Close and Loose Associations in Keyword Search from Structural Data

Close and Loose Associa ions in Keywo d Sea ch om S uc u al Da a Johanna Vainio Uni e si y o Tampe e Kansle in inne 1 FI-33014 Uni e si y o Tampe e, Finland s.johanna. ainio@s a .u a. i Ma ko Junkka i Uni e si y o Tampe e Kansle in inne 1 FI-33014 Uni e si y o Tampe e, Finland ma ko.junkka i@u a. i Jaana Kekäläinen Uni e si y o Tampe e Kansle in inne 1 FI-33014 Uni e si y o Tampe e, Finland jaana.kekalainen@u a. i ABSTRACT Keywo d sea ch o e s uc u al da a enables use s o seek in o ma ion om da abases wi hou knowing he s uc u e o da a o mas e ing ac ual que y languages like SQL. In a keywo d que y, da a i ems o ex a ibu es a e ma ched o he keywo ds and he esul o a que y is ypically a se o g aphs consis ing o connec ed uples. The esul should be anked which means ha he ex a ibu es and connec ions mus be sco ed and combined. Typically, he leng h o a connec ion is he main c i e ion in anking he connec ions, i.e. sho e connec ions a e sco ed highe han longe ones. The leng h o a connec ion is usually based on he o eign key e e ences bu hei di ec ion has ecei ed less a en ion. A he concep ual le el, ca dinali y cons ains co espond o o eign key e e ences o hei combina ion. In he p esen pape , we in es iga e he e ec o he combina ions o ca dinali y cons ains on he esul o a keywo d sea ch. We ind ha he combina ion o ca dinali y cons ain s indica es how close he associa ion be ween keywo ds is. We also show ha he Minimal To al Joining Ne wo k o Tuples (MTJNT) p inciple loses seman ic connec ions o agmen s he esul s o a keywo d sea ch om ela ional da abases. CCS Concep s •In o ma ion sys ems •In o ma ion sys ems ~ Rela ional da abase model • In o ma ion sys ems ~ En i y ela ionship models • In o ma ion sys ems ~ Da abase que y p ocessing Keywo ds Keywo d que ies o e s uc u ed da a; associa ions in ela ional da abases; ca dinali y cons ain s; ER model. 1. INTRODUCTION Keywo d sea ch enables end-use s o sea ch da a om ela ional da abases wi hou knowledge o he syn ax o a que y language o he s uc u e o he da a. Howe e , keywo d sea ch in ol es ambigui y and aises new challenges. T adi ionally, ambigui y is associa ed wi h he na u e o keywo d sea ch, i.e. ma ching sea ch keys o documen con en s is mo e o less uzzy. In he con ex o s uc u ed da a, he na u e o ela ionships among en i ies and ex a ibu es may also a ec di e en kinds o seman ic in e p e a ions. Namely, en i ies may be associa ed wi h each o he ia di e en kinds o ela ionships. In he p esen pape , we i s s udy which kinds o se ings he concep ual associa ions o a seman ic da a model se e o he connec ions o en i ies. Then, we analyze hei oles in keywo d sea ch in ela ional da abases. In in o ma ion e ie al, keywo d sea ch inds documen s ha con ain all o some o he keywo ds and anks he documen s acco ding o he s a is ical p ope ies o hei wo ds. The e is no need o sol e how documen s con aining he keywo ds a e connec ed. In con ex o ela ional da abases, keywo d sea ch can be used o ind he op anked connec ions o uples ha con ain all o some o he keywo ds. To p oduce anking, uples ha con ain a keywo d a e e ie ed, and connec ions be ween hese uples a e p oduced. A connec ion o uples may, o example, be a minimal o al joining ne wo k o uples (MTJNT) [4] o S eine ee [1] [2]. The e a e also di e en app oaches how o ank he p oduced connec ions. Ranking can be based, simply, on he