英文通過 AI 驅(qū)動的文本挖掘識別綠色職位的擴(kuò)展清單

PublicDisclosureAuthorizedPublicDisclosureAuthorized

PolicyResearchWorkingPaper10908

IdentificationofanExpandedInventory

ofGreenJobTitlesthroughAI-DrivenTextMining

Micha?PalińskiGüne?A??k

TomaszGajderowicz

MaciejJakubowski

Ef?anNas?zen

DhushyanthRaju

WORLDBANKGROUP

SocialProtectionandJobsGlobalPracticeSeptember2024

PolicyResearchWorkingPaper10908

Abstract

Thisstudyexpandstheinventoryofgreenjobtitlesbyincorporatingaglobalperspectiveandusingcontemporarysources.Itleveragesnaturallanguageprocessing,specificallyaretrieval-augmentedgenerationmodel,toidentifygreenjobtitles.Theprocessbeganwithasearchofacademicliter-aturepublishedafter2008usingtheofficialAPIsofScopusandWebofScience.Thesearchyielded1,067articles,fromwhich695uniquepotentialgreenjobtitleswereidenti-fied.Theretrieval-augmentedgenerationmodelusedtheadvancedtextanalysiscapabilitiesofGenerativePre-trained

Transformer4,providingareproduciblemethodtocatego-rizejobswithinvariousgreeneconomysectors.Theresearchclusteredthesejobtitlesinto25distinctsectors.Thiscatego-rizationalignscloselywithestablishedframeworks,suchastheU.S.DepartmentofLabor’sOccupationalInformationNetwork,andsuggestspotentialnewcategorieslikegreenhumanresources.Thefindingsdemonstratetheefficacyofadvancednaturallanguageprocessingmodelsinidentifyingemerginggreenjobroles,contributingsignificantlytotheongoingdiscourseonthegreeneconomytransition.

ThispaperisaproductoftheSocialProtectionandJobsGlobalPractice.ItispartofalargereffortbytheWorldBanktoprovideopenaccesstoitsresearchandmakeacontributiontodevelopmentpolicydiscussionsaroundtheworld.PolicyResearchWorkingPapersarealsopostedontheWebat

/prwp.Theauthorsmaybe

contactedatsnasozen@anddraju2@.

ThePolicyResearchWorkingPaperSeriesdisseminatesthefindingsofworkinprogresstoencouragetheexchangeofideasaboutdevelopmentissues.Anobjectiveoftheseriesistogetthefindingsoutquickly,evenifthepresentationsarelessthanfullypolished.Thepaperscarrythenamesoftheauthorsandshouldbecitedaccordingly.Thefindings,interpretations,andconclusionsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelopment/WorldBankanditsaffiliatedorganizations,orthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.

ProducedbytheResearchSupportTeam

IdentificationofanExpandedInventoryofGreenJobTitlesthroughAI-DrivenTextMining

Micha?Paliński

Güne?A??k

TomaszGajderowicz

MaciejJakubowski

Ef?anNas?zen

DhushyanthRaju

Keywords:AI,textmining,occupationalclassification,greenjobs,greeneconomyJELcodes:J23,Q52,O14

Paliński:UniversityofWarsaw,Warsaw,m.palinski@.pl.

A??k:TOBBUniversityofEconomicsandTechnology,Ankara,gunesasik@.

Gajderowicz:UniversityofWarsaw,Warsaw,tgajderowicz@.pl.

Jakubowski:UniversityofWarsaw,Warsaw,mjakubowski@.pl.

Nas?zen:WorldBank,Ankara,snasozen@.

Raju:WorldBank,Washington,DC,draju2@.

WethankBerfu?opurforresearchassistancewiththeliteraturesearch.WearealsogratefultoBurakBask?n,PaoloBelli,AhmetKurnaz,RenéLeónSolano,andAivinVicquierraSolatorioforusefulcomments.

2

1.Introduction

Theescalatingimpactsofclimatechangeunderscoretheurgencyofagreentransition—apivotalshifttowardsustainablepracticesthatisessentialforourplanet’sfuture.Thistransitionisexpectedtoacceleraterapidly,necessitatingthatpolicymakersanalyzeitsimpactsonnationallabormarketsanddevelopeffectivestrategiestonavigatetheevolvinglandscape.Understandingthescopeandnatureofgreenjobsiscrucialforinformingpublicpolicy,enablinggovernmentsandorganizationstodeveloptailoredandtargetedstrategiesforeducation,training,andemploymenttosupportasustainableeconomy.

Worldwide,themostwidelyusedsourceofgreenjobtitlesistheGreenOccupationslist,constructedbytheU.S.DepartmentofLabor’sOccupationalInformationNetwork(O*NET)in2009(Dierdorffetal.2009).O*NET’soriginalapproachinvolvedreviewingpublicationscoveringawidearrayofworkplacetopicspertinenttothegreeneconomy.Inassessinggreenjobs,researchpredominantlyemploystwomethods:top-downapproaches,whichcategorizeentiresectorsorindustriesasgreen,andbottom-upapproaches,whichfocusonspecificoccupations,defininggreenjobsbasedonthegreennatureofthetasksorskillsassociatedwiththoseroles(Valeroetal.2021).O*NET’sclassificationisthemostoftenusedsourceforoccupationretrievalinthebottom-upapproachtogreenjobsanalysis(OECD2023).

ThegreenjobtaxonomydevelopedbyO*NEThasbeeninstrumentalinshapingquantitativeresearchonthegreeneconomy.IntheUnitedStates,itsimpactisreflectedinstudiesbyresearcherssuchasConsolietal.(2016),Poppetal.(2020),Vona,Marin,andConsoli(2019),andVonaetal.(2018).Thetaxonomyhasalsobeenadaptedforvariousregions,includingtheEuropeanUnion(BowenandHancké2019),theNetherlands(Elliottetal.2021),theUnitedKingdom(Valeroetal.2021),OrganisationforEconomicCooperationandDevelopment(OECD)membercountries(OECD2023),VietNam(Doanetal.2023),andArgentina(delaVega,Porto,andCerimelo2024).Afterclassifyingoccupationsasgreen,studiesdelveintothespecificskillsandtasksrequiredforgreenjobs,analyzetrendsingreenjobcreationanddistribution,andassessthebroadereconomicimpacts,suchasproductivity,innovation,andgrowth,associatedwiththegreentransition.

However,twomainissuesmakeO*NETlessrelevantforworldwideuse,particularlyforgreenjobs.First,O*NETwasbuiltin2009,withthelastmajorrevisionofthetaxonomycompletedin2011(Dierdorffetal.2011)andtheassociatedreferencebooklastupdatedin2013(O*NET2013).Theliteratureongreenjobshasexpandedsignificantlysince2009.Second,O*NETisdesignedfortheU.S.labormarket,identifyingtaskswithinoccupationsbasedontheU.S.context.Thetasksandskillsrequiredtoperformthesejobsdependontheproductiontechnology,whichcandiffersignificantlybetweentheUnitedStatesandothereconomies,suchaslow-andmiddle-incomecountries.

Ourstudyaimstoexpandtheinventoryofgreenjobtitlesbyintegratingaglobalperspectiveandincorporatingcontemporarysources.Ourliteraturereviewcomprisedasearchforarticlespublishedafter2008usingScopusandWebofScience—twoleadingbibliographicdatabasesthatarewidelyusedbytheacademiccommunityforaccessinganextensiveglobalcollectionofpeer-reviewedpublicationsacrossvariousdisciplines(ZhuandLiu2020).Theyear2008markeda

3

criticaljunctureinthedialogueongreenjobs,withthefirstexplicitdefinitionoftheconcept(Stanef-Puic?etal.2022).

TheconstructionofataxonomylikeO*NETtypicallyinvolvesqualitativecodingtoidentifyjobtitleswithinagreencontext,amethodthatislabor-intensiveandtime-consuming.However,thereisagrowingtrendtowardusingnaturallanguageprocessing(NLP),oftenaugmentedbyexpertreview,asapotenttoolforjobidentificationandcategorizationacrossvariouscontexts,includingthegreeneconomy(Chiarelloetal.2021;Decorteetal.2021;Li,Sunetal.2020;Papoutsoglouetal.2022).Asignificantexampleisthe2022initiativebytheEuropeanCommission,whichemployedtheBidirectionalEncoderRepresentationsfromTransformers(BERT)NLPalgorithmalongsidemanuallabelingtoidentifygreenconcepts(skillsandknowledge)withintheEuropeanSkills,Competences,Qualifications,andOccupationsclassification(EC2022).

AlignedwiththisNLP-drivenmethodologicalevolution,ourresearchutilizedanadvancedartificialintelligence(AI)pipeline,specificallytheretrieval-augmentedgeneration(RAG)model(Lewisetal.2020),toidentifygreenjobtitlesinacademicliterature.Thistechnologyenabledtheexaminationofasubstantiallylargersetofliteraturethanmanualmethodscouldaccommodate.RAGstandsoutasaneffectiveNLPapproach,mergingthebenefitsofretrieval-andgenerative-basedAImodels,therebyaddressingprevalentissuesinbasicgenerativeAI,suchashallucinationsandthelackofdomain-specificknowledge(Gaoetal.2023).Importantly,ourapproachisreproducible,allowingthelistofgreenjobstobeupdatedastheliteratureonthegreentransitionexpandsinthefuture.

OursearchoftheacademicliteraturepublishedbetweenJanuary2009andApril2024,whenweconductedthesearch,ultimatelyyielded1,067articlesforanalysis.Wefoundthattheacademicliteratureonthegreentransitionhassignificantlyexpandedoverthepast15years,bothinthenumberofarticlesandinthediversityofrepresentedcountriesandregions.In2009,therewereonly44articlesonthegreentransition.By2023,thisnumberhadincreasedto162.In2009,articlesalmostexclusivelycoveredtheUnitedStates,Canada,China,andcountriesintheEuropeanUnion.By2023,thecoveragehadexpandedtoincludeEurope,theCaucasus,SoutheastAsia,andAfrica.

Weidentified695uniquegreenjobtitlesfrom105articles(10percentofthe1,067articles).ComparingourlistofgreenjobswiththoseidentifiedinO*NET,wefoundthat17percentofthejobtitlesmatchedperfectlyoralmostperfectlywithO*NET,whilewealsoidentifiedpotentiallynewtitlesthroughlessprecisematches.

OurstudydemonstratesthatAI-basedmodelscanaddresscapacitychallengesinidentifyingqualitativeinformationfromalargeandexpandingbodyofliterature,despitesomelimitations.FutureresearchandpracticeshouldfocusonrefiningtheseAI-drivenmethodsandintegratingadditionalinformationsourcestocontinuouslyupdateandexpandtheinventoryofgreenjobtitlesastheliteratureonthegreentransitionevolves.

4

2.Approach

IdentificationofRelevantLiterature

InApril2024,weconductedasearchoftheliteraturepublishedsinceJanuary2009,usingtheofficialapplicationprogramminginterfacesofScopusandWebofScience.Oursearchstrategyinvolvedkeywordcombinationspreviouslyvalidatedinthreesystematicliteraturereviewsrelatedtogreenjobs(ApostelandBarslund2024;KozarandSulich2023;Stanef-Puic?etal.2022).Thekeywordcombinationsincluded“greenjob(s),”“greenoccupation(s),”“greenemployment,”“sustainablejob(s),”“sustainableoccupation(s),”“greentransitionjob(s),”and“green-collarjob(s).”(Box1liststhesearchqueries.)Thesekeywordsweresearchedwithintitles,abstracts,authorkeywords,and“topics”asreferencedinthedatabases.Toensurecredibleresults,werestrictedourfocustopeer-reviewedpublications,specificallyarticlesandreviews(hereafterreferredtoas“articles”).

