Knowledge Unit Relation Recognition Based on Markov Logic Networks

Transcription

1 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER Knowledge Unt Relton Recognton Bsed on Mrkov Logc Networks We Wng 1, 2, We We 2, Je Hu 2, Juntng Ye 1, nd Qnghu Zheng 1 1. School of Electronc nd Informton Engneerng, X n Jotong Unversty, X n, Chn 2. School of Computer Scence nd Engneerng, X n Unversty of Technology, X n, Chn Eml: wellwng@126.com Astrct Knowledge unt (KU) s the smllest ntegrl knowledge oject n gven domn. Knowledge unt relton recognton s to dscover mplct reltons mong KUs, whch s crucl prolem n nformton extrcton. Ths pper proposes knowledge unt relton recognton frmework sed on Mrkov Logc Networks, whch comnes prolstc grphcl models nd frst-order logc y ttchng weght to ech frst-order formul. The frmework s composed prncplly of structure lernng, rtfcl dd or delete formuls, weght lernng nd nferrng. Accordng to the semntc nlyss of KUs nd ther reltons, ground predcte set s frst extrcted. Next, the ground predcte set s nputted nto structure lernng module to cheve weght formul set. Then, n order to overcome lmttons of structure lernng, the weght rule set s dded or deleted y humn. The new weght formul set s turned nto weght lernng module to cqure the lst weght formul set. Fnlly, knowledge unt reltons re recognzed y nferrng module wth the lst weght formul set. Experments on the four dt sets relted to computer domn show the utlty of ths pproch. The tme complexty of structure lernng s lso nlyzed. Index Terms Knowledge Unt; Relton Recognton; Mrkov Logc Networks; Knowledge Nvgton I. INTRODUCTION Knowledge unt (KU) s the smllest ntegrl knowledge oject n gven domn, such s defnton, theorem, rule, or lgorthm [1-5]. It s often emedded n vrety of sem-structured or unstructured texts such s TXT, DOC nd HTML n lner wy. Such orgnzton of knowledge sme s trdtonl pper med, s dffcult to ndcte reltons mong KUs. Knowledge unt relton s n mplct dependence relton etween two KUs. In computer network, preorder s eng n the KU Defnton of computer network nd the KU Clssfcton of computer network, whch denotes the ltter s successor to the former n content; nlogy les n the KU Defnton of TCP nd the KU Defnton of UDP, whch denotes oth hve correspondng reltonshp n functon or content. Knowledge unt relton recognton (KURR) s to mne these hdden reltons from KU set, whch cn e very useful for pplctons such s knowledge orgnzton nd knowledge nvgton. In knowledge orgnzton, knowledge cn e orgnzed nto multlyer nd mult-prtcle. From top to ottom, the frst lyer s concepts nd concept reltons; the second lyer s KUs nd KU reltons; the ottom lyer s knowledge documents nd metdt. In knowledge nvgton, when one hs lerned the KU Defnton of TCP, we cn gude her/hm to the KU Defnton of UDP sed on KU relton network. Thus one wll understnd these two KUs deeply [3]. To the est of our knowledge, reserch nto KURR ws stll n prelmnry stge. Wng et l. select nformton of term, type, dstnce, KU relton level nd document level s fetures to represent cnddte relton nstnces. They use Support Vector Mchnes (SVM) s clssfer to recognze KU reltons from four dt set of computer domn. The expermentl results show tht term, type nd dstnce fetures re more effectve [3]. However, the recll of nlogy recognton s lower. Wng et l. employ relton Gussn processes (RGP) nd ntegrted nformton of terms, semntc type, poston of knowledge elements nd relton grph structure to predct KU nlogy relton. The expermentl results show tht the recll nd F1-mesures ovously mprove [4]. Chen et l. tret the KURR tsk s sequence lelng prolem. Vrous fetures ncludng terms, semntc type, dstnce nd context nformton re ncorported to represent cnddte relton nstnces. Expermentl evluton shows tht the method cheves etter performnce. It lso ndctes tht Condtonl Rndom Felds (CRFs) outperform other prolstc models.e. Hdden Mrkov Model (HMM) nd Mxmum Entropy (ME) [5]. Lu et l. fnd two fetures of KU, whch re the dstrutonl symmetry of the domn terms nd the locl nture of the lernng-dependency. Ther method conssts of three stges, (1) Buld document ssocton reltonshp y clcultng the dstrutonl symmetry of the domn terms. (2) Generte the cnddte relton nstnces y mesurng the loclty of the dependences. (3) Use clssfcton lgorthm to dentfy the lernng-dependency of cnddte relton nstnces. The lernng-dependency s the preorder relton. The method extrcts the lernng-dependency effcently [1-2]. Ths pper proposes KURR frmework sed on Mrkov Logc Networks (MLN), whch comnes prolstc grphcl models nd frst-order logc y ttchng weght to ech frst-order formul. The frmework s composed prncplly of structure lernng, do: /jnw

2 2418 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER 2014 rtfcl dd or delete formuls, weght lernng nd nferrng. Accordng to the semntc nlyss nd sttstcs of KUs nd ther reltons, seventeen predctes re defned, ncludng thrteen evdence predctes nd four query predctes. The method of formul extrcton s lso dscussed to remedy the defects of structure lernng. The rest of the pper s orgnzed s follows. Secton II gves defnton of KU, KU relton nd MLN. Secton III descres the frmework of KURR. Expermentl results re dscussed n Secton IV. Secton V presents the conclusons. II. BASIC NOTIONS In ths secton, severl notons relted to KURR re gven. A. KU KU re utomtclly extrcted from textul knowledge documents, whch s our pror reserch work [6]. A forml defnton of KU s s follows [3]: Defnton 1. A KU s quntuple, ke : d, T, CT,st,tc (1) where ) d s the sequence numer of the KU. Its vlue s ssgned y the order of extrcton. In ths pper, d s used to dentfy the KU. ) T s the term set of the KU. c) CT s the core term set, CT T. So-clled core term s the term tht cn ndcte semntc suject or descrptve oject of the KU. d) st s the semntc type of the KU. st { defnton, descrpton, ttrute, clssfcton, method, structure, dstncton, exmple, evoluton}. e) tc s the text content of the KU. Once KUs hve een extrcted from text document, vlues of d, T, CT,st nd tc re cqured. Defnton 2. Let KURT { preorder, nlogy, llustrton, undefnedr } e the type set of KU relton. Let ( ku, ku, kurt) denote relton nstnce, where ku, ku KU KU,, KU s the set of KU nd kurt KURT. When ) kurt = preorder, t ndctes tht ku s successor to ku n content; ) kurt = nlogy, t ndctes tht ku s prllel or correspondng to ku on certn spect; c) kurt = llustrton, t ndctes tht ku s n explnton or exmple for ku ; d) kurt = undefnedr, t ndctes tht preorder, nlogy nd llustrton do not exst etween ku nd ku. For nstnce, ( 513,515, preorder), ( 202,203, llustrton), ( 202, 208, nlogy) nd ( 515,202, undefnedr) cn e derved from KUs n tle I. B. MLN Mrkov Logc s sttstcl reltonl lernng lnguge sed on frst-order logc nd Mrkov Networks. Mrkov Logc cn extend frst-order logc to llow formul to e weghted rther thn e strctly true or flse; tht s, frst-order logc formul cn e volted wth penlty. A MLN, then, s set of these weghted frst-order logc formul [7]. Defnton 3 [8]. A Mrkov logc network L s set of prs ( F, ), where F s formul n frst-order logc nd s rel numer. Together wth fnte set of constnts C { c1, c2,, c C }, t defnes Mrkov network M LC, s follows: ) M LC, contns one nry node for ech possle groundng of ech predcte pperng n L. The vlue of the node s 1 f the ground tom s true, nd 0 otherwse. ) LC, M contns one feture for ech possle groundng of ech formul F n L. The vlue of ths feture s 1 f the ground formul s true, nd 0 otherwse. The weght of the feture s the ssocted wth F n L. The prolty dstruton over possle worlds x specfed y the ground network s clculted y (2), where n ( x ) s the numer of true groundngs for F n x nd Z s the prtton functon tht s used to mke the summton of ll possle groundngs dds up to one [9-10]. 1 P( X x) exp( n( x)) (2) Z III. METHOD In ths secton, KURR frmework sed on MLN s presented. As shown n fgure 1, the frmework cn e decomposed nto fve mjor modules: optonl formul set, structure lernng, rtfcl dd or delete formuls, weght lernng nd nferrng. Frst, ccordng to sttstcl nlyss of KU nd KU reltons, some ovous rules re proposed y hnd. These rules re further trnsform to frst-order logc formuls, whch form formul set. Next, ground predctes nd formuls re nput nto structure lernng module nd the weght formul set s generted [11-12]. The structure lernng s mportnt to MLN [13-15]. The effcency of structure lernng s currently need to e mproved [16]. The weght formul set cn e used for nferrng drectly. In our frmework, rtfcl dd or delete formuls module nd weght lernng module re set. The rtfcl dd or delete formuls revses the formul y hnd nd pror knowledge. Weght lernng estmtes the weght ssocted wth ech formul of gven structure. These two modules ll produce new weght formul set. Fnlly, KU reltons re cqured y nferrng.

3 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER TABLE I. EXAMPLES OF KU d T CT st tc computer network, telecommuncton system, communctng, shrng resources computer networks, Locl Are Network, LAN, Metropoltn Are Network, MAN, Wde Are Network, WAN User Dtgrm Protocol, UDP, protocols, Internet protocol sute, connectonless, trnsport lyer protocol, exchnges pckets, error recovery, gurnteed delvery, cknowledgment UDP, voce over IP, VoIP, retrnsmt, pckets Trnsmsson Control Protocol, TCP, connecton-orented protocol, Internet Protocol, IP, TCP/IP, relle communcton, full-duplex, process-to-process connectons computer network computer networks User Dtgrm Protocol, UDP UDP Trnsmsson Control Protocol, TCP defnton clssfcton defnton exmple defnton A computer network s multple computers connected together usng telecommuncton system for the purpose of communctng nd shrng resources. Bsed on ther scle, computer networks cn e clssfed s Locl Are Network (LAN), Metropoltn Are Network (MAN), nd Wde Are Network (WAN). User Dtgrm Protocol (UDP), one of the core protocols of the Internet protocol sute, s connectonless, trnsport lyer protocol tht exchnges pckets wth mnml error recovery servces nd wthout gurnteed delvery or cknowledgment. UDP s wdely used for stremng udo nd vdeo, voce over IP (VoIP) nd vdeoconferencng, ecuse there s no tme to retrnsmt erroneous or dropped pckets. Trnsmsson Control Protocol (TCP) s the connecton-orented protocol ult on top of Internet Protocol (IP) nd s nerly lwys seen n the comnton TCP/IP (TCP over IP). It dds relle communcton nd provdes full-duplex, process-to-process connectons. Fgure 1. MLN-sed knowledge unt relton recognton frmework In fct, mny MLN sed pproches do not nclude structure lernng ecuse of the effcency [9] [17-19]. A. Predcte nd Ground Predcte Extrcton Predcte symols represent reltons mong ojects n the domn (e.g., Preorder) or ttrutes of ojects (e.g. Defnton). A ground predcte s n tomc formul ll of whose rguments re ground terms (e.g. Preorder (513, 515)). Accordng to the defnton of KU nd KU relton type, ntegrtng wth our nlyses, seventeen predctes re defned, ncludng thrteen evdence predctes nd four query predctes. 1) Evdence Predctes Defnton (ID): The type of ID KU s defnton. Descrpton (ID): The type of ID KU s descrpton. Attrute (ID): The type of ID KU s ttrute. Clssfcton (ID): The type of ID KU s clssfcton. Method (ID): The type of ID KU s method. Structure (ID): The type of ID KU s structure. Dstncton (ID): The type of ID KU s dstncton. Exmple (ID): The type of ID KU s exmple. Evoluton (ID): The type of ID KU s evoluton. CSuStrng (ID1, ID2): The core term of ID1 KU s the sustrng of the core term of ID2 KU. KUSmeType (ID1, ID2): The type of ID1 KU nd ID2 KU s sme. CSme (ID1, ID2): The core terms of ID1 KU nd ID2 KU re sme. CBelong (ID1, ID2): The core term of the ID1 KU ppers the term set of ID2 KU nd do not pper the core term set of ID2 KU. 2) Query Predctes Preorder (ID1, ID2): The relton type of ID1 nd ID2 KU s preorder. Anlogy (ID1, ID2): The relton type of ID1 nd ID2 KU s nlogy. Illustrton (ID1, ID2): The relton type of ID1 nd ID2 KU s llustrton. UndefnedR (ID1, ID2): The relton type of ID1 nd ID2 KU s undefnedr. The vlues of evdence predctes s 1 or 0. The vlues of query predctes s 1 or 0 n trnng phse nd prolty n test phse. 3) Ground Predcte Extrcton The KUs re stored n KU dtse wth (1). The evdence predctes cn trnsform nto ground predcte y scnnng KU dtse. For evdence predctes of one rgument (e.g. Defnton (ID)), ground predctes cn e cqured y one loop. But for evdence predctes of two rguments (e. g. CSuStrng (ID1, ID2)), every two KUs re potentl ID1 nd UD2 nd the clcultons re huge. Lu et l. fnd the locl nture of the lernngdependency [1-2]; tht s, the two KUs tht hve relton re locl. Thus, we cluster the text documents [20-21]. The ID1 nd ID2 prs re generted mong those documents n the sme cluster. For query predctes, cnddte relton nstnces re lso generted mong those documents n the sme cluster. Then, we mnully nnotted the type of KU reltons mong cnddte relton nstnce set. The nnottng work ws conducted s follows: We hred 24 undergrdute students n ther junor yer or senor yer from the computer scence deprtment. They were sked to log on to We-sed nnottng system, where they were gven the nnotton ssgnments. They were sked

4 2420 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER 2014 to nnotte the type of relton y usng ther own knowledge ckground nd ddtonl textook resources. The work lsted 6 months. We creted the dt set tht covers the four courses fter we doule checked the students work [1]. B. Rule Extrcton Formul set s optonl. In order to the performnce of KURR, some rules re extrcted y hnd. These rules re esy to trnsfer nto formuls. A exmple of rule s s follows. f the core terms of ID1 nd ID2 KU re sme, nd the type of ID2 KU s exmple nd the type of ID1 KU s defnton, the relton type of ID1 nd ID2 KU s llustrton. The frst-order logc s: x y, CSme ( xy, ) Exmple ( y) Defnton ( x) Illustrton ( xy, ) The 202 nd 203 KU n tle I meet ths rule. We nlyze the dt set nd hope to cqure vlule rules. Algorthm 1 s the sttstcl nlyss lgorthm. The percentge of KU relton whch meets certn evdence predcte nd the type of ID1 nd ID2 KU cn e cheve y Algorthm 1. Algorthm 1. Sttstcl nlyss lgorthm Input: Dt set of KUs nd KU reltons Output: Percentge of KU relton whch meets certn evdence predcte; The type of ID1 nd ID2 KU Defnton: KU: Knowledge unt KUR: Knowledge unt relton ReltonDt: knowledge unt reltonl dt sets EP: Evdence predcte EPS: Evdence predcte set Clque,j : represents one of four type reltons, j represents the th type relton n the j set Relton (ID1, ID2): represents one of four type reltons, ncludng preorder, nlogy, llustrton, undefnedr Steps For Relton (ID 1j, ID 2k ) ReltonDt For Relton one of four reltons Whle(ID 1j s fxed vlue) ID 2k nd ID 1j KU sve nto Clque,j Lookng for new ID 1j For EP EPS By trversng Clque,j (j chngng vlues), chves percentge of ll the knowledge unt EP tht re unoccuped n the totl knowledge unt of Clque,j For KUR Relton (ID1,ID2) Get the type nformton of ID1 nd ID2 The end of the lgorthm Seven rules re gned y the results of Algorthm 1. C. Structure Lernng Usng serch strtegy, Structure lernng costs huge tme nd memory, whch s n mportnt fctor restrctng the development of MLN. Thus, structure lernng ecome key reserch drecton n MLN [22]. Structure lernng method currently hs een further mproved, ut the core de s modelng v ground predctes of the known world to fnd possle frst-order formuls (usully n the form of norml form). The formuls re ssumed to e held nd the current world mxmum pseudo-lkelhood prolty (.e., the exstence posslty of the world) s clculted under the KB (Knowledge Bse, collecton of frst-order formuls). The current world mxmum pseudo- lkelhood vlue of the lrgest KB (correspondng to the structure of MLN) s gned y loop [23-25]. Structure lernng strt from ll sngle predctes s useful [14]. A. Dt Set IV. EXPERIMENT We downloded 1205 documents relted to four courses of computer domn from the Internet. Then, 4,052 KUs re extrcted nd 6,252 KU relton nstnces re nnotted. Tle II s the dstruton of KU relton nstnces. TABLE II. THE DISTRIBUTION OF KU RELATION INSTANCES Course preorder nlogy llustrton undefnedr Computer network Computer orgnzton Prncple of dtse Computer rchtecture The MLN system Alchemy s employed to lern nd nfer [26]. Precson (P), Recll (R), nd F1 Mcro-verged (F1) mesure re dopted to evlute the performnces of KURR. B. Experment Process 1) Structure Lernng Input: The ground predcte set nd the formul set. Output: The weght frst-order logc formul set. For nstnce, Illustrton(1,2) v Defnton(2) v!defnton(1) Descrpton: the vlues of s weght of the formul, followed y the formul of the structure lernng. nstructon:./lernstruct - trn.mln -o struct.mln -t nput.d -ne Preorder, Anlogy, Illustrton, UndefnedR -mnwt cchesze tghtmxiter 100 -emsze 3 -mxnumpredctes 3 -loosemxiter 2 -numclusesreevl 4 Structure lernng uses serch strtegy nd tkes gret del of tme nd memory. 2) Weght Lernng Input: The formul set of structure lernng nd rtfclty. Output: The weght frst-order logc formul set. nstructon:./ Lernwts-cw CSuStrng KUSmeType CSme CBelong, Defnton, Descrpton, Attrute, Clssfcton, Method, Structure, Dstncton, Exmple, Evoluton - C_wtLernng.mln -t the nput.d -m -perodc -d -o rules wtlerned.mln the the -ne Preorder, Anlogy, Illustrton, UndefnedR noadduntcluses -dmxhour 24

5 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER ) Inference Input: The weght frst-order logc formul set nd ground predcte set of the test dt set. Output: The prolty of query predctes. nstructon:./nfer -cw CSuStrng KUSmeType CSme CBelong, Defnton, Descrpton, Attrute, Clssfcton, Method, Structure, Dstncton, Exmple, Evoluton - B_wtLernng.mln -e nput.d -r pro.txt -q Preorder, Anlogy, Illustrton, UndefnedR. C. Structure Lernng to Weght Lernng In structure lernng, there re two crucl prolems. Frst, the runnng tme s too long. When the KU sze s 500, structure lernng process tkes 3-6 dys. Second, the result s not stsfyng. The formul of structure lernng s mnly sngle predcte (greter thn 90%). Ths result s not prtculrly menngful. Therefore, we dopt mnly weght lernng nd nferrng to ccomplsh KURR tsk. Structure lernng s employed to gn frst-order logc formul set n smll sze KU dt set. These formul set s thrown nto weght lernng together wth formul set of rtfclty. D. Results nd Dsscusson We conducted threefold cross vldton. Thus, ll the results reported here re those verged over three trls. The performnce of KURR s shown n tle III. The F1 of preorder, llustrton nd undefnedr re thn 88.20% on four courses whle the mxml F1 of nlogy s 57.62%. Our method s effectve for preorder, llustrton nd undefnedr relton recognton. Tle III lso ndctes tht the R of preorder, llustrton nd undefnedr relton s hgher thn the P, especlly preorder nd llustrton, whose verge of R-P s 9.13% nd 6.05% respectvely. They show tht our predctes re not perfectly ccurte for recognzng these three types of reltons. By contrres, the P of nlogy relton s hgher thn the R. Ther mnml P-R s 48.54%. They show tht our predctes re too strct for recognzng nlogy relton nd mss mny nlogy relton nstnces. How to fnd new predctes nd mprove the recll of nlogy relton s our urgent work. E. The Runnng Tme of Structure Lernng Here, the runnng tme of structure lernng s refly dscussed. The vrles s only the numer of KU. Consderng runnng tme nd dstruton, the followng runnng tme of ten dt szes re smpled, 2, 5, 10, 20, 50, 100, 200, 300, 400, 500. The lgorthm s sed smplng, so the results s certn rndomness. To vod the nfluence of rndomness, the verge runnng tme s dopted y three smples. The results of three smples s showed n tle IV. The computer s Intel E5410, Dul Core, 2.33GHz nd 2G RAM. Through the verge runnng tme of structure lernng n the tle IV, the tme curve shown n Fgure 2. Cftool of Mtl s employed to get the ftted curve nd fttng formul wth 95% confdence ounds s s follows. T( n) 8626*exp( n) (3) TABLE III. THE PERFORMANCE OF KURR (%) Course preorder nlogy llustrton undefnedr P R F1 P R F1 P R F1 P R F1 Computer network Computer orgnzton Prncple of dtse Computer rchtecture TABLE IV. THE RUNNING TIME OF STRUCTURE LEARNING (TIME UNIT: SECOND) Dt sze Averge Fgure 2. The tme curve of structure lernng

6 2422 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER 2014 Goodness of ft: SSE: 1.682e+09 R-squre: Adjusted R-squre: RMSE: 1.45e+04 From the (3), when n s 1000,.e. the numer of the KU s 1000, the runnng tme of structure lernng s e+06 seconds, whch s out 103 dys. V. CONCLUSION KURR s to dscover ntrnsc nd hdden reltons from KU set. In ths pper, we present frmework sed on MLN to recognze KU relton. The frmework s composed prncplly of structure lernng, rtfcl dd or delete formuls, weght lernng nd nferrng. Seventeen predctes re defned nd seven rules re extrcted. We conducted experments to evlute the performnce of the frmework, predctes nd rules for KURR on four dt sets relted to computer domn. The results show tht the frmework cn cheve compettve performnce of preorder, llustrton nd undefner relton recognton. We wll extend our study n the followng drectons: ) We suggest usng structure lernng to cqure frst-order logc formuls. But the tme complexty s hgh nd the frst-order logc formuls of structure lernng s not prtculrly menngful. We should study the lgorthm of structure lernng nd mprove the effcency of lernng. ) The performnce of nlogy relton recognton s not stsfyng. Anlogy relton s symmetrc. Two KUs n n nlogy relton nstnce re usully smlr n structure. We wll ncorporte ths nformton nto predctes. ACKNOWLEDGMENTS We would lke to thnk the nonymous revewers for ther vlule comments. Ths progrm s supported y Scentfc Reserch Progrm Funded y Shnx Provncl Educton Deprtment (Progrm No. 2010JK723),Nturl Scence Bsc Reserch Pln n Shnx Provnce of Chn (Progrm No. 2012JM8047), the Speclzed Reserch Fund for the Doctorl Progrm of Hgher Educton of Chn (Progrm No ),Ntonl Nturl Scence Foundton of Chn (Progrm No , ), Ntonl Hgh-Tech Reserch nd Development Pln of Chn (Progrm No. 2008AA01Z131) nd Scence nd Technology Project of X'n (Progrm No. CX12629). Ths work s lso supported y the Chn Postdoctorl Scence Foundton (Progrm No. 2013M542370). REFERENCES [1] J. Lu, L. Jng, Z. Wu, Q. H. Zheng, nd Y. N. Qn, Mnng lernng-dependency etween knowledge unts from text, The VLDB Journl, vol. 20, No. 3, pp , [2] J. Lu, L. Jng, Z.H. Wu, Q. H. Zheng, nd Y. N. Qn, Mnng preorder relton etween knowledge unts from text, Proceedngs of the 2010 ACM Symposum on Appled Computng, pp , [3] W. Wng, Q.H. Zheng, J. Lu, Y. Y. Chen, nd P. F. Tng, Explotng vrous nformton for knowledge element relton recognton, Proceedngs of the IEEE Interntonl Conference on Grnulr Computng, pp , [4] W. Wng, Q. H. Zheng, nd Y. Y. Chen, Knowledge element nlogy relton recognton usng text nd grph structure, Proceedngs of the Interntonl Conference on Nturl Lnguge Processng nd Knowledge Engneerng, pp , [5] Y. Y. Chen, Q. H. Zheng, W. Wng, nd Y. Chen, Knowledge element relton extrcton usng condtonl rndom felds, Proceedngs of the 14th Interntonl Conference on Computer Supported Coopertve Work n Desgn, pp , [6] X. Chng nd Q. H. Zheng, Knowledge element extrcton for knowledge-sed