From: KDD-96 Proceedings. Copyright © 1996, AAAI (www.aaai.org). All rights reserved. Undiscovered Public Knowledge: a Ten-Year Update Don R. Swanson* and Neil R. Smalheiserf .+“. . . or“WI ...- 11-. urvrsion numaruues, wnrversitt of Chicago, 1010E. 59th St.,Chicago,IL 60637;swanson@kiwi.uchicago.edu tDepartmentof Pediatrics,Universityof Chicago, 5841 S. Maryland Ave., Chicago,IL 68637;sma2@midway.uchicago.edu Abstract Two literatures or setsof articles are complementaryif, consideredtogether, they can reveal useful information of scientik interest not apparentin either of the two setsalone. Of particular interest are complementaryliteratures that are also mutually isolated and noninteractive (they do not cite each other and are not co-cited). In that case,the intriguing possibility akrae that n&wd “s. hv Lyww u-c thm “‘1 &tfnrmnt;nn YLL”I&.L.sU”4L 6uy’ u, mwnhXno b..S..“Y.Ayjthem .a.-** 4.Y nnvnl ..u. -... During the past decade,we have identified seven examplesof complementary noninteractive structuresin the biomedical literature. Each structure led to a novel, plausible, and testablehypothesis that, in severalcases,was subsequently corroboratedby medical researchersthrough clinical or laboratory investigation. We have also developed,tested,and described a systematic,computer-sided approachto iinding and identifying complementary noninteractive literatures. Specialization, Fragmentation, and a Connection Explosion By someobscurespontaneous processscientistshave respondedto the growth of scienceby organizingtheir work into soecialties, -r- -~ thus permittingeachindividualto focuson a smallpart of the total literature. Specialties that grow too large tend to divide into subspecialties that havetheir own literatureswhich, by a processof repeated splitting,maintainmore or lessfixed and manageablesize. As the total literaturegrows,the numberof specialties,but not in generalthe sizeof each,increases(Kochen,1963; Swanson,199Oc). But the unintendedconsequence of specializationis fragmentation.By dividing up the pie, the potential relationshipsamongits piecestendto be neglected. Althoughscientificliteraturecannot,in the long run, grow disproportionatelyto the growth of the communitiesand resourcesthat produceit, combinationsof implicitlyrelatedsegmentsof literaturecan grow much fasterthan the literatureitself and can readily exceedthe capacityof the communityto identify and assimilatesuchrelatedness (Swanson,1993). The signilicanceof the “information explosion”thusmay lie not in an explosionof quantityper se,but in an incalculablygreatercombinatorialexplosion of unnoticedand unintendedlogical connections. The Significance of Complementary Noninteractive Literatures If two literatureseachof substantialsize are linked by argumentsthat they respectivelyput forward -- that is, are “logically”related,or complementary-- one would expect to gain usefuiinformationby combiningthem. For example,supposethat one (biomedical)literature establishesthat someenvironmentalfactor A influences certaininternalphysiologicalconditionsand a second literatureestablishesthat thesesamephysiological changesinfluencethe courseof diseaseC. Presumably, then,anyonewho readsboth literaturescould conclude that factor A might influencediseaseC. Under such --->!L--- -f -----l-----ry-?r-. --- --.---,a ?1-- _----_I rl-conamons or comptementdnty one woum dtso expect me two literaturesto refer to eachother. If, however,the two literatureswere developedindependentlyof one another, the logical linkageillustratedmay be both unintendedand unnoticed.To detectsuchmutualisolation,we examine the citation pattern. If two literaturesare “noninteractive” = that ulyc ir1U) ifa. thmv “W, hnvm na6L.V n~.rer ..Y.“. fnr ,“a odAnm\ vva&“..n] kppn “W.. &ml UluIu together,and if neithercites the other,- then it is possiblethat scientistshavenot previouslyconsidered both iiteraturestogether,and so it is possiblethat no one is awareof the implicit A-C connection.The two conditions,complementarilyand noninteraction,describe a model structurethat showshow useful informationcan remainundiscoveredeventhoughits componentsconsist of public knowledge(Swanson,1987,199l). Public Knowledge / Private Knowledge Thereis, of course,no way to know in any particularcase whetherthe possibilityof an AC relationshipin the above modelhasor hasnot occurredto someone,or whetheror not anyonehasactuallyconsideredthe two literatureson A andC together,a private matterthat necessarily remainsconjectural.However,our argumentis based only on determiningwhetherthereis any printed evidence to the contrary. We are concernedwith public ratherthan Data Mining: Integration Q Application 295 private knowledge -- with the state of the recordproduced rather than the state of mindof the producers(Swanson, 1990d).Thepoint of bringing together the ABand BC literatures, in anyevent, is not to "prove"an AClinkage, (by consideringonlytransitive relationships)but rather call attention to an apparentlyunnoticedassociationthat maybe worthinvestigating. In principle anychain of scientific, includinganalogic,reasoningin whichdifferent links appearin noninteractiveliteratures maylead to the discoveryof newinteresting connections. "Whatpeople know"is a common understandingof whatis meantby "knowledge". If takenin this subjective sense, the idea of "knowledge discovery"could mean merelythat someonediscoveredsomethingthey hadn’t knownbefore. Ourfocus in the present paper is on a secondsense of the word"knowledge",a meaning associatedwith the productsof human intellectual activity, as encoded in the public record,rather than with the contents of the humanmind.This abstract worldof human-created"objective" knowledgeis opento explorationanddiscovery,for it cancontainterritory that is subjectively unknown to anyone(Popper, 1972). Our workis directedtowardthe discoveryof scientificallyuseful informationimplicit in the public record,but not previously madeexplicit. Theproblemweaddress concernsstructureswithinthe scientific literature, not withinthe mind. The Process of Finding Complementary Noninteractive Literatures Duringthe past ten years, wehavepursuedthree goals: i) to showin principle hownewknowledgemightbe gained by synthesizinglogically- related noninteractive literatures; ii) to demonstrate that suchstructuresdo exist, at least withinthe biomedical literature; andiii) to develop a systematicprocessfor findingthem. In pursuit of goal iii, wehavecreatedinteractive softwareanddatabasesearchstrategies that can facilitate the discoveryof complementary structures in the publishedliterature of science. Theuniverseor searchspace underconsiderationis limited only by the coverage of the majorscientific databases,thoughwehavefocused primarily on the biomedicalfield and the MEDLINE database(8 millionrecords). In 1991,a systematic approachto finding complementary structures was outlined andbecamea point of departurefor software development(Swanson,1991). The systemthat has now taken shapeis basedon a 3-wayinteraction between computersoftware, bibliographic databases, and a human operator. Taeinteraction generatesinformationstructtues that are usedheuristicallyto guidethe searchfor promisingcomplementary literatures. Theuser of the systembeginsby choosinga question 296 Technology Spotlight or problem area of scientific interest that can be associated with a literature, C. Elsewherewedescribe andevaluate experimentalcomputersoftware, whichwecall ARROWSMITH (Swanson& Smalheiser, 1997), that performstwo separate functionsthat can be used independently.Thefirst functionproducesa list of candidatesfor a secondliterature, A, complementary to C, fromwhichthe user can select one candidate(at a time) an input, alongwith C, to the secondfunction.This first function can be consideredas a computer-assistedprocess of problem-discovery, an issue identified in the AI literature (Langley,et al., 1987;p304-307). Alternatively, the user maywishto identify a secondliterature, A, as a conjectureor hypothesisgeneratedindependentlyof the computer-produced list of candidates. Ourapproachhas beenbasedon the use of article titles as a guideto identifying complementary literatures. As indicatedabove,our point of departurefor the second functionis a tentative scientific hypothesisassociatedwith twoliteralxtres, AandC. Atitle-word searchof MEDLINE is usedto create twolocal computertitle-files associatedwith A andC, respectively.Thesefiles are used as input to the ARROWSMITH software, which then producesa list of all wordscommon to the two sets of titles, exceptfor wordsexcludedby an extensivestoplist (presently about5000words).Theresulting list of words providesthe basis for identifyingtitle-wordpathways that mightprovide clues to the presenceof complementary argumentswithinthe literatures corresponding to Aand C. Theoutputof this procedureis a structuredtitledisplay(plusjournalcitation), that servesas a heuristic aid to identifyingword-linked titles andservesalso as an organizedguideto the literature. Seven Examples of Literature.Based Knowledge Synthesis Theconcept of "undiscoveredpublic knowledge"based on complementary noninteractiveliteratures was introduced,developed,andexemplifiedin (Swanson, 1986a, 1986b).Since 1986, wehave described six more examples,each representinga synthesis of two complementary literatures in whichbiomedical relationshipsnot previouslynotedin print werebroughtto light. Wedescribe also the hypothesesto whichthey have led, andthe strategies wehavefollowedin findingand identifying these structures (Swanson 1988,1989a,1989b, 1990a; Smalheiser& Swanson1994, 1996). Weidentify these exampleshere in terms of A, B, andC, wherein associationsbetweenAand B are foundin oneliterature andassociationsbetweenB andC in anotherliterature, leadingus to drawcertain inferencesabouta previously mreportedassociation betweenAand C. In most cases weanalyzedmultiple B-termsfor a given Aand C. In the fo l l o w i dn eg s c ri p tiw oeind e n tiofyn i yA a n dC , a n da fe w o f th em o rei m p o rtaBn-ct o n n e c tiao l nosnw, gi thth em a i n c o n c l u soi roh ny p o th etos wi sh i c wh ew e rel e d O . th e r a u th ohrsa v re e v i e w e dx ,te n d oe rda ,s s e s th s ei sdw o rk (C h e n1 ,9 9 D 3 ;a v i e1s 9, 8 G 9 ;a rfi e l1d 9, 9 4G ;o rd o&n L i n d s a1 y9, 9 L6 e, s k1, 9 9 1S p; a s si enpr, re s s ). I __ E x a m p l 1e ,1 9 8 6 D: i e ta ryF i s hO i l s( A ) a n d R a y n a u dD’si s e a s(C e) D i e ta ry fi s ho i l s(e s pe .i c o s a p e n taa ec ni dol e)i ca to d c e rta bi nl o oadn dv a s c u cl ah ra n g( eBjsth a at res e p a ra te i y k n o w ton b eb e n e fi ctoi ap la ti e nwtsi thR a y n a uddi ’s sease. O n eB -l i n k a gfoer e, x a m pwl ea ,sd: i e ta ry e i c o s a p e n taa ec ni cdoai ncd e c re ab sl oe ovdi s c o s (B); i ty a b n o rm ah li lgybh l o o vdi s c o s hi tya bs e eren p o rtei nd p a ti e nwtsi thR a y n a uddi ’s s e a s( Se w. a n s o1 n9 ,8 6 a , 1 9 8 6 1b 9, 8 7T).h ei n fe re nthc ea fit s ho i l sm a yb e n e fi t R a y n a pu ad ti e nmtsa yb ere g a rd ae sda s u c c e s s fu l p re d i c ti tw o no;y e a rs a fte pr u b l i c a oti of thn ea b o v e a n a l y sthi sefi, rs ct l i n i ctria al dl e m o n s tras tiu nc gah b e n e fi cei ffe a l cot f fi s ho i lw a sre p o rteb dy m e d i c a l re s e a rc h(cei rs te ad n d i s c u s isneSdw a n s o1 n9 ,9 3 ). E x a m p l2e,1 9 8 8 M: a g n e s i uDme B c i e n c( Ay ) a n d M i g ra i n (C e ). B c o n s i os ts fe l e v ei nnd i reccot n n e c ti wo hn isc, lhe dto a p re d i c titho an m t a g n e s di ue mfi c i e nmci yg hb tei m p l i c a te d i nm i g ra i hn ee a d a c (T h ew. os u c lhi n k a gaeres fo, r e x a m p ml ea: g n e s ci ua m ni n h i bsi pt re a d idnegp re s s ii no n th ec o rte xa, n ds p re a d idnegp re s s m i oanyb ei m p l i c ai te nd m v ”“.--’ m i n e n---tt2 ----, c h .r rn .ri n rn & fi & ~ ~ ra & h a v e & n “_i o“” ----xom.de ------ u s e ad sa m o d eo lfe p i l e p as yn ,de p i l e phs ay bs e e n a s s o c i awtei thdm i g ra i n( Se w) a n s o1 n9 ,8 8 ,1 9 8 9S bi n).c e p u b l i c a oti of thn a at n a l y smi so ,reth a n1 2d i ffe re gn ro t ups o f m e d i crea sl e a rc hh ea rs v re e p o rtea ds y s te moi rcl o c a l m a g n e s di ue m fi c i e ni ncmy i g ra i on rea fa v o ra brel es p o n s e o fm i g ra i pn ae ti e ntotsd i e ta sryu p p l e m e n wtai th ti o n m a g n e s (c i ui m te ad n dd i s c u s isneS dw a n s o1 n9 ,9 3 ). E x a m p l3e ,1 9 9 OA: r g i n i n (eA ) a n dS o m a to m e dCi n (C j . B c o n s i os ts ffi v ep h y s i o l oa gs isco c i a til eo an ds i to n gth e i n fe re nthc ae ot ra l lay d m i n i s tea re rg di n i m n ea yi n c re a s e b l o ol de v e ol sf s o m a to m eC d(thi nel a ttebr e i nkgn o wton h a v aen u m b oe frb e n e fi cei ffe a l c ts F). o er x a m p l e , i n fu s ea drg i n i sn tiem u l a thte esre l e a os feg ro w th h o rm o na en ,dth el a tte i rntu rni sk n o w ton i n c re absl eo o d l e v e ol sf s o m a to m eC d( Si nw a n s o1 n9 ,9 O aO ).u r i n fe re n cl ee dsto a p ro p o sb aym l e d i crea sl e a rc htoe rs c o n d uacc tl i n i ctria al l(d i s c u s fus erthd ei rn ( S w a n s o n , 1 9 9 3 )). E x m i e 4 , i 9 9 4 :D i e ta ryh k g n e s i u (mA j a n d N e u ro l o g Di ci s e a s(C e j. E n d o g e nmo au gs n e s ii ou nmps l a ay k e yro l ei nre g u l a ti n g e x c i to to x i mc ietyd i a bteydth eN M D Are c e p to (B). r E x c i to to x i icni tu ty rni sth o u g toh th a v ae ni m p o rtaron lt e i n v a ri o un se u ro l odgi isce a sWe ses. u g g e sthteadth t e p o s s i be lffe e cot f m a n i p u l aetixnogg e n (e o u.gsd. i e ta ry ) m a g n e s oi unbmra i fu n n c ti oo nrn e u ro l odgi isce ams e ri ts i n v e s ti g a( tiS m o na l h e i&s Sc wr a n s o1 n9 ,9 4 ). _ --- - ” E x a m p i5e , I9 9 5 :In d o m e th a c( iAn) a n dA i z h e i m e r’s D i s e a s(C e) In th i se x a m pthl eeA, a n dC l i te ra tu re a resn e i th de ir s j o i n t n o nr o n i n te ra c In ti vd ee. e thd ,e rei sc l i n i caanl d e p i d e m i o leovgi idce nthc aeit n d o m e th m a ca i yhn a v ae p ro te c tiev ffe e cat g a i nAsl tz h e i m ed r’s i s e a sHeo.w e v e r, w efo u n cd e rta i n d i reacst s o c i a ti(B) o nb se tw e th e ne tw o l i te ra tu re th as wt e ren o m t e n ti o ni nethd ed i re c(A-C t ) l i te ra tu orere, l s e w h eOren .eo f th e sBe-re l a ti o n si hn i p s p a rti c u il na dr i c a th te ad it n d o m e th ab ce icna, uosf iets * “L -*:“.