Disambiguation of Inventors, USPTO

Transcription

1 Disambiguation of Inventors, USPTO Guan-Cheng Li University of California, Berkeley College of Engineering University of California, Berkeley Fung Technical Report No September 17, Blum Hall 5580 Berkeley, CA (510)

2 The Coleman Fung Institute for Engineering Leadership, launched in January 2010, prepares engineers and scientists from students to seasoned professionals with the multidisciplinary skills to lead enterprises of all scales, in industry, government and the nonprofit sector. Headquartered in UC Berkeley s College of Engineering and built on the foundation laid by the College s Center for Entrepreneurship & Technology, the Fung Institute combines leadership coursework in technology innovation and management with intensive study in an area of industry specialization. This integrated knowledge cultivates leaders who can make insightful decisions with the confidence that comes from a synthesized understanding of technological, marketplace and operational implications. Lee Fleming, Faculty Director, Fung Institute Advisory Board Coleman Fung Founder and Chairman, OpenLink Financial Charles Giancarlo Managing Director, Silver Lake Partners Donald R. Proctor Senior Vice President, Office of the Chairman and CEO, Cisco In Sik Rhee General Partner, Rembrandt Venture Partners Fung Management Lee Fleming Faculty Director Ikhlaq Sidhu Chief Scientist and CET Faculty Director Robert Gleeson Executive Director Ken Singer Managing Director, CET Copyright 2013, by the author(s). All rights reserved. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission. 130 Blum Hall 5580 Berkeley, CA (510)

3 Abstract: Since the USPTO does not require unique identifiers in the application process, it can be hard to search for a specific inventor and his or her patents, especially if the inventor s name is common or has multiple forms. Ad-hoc disambiguation methods based on thresholds and string comparison matching are common but vulnerable to bias and idiosyncratic to a training set. To tackle the problem, I employ a large scale-clustering algorithm, adaptive K-means, which automates the disambiguation. The algorithm considers statistical correlations between each inventor name pair, and learns from the entire USPTO dataset regarding when that pair needs to be lumped, i.e., be treated as the same person, thereby being assigned one unique identifier, or to be split, i.e., be treated as different persons, hence being assigned two different identifiers. Our goal is to have a robust engine that will automatically disambiguate the entire USPTO inventors on a weekly basis, upon each issuance of new patents every Tuesday, and to make the disambiguated dataset available to interdisciplinary researchers who rely on this data. This work is supported by the United States Patent and Trademark Office And the American Institutes for Research. 130 Blum Hall 5580 Berkeley, CA (510)

4 Algorithm* Todisambiguateinventors,weconsiderseveralpatentattributesincludingthepublishedname, patenttechnologyclass,city,his/herco9inventornames,andassignee.todisambiguate,we meantoassignuniqueidentifierstoeachinventoracrosstheusptopatentdatabase. Step1:Vectorization Wedefineadocumentunitofapatentasacollectionofthatpatent sattributes.werepresent thatunitasanunorderedcollectionofkeys.theoccurrenceofeachkeyisusedasafeaturefor trainingaclassifier.webuildadocument9by9keyincidencematrix,whichissparsebecausea patentcannotuseallthekeys.forexample,ifapatenthasthreeinventors,hasaprimaryclass, hasoneassignee,hasacity,thenthesumoftherowofthatpatentis =6.ifapatentis filedbyaloneinventor,thenthesumoftherowofthatpatentis =3.toillustrate,see thebelowfigure. SupposeJohnDavisfiledfivepatents,A,B,C,D,E.Here,letJohnDavisbeacolumnthatis depictedbyablackverticaltab.welookcloseratthefiverowsthathave 1 salongthatcolumn, namelytheindexofthepatentsbeingfiledbyjohndavis. Havingformedamatrix,wecancomputecorrelationsbetweenrows(patents)andcolumns (inventors). Step2:Distancemeasurements Distancemeasurementscanbecomputedinanumberofways,e.g.,Euclideandistance, Manhattandistance,Chebyshevdistance,andMahalanobis.Theyallowforevaluationofthe

5 distancethataninventorvariesfromtheotherinventor.here,iadopttheeuclideandistance. Step3:Splittingthesameinventornamebyinter9clustering Certaincommonnamescanrepresentmultiplepersons,e.g.,JohnDavis.If,infact,thereare fourdifferentsuchpeopleacrosstheusptodataset,weshouldpreparefouruniqueidentifiers forjohndavis. Weexaminetheinventorblockandextractalistofuniqueinventornames.Toinitialize,we treateachindividualinventorasadistinctperson,e.g.,byassigninginitialidentifiers.then,we clustereachinventorblockthatiscenteredbytheinventornames,i.e.,johndavisasan example,byapplyingthek9meansclusteringalgorithmbasedonabottom9upapproach.when splitting,wecalculatethenewmeanstobethecentroidsoftheobservationsinthenew clusters. Wedeterminetheclustersizebyminimizingthesumofeachclusters sumoftheinter9cluster s squareddistances.inthisexample,thefivenames,johndavis,aregroupeddowntofour clusters.whenthesumnolongerdecreases,wehavereachedalocalminimum.inotherwords, increasingtheclusternumbertofivewouldn tdecreasetheobjectivefunctionandhencewe stopatfourclustersforjohndavis. Step3:Lumpingdifferentinventornamesbyintra9clustering Oncesplittingisdone,weassignthecentersofeachclusteruniqueidentifiersandaugmentthe

