O# the gro%ing,uantity o# educational resources, such as, the Multimedia 0ducational 2esources / eaching M02LO /! 345 repository, contains 67,888 resources approximately$., it is re,uired to build a more appropriate one$ +hen recovery #unction does not #ind a LO, %hich #ul#ills user&s re#uirements or it %a
O# the gro%ing,uantity o# educational resources, such as, the Multimedia 0ducational 2esources / eaching M02LO /! 345 repository, contains 67,888 resources approximately$., it is re,uired to build a more appropriate one$ +hen recovery #unction does not #ind a LO, %hich #ul#ills user&s re#uirements or it %a
O# the gro%ing,uantity o# educational resources, such as, the Multimedia 0ducational 2esources / eaching M02LO /! 345 repository, contains 67,888 resources approximately$., it is re,uired to build a more appropriate one$ +hen recovery #unction does not #ind a LO, %hich #ul#ills user&s re#uirements or it %a
IPN Mexico City, Mexico aramirezar@ipn.mx, eacostag@ipn.mx ABSTRACT Several approaches have been proposed to recommend the most suitable Learning Object LO! according to student"s pro#ile$ It gauges the similarity bet%een the student&s pro#ile and metadata o# stored LOs$ ' recovery #unction is used in most o# these alternatives( the highest value means a high degree o# matching %ith the user&s needs$ )sually the user has the option to pic* out a LO o# the ran*ed list to loo* through it$ +hen recovery #unction does not #ind a LO, %hich #ul#ills user"s re,uirements or it %as ran*ed too lo% to be considered -suitable., it is re,uired to build a more appropriate one$ /he combination o# t%o or more LOs adds ne% content to the ne% LO so it might be closer to the user&s needs$ /his paper sho%s an approach to do so %hich is incremental because let us gather *no%ledge by means o# merging LOs$ KEYWORDS 01learning, learning object, merging$ !TROD"CTO! /he tas* o# manually revie%ing Learning Objects LO! is enormous because o# the gro%ing ,uantity o# educational resources, such as, the Multimedia 0ducational 2esources #or Learning and Online /eaching M02LO/! 345 repository, contains 67,888 resources approximately$ In addition to this, a Learning Object 2epository LO2! is developed to retrieve LOs in a use#ul %ay and it also has the capability to return a large number o# hits$ )sers #ace the problem o# deciding %hich resources might be closer to their needs, there#ore LOs are expected to be rated by using an automatic method$ /his ran*ing o# LO has a strong in#luence on ho% the user pic*s out a LO #or bro%sing it$ 9or instance, a long list o# ran*ed LOs cannot be bro%sed completely by a user$ Moreover, a manual classi#ication could bias the score o# a LO and it could not match %ith user&s pre#erences because it depends on the human evaluator context$ It is common that most o# search engines ran* a LO as #ollo%s 3:75 3::5; By giving an example of a L By setting search fields By giving a !ser"s profile 'lthough many o# these search engine use a recovery #unction %hich can deal %ith numeric or categorical #ields, both *ind o# #ields cannot be handled by most o# them$ Ran#ing by gi$ing an e%am&le o' a LO In this approach, the user supplies an example o# the desired LO then( a recovery #unction compares the example to LOs stored in LO2, and returns the one that %as ran*ed as the highest$ In the case o# SM0/0 3:<5, it measures similarity bet%een metadata o# the example and metadata o# stored LOs$ Since it is similar to the algorithms used #or early %eb search engines, it has a handicap$ /he measure %ill be slant i# the number o# a term, appearing in metadata, increases 3:5$ Other %or* %ith a di##erent approach is presented in 3:65 %hich ran*s a LO building and training a neural net%or*$ /he method to select LOs can be summari=ed in three steps :! the di##erence bet%een #eatures o# supplied and stored LO is calculated >! the model is trained to get a good accuracy in ran*ing 6! the model is applied to a ran*ing tas*$ ?ue to the computation done in the #irst step, a pair1%ised tas* is carried out in order to compare the example to each LO$ It rises the time o# computation %hen a large number o# objects belong to the search set$ @o%ever, the previous approaches have a disadvantage because it is re,uired that the user is ac,uainted %ith the topic$ 9or instance, a #reshman o# computer sciences, %ho has never covered topics related to advanced programming, might have trouble %ith giving an example o# such subject$ Ran#ing by setting searc( 'iel)s ' common method is to supply a search criterion, %hich is composed o# *ey%ords such as title, language, material type, audience etc$ 'lthough, in#ormation retrieval techni,ues are used #or calculating the degree o# similarity 3:53>5 as it %as mentioned in the previous section, the main di##erence is that *ey%ords do not represent any object so there is not a semantic relationship among them$ /his approach has become popular not only on the LO search but also on documents$ ' combination o# a *ey%ord1based search and a #eedbac* ,uery is used in 3:A5 3>85 3>:5$ Instead o# counting the number o# %ords in metadata, the search engine gathers the citations o# authors and ho% many times their LOs have been do%nloaded$ Ran#ing by gi$ing a user*s &ro'ile In 3:85 3::5, the authors propose t%o metrics based on an analysis o# the LOs that have been bro%sed previously by users$ /he #irst metric measures the pre#erence o# a user, computing the number o# times that a given value is presented in the given #ield in the metadata$ /he second metric counts the number o# LOs shared by users in order to obtain their similarity$ 9rom this point o# vie%, t%o users have a similar pro#ile i# they loo* through the same LOs$ 30scribir texto5 ' pedagogical model in combination to the Learning Object Metadata LOM! 3<5 is used in 3>>5 to split the pro#ile and support the student learning pre#erence$ In the #irst stage, the search engine uses the topic parameter to determine %hich LOs are considered #or the #ollo%ing step in the ran*ing process$ In the second stage, LOs are #iltered again, comparing the student&s pro#ile %ith metadata o# a LO$ 's a result, the LOs matching %ith the pro#ile are sho%ed to the student in a list$ It is clear that the ran*ing process selects LO"s randomly because there is no control on ho% they are retrieved along the t%o stages per#ormed by the system$ Conse,uently, the user cannot *no% %hat is the most relevant LO in the #inal subset$ In this paper, %e propose an algorithm to merge LOs %ith no human interaction$ /he user"s needs, expressed as a ,uery, drive the selection o# redundant items along the process o# merging$ Our approach becomes #lexible since it can use any 2ecovery 9unction 29! to select an item o# out the available ones$ 'dditionally, %e propose a novel measure based on con#usion %hich ta*es into account the context o# the LOs to be evaluated$ In the remaining sections, some de#initions are sho%ed, the pac*ing o# learning objects is brie#ly explained, the algorithm #or merging learning objects is described and #inally, the results and conclusion are presented$ DE+!TO!S In a ran*ing process, many phases might be involved including the tas* o# gauging the matching degree o# a LO %ith the user&s ,uery B!