OursearchapproachdifferssignificantlyfromthatofO*NET.WhileO*NETmeticulouslyindexedandcategorizedsourceswithinitsreferencebook,thespecificsofitsselectionprocessaresparinglydescribed.Thereisnodetailedinformationonspecifickeywordsormethodsusedingatheringthearticles.Itinvolvedcollectingandreviewingmorethan60publications,includingacademicjournals,commissionedreports,industrywhitepapers,andgovernmenttechnicalreports.Additionally,O*NETconductedasubstantialreviewofvariousinternetsourcesrelatedtothegreensector’sworkforce(O*NET2013).

IdentificationofGreenJobTitlesintheLiterature

WeusedtheRAGmodeltosimulatetheworktraditionallyperformedbyresearchassistantswhowouldmanuallytaggreenjobtitleswithinarticles.Thismanualtaggingprocess,extendingover1,000pagesacrossthearticlesintheanalysisset,isresource-intensiveandsusceptibletoerrorsfromhumanoversight,cognitivebiases,andheuristicshortcuts.Incontrast,theRAGmodeloffersarobustandconsistentapproach.

AsignificantadvantageofusingtheRAGmodelisthereproducibilityoftheresults.ByutilizingseedparametersavailableintheOpenAImodels,specificallytheGenerativePre-trainedTransformer4(GPT-4)-0125-previewmodel,weensuredthatourresultswerereplicable,providingadegreeofconsistencythatmanualtaggingstrugglestoachieve.Althoughthemodelscannotbeentirelydeterministicduetotheirinherentstochasticnature,theuseofseedparametershelpstoensurethattheresultsarehighlyconsistentacrossmultipleruns(Anadkat2023).Furthermore,theadvancednaturallanguageunderstandingcapabilitiesoftheGPT-4modelenabledanuancedanalysisofthecontextinwhichjobtitlesarediscussedinthearticles.Thisisparticularlyvitalinouranalysisset,wheregreenandnongreenjobsareoftenmentionedinthesamearticles.Themodel’sabilitytodiscernthecontextandclassifyjobtitlesaccordinglyisasubstantialimprovementoverolderNLPapproaches,suchaslesscapableembeddingmodelslikeBERTorfullysupervisedmethodslikenamedentityrecognition(NER),whichmightnotcapturesuchsubtletiesornuances.

5

EmployingtheRAGmodel,weusedembeddingmodelstoidentifyrelevantsectionswithinarticles(chunks)thatdiscussedspecificjobtitles.WeusedOpenAI’smostcapabletext-embedding-3-largemodelwith3,072dimensionsintheembeddingprocess.Whilechunkingisoftenemployedtocircumventthecontextwindowlimitationsofcertainmodels,ourapplicationofGPT-4,whichboastsanexpansivecontextwindowof128,000tokens(comparableto96,000words),wasnothinderedbysuchconstraints.Instead,thedecisiontochunktextinouranalysiswasdictatedbythefactthatchunkingsignificantlyimprovestherelevanceofretrievedcontentasitdecreasesnoiseintheembeddedtext(Yepesetal.2024).Next,weemployedtheGPT-4modeltoscrutinizesegmentsofarticleswherejobtitleswerementioned,aimingtoinferfromthecontextwhethertheauthorsclassifiedtheserolesasexamplesofgreenjobs.Awareoftheseveralcompetingdefinitionsof“greenjobs”intheacademicliterature(Stanef-Puic?etal.2022),werefrainedfromadheringtoanysingulardefinition.Instead,wedirectedthemodeltodetermineiftheauthorsconsideredthatthesejobsweregreen,suchaswhethertheywerediscussedwithintherealmsofthegreeneconomy,sustainability,orclimatechangemitigation.Wepurposefullydidnotexposethemodeltoanypreestablishedclassificationsofgreenjobstopreventprimingeffectsandpromoteanunbiasedevaluationbasedoncontext.WepresentamoredetaileddescriptionoftheRAGmodel’spipelinestagesintheappendix.

ThegenerativecapabilitiesoftheAIwerespecificallyharnessedinthefinalstageoftheRAGmodelimplementationprocess(figure1).WhiletheAIpossessesextensiveknowledgefromitstraining,ourmodelstrategicallyrefrainsfromusingthisknowledge.Themodel’sgenerativefunctionsarenotemployedtointroduceorinferinformationfromitstrainingbutrathertointerpretandanalyzethetextthatispresentedtoit.Whenthemodelidentifiespotentialsectionsofthetextthatmightdiscussgreenjobs,itleveragesitsnaturallanguageunderstandingcapabilitiestoanalyzethegiventext.Thegoalistoascertainwhethertheauthorsofthearticlesareindeedmentioningspecificjobtitlesandifthesetitlesarediscussedwithinthegreencontext.

ThemodelweusedhascommonalitieswithNER,aprocessinNLPthatinvolvesidentifyingandcategorizingkeyinformation(entities)intext(Li,Shietal.2020).Entitiescouldbenamesofpeople,companies,locations,andsoforth.Ourworkparallelsthisapproachbyidentifyinggreenjobtitleswithintext.IllustratingtheprogressioninNLP,researchhasshownthateventheolderGPT-3modelcouldmatchtheperformanceoffullysupervisedNERbaselines(Wangetal.2023).Zhouetal.(2023)demonstratethattheLargeLanguageModelMetaAI,alargelanguagemodel(LLM),significantlyoutperformssupervisedNERmodels,asevidencedbyasubstantialmarginintheF1score,ameasureofatest’saccuracy.Thiscomparisonspanned43datasetsencompassingninevarieddomains.Similarly,Monajatipooretal.(2024)demonstratethatinthebiomedicalfield,GPT-4outperformstraditionalNERmodels.

EmpiricalstudiescomparingGPTmodelstoearliertextminingmethods,suchasBERT,remainlimited.Comparedtofine-tunedBERTmodels,GPT-3hasexhibitedsuperiorperformanceintextclassificationtasksinrelatedcontexts(LigaandRobaldo2023;PawarandMakwana2022).However,GPT-4,whenemployedinazero-shotsetting,significantlyoutperformedthebaseBERTmodelbutwasoutperformedbyfine-tunedBERTmodelsinspecifictaskssuchasproteinsequenceidentification(Rehanaetal.2023).Despitethesefindings,nostudieshavebeenidentifiedthatdirectlycomparetheperformanceofthesemodelsinacontextsimilartoours,whereaccurateclassificationishighlycontingentonthesurroundingcontext,suchasdistinguishing

6

betweengreenandnongreenjobs.WepositthatGPTmightoutperformBERTinthiscontextbecauseitsmorecomplexarchitecture,largernumberofparameters,andabilitytohandlelongercontextlengthslikelyenableittobetterdifferentiatenuanced,context-dependentinformation,suchasclassifyingjobsasgreen.

Akeyfeatureneededinsuchexercisesisvalidatingtheoutputofthemodel.Infieldslikebiology,medicine,law,programming,orfinance,standardizedbenchmarksexisttomeasuretheefficacyofLLMsasNERtools(Zhouetal.2023).However,forourpurposes,suchbenchmarksareunavailable.Toensurethevalidityofourresults,weundertooktwotypesofchecks.First,wespecificallyfocusedonarticlesforwhichthemodeldidnotidentifyanygreenjobtitlestocheckforfalsenegatives.Thissituationisrelativelycommonsinceauthorsmightdiscussgreensectorsoftheeconomywithoutexplicitlymentioningjobtitles.Tothisend,werandomlyselectedasampleofsucharticlestoreviewmanually,ensuringthattheabsenceofidentifiedgreenjobtitleswasconsistentwiththecontentofthearticles.Second,weconductedareviewofallthearticlesinwhichthemodelidentifiedjobtitles.Thisstepwastocheckforfalsepositives—thatis,erroneouslyclassifyingnongreenjobtitlesasgreen—andtodetectanyinstancesofhallucinationswherethemodelmightgeneratenonexistentjobtitles.

Togainabetterunderstandingofthecontextinwhichgreenjobsareanalyzed,wemappedthearticlesmentioninggreenjobtitlestoeconomicactivities,basedontheInternationalStandardIndustrialClassificationofAllEconomicActivities(ISIC)classificationscheme.ISICisaUnitedNationssystemforclassifyingeconomicdataaccordingtoindustry.Forthismapping,weprovidedChatGPT-3.5withtheISICclassification,includingdescriptionsofallactivities,andpromptedittofindthebesttop-levelmatchesforallarticles.

Wealsoidentifiedthegeographicalcoverageofalltheretrievedarticles.Forthis,wepromptedChatGPT-3.5toretrievecountriesmentionedasthebasisforanalysisintheabstractsandtitlesofallthearticles.Ifnocountrieswerementioned,weassumedthatthearticlehadaglobalperspective.Next,weusedPython’spycountrypackage(Theune2024)forfuzzymatchingofthecountrynameswithofficialISOcountrycodesandidentifiedthecontinentsofthecountries.Thisapproachallowedustoillustratethegloballandscapeofgreenjobsresearch.

MatchingofIdentifiedGreenJobTitleswithO*NET

WeemployedembeddingmodelingtorepresentboththejobtitlesweidentifiedandthosefromO*NETas3,072-dimensionalvectors,enablingasystematiccomparison.Forthistask,weusedthetext-embedding-3-largemodel.Byutilizingcosinesimilarity,arecommendeddistancemeasureforthismodel,weidentifiedtheclosestmatchesbetweenouridentifiedjobtitlesandthoseinO*NET.CaseswherethejobtitlesshowedonlyminimalsimilarityindicatedpotentialnewgreenjobtitlesthatwerenotyetrecognizedinO*NET.Thematchingprocessforthissteppresentedasignificantchallengeforthemodelbecauseitoperatedwithminimalcontextthatincludedonlythejobtitlesthemselves.HadwebeenabletoutilizedetailedtasksandskillsrelevanttothesejobsalongsidethejobdescriptionsfromO*NET,wecouldhaveachievedamoreinformedandaccuratematchingprocess.However,thenatureofthearticlestypicallydoesnotlenditselftoasystematicdiscussionofjobroles,includingspecifictasksandskills.

7

Wealsomappedthegreenjobtitlesintomajorgreeneconomysectorsthroughclusteringbasedontheirsemanticsimilarity.WeusedjobtitleembeddingsandappliedUniformManifoldApproximationandProjectionforDimensionReduction(UMAP)(McInnes,Healy,andMelville2018),followedbyHierarchicalDensity-BasedSpatialClustering(HDBSCAN)(McInnes,Healy,andAstels2017).Fine-tuningthesetechniqueswasessentialtoachievemeaningfulresults.

WeconfiguredUMAPwith10neighborstobalancelocalandglobalstructure,andaminimumdistanceof0.1tocontrolthedensityofpointpacking,ensuringthatlocaldetailwaspreserved.Thisconfigurationmaintainssimilarityamongnearbypoints(localstructure)whilegroupingclustersofsimilarpointstogether(globalstructure).

Forclustering,weusedHDBSCANwithaminimumclustersizeof10toensuresignificanceandaminimumsamplesizeoffourtodefinethenumberofpointsrequiredtoformadenseregion.

3.Results

GreenLiterature

Oursearchyieldedatotalof1,367articles,withScopuscontributing991articlesandWebofSciencecontributing376articles(table1).WeusedtheDigitalObjectIdentifierandInternationalStandardSerialNumbertocross-checktheuniquenessofthearticlesacrossthetwodatabases.Wefoundthat88percentofthearticlesindexedinWebofSciencewerealsoindexedinScopus.Consequently,weintegratedtheuniquearticlesfromWebofScience—thosenotfoundinScopus—toarriveatourunique“analysisset”of1,067articles.Inourensuinganalysis,weusedthefulltextsof567articlesforwhichwewereabletoretrievePortableDocumentFormat(PDF)filesandtheabstractsfortheremaining500articles.Whilemostofthepublicationsarearticles,with915fromScopusand353fromWebofScience,weretrievedadiversesetofpublications,includingconferenceproceedings,bookchapters,andothermaterials(table2).