6 matrixcolumn9wisebytheiruniqueidentifiers,asdepicted Bylumping,wemeantomergenamingaliasesintoonepersonif,infact,theyshouldbeone personbasedonwhattheotherfactors,suchasco9inventor,assignee,class,orcity,suggest. ThisstepisdesignedtotreatJimandJames,BobandRobert,or,ArnaudGourdol,ArnaudP Gourdol,ArnaudPJGourdol,andArnoGourdolassamepersons.Thealgorithmassigns inventorstothenearestclusterbydistance.here,ifaninventornotinaclusterisdetermined bythealgorithmtobelumpedintoanothercluster,therearethreepossibilitiesforthenaming match. (1) Thelastnamesagree,thefirstnamesdisagree,andthefirstletterofthemiddlenames agree(ifany),forexample,jimandjames.lumpingisperformed. (2) Thelastnamesdisagree,thefirstnamesagree,andthefirstletterofthemiddlenames agree(ifany),duetomarriageandlastnamechange.lumpingisperformed. (3) Boththelastnamesandfirstnamesdisagree.Namechangeistheonlyreasonforthemto bethesameperson.withoutpersonallyknowingoneoftheinventors,itisimpossibletotell. Lumpingisnotperformed. Step4:Blockingandtie9breaking ThegoaloftheautomatedK9meanalgorithmistoproduceahighclusterqualitywithhigh intra9clustersimilarityandlowinter9clustersimilarity.thisobjectivefunctiontobeoptimizedis expressedas:

7 Ø Q representstheclusterquality.ifthequalityis0orlower,thentwoitemsofthesamecluster are,onaverage,moredissimilarthanapairofitemsfromtwodifferentclusters.ifthequality ratingiscloserto1,itmeansthattwoitemsfromdifferentclustersareentirelydissimilar,and itemsfromthesameclusteraremoresimilartoeachother.thiswillalsoresultinadenser k9mean. Generalization* ThevectorizationofentireUSPTOdatabaseallowsforthedisambiguationprocesstobegeneric. Thiswillalsofacilitatedisambiguationofattorneynamesbyexaminernamesbysimply augmentingthematrixtotherightbyattorneyblocksorexaminerblocks,ortags. Results*for*Download* Thedisambiguationengineprocessedatotalof10,708,200inventornamesacross5,021,243 patents,andidentified3,421,276uniqueinventors.download 1975 July16, Accuracy*Assessment* Weassessaccuracybymeasuringlumping!LandsplittingSerrors. Lumpingoccurswhen distinctinventorsareincorrectlyidentifiedasone.splittingoccurswhenoneinventoris identifiedasmultipleinventors.inthepresentmethod,twoormoreinventorsinthesame clusterconstitutesalumpingerror;oneinventorintwoormoreclustersconstitutesasplitting error. Inordertoestimatetheerrorratesinthetwoclusteringsolutions,wecomparedoureffortstoa manuallycurateddataset.theoriginaldatasetwasasampleof95usinventors(1333 inventor9patentinstances)drawnfromtheengineeringandbiochemistryfields,withcurrentor previousacademicaffiliations.astheseareeminentacademics,thisdatabaseoversamples prolificinventors.thepatentswithinthebenchmarkdatasetwerefirstidentifiedfrom inventors CVs.Weattemptedtocontactallinventorsinthedataset,via andthenphone, inordertovalidateourdisambiguationoftheirpatents.wealsocross9checkedourresultswith onlineresourcesandhumanpatternrecognition.thelatestversionofthecomparisonfileis downloadableat:

8 (BenchmarkfileV7) Theactualcomparisonresultsareat: Foreachinventorinthisstandardweidentifiedtheirsplitrecords(thatfailedtomaptohis/her largestcluster).thetotalnumberofsplitrecordsdividedbythetotalnumberofrecordsinthe standardyieldsoursplittingstatistic.similarly,foreachclusterinthestandard,weidentified lumpedrecords(thatdidnotbelonginthelargestsub9clusterbyasingleinventorinthe standard.)thetotalnumberoflumpedrecordsdividedbythetotalnumberofrecordsinthe standardyieldsourlumpingstatistic.basedonthisbenchmark,splittingandlumpingerrorsare 11/558=1.97%and27/585=4.62%.