheISO639-3setoflanguage codes.TheISO639-3isaninternationalstandardoflanguagecodesusedbymanylinguists andappliedinmanycross-linguisticdatabases.Inafewcasestherearenotclearone-to-one mappings;insucheventsthesocalledmacrolanguage-codesoftheISO639-2havebeenused instead(thisconcernsAlbanianwhichhasbeenmarkedas[sqi],Arabic[ara],Dinka[din], Kurdish[kur],Latvian[lav]andYiddish[yid]).Theissueoflanguagenames,languageclassifi- cationandISO-setsisverycomplicated;formoreonthistopicsee[27]and[28].Inourset someofthesecomplicationsaremadeparticularlyevidentbythelanguages“Bosnian”[bos], “Serbian”[srp]and“Croatian”[hrv]whichallareseenasdifferentlanguagesbytheEthnolo- gueandGlottolog,butareperceivedasonelanguagebysomeotherlinguistsandnationalcen- suses.Weacknowledgetheseproblems,solvingthemishoweverbeyondthescopeofthis study. Theissueoflanguagenamesandcodesalsobecomesaconcernwhenwecompareourdata withotherdatabases.Wecomparedtheconfusiondatafromthegamewiththesimilarities anddifferencesbetweenthelanguages’phonemeinventories.Dataonthephonemeinvento- riesofthelanguagesweretakenfromthePhoneticsInformationBaseandLexicon-database (PHOIBLE)[29],whichisitselfacompilationofinformationfromotherdatabases.PHOIBLE alsoemploysISO639-3,butunfortunatelynotalllanguagesoftheGLGsamplehadadirect correspondenceinPHOIBLE.Inthecaseofthemacro-languagesof639-2weusedthesub- varietyof639-3thatwaspresentinPHOIBLE.Inaveryrestrictedsetofcasesweusedaneigh- bouringlanguagevarietywhentherewasnotaclearmapping;thisconcernsonlyMalay([zlm] inGLG,whichwaslinkedto[zsm]inPHOIBLE),Saami[sme−>sma]andKhmer[khm−> kxm].Inotherwords,forthesethreelanguageswedidnotcomparetheexactsamevarieties acrossourGLGsampleandthesampleofPHOIBLE.Wedidhoweverhaveaccesstoavery closelyrelatedvarietyandwebelievethatthiswillhavelittle,ifany,effectontheresults. Phonemeinventoriesfor60ofthelanguagesincludedintheGLGwerefoundinthePHO- IBLEdatabase.PHOIBLEincludesdatafromtheSPA(StanfordPhonologyArchive),UPSID (UCLAPhonologicalSegmentInventoryDatabase)andothercollections.Formanyofour languages,therewasmorethanoneanalysisoftheinventoryavailableinPHOIBLE.Thedif- ferentdatabasesthatmakeupPHOIBLEandprovidedifferentanalysesofthesamelanguage differintheirdesign,renderingdirectcomparisonsovertheminappropriate.Languageswere PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 9/35 Whyaresomelanguagesconfusedforothers? onlycomparedifthereexistedinventoriesforeachofthelanguagesfromcamefromthesame source(UPSIDwasonlycomparedwithUPSID,SPAonlywithSPA).Intotaltherewere31 languagesforwhichwecouldmeasurethedistancebetweeneachpairinthismanner.For eachlanguagepair,thedistancebetweenthemwascalculatedas1minustheproportionof phonemesincommon.Tonesegmentswereexcludedinallcomparisonssincetheywerenot representedinUPSID. Theplayers TheonlyinformationabouttheplayerswhichthegameelicitsistheirIP-address,whichwe usetoinferwhatcountrytheyareplayingfrom(theIPaddressesweremadeanonymousin thepubliclyavailabledata).Wemaketheassumptionthattheplayeridentifieswiththiscoun- tryinsomemanner(thoughsomemaybevisiting,travelling,tunnellingtheirtrafficthrough othercountriesviaVPNorsimilartechnology,etc.).Thedataincludesresponsesfromanesti- mated767,000uniqueIPaddresses(GoogleAnalyticsestimates964,000uniqueuseridentities, meaningthatprobablymanydifferentpeopleplayonthesamecomputer).Playerscanplay multipletimes,butforsimplicity,weassumethatmostplayersarecasualplayerswhodonot playoftenenoughtolearnfromplayingonly. Furthermore,weassumethatmostplayershaveaknowledgeofEnglish,atleastsufficiently tofindthegameinthefirstplaceandunderstandtheinstructions.Theplayersarealsomost likelycomputer-literateandshareaninterestinlanguages,orareinterestedingamesofthis kindwhereonemeasuresknowledgeoftheworldand/orlinguisticsandcompetesagainst friends.Thegamehasbeenextensivelysharedoncertainblogs,forumsandsocialmedia, whereweexpecttheretobealargenumberoflanguageenthusiasts/nerds.Theplayersofthe GLGcomefromallovertheworld,butcertainareasareoverrepresented.Table1displays numberofguesses(trials)fromeachcontinent. Measuringconfusionbetweenlanguages Therawdataproducedbythegameisalistofplayerresponses,alongwithinformationonthe optionsavailabletotheplayerineachtrial,aswellasthecountryoforiginoftheIP-address. Weusethisdatatocalculatetheprobabilityofonelanguagebeingconfusedforanother,inthe followingway:givenatargetlanguageinthespeechsample,andanumberofalternativelan- guagenamestochoosefrom,aplayerselectsonelanguageastheirguess.Fromalargesample ofgames,weknowtheconditionalprobabilityofguessingaparticularlanguageG,giventhat theplayeractuallyheardaparticulartargetlanguageT—thatis,thechanceofaplayerconfus- ingTforG.ThisistheproportionoftimesaplayeractuallyguessedGwhenhearingT Table1.LocationsofguessesbycontinentbasedonIP-address.Notethatwhiletherearenoindigenous peopleofAntarctica,thereareinfactresidentsthere,varyingfrom1,100to4,400duringtheyear[30]. ContinentofIP-address Numberofguesses Europe 7,963,630 NorthAmerica 5,980,767 Asia 841,609 Pacific 364,390 SouthAmerica 356,390 Africa 74,032 Antarctica 11 Total 15,580,829 https://doi.org/10.1371/journal.pone.0165934.t001 PLOSONE|https://doi.org/10.1371/journal.pone.0165934 April5,2017 10/35
Description: