Identifying Basic Level Entities in a Data Graph

Size: px

Start display at page:

Download "Identifying Basic Level Entities in a Data Graph"

Allyson Judith Norton
5 years ago
Views:

1 School of Computing FACULTY OF ENGINEERING Idntifying Basic Lvl Entitis in a Data Graph Expanding usr s knowldg whil xploring data graphs By Marwan Al-Tawil Suprvisors Vania Dimitrova Brandon Bnntt

2 Outlin of th talk Introduction Problm and Focus - What maks som ntitis mor important to provid bttr paths? - How to dvlop automatic ways to idntify knowldg anchors? Idntifying Basic Lvl Entitis Exprimntal Study Exprimntal Rsults Conclusion and Futur work

Introduction W usually xplor data to larn nw things Oftn, larning

Wb is our first choic [Marchionini 2006] Th Wb is bcoming powrful with

graphs motivats this rsarch Cognitiv structurs may not match th

3 Introduction W usually xplor data to larn nw things Oftn, larning involvs xploring larg amount of data and dciding what to xplor nxt Th Wb is our first choic [Marchionini 2006] Th Wb is bcoming powrful with linkd data graphs [Brnrs-L 2009] Two dvlopmnts in xploration ovr data graphs motivats this rsarch Cognitiv structurs may not match th smantic structurs Entitis and paths ar diffrnt Usrs hav diffrnt cognitiv structurs

4 Problm & Focus Aiding usrs xploration ovr data graphs to xpand knowldg Srndipitous larning occurs whil moving btwn individual ntitis. Howvr, not all paths ar bnficial for knowldg xpansion. Som ntitis ar important for knowldg xpansion (knowldg anchors). Idntifying ths ntitis is important to nudg th usr through bttr baths.

5 Problm & Focus Two important qustions w nd to answr: What maks som ntitis mor important to provid bttr paths? How to dvlop automatic ways to idntify knowldg anchors?

6 Problm & Focus What maks som ntitis mor important to provid bttr paths? Exploratory study Datast : MusicPinta Instrumnts - 71k prformancs (albums, tracks) - 188k artists - 2.4M ntitis and 38M tripls.

7 Problm & Focus What maks som ntitis mor important to provid bttr paths? Exploratory study Thr stratgis wr xamind (Dnsity, Familiarity, Unfamiliarity). Figur Obsrvations: - Cntral ntitis with many subclasss ar good potntial anchors - Rcognition is a ky nablr for knowldg xpansion - Encourag connctions to discovr nw ntitis linkd to rcognizd ons Subsumption for maningful larning Familiar and basic ntitis ar usd as anchors to subsum nw knowldg into usrs cognitiv structur. Basic ntitis ar dlibratly introducd prior bringing nw knowldg.

8 Problm & Focus Two important qustions w nd to answr What maks som ntitis mor important to provid bttr paths? How to dvlop automatic ways to idntify knowldg anchors?

9 Problm & Focus How to dvlop automatic ways to idntify knowldg anchors? W utiliz th Cognitiv Scinc notion of basic lvl objct introducd by Rosch. Catgory objcts that ar commonly usd in our daily lif (.g. Chair and Dog). Catgory objcts that carry th most information, possss th highst catgory cu validity, and ar, thus, th most diffrntiatd from on anothr. W adopt two approachs Distinctivnss Idntifis most diffrntiatd ntitis whos cus linkd to its mmbrs, and not linkd to othr ntitis Homognity Idntifis ntitis whos mmbrs hav high similarity valus. (i.. complmntary with th distinctivnss) v4 v1 v7 v5 v8 v2 v9 v6 v10 v3

10 Idntifying Knowldg Anchors Prliminaris A Data Graph is a labld dirctd graph V { v1, v2,..., vn} E {, 2,..., 1 m is a finit st of vrtics } is a finit st of dgs P p, p,..., p } is a st of tripls v s, i, vo { 1 2 k Entitis st V in th graph can b dividd into: Catgory ntitis and laf ntitis Rlationship typs; Hirarchical: dnot catgory mmbrship btwn th Subjct and Objct ntitis (.g. ; dctrms:subjct) Domain-spcific: othr than hirarchical rlationships (.g. prformancs) Candidat ntitis ar catgory ntitis DG V, E, P Carnatic music instrumnt Jaakko dctrms: subjct Music Ontology: Instrumnt Instrumnt Violin Kmnch Fiddl String Instrumnt Pluckd string instrumnt Lap Stl Guitar Stl Guitar Guitar Dobro Abstract domain ntity Rsonator Guitar Catgory ntitis Laf ntitis

11 Idntifying Knowldg Anchors Algorithm I: Distinctivnss Mtrics adoptd from formal concpt analysis.g. =dctrms:subjct ntity Blls1 ntity 2 ntity 3 Tubular blls Carillon Cowbll Gankogui Handblls Dctrms: subjct ntity n Dctrms: subjct Pitchd prcussion Drum kit componnts Hand prcussion Idiophons Marching prcussion Mlodic prcussion Orchstral prcussion Blls Sports paraphrnal ia Swiss musical instrumnts Unpitchd prcussion Latin prcussion N := 3 M := 4

12 Idntifying Knowldg Anchors Algorithm II: Homognity Mtrics us st basd similarity mtrics Xylophon ntity 1 ntity 2 ntity 3 Clsta ntity n Crotals Gamlan Gmlan Glocknspi l Marimba Vibraphon Mtallophon Amadinda Balafon Bamboo Angklung Mtal Angklung v1 v2 v3 v4 v5 v6 v7 v8 v9 v v1, v3 v {v3} I v3 U, v1 v3

13 Problm & Focus Two important qustions w nd to answr What maks som ntitis mor important to provid bttr paths? How to dvlop automatic ways to idntify knowldg anchors? How to valuat?

14 Exprimntal Study Exprimntal Stup W follow arlir Cognitiv Scinc studis (fr naming tasks 10s). Imags of all taxonomical ntitis linkd via wr prsntd - Eight survys prsntd 256 laf ntitis. - Two survys prsntd 108 top/catgory ntitis. Bnchmarking Sts Idntifid Accuracy and frquncy wr usd St 1: accurat naming of a catgory ntity(parnt) whn laf ntity is sn - Violotta Violin. St 2: accurat naming a catgory ntity with its xact nam, childrn or parnt. - Violin Violin. - Fiddl Violin. - Pluckd String Inst String Inst Strong Anchors (St1 St2) {Accordion, Bll, Bouzouki, Clarint, Drum, Flut, Guitar, Harmonica, Harp, Saxophon, String instrumnt, Trumpt, Violin, Xylophon} Wak Anchors (St1 St2) {Accordion, Banjo, Bll, Bouzouki, Cllo, Clarint, Drum, Elctric piano, Flut, Gong, Guitar, Harmonica, Harp, Lut, Lyr, Organ, Rcordr, Saxophon, String Instrumnt, Trombon, Trumpt, Tuba, Violin, Xylophon}

15 E D H Exprimntal Rsults Quantitativ Analysis For ach mtric, w aggrgat (union) ntitis for hirarchical rlationships. Idntify a cut-off thrshold point (normaliz valus and tak 60 th prcntil). Prcision valus wr poor ( for StrongAnchors & for WakAnchors) Inspcting th Fals Positiv ntitis, w noticd two rasons for th poor prcision. - Picking ntitis with a low numbr of subclasss [ SN v 1 (1/ { v : v v} ) ] - Rturning FP ntitis which had long labl nams [us wightd mdian of labls] Baslin: using Strog Anchors using WakAnchors Prcision rsults wr improvd noticably (lowst valu 0.41 to highst valu 0.62).

16 E D H Exprimntal Rsults Hybridization Analyz algorithms prformanc at diffrnt taxonomical lvls. - Eight taxonomical lvls - First and last lvls ar xcludd (Figur) Two huristics: - Us hirarchical Jaccard for most spcific catgoris in th graph (FP ntitis wr rich in domain-spcific rlationships) - Tak th majority voting for all othr taxonomical lvls. (Most of th ntitis at th middl and top taxonomical lvl will b wll rprsntd in th graph hirarchy and may includ domain-spcific rlationships) Bnchmarking sts Prcision Rcall StrongAnchors WakAnchors

17 E D H Exprimntal Rsults Qualitativ Analysis of Hybridization xamin FP and FN ntitis Obsrvations Missing basic lvl ntitis du to unpopulatd aras in th data graph - Non of th mtrics pickd FN ntitis (Cllo, Banjo) that blongd to th bottom quartil of th taxonomy (small numbr of subclasss and limitd prformancs). W argu that ths ntitis would tak th usr to dad-nds with unpopulatd aras which may b confusing for navigation Slcting ntitis that ar suprordinat of basic lvl ntitis - Th FP includd ntitis, such as which ar wll prsntd in th graph hirarchy (.g. Rds has 36 subclasss linkd to 60 DBpdia catgoris). - Also, thir mmbrs participat in many domain-spcific rlationships (.g. Rds mmbrs ar linkd to 606 prformancs). W argu that such ntitis could b sn as good picks bcaus thy can provid navigation bridgs to rach basic lvl ntitis

18 E D H Contribution W uniquly provid formal dscription of mtrics and corrsponding algorithms for idntifying knowldg anchors in a data graph. Implmntation of th algorithms to idntify basic lvl ntitis in a data graph in th music domain. Th prformanc of th algorithms is xamind using a bnchmarking st with basic lvl ntitis idntifid by humans. This work has bn accptd in HT2016 confrnc.

19 Conclusion and Futur work Conclusion W dvlop algorithms for idntifying knowldg anchors in a data graph. Knowldg anchors can provid bridgs to subsum nw knowldg. Our approach can b validatd with Crowdsourcing in othr domains. Our approach can b usful for th cold start problm. Our approach can b also applid to ontology summarization for capturing a lay prson s viw of th domain. Futur work Utiliz th algorithms to gnrat navigation paths by following Ausubl s subsumption stratgis for maningful larning. Apply th algorithms in anothr domain.

20 Thank you, Qustions? Marwan Al-Tawil Contact:

Department of Humanities & Religious Studies Assessment Plan (REV 6/16)

Department of Humanities & Religious Studies Assessment Plan (REV 6/16) Dpartm of Humanitis & Rligious Studis Plan (REV 6/16) Larning Goals Outcoms 1. Knowldg of Human Culturs: Studs Humanitis & Rligious Studis should b abl to dmonstrat knowldg of human culturs, thir valus