numbe o joins o a connec ion o a ibu e, uple o edge le el sco es o combina ions o hem. [6, 7, 8] Di e en connec ions may con ain di e en amoun o in o ma ion and di e en in e p e a ions, e en be ween he same keywo d uples. The e o e, he sho es connec ion is no always he bes ; a longe pa h may be mo e app op ia e [5, 6]. We d aw his conclusion by analyzing he closeness and looseness o concep ual associa ion. This dimension is based on he ca dinali y cons ains ha appea in he connec ions o en i ies. 2. CONCEPTUAL ASSOCIATIONS Seman ic da a models a e concep ual me hods o ep esen ing concep s and he ela ionships among hem. The ER model is he mos common seman ic da a model and i s p incipal p imi i es a e he en i y ype, a ibu e, and ela ionship ( ype) be ween he en i y ypes. A ela ionship in ol es a ca dinali y cons ain ha may be 1:1, 1:N, N:1 o N:M. A cons ain de e mines how many ins ances a e pa icipa ing in he ela ionship a he ex ensional (ins ance) le el. Le he ER schema o Figu e 1 illus a e his. The schema is a agmen om [3] bu no a ibu es a e ep esen ed. The example con ains ou en i y ypes (DEPARTMENT, EMPLOYEE, DEPENDENT and PROJECT) and ou ela ionships among hem. In he example, se e al employees may wo k o a depa men and an employee wo ks in one depa men . An employee may ha e se e al dependen s and a dependen has one employee as a gua dian. Fu he mo e, an employee may wo k on se e al p ojec s and a p ojec may ha e se e al employees. Finally, a depa men may con ol se e al p ojec s and a p ojec is con olled by a single depa men . 2017, Copy igh is wi h he au ho s. Published in he Wo kshop P oceedings o he EDBT/ICDT 2017 Join Con e ence (Ma ch 21, 2017, Venice, I aly) on CEUR-WS.o g (ISSN 1613-0073). Dis ibu ion o his pape is pe mi ed unde he e ms o he C ea i e Commons license CC-by-nc-nd 4.0 Figu e 1: ER- schema Figu e 1 illus a es ha an employee and a depa men may be in associa ion in wo ways. Fi s , an employee wo ks o a depa men and second, (s)he wo ks on a p ojec con olled by he depa men . The i s al e na i e in ol es one ela ionship and he second wo ela ionships, i.e. he i s pa h is sho e bu he longe one con ains mo e in o ma ion because i also de e mines in which p ojec he employee is wo king. This is an essen ial issue in keywo d sea ch om s uc u al da a whe e he esul connec ions should be anked. In o he wo ds, i we wan o emphasize access o mo e in o ma ion a longe connec ion should be anked be o e sho e connec ions. Howe e , usually longe connec ions a e lowe in a esul lis o no in he esul s lis a all. This is jus i ied, because longe connec ions en ail mo e ambigui y han sho e ones and may lose associa ions be ween en i ies. Howe e , he le el o he ambigui y a connec ion in ol es can be examined, and hus, dec eased and con olled. Nex we in es iga e how he le el o he ambigui y can be de e mined based on he ca dinali y cons ain s o he ER-model. An en i y ype in ol es a se o en i ies whe eas ela ionships de e mine how he en i ies can be connec ed o each o he . In he p esen pape , we concep ualize ha a close connec ion be ween en i ies means ha hey a e associa ed wi h each o he unambiguously h ough hei ela ionships. Table 1 con ains a sample o immedia e and ansi i e ela ionships be ween en i y ypes. Rela ionships 1 and 2 ep esen a si ua ion whe e wo en i y ypes and he co esponding en i ies a e connec ed immedia ely. In he immedia e ela ionships, he e is no ambigui y in he seman ics o he connec ions, i.e. he co esponding en i ies a e closely associa ed o each o he . A ansi i e ela ionship con ains mo e han one immedia e ela ionships, i.e. he co esponding en i ies a e connec ed o each o he ia a middle en i y. T ansi i e ela ionship 3 consis s o wo immedia e ela ionships bo h ha ing he ca dinali y cons ains 1:N. In o he wo ds, o one depa men he e a e se e al employees and o each employee he e may be se e al dependen s, bu no ice e sa. This means ha he e is a ansi i e 1:N ela ionship be ween he en i y ypes depa men and dependen . In o he wo ds, he connec ion is (in e se) unc ional. We in e p e bo h in e se unc ional connec ions, only 1:N ela ionships, and unc ional connec ions, only N:1 ela ionships, as unc ional. This is because a connec ion can be ep esen ed in bo h di ec ions, i.e. he connec ion 3 in Table 1 can be ep esen ed om dependen o depa men (dependen N:1 employee N:1 depa men ) as well. A unc ional ela ionship may also con ain 1:1 ela ionships. The e o e, we de ine ha i X1,Y1,…,Xn,Yn ep esen s he ca dinali y cons ain s o a ansi i e ela ionships such ha i  {1, …, n} holds ha Xi = 1 o i  {1, …, n} holds ha Yi = 1 hen he ela ionships is unc ional. In gene al, he immedia e ela ionships and ansi i e unc ional ela ionships de e mine a close connec ion be ween en i ies a he ex ensional le el. Table 1. Rela ionships and hei ca dinali ies in he ER schema Rela ionship Ca dinali y 1 depa men – employee depa men 1:N employee 2 p ojec – employee p ojec N:M employee 3 depa men – employee –dependen depa men 1:N employee 1:N dependen 4 depa men – p ojec – employee depa men 1:N p ojec N:M employee 5 p ojec – depa men – employee p ojec N:1 depa men 1:N employee 6 depa men – p ojec – employee – dependen depa men 1:N p ojec N:M employee 1:N dependen T ansi i e ela ionship 4 consis o 1:N and N:M ela ionships espec i ely. This means ha one depa men can be associa ed wi h employees h ough se e al p ojec s. In o he wo ds, employees ha wo k on a p ojec con olled by a depa men may o may no wo k in he depa men . Fo he eason ha he e a e wo kinds o seman ic in e p e a ions, his kind o ansi i e ela ionship may cause a loose connec ion a he ex ensional le el. T ansi i e ela ionship 5 con ains wo immedia e ela ionships ha ing ca dinali y cons ain s N:1 and 1:N espec i ely. This is called a ansi i e N:M ela ionship because se e al en i ies o he s a en i y ype may be connec ed o se e al en i ies o he end en i y ype ia a middle en i y. This kind o ela ionship causes a mo e ambiguous in e p e a ion on he connec ion o en i ies. Namely, an employee is associa ed wi h a p ojec al hough s(he) may no wo k on i , i.e. an employee is associa ed wi h e e y p ojec a depa men is con olled. The e o e, a ansi i e N:M ela ionships may also cause a loose connec ion be ween en i ies. In gene al, le X1,Y1,…,Xn,Yn, whe e X1 ≠ 1 and Yn ≠ 1, ep esen he ca dinali y cons ain s o a ansi i e ela ionships, hen he ela ionship is N:M ansi i e. Connec ion 6 con ains h ee immedia e ela ionships. The i s ela ionship possesses he 1:N cons ain and he las 1:N cons ain . Howe e , his is no ansi i e 1:N ela ionship because i con ains a ansi i e N:M ela ionship as a pa o i . The e o e, i allows loose connec ions a he ex ensional le el. Nex we demons a e close and loose connec ions a he da abase le el and hei e ec s on keywo d sea ch in ela ional da abases. 3. ASSOCIATIONS IN RELATIONAL DATABASES Roughly speaking, an ER-schema is implemen ed in ela ional da abases such ha o each en i y ype a ela ion is implemen ed. Fo each 1:N ela ion a o eign key is inse ed o he N-si e. Fo each N:M ela ionships a middle ela ion is o med. This ela ion con ains he o eign keys om bo h he pa icipa ing en i y ypes ( ela ions in RDB). A o eign key cons ain is ypically ep esen ed as an a ow om a o eign key o he ela ed p ima y key. The DEPARTMENT EMPLOYEE PROJECT WORKS ON WORKS FOR CONTROLS 1 N M N N 1 DEPENDENT DEPENDENTS 1 N da abase schema and da abase ins ance o Figu e 1 is ep esen ed in Figu e 2. A ibu es a e now ep esen ed. DEPARTMENT ID D_NAME D_DESCRIPTION d1 Cs The main opics o eaching a e p og amming, da abases and XML. d2 in The main opics o eaching a e in o ma ion e ie al and XML. d3 his o y The main opics o eaching a e his o y o Scandina ian. PROJECT ID D_ID P_NAME P_DESCRIPTION p1 d1 DB- p ojec Di e en da a models a e in eg a ed, such as ela ional, objec and XML p2 d2 XML and IR XML o e s a no a ion o s uc u ed documen s. p3 d2 IR ask Task based in o ma ion e ie al WORKS_FOR ESSN P_ID HOURS e1 p1 40 e2 p3 56 e3 p2 70 e4 p3 60 EMPLOYEE SSN L_NAME S_NAME D_ID e1 Smi h John d1 e2 Smi h Ba ba a d2 e3 Mille Melina d1 e4 Walke John d2 DEPENDENT ID ESSN DEPENDENT_NAME 1 e3 Alice 2 e3 Theodo e Figu e 2. Da abase schema and ins ance A keywo d sea ch ypically ocuses on a ibu e alues. A keywo d may ma ch he whole a ibu e alue o a wo d in a ex a ibu e. Le us conside a sample keywo d sea ch Smi h XML “Smi h” ma ches wo i s employees whe eas “XML” ma ches wo p ojec s and wo depa men s. Connec ions 1 – 7 in Table 2 ep esen s some o he connec ions o he keywo d que y “Smi h XML” in he RDB in Figu e 2. John Smi h is associa ed wi h XML h ough di e en connec ions. The sho es and he longes connec ions a e be ween an employee and a depa men as shown in Table 1. John Smi h is also associa ed wi h XML h ough he p ojec by he connec ions ha ing wo s eps (connec ions 2 and 3 in Table 2). Howe e , WORKS_FOR is a middle ela ion and he leng h o he connec ion would be one i he concep ual schema we e ollowed. In o he wo ds, in concep ual app oach middle ela ions should no be aken in o accoun when calcula ing he leng h o a connec ion. Table 2. Connec ions in he RDB and leng hs o he connec ions in he RDB and he ER connec ion leng h in RDB leng h in ER 1 d1(XML) – e1(Smi h) 1 1 2 p1(XML) – w_ 1 – e1(Smi h) 2 1 3 p1(XML) – d1(XML) – e1(Smi h) 2 2 4 d1(XML) – p1(XML) – w_ 1 – e1(Smi h) 3 2 5 d2(XML) – e2(Smi h) 1 1 6 p2(XML) – d2(XML) – e2(Smi h) 2 2 7 d2(XML) – p3 – w_ 2 – e2(Smi h) 3 2 8 d1 – e3 – 1(Alice) 2 2 9 d2 – p2 – w_ 3 – e3 – 1(Alice) 4 3 In a schema (in ensional) le el, connec ions 1 and 2 ha e a close associa ion and connec ions 3 and 4 ha e a loose associa ion be ween he en i ies. Howe e , in an ins ance le el, also connec ions 3 and 4 ha e a close associa ion be ween he en i ies. The connec ions can be ead as ollows: 1) “employee e1(Smi h) wo ks o depa men d1(XML)” 2) “employee e1(Smi h) wo ks on a p ojec p1(XML)” 3) “employee e1(Smi h) wo ks o depa men d1(XML), ha con ols p ojec p1(XML)” 4) “employee e1(Smi h) wo ks on p ojec p1(XML), ha is con olled by depa men d1(XML). In his case employee e1 wo ks on p ojec p1 as associa ed in connec ion 3 and employee e1 wo ks o depa men d1 as associa ed in connec ion 4, bu his canno gene ally be assumed wi hou in es iga ing o he connec ions. This is illus a ed nex . The closes and longes associa ion be ween Ba ba a Smi h and XML ela es o he desc ip ion o he depa men because she wo ks in a depa men ha ma ches XML (connec ions 5 and 7 in Table 2). I is wo h no ing ha Ba ba a is also associa ed wi h p ojec p2 in connec ion 6 al hough she does no wo k in i . This is because he connec ion con ains N:1 and 1:N ela ionships. In o he wo ds his connec ion gi es b oade in e p e a ion and p ojec p2 and employee e2 (Ba ba a Smi h) a e in a loose associa ion. I he ank o connec ions 1 - 7 we e based on he leng h o he connec ion in RDB, he bes connec ions a e 1 and 5 and he wo s connec ions a e 4 and 7. I he leng h o he ER-model we e ollowed and he close associa ions we e emphasized, he bes connec ions a e 1, 2 and 5 and he wo s connec ions a e 3 and 6. In he la e app oach connec ions 4 and 7 ha e a be e ank because hey do no lose he close associa ion (in he schema le el), i.e. he employee wo ks in he depa men and in he p ojec he connec ion includes. Connec ions 8 and 9 in Tables 2 and 3 co espond o ela ionships 5 and 6 in Table 1. Connec ion 8 has a close associa ion and connec ions 9 has a loose associa ion be ween en i ies in bo h he schema and ins ance le els. A commonly used app oach o o m connec ions is Minimal To al Joining Ne wo k o Tuples (MTJNT) [4] o S eine ee [1] [2]. In he MTJNT app oach e e y keywo d exis s in a leas one uple o he joining ne wo k. I is no possible o emo e any uple om he joining ne wo k wi hou losing MTJNT. The MTJNT app oach e u ns minimally connec ed uples ha s ill con ain e e y keywo d. This app oach can lose some meaning ul uples ha a e associa ed o keywo d que ies and MTJNTs. In he p e ious example connec ions 3, 4, 6 and 7 a e los , i he MTJNT app oach we e ollowed. Table 3. Connec ions and ela ionships o he connec ions in he RDB Connec ion Connec ion wi h ela ionships 1 d1(XML) – e1(Smi h) d1(XML) 1:N e1(Smi h) 2 p1(XML) – w_ 1 – e1(Smi h) p1(XML) 1:N w_ 1 N:1 e1(Smi h) 3 p1(XML) – d1(XML) – e1(Smi h) p1(XML) N:1 d1(XML) 1:N e1(Smi h) 4 d1(XML) – p1(XML) – w_ 1 – e1(Smi h) d1(XML) 1:N p1(XML) 1:N w_ 1 N:1 e1(Smi h) 5 d2(XML) – e2(Smi h) d2(XML) 1:N e2(Smi h) 6 p2(XML) – d2(XML) – e2(Smi h) p2(XML) N:1 d2(XML) 1:N e2(Smi h) 7 d2(XML) – p3 – w_ 2 – e2(Smi h) d2(XML) 1:N p3 1:N w_ 2 N:1 e2(Smi h) 8 d1 – e3 – 1(Alice) d1 1:N e3 1:N 1(Alice) 9 d2 – p2 – w_ 3 – e3 – 1(Alice) d2 1:N p2 1:N w_ 3 N:1 e3 1:N 1(Alice) The associa ion o he keywo d que y in connec ion 4 is al eady implici ly isible o he use in connec ions 1 and 2. Howe e , in ha case we ha e o assume ha he use b owses h ough hese wo answe s and disco e s he associa ion om answe s. Fu he , i is no always he case ha he associa ion is implici ly isible in he o he e u ned associa ions as is he case in connec ion 7. 4. DISCUSSION AND CONCLUSIONS We ha e in es iga ed he e ec s o he ypes o connec ions on he esul s o keywo d que ies o e s uc u al da a. We conside ed how ca dinali y cons ain s a ec he anking o que y esul s. We no iced ha ca dinali y cons ains can be u ilized o in e he looseness o an associa ion. A loose associa ion gi es a mo e ex ensi e esul o a keywo d que y because en i ies ( uples) a e associa ed o each o he h ough a mo e gene al en i y o se e al en i ies. The closeness o a connec ion a he ex ensional le el can pa ly be in e ed om he ca dinali y cons ain s o he ER model. Immedia e and ansi i e unc ional ela ionships ensu e he close connec ion be ween he co esponding en i ies. Ins ead, o he combina ions allow close o loose connec ions be ween pa icipa i e en i ies. Fo example, in a ansi i e N:M ela ionship se e al en i ies may be connec ed o each o he h ough a mo e gene al en i y and he seman ics o he ela ionship is ague. Howe e , u he s udies a e needed o in es iga ing how ou indings could be u ilized in anking he esul connec ions. One c i e ion could be he numbe o ansi i e N:M ela ionships in a connec ion. A mo e p ecise app oach could be achie ed by analyzing he ac ual numbe o pa icipa ing en i ies ( uples) in a da abase ins ance. We also p oposed ha he leng h o connec ions should be based on he ela ionships a he concep ual le el because he N:M ela ionship co esponds o a concep ual ela ionship. Mo eo e , 1:N o N:1 ela ionship can be implemen ed by a middle ela ion. By using concep ual ela ionships he leng h o connec ions does no depend on implemen a ion issues o his kind. The esul s o a keywo d sea ch may p oduce se e al pa hs be ween uples and hey should be anked based on hei assumed ele ance. One widely used indica ion has been he leng h o he pa h, i.e. he sho es pa hs a e ypically assumed o be mo e ele an han longe pa hs. Howe e , longe pa hs may con ain mo e in o ma ion han sho e pa hs and sho e pa hs may chop a seman ic connec ion be ween en i ies o ex a ibu es/documen s. The e o e, he e should be an al e na i e whe e he use could selec longe pa hs, i s/he is in e es ed in la ge con ex o ma ched alues o documen s. 5. REFERENCES [1] Adi ya, B., Bhalo ia, G., Chak aba i, S., Hulge i, A., Nakhe, C., Pa ag, P. and Suda shan, S. BANKS: B owsing and Keywo d Sea ching in Rela ional Da abases. In P oceedings o he 28 h In e na ional Con e ence on Ve y La ge Da a Bases. VLDB ’02. VLDB Endowmen , 2002, 1083-1086. [2] Be gamaschi, S., Gue a, F. and Simonini, G. Keywo d Sea ch o e Rela ional Da abases: Issues, App oaches and Open Challenges. In Fe o, N. ed. B idging Be ween In o ma ion Re ie al and Da abases: PROMISE Win e School 2013. Re ised Tu o ial Lec u es. Sp inge Be lin Heidelbe g, Be lin, Heidelbe g, 2014, 54-73. 2002, 1083- 1086. [3] Elmas i, R. and Na a he, S. B. Fundamen als o Da abase Sys ems, Fou h Edi ion. Addison-Wesley Longman Publishing Co., Inc, Bos on, MA, USA, 2003. [4] H is idis, V. and Papakons an inou, Y. Disco e : Keywo d Sea ch in Rela ional Da abases. In P oceedings o he 28 h In e na ional Con e ence on Ve y La ge Da a Bases. VLDB ’02. VLDB Endowmen , 2002, 670-681. [5] Ka ga , M., An, A., Ce cone, N., God ey, P., Szlich a, J. and Yu, X. MeanKS: Meaning ul Keywo d Sea ch in Rela ional Da abases wi h Complex Schema. In P oceedings o he 2014 ACM SIGMOD In e na ional Con e ence on Managemen o Da a.. SIGMOD ’14. ACM, New Yo k, NY, USA, 2014, 905-908. [6] Ka ga , M., An, A., Ce cone, N., God ey, P., Szlich a, J. and Yu, X. Meaning ul keywo d sea ch in ela ional da abases wi h la ge and complex schema. In Anonymous 2015 IEEE 31s In e na ional Con e ence on Da a Enginee ing. 2015, 411-422. [7] Li, L., Pe schula , S., Tang, G., Pei, J. and Luk, W. E icien and E ec i e Agg ega e Keywo d Sea ch on Rela ional Da abases. In .J.Da a Wa ehous.Min., 8, 4 (oc 2012), 41-81. DOI=10.4018/jdwm.2012100103. [8] Zeng, Z., Bao, Z., Lee, M. L. and Ling, T. W. Towa ds An In e ac i e Keywo d Sea ch o e Rela ional Da abases. In P oceedings o he 24 h In e na ional Con e ence on Wo ld Wide Web. ACM, New Yo k, NY, USA, 2015, 259-262. .