“-.z“..& - z ”.:L --..,.a . .-a ”..“-^ ..d .“,“. x L *,*1 -L ,:-,, i 3 i i i i -L ;u u n w z rg ~ ~ Z L S ;U V IL ~W, U IU a u w m ta y i l l ~ c w . ~ ‘U L IIG U L IG I p a ti e nb tsye x a c e rb actionggn i tidvyes fu n c ti Bo en c. a u s e th i sp o s s i b i al iptyp a re nh tla ydn o bt e e pn re v i o u s l y re p o rtewde,b ro u g iht tot th ea tte n ti o nf n e u ro s c i e n ti s ts ( S m a l h e i&s Se wr a n s o1 n9 ,% ). E x a m p l6e ,1 9 9 5 E: s tro g e n( A ) a n dA l z h e i m e r’s D i s e a s(C e) A s i ne x a m p5l ,th e eA a n dC l i te ra tu re a resi n te ra c ti v e ; e s tro g re e np l a c e mtheenrat pi syre p o rtetodb ea s s o c i a te d w i tha i o w ei rn c i d e no cf Ael z h e i m ed r’s i s e a bs ue th t, e m e c h a n oi sf smu c ah ne ffe ci tsu n k n o wWn .ere p o rte d s e v e ra p rel v i o u s l y -u n i n v eB s-re ti gl aa titeodn s h i p s , p a rti c u l ao rln yei n v o l v ianngti o x i daacntit v i tyth ,a t a p p e a to remd e rii tn v e s ti g aatispo on s s i be lxep l a n a tio of n s th i si n tri g u ire n gl a ti o n s( hS m i pa l h e idszSe wr a n s oi nn , p re s s ). E x a m p l 7e ,1 9 9 6 P: h o s p h o l i p a (sAe) as n dS l e e (C p ) T h eA a n dC l i te ra tu re a resd i s j o iannt dn o n i n te ra c ti v e , b u it m p i i c ire tl yl a tethd ro u ga hs e ot f s u b s ta n(nc eo sta b i y i n te rl e u lk=i nD F tu, m onr e c ro sfai cs toar n d e n d o to x i n /l i p o p o l y s wa ch ci chahareriwdeel kl) n o wbn o th to p ro m o stel e eapn dto s ti m u l ao te n eo rm o re p h o s p h o l i pTahs iess stu. d iyd e n ti fiael di s ot f a g e n ts w h o seeffe c ts o ns l e eapx e s p e c i al i lkl ey ltoy i n v o l v e p h o s p h o l i pa an sdseusg; g e sstee vd e ra s tra l i g h tfo rw a rd e x p e ri m eten statsol f o u hr y p o th eths ai spt h o s p h o l i p a s e s m a yb ei n v o l v iende n d o g e np oa uthsw athy sa re t g u l a te s l e e( Sp m a l h e i&s Se wr a n s oi nnp, re p a ra ti o n ). D a ta M i n i n g :In te g ra ti o n6 r A p p l i c a ti o n 297 Comment The objectsof studyin the work summarizedhereare complementarystructureswithin the scientificliterature. The recognitionof meaningfulassociationsandultimately that of complementarityrequirea high level of subject expertise.The unruly problemsof meaningwithin the naturallanguageof titles and abstractspresentserious obstaclesto more fully automatingthis processof knowledgediscovery. Our computeraidsare therefore designedto enhanceand stimulatehumanability to see connectionsandrelationships.Theseaidsnecessarily derivefrom the immensedatabasesthat providethe routes of intellectualaccessto the literature. Our goal thusfar hasbeento producea working practicalsystemthat yields immediateresultsin furtheringthe aims of biomedical re-w&. and which ueneratfs && .,_-___ at __ the ---’ came ____-’time _____o-------1 , --- atd problemsthat contributeto understandingliterature-based scientiticdiscovery. References ChenZ. 1993.Let documentstalk to eachother:A computermodel for connectionof short documents,The Journal of Documentation 49(1)&l-54. Davies,R. 1989.The Creationof New Knowledgeby InformationRetrievaland Classification.TheJournal of n,.r..-“..,“4c.-... -t.t\~~.~ A<fA\dYl2 2n.l Y”Lurrc;,~K4Lc”r, I .FJ”l. Garrield,E. 1994.Linking literatures:An intriguing use of the Citation Index, Current Contents#21,3-5. Gordon,M.D. & Lindsay,R. K. 1996. Toward discovery supportsystems:A replication,reexamination,and extensionof Swanson’swork on literaturebased discoveryof a connectionbetweenRaynaud’sand fish oil. Journal of the American Societyfor Information Science 47: 116128. Kochen,M. 1963.On naturalinformationsystems: pragmaticaspectsof informationretrieval.Methodsof &formation in Medicine 2(4): 143-147. Langley,P., Simon,H. A., Bradshaw,G. L., andZytkow, J. M. 1987. Scientific Discovery. Computational Explorations of the Creative Process. Cambridge,Mass.: MIT Press Lesk, M. 1991.SIGIR ‘91: The More ThingsChange,the More They Stay the Same.In: SIGIR FORUM. 25(2}:4-7, ACM Press. Popper,K. R. 1972.Objective KnowledgeOxford. Smalheiser,N.R. and Swanson,D. R. 1994. Assessinga gap in the biomedicalliterature:magnesiumdeficiency and neurologicdisease. Neurosci Res Commun 15(l): 1-9. Smalheiser,N.R. and Swanson,D. R. 1996. Indomethacinand Alzheimer’sDisease.Neurology 46583. Smalheiser,N.R. and Swanson,D. R. (in press)Linking 298 Technology Spotlight Estrogento Alzheimer’sDisease:An informatics approach.Neurology Spasser,M. (in press)The enactedfate of undiscovered public knowledge.Journal of the American Societyfor Information Science. Swanson,D. R. 1986aUndiscoveredpublic knowledge. 1 ih*m.., Cbt.nrtarlr~ <M3\- 1 N-1 1a”. Q Y.V, w, sucor ‘CR‘, .Jv\a,.rva-r Swanson,D. R. 1986b.Fish Oil, Raynaud’sSyndrome, and UndiscoveredPublic Knowledge.Perspectivesin Biology and Medicine 30(1):7-18. Swanson,D. R. 1987.Two Medical Literaturesthat are Logically but not BibliographicallyConnected.Journal of the American Societyfor Information Science38(4):228- 233. Swanson,D. R. 1988.Migraine and Magnesium:Eleven NeglectedConnections.Perspectivesin Biology and Medicine 31(4):526557. Swanson,D. R. lY8Ya Online Searchfor LogicallyRelatedNoninteractiveMedical Literatures:A Systematic Trial-and-ErrorStrategy.Journal of the American Society for Information Science40:356-358. Swanson,D. R. 1989b.A SecondExampleof MutuallyIsolatedMedical LiteraturesRelatedby Implicit, UnnoticedConnections.Journal of the American Society for Information Science40~432-435. Swanson,D. R. 199Oa.SomatomedinC and Arginine; Implicit ConnectionsBetweenMutually-Isolated Literatures.Perspectivesin Biology and Medicine 33(2):157-186. Swanson,D. R. 199Ob.Medical Literatureas a Potential Sourceof New Knowledge. Bulletin of the Medical Library Association 78(1):29-37. Swanson,D. R. 199Oc.IntegrativeMechanismsin the Growth of Knowledge A legacyof ManfredKochen. It&ormation Processing & Management26(1):9-16. Swanson,D. R. 199Od.The absenceof co-citationas a clue to undiscoveredcausalconnections,in Borgman,C. L., ed. Scholarly Communicationand Bibliometrics. r*n ,Q)c)x*^__. ¶...-. l-L.-l- e-3A- “p-m IUDI. m-c* lLY-IJI.lYGWUUTY~am,L~;3d~G Swanson,D. R. 1991.ComplementaryStructuresin Disjoint ScienceLiteratures.In SIGIR91 Proceedingsof the FourteenthAnnual Internclrional ACMISIGIR Conferenceon Researchand Developmentin Information Retrieval Chicago,Ott 13-16,199l ed. A. Bookstein,et. al. New York: ACM, p. 280-9. Swanson,D. R. 1993.Interveningin the Life Cyclesof ScientificKnowledge,Library Trends41(4):606-631. Swanson,D. R. and Smalheiser,N. R. 1997.An interactivesystemfor finding complementaryliteratures:a stimulusto scientificdiscovery. Forthcoming.