$ Next, %e %ill introduce the de#inition o# Recovery function: #efinition $% Recovery f!nction &R'( Let LO a given learning object and B a given user"s ,uery$ +e de#ine a recovery #unction 29! as; RF(Q, LO)=m Notice that m 38, :5 %hich means that a LO evaluated %ith a value near to : is closer to Q than other one %ith a lo%er value$ 9rom this de#inition %e can see that 29 is not limited to use an especial techni,ue or measure but its outcome must be normali=ed$ /he computation o# confusion 3C5 is sensitive to the context inasmuch as it uses a metric o# hierarchical elements$ It means that, although the elements in a hierarchy are the same, a di##erent topology changes the metric$ It %ill be explained #urther$ +e introduce the de#inition o# confusion and some examples #or completes, a deep explanation can be #ound in 3C5$ #efinition )% *onf!sion in !sing r instead s, for a hierarchy + Let r, s symbolic values, H a hierarchy and r, s H , then the conf!sion in using r instead o# s, %ritten conf(r, s), is; conf(r, r)=conf(r, s)=0 if s is any ascendant of r. conf(r, s)=1+conf(r, fater!of(s)) oter"ise #$am%&e 1: /he hierarchy H1 in 9ig$ : depicts the organi=ation o# statistical methods$ i# someone as*s you #or 'statistica& test( and you give 'inde%endent one sam%&e t test( conf('inde%endent one sam%&e t test(, 'statistica& test()=0 but i# you are as*ed #or 'inde%endent one sam%&e t test. and you give 'students t test( conf( 'students t test(, 'inde%endent one sam%&e t test()=1$ 'ccording to the de#inition >, confusion is computed counting descending lin*s #rom r to s$ #$am%&e ): 9rom 9ig$ :$ %e can see that -e$%erimenta&( is a %rong ans%er #or -*nde%endent t"o sam%&e t test( because conf('e$%erimenta&(, (*nde%endent t"o sam%&e t test()=+ /he normali=ed version o# con#usion is 365; ,O-F(r, s) =conf(r, s). +here is the height o# the hierarchy used #or computing conf(r, s)$ /he ,O-F (r, s) #or example : and > are 8, :D6 and : respectively$ 's %as mentioned previously, the context o# con#usion is given by a hierarchy$ 0ven thought, confusion let us gauge the similarity bet%een t%o concepts( slight modi#ication in the hierarchy topology leads changes in its outcome$ /his measure depends completely on hierarchy$ #efinition ,% Learning b-ect &L( ' learning object is a digital entity designed #or letting students accomplish their educational goals, also it has the #ollo%ing #eatures; it is reusable, interoperable, modular, accessible and portable 3A5 3:>5 3:E5 3:45$ 'ig!re $% A hierarchy H1 of statistical methods. LEAR!!, OB-ECT .ACK!, /o ensure the LO #eatures, LO designers use standards as Instructional Management Systems IMS! 375, Sharable Content Object 2e#erence Model SCO2M! 3:C5 and Learning Object Metadata LOM! 3<5$ Our approach #ocus on SCO2M since it is hierarchical organi=ed and uses LOM in combination %ith IMS, all o# them %ritten in FML$ /he content organi=ation see 9ig$ >!, in SCO2M, is a hierarchical map that de#ines the intended use o# the content through activities$ 0ach item can be nested into another one( a lea# item has no children nested items! thus it is lin*ed to a resource such as a %ebpage$ 'ig!re )% Content organization of a LO packaged using SCORM. Our approach parses this content organi=ation to a hierarchy, as it is sho%ed in 9ig$ :, ta*ing into account that only lea# items have resources$ MER,!, LEAR!!, OB-ECTS Similarity among to&ics o' LOs In the process o# merging learning objects, it is expected that ran*ed LOs are closely related to the ,uery topic hence similar concepts belong to di##erent LOs$ /o #ind the most similar concept ,/ H / to ,a H a %e use the #unction sim(,a, H) 3E5 described belo%$ Let 0a the ascendant node o# ,a #ather o# ,a! and 0/ the ascendant node o# ,/ #ather o# ,/!