lernng resources orgnzton, Proceedngs of the 6th nterntonl conference on Advnces n we sed lernng, pp , [7] C. Kennngton nd D. Schlngen, Stuted ncrementl nturl lnguge understndng usng Mrkov Logc Networks, Computer Speech nd Lnguge, vol. 28, no. 1, pp , [8] M. Rchrdson nd P. Domngos, Mrkov logc networks, Mchne Lernng, vol. 62, no. 1-2, pp , [9] D. Fernández, S. Mrn, J. Lldós, nd A. Fornés, Contextul word spottng n hstorcl mnuscrpts usng Mrkov logc networks, Proceedngs of the 2nd Interntonl Workshop on Hstorcl Document Imgng nd Processng, pp , [10] H. Ppdopoulos nd G. Tznetks, Explotng structurl reltonshps n udo musc sgnls usng Mrkov Logc Networks, Proceedngs of the IEEE Interntonl Conference on Acoustcs, Speech nd Sgnl Processng, pp. 1-5, [11] Q. T. Dnh, M. Exryt, nd C. Vrn, Genertve structure lernng for Mrkov logc networks sed on grph of predctes, Proceedngs of the Twenty-Second nterntonl jont conference on Artfcl Intellgence, vol. 2, pp , [12] R. A. Ross, L. K. McDowell, D. W. Ah, nd J. Nevlle, Trnsformng grph dt for sttstcl reltonl lernng, Journl of Artfcl Intellgence Reserch, vol. 45, no.1, pp , [13] S. Kok nd P. Domngos, Lernng Mrkov logc network structure v hypergrph lftng, Proceedngs of the 26th Annul Interntonl Conference on Mchne Lernng, pp , [14] S. Kok nd P. Domngos, Lernng the structure of Mrkov logc networks, Proceedngs of the 22nd nterntonl conference on Mchne lernng, pp , [15] L. Mhlkov nd R. J. Mooney, Bottom-up lernng of Mrkov logc network structure, Proceedngs of the 24th nterntonl conference on Mchne lernng, pp , [16] S. Kok nd P. Domngos, Lernng Mrkov Logc Networks Usng Structurl Motfs, Proceedngs of the 27th nterntonl conference on Mchne lernng, pp , [17] P. Sngl nd P. Domngos, Entty Resoluton wth Mrkov Logc, Proceedngs of the 6th IEEE Interntonl Conference on Dt Mnng, pp , [18] S. Sorower, T. G. Detterch, J. Ro Dopp, W. Orr, P. Tdepll, nd X. Fern, Invertng Grce s Mxms to

7 JOURNAL OF NETWORKS, VOL. 9, NO. 9, SEPTEMBER Lern Rules from Nturl Lnguge Extrctons, Proceedngs of Neurl Informton Processng Systems, pp , [19] E. Y. H, J. P. Rowe, B. W. Mott, nd J. C. Lester, Gol recognton wth mrkov logc networks for plyer-dptve gmes, Proceedngs of the Seventh AAAI Conference on Artfcl Intellgence nd Interctve Dgtl Entertnment, pp , [20] T. Toht, W. Musjn, nd A. Hmdull, Effcent Term Extrcton nd Indexng Approch n Smll-Scle We Serch of Uyghur Lnguge, Journl of Multmed, vol. 8, No. 5, pp , [21] H. J. J, S. F. Dng, H. Zhu, F. L. Wu, nd L. N. Bo, A Feture Weghted Spectrl Clusterng Algorthm Bsed on Knowledge Entropy, Journl of Softwre, vol. 8, No. 5, pp , [22] T. Khot, S. Ntrjn, K. Kerstng, nd J. Shvlk, Lernng Mrkov Logc Networks v Functonl Grdent Boostng, Proceedngs of the IEEE 11th Interntonl Conference on Dt Mnng, pp , [23] P. Sngl nd P. Domngos, Dscrmntve trnng of Mrkov logc networks, Proceedngs of the 20th ntonl conference on Artfcl ntellgence, pp , [24] P. Domngos, S. Kok, H. Poon, M. Rchrdson, nd P. Sngl, Unfyng logcl nd sttstcl AI, Proceedngs of the 21st ntonl conference on Artfcl ntellgence, vol. 1, pp. 2-7, [25] Y. Xe, B. Luo, R. B. Xu, nd S. B. Chen, Smooth Hrmonc Trnsductve Lernng, Journl of Computers, vol. 8, No. 12, pp , [26] S. Kok, M. Sumner, M. Rchrdson, P. Sngl, H. Poon, et l., The Alchemy system for sttstcl reltonl AI, Techncl report, Deprtment of Computer Scence nd Engneerng, Unversty of Wshngton, Settle, WA, We Wng ws orn n He receved hs B.Sc., nd M.Sc. from North Unversty of Chn n 1997 nd X'n Jotong Unversty of Chn n 2003, respectvely. Hs reserch nterests re text mnng nd mchne lernng.