Boththenumberandgeographicalspreadofthearticlesongreenliteraturehaveexpandedsince2009.In2009,therewere44articles,andin2023,therewere162articles(figure2).Moreover,thearticlesin2009almostexclusivelycoveredNorthAmerica(table3).However,thescopegraduallydiversified.Notably,by2015,thenumberofarticlescoveringEuropeancountrieshadsurpassedthosecoveringNorthAmericancountriesincumulativeterms;by2022,thesamehadoccurredforAsiancountries.

Intheearlyyearsofouranalysisperiod,therewerenoarticlesfocusingspecificallyongreenjobsinSouthAmericaandhardlyanyinAfricaorOceania.Thissituationhaschangeddramatically,withsubstantialincreasesinthenumberofarticlescoveringthesecontinentsovertime.Thistrendindicatesthatresearchongreenjobsisbecomingincreasinglyglobalized,encompassingabroaderrangeofgeographicalsettings.

GreenJobTitles

Weinitiallyidentified799potentialgreenjobtitlesusingtheRAGpipeline(seefileinGitHubforthelistoftitles).Weexcluded66titles,astheywerenotjobtitlesbutreferredtogreenactivities.

8

Forexample,inthestudybyAfolabietal.(2018),whilethemodelcorrectlyidentified“environmentalcompliancespecialist”asagreenjobtitle,italsoincorrectlytaggedactivitieslike“solarpanelmanufacturing”or“reductionofwaterusageon-site”asjobtitles.Thefollowingisthedirectcitationfromwhichthemodelinferredthesetitles:“Thedatarevealedareasthatarepeculiartotheprovisionofgreenjobsintheconstructionsectorsuchassolarpanelmanufacturing(...)”(Afolabietal.2018,2).Itseemsthatthemodelerroneouslyassumedthatthephrase“greenjobsintheconstructionsectorsuchas(...)”wasintroducingalistofjobtitles.ThismisinterpretationillustratesthetypesofheuristicsthemodelmightemployandemphasizestheimportanceofqualitycheckswhileworkingwithLLMsintheircurrentstateofdevelopment.Next,weexcluded21titlesbecausetheyweretoobroad,suchas“technicians”and“engineers.”Thisindicatesthatdespiteinstructingthemodelinourpromptstoreturnonlyspecificjobtitles,roles,oroccupations,iterroneouslyproducednonspecificresultsinalimitednumberofinstances.Finally,westandardizedalljobtitlestotheirsingularformsandremovedduplicates.Thisprocesseliminatedanadditional17jobtitles,resultinginatotalof695uniquegreenjobtitles.

Ouranalysisrevealedthatthe695jobtitlesappearinfrequentlyacrossthe1,067articles,withjust105articles(10percentofthetotalanalysisset)mentioningoneormoregreenjobtitles.Additionally,jobtitlesweremorefrequentlyidentifiedinfulltexts,withonly19jobtitlesfoundinabstracts.

Theglobalperspectiveissignificant,with28articlesreferencinggreenjobtitlesinternationally(table4).TheUnitedStatesfollowsclosely,withmentionsin20articleswithgreenjobtitles.TheEuropeanUnioniswell-representedwith13mentions.OthernotablecountriesincludeBrazil,China,andSpain,eachwith6mentions.Intotal,wefound40countriesmentionedinthearticleswithgreenjobtitles,although19countrieswerementionedonlyonce.

Overone-thirdofthejobtitlesareinengineering,anotherone-fiftharetechnician-leveljobs,andasignificantshareincludesjobsinbusinessandadministration,suchaspolicyspecialists.Severalothertitlesareintheareasofbuildingand