$ /here are #our cases #or see*ing ,/; Case :$ ,a matchs %ith ,/ and 0a matchs %ith 0/$ In this case 0/ can be the #ather, grand#ather or great #ather o# ,/$ sim returns ,/ and the value : %hich means ,/ is similar to ,a$ Case >$ /he parent 0/ is similar to 0a but 0/1s children do not match %ith ,a$ Calling recursively 0/1= sim(0a), the #unction try to con#irm that 0/ is an ancestor o# ,a$ I# the 0/1 returned is the root o# @b, the algorithm concludes unsuccess#ully returning the value o# 8$ Other%ise sim is #ocused on #inding o# ,/1, a child o# 0/$ /hree cases are possible; a$ Most o# the metadata o# ,/1 coincide %ith metadata o# ,a$ I# ,/1 has children, they must be chec*ed %ith children o# ,a using sim recursively$ I# their children have an appropriate level o# matching the algorithm returns ,/1 and the degree o# coincidence bet%een metadata$ b$ Other%ise, the brother o# 0/ and their descendants are compared recursively %ith ,a, until a reasonable level, using sim$ I# it is success#ul( the degree o# coincidence o# metadata is returned$ I# a#ter this search, no concept is #ound then adds ,a to 0/ o# resulting LO$ c$ I# sim #ails to #ind ,1/, an arbitrary but reasonable ans%er is a child o# the brother o# 0/( tha is, ,/1 is a cousin o# Ca$ Case 6$ ,a matchs %ith ,/ but their #athers do not$ /o determine i# ,a is similar to ,/, most o# metadata in ,a should coincide %ith most o# those in ,/$ Moreover most o# children o# ,a coincide 30scribir texto5 using sim! %ith most o# children in ,/$ /he value returned is the #raction o# metadata and children o# ,/ matching %ith those o# ,a$ I# the previous test #ails, sim is not able to #ind a similar concept and return 8$ Case E$ Neither ,/ nor 0/ %as #ound$ /he #unction concludes unsuccess#ully returning 8$ Algorit(m 'or merging learning object /he main issue o# merging LOs is detecting redundant items$ /he *ey ideas to tac*le it are to search #or similar items by means o# sim #unction and add the most relevant one to the ne% LO$ /he relevance o# t%o concepts is computed by a 29 comparing the user&s ,uery B to a LO or its items individually there#ore merging t%o LOs might have a di##erent outcome depending on B$
0ven though our approach is par1%ised, it can be generali=ed applying the outcome o# the Merging Learning Object algorithm MLO! as an input o# itsel#$ 9rom this point o# vie%, merging three LO is done by MLOMLOLO:,LO>!,LO6!$
+e describe the algorithm as #ollo%s; Let Ha, H/ hierarchical representations o# t%o LO see 9ig$ >! and Hc the resulting LO$ MLO @a,@b,@c! :$ merge@a,@b,@c! >$ merge@b,@c,@c! /he #unction mer2e is the core o# the algorithm$ It travels @a in breadth1#irst order trying to #ind similar concepts bet%een the t%o LOs. Merge@a,@b,@c! : 9or each ,a H a > CbG simCa,@b! 6 I# Cb is null E add Ca to @c 7 else C i# 29B,Ca,!H 29B,Cb! < mergeCa,Cb,Ca! 4 mergeCb,Ca,Ca! A add Ca to @c :8 else :: mergeCb,Ca,Cb! :> mergeCa,Cb,Cb! :6 add Cb to @c :E end i# :7 end i# :C end #or I# sim #inds a ,/, it evaluates ,a and ,/ then the best item is chosen according to 29 line C!$ Since Ca and Cb are a sub hierarchy see 9ig$ 6a$!, it is necessary merge them line <, 4, :: and :>! see 9ig$ 6b$!$ 9inally, the item is added to @c %ithout children i# it does not belong to Hc yet$ /he #irst invocation o# merge, in MLO, adds mostly concepts #rom Ha to Hc as the case, , may be #rom Ha or H/ i# it has a similar concept! but it does not include concepts in H/ that are not in Ha$ /his problem is solved in the second invocation$ Reco$ery 'unction base) on con'usion measure /he recovery #unction based on con#usion RFconf #or short, measures the total con#usion among items in %hich a term o# the user&s ,uery appears$ 9ormally RFconf(Q,&o! is de#ined; Let Q a user&s ,uery %ith ti terms i 1.., n , LO a learning Object, itemi LO an item %here a term ti Q appears and &o LO a 2iven item$
RFconf(Q,&o)=1.n 3i n ,O-F(&o,itemi) Ienerali=ing the above #ormula #or a learning object; RFconf(Q,LO)=1.m34 m RFconf(Q,&o4) /he generali=ation computes the average con#usion o# all items that con#orm the LO according to the user"s ,uery$ a! b! 'ig% ,a% Similar items from to different LOs are detected !y sim. 'ig% ,b% "he process of merging to similar items is depicted# merge adds $propertie%s from Ca and decides that $interpretation% from C! is the most appropriate to !e add. &inally $assumption% is added as ell.
#$am%&e +: 9rom 9ig$ 6a$ +e can see that interpretation is redundant in both LOs so only one must be chosen$ Jou give the ,uery BG.properties, applications, example.$ /he items %here the terms -properties., -engineering. and -example. appear are labeled e,ually in LOa and LOb #or simplicity$ 0valuating RFconf(Q,&o) for -inter%retation. #rom LOa$ RFconf(Q, ( inter%retation() = 1.) ( ,O-F ( 'inter%retation(, (%ro%erties( ) + ,O-F( 'inter%retation(, ( engineering() ) = 1.) ( 1.) + ).) ) = +.5 0valuating RFconf(Q,&o) for -inter%retation. #rom LOb$ RFconf(Q, (inter%retation() = ,O-F ( 'inter%retation(, ( example( ) = 1.) +e are interested in the term closer to the ,uery, in other %ords, %e search the item that minimi=e the confusion in a given context$ In accordance %ith the results -interpretation. #rom LOb is selected$ /he RFconf is not term1dependent as recovery #unction based on KSM$ /his is due to the RFconf gauges ho% #ar the target item is #rom items matching %ith ,uery terms$ RES"LTS In this section, %e present the result o# merging learning objects LOa and LO/ automatically$ /he t%o learning objects %ere pac*ed using SCO2M and merged to produce the result LOc$ /he title and metadata o# each item %ere examined in order to determine their similarity by Kector Space Model techni,ue KSM! 3:5$ /%o 29 %ere tested, the #irst one is based on KSM and the second one on con#usion metric explained in the previous section$ /o compare the error o# our approach the same source LO, merged by MLO, %ere #used manually$ /he error %as computed by #ollo%ing #ormula$ #rror= (num/er of re%eated items in LOc) . ( itemsLO). +here LO, is the result o# merging LOa and LO/ and itemsLO is the number o# items belonging to LO merged manually$ /he LOs manually merged %ere considered as a correct outcome$ /he human merger only has access to the metadata o# the item so our experiments %ere #air$ /he 9ig$ E sho%s the results o# #using eight di##erent LOs #rom sundry topics$ /he second column o# the table presents the number o# repeated items in the LOc$ ' repeated item #alse1negative detection! means that MLO %as not able to detect similar concepts( conse,uently both %ere added instead o# only one$ +e ,uanti#y the number o# items #rom a LO the correct #usion! in order to sho% that MLO avoids a #alse1positive detection$ In this case, the number o# items in LOc is less than the number o# items in LO$ /he percentage o# items #rom LOa and LOb that have metadata is sho%ed in the column number #ive$ CO!CL"SO! A!D +"T"RE WORK /he MLO algorithm is a valuable tool because it lets us accumulate *no%ledge by means o# #using LOs #rom similar topics$ /he 29 plays a *ey role in selecting the most similar item to user&s ,uery there#ore the resulting LO could be closer to user &s needs$ /he success o# our approach depends on the metadata and the title given to each item because they are essential in the comparison carried out by MLO$ /he lac* o# metadata is made up #or the label o# the items there#ore the LO %ith a lo% rate o# metadata %ere merged success#ully$ It is clear that MLO needs a search engine to provide a ran*ed list o# LO$ It guarantees that all LOs belong to the same topic$ 30scribir texto5 /he algorithm needs to be proved in a large amount o# complex LOs such as a #ull course %here items are nested deeply$ )n#ortunately the LOs that #ul#ill these #eatures are scarcely available on the %eb$ 'n evaluation carried out by learners is included in a #uture %or* due to hitherto the number o# LOs o# a speci#ic topic are not enough to develop a #air test$ Since MLO relies on sim and mer2e #unctions #or automatic merging, it biases the process o# selecting similar items according to user&s ,uery$ On the other hand a human #usion %ould choose one out o# t%o items in an arbitrarily %ay$ In addition to this, MLO can merge several LOs as %as described$ +e plan to explore other directions such as using the same 29 in the MLO algorithm and search engine to veri#y the hypothesis that the merged LO is ran*ed as the highest one$ 9or #uture %or* %e are interested in not only merging LOs but also building them #rom a conceptual representation provided by the user$ .opic Repeated items in LOc /tems in LOc /tems in the correct LO 0 of items 1ith metadata Error LOa LO! @uman body : :C :7 :4 8 8$8C Industrial 2evolution > >C >E 6A$: 46$6 8$84 Linear 2egression 8 :8 :8 8 8 8 Psychoanalysis 8 >: >: 8 >8 8 Student"s t test > :> :8 8 8 8$> 'ig!re 2% "he results of merging LOs automatically and manually. 209020NC0S [1] 2$ '$ Lae=a1Jates and L$ 2ibeiro1Neto$ Modern In#ormation 2etrieval$ 'ddison1+esley Longman Publishing Co$, Inc$, Loston, M', )S', :AAA$ [2] S$ Cost and S$ Sal=berg$ ' %eighted nearest neighbor algorithm #or learning %ith symbolic #eatures$ Mach$ Learn$, :8;7<1<4, Manuary :AA6$ [3] '$ Iu=man1'renas, '$1?$ Cuevas, and '$ Mimene=$ /he centroid or consensus o# a set o# objects %ith ,ualitative attributes$ 0xpert Syst$ 'ppl$, 64;EA841EA:A, May >8::$ [4] '$ Iu=man1'renas and M$ M$ Olivares1Ceja$ Measuring the understanding bet%een t%o agents through concept similarity$ 0xpert Syst$ 'ppl$, 68;7<<17A:, May >88C$ [5] Instruccional Management System IMS! >886!$ http;DD%%%$imsglobal$orgD$ [6] S$ Levach*ine and '$ Iu=man1'renas$ @ierarchies measuring ,ualitative variables$ In CICLing, pages >C>1 ><E, >88E$ [7] Learning Object Metadata LOM! >88>$ http;DDltsc$ieee$orgD%g:>D>88>8C:>1#inal1lom1dra#t$html$ [8] M02LO/$ http;DD%%%$merlot$orgD$ [9] M$ N$ Moste#aoui, S$ O$ and I$ M$ Iiaglis$ 'dvances in )bi,uitous Computing; 9uture Paradigms and ?irections$ III Publishing, >884$ [10] F$ Ochoa and 0$ ?uval$ )se o# contextuali=ed attention metadata #or ran*ing and recommending learning objects$ In Proceedings o# the : st international %or*shop on Contextuali=ed attention metadata; collecting, managing and exploiting o# rich usage in#ormation, C'M' P8C, pages A1:C, Ne% Jor*, NJ, )S', >88C$ 'CM$ [11] F$ Ochoa and 0$ ?uval$ 2elevance ran*ing metrics #or learning objects$ I000 /rans$ Learn$ /echnol$, :;6E1E4, Manuary >884$ [12] P$ 2$ Polsani$ )se and abuse o# reusable learning objects$ Mournal o# ?igital In#ormation, 6E!, 8> >886$ [13] L$ 2igutini, /$ Papini, M$ Maggini, and M$ Lianchini$ ' neural net%or* approach #or learning object ran*ing$ In Proceedings o# the :4th international con#erence on 'rti#icial Neural Net%or*s, Part II, IC'NN P84, pages 4AA1 A84, Lerlin, @eidelberg, >884$ Springer1Kerlag$ [14] @$ 2yu and ?$ Parsons$ Innovative Mobile Learning; /echni,ues and /echnologies$ In#ormation Science 2e#erence 1 Imprint o#; III Publishing, @ershey, P', >884$ [15] M$ San=12odrQgue=, and M$ ?odero, and S$ SRnche=1'lonso$ 2an*ing learning objects through integration o# di##erent ,uality indicators$ Learning /echnologies, I000 /ransactions on, 6E!;674 16C6, oct$1dec$ >8:8$ [16] Sharable Content Object 2e#erence Model SCO2M! >88E, http;DD%%%$adlnet$govDtechnologiesDscormDde#ault$aspx$ [17] SM0/0$ http;DD%%%$smete$orgDsmeteD$ [18] '$ Kinha$ 2eusable learning objects; theory to practice$ SIICS0 Lull$, 6<;E:61E:6, Mune >887$ [19] N$ J$ Jen, 9$ 9$ @ou, L$ 2$ Chao, and /$ O$ Shih$ +eighting and ran*ing the e1learning resources$ In Proceedings o# the >88A Ninth I000 International Con#erence on 'dvanced Learning /echnologies, IC'L/ P8A, pages <8:1<86, +ashington, ?C, )S', >88A$ I000 Computer Society$ [20] N$ J$ Jen, /$ O$ Shih, L$ 2$ Chao, and B$ Min$ 2an*ing metrics and search guidance #or learning object repository$ I000 /rans$ Learn$ /echnol$, 6;>781>CE, Muly >8:8$ [21] N$ J$ Jen, /$ O$ Shih, and B$ Min$ ' ne% paradigm o# ran*ing and searching in learning object repository$ In Proceedings o# the second 'CM international %or*shop on Multimedia technologies #or distance learning, M/?L P:8, pages :1C, Ne% Jor*, NJ, )S', >8:8$ 'CM$ [22] L$ '$ M$ Naina, M$ 9$ 2odrigues, Mr, and I$ Lressan$ 'n approach to design the student interaction based on the recommendation o# e1learning objects$ In Proceedings o# the >4th 'CM International Con#erence on ?esign o# Communication, SII?OC P:8, pages >>61>>4, Ne% Jor*, NJ, )S', >8:8$ 'CM$ 30scribir texto5