Enhanced e-Learning Experience by Pushing the Limits of Semantic Web Technologies

Abstract

We investigate a novel approach to e-Learning using Semantic Web technologies and aiming to optimise the learners' experience incorporating pedagogical strategies into the learning process. The increasing availability of Semantic Web based educational resources and the establishment of open metadata standards like IEEE LOM pave the way for enhanced e-Learning systems that support personalised learning based on a reasoning framework and formal ontologies. We use ontologies to describe learning objects and the learner state, and de ne pedagogical recommendation axioms that specify which learning objects are best suited for a particular learner in a speci c situation. Recent pedagogical ndings suggest that the individual learning can be optimised by means of guidance through learning pathways, i. e., a particular order in which learning objects have to be studied. To this end, we o er an OWL model for learning pathways as structured sequences. We show the strengths and limits of OWL and propose a solution to overcome problems such as handling of soft constraints and ranking of result sets. The calculation for ranking is based on individual weights for speci c contextual as well as learner pro le features and considers the degree of match between learning objects and learner's needs. The validity of our approach is shown by means of real-life course material, i.e. a complete course on the Philosophy of Didactics, implemented as a prototype system and interfaced to variouss standard Learning Management Systems.

-

Today, sharing and designing course materials is facilitated by the availability of international standards that allow the integration of freely available resources. However, to make learning more interesting for an individual student, enhanced user adaptation is required. This is a key aspect of advanced Technology Enhanced Learning (TEL) systems as they are expected to be used by several learners without assistance of a human tutor.

We propose an ontology-based approach for user adaptation based on a set of learner attributes (e. g., age, gender, learning speed) and the course material being annotated with essential metadata. Learning Objects (LO s)5 are de ned as small, self-contained, reusable units of learning and generally form part of a course. More speci cally, adaptation takes place according to the learners' behaviour, performance, pro le, learning history, contextual situation (the socalled Didactic Factors ), and the speci ed learning goal along with the learning pathway settings, resulting in a personalised recommendation. This kind of adaptation was examined before in terms of its potential to provide a personalised and improved learning pathway o ered to students together with a customised guidance along the learning pathway [ 26 ].

Ontologies provide a uniform model for the di erent aspects of a learning process that can be conceptualised as the navigation through a network of available LO s, thus requiring techniques for dynamic and adaptive sequencing of LO s. One major contribution of this paper is an enhanced OWL model for e-Learning that is used in tandem with an OWL reasoning framework. It is also applicable for dynamic courseware generation based on more abstract concepts, e. g., the type of the Learning Object (i. e., Knowledge Type). This is particularly useful in web-scale e-Learning settings, where the set of LO s is generally very large, resulting in a high number of pathway relations between them, and thus extensive manual annotation e ort.

However, two important issues that need to be addressed clearly show the limits of purely knowledge-based semantic approaches, namely ranking of result sets and handling of soft constraints. The former problem is particularly important when dealing with MOOCs (Massive Open Online Courses) that often deliver a large number of hits. In this case, a prioritisation of LO s w.r.t. an appropriate ranking scheme is required that re ects the relative importance of LO s for a particular learner. On the other hand, when dealing with closed learning environments, it can also happen that a complete satisfaction of all constraints might not be possible. Therefore, we propose an extension of the ontology-based framework which comprises the calculation for ranking LO s based on individual weights for speci c Didactic Factors and considering the degree of match between LO s and learner's needs.

In our work, we focus on the recommendation of LO s considering all required features are given. In our setting, they will be provided by a standard Learning Management Systems (LMS) (e. g., Moodle, Clix, ILIAS). In our application scenario, learners are o ered navigational recommendations, yet, they are free to choose whether or not to follow them.

In the remainder of this paper, Sect. 2 discusses related work in the eld of ontology-based frameworks for personalised e-Learning with a critical review. Section 3 describes our recommendation approach, including an extended module for integrating ranking on top of the reasoning process. Section 4 depicts implementation and validation aspects, before Sect. 5 concludes the paper. 5 In this paper we also talk about Knowledge Objects.

State of the Art: Ontology-based Frameworks for e-Learning

Various ontological frameworks were proposed for e-Learning [ 15, 14, 28, 9, 24, 5, 17 ], focusing on incorporation of open standards and personalised recommendations [ 7, 23 ].

Knowledge integration in most cases considers the metadata of the learning resources, complemented by a domain model and a learner model [ 15, 14 ]. In our work, we explicitly refrain from integrating domain ontologies for the course content since these resources are often not readily available, and building up an ontology from scratch generally requires enormous human e ort. We therefore claim that as few assumptions as necessary should be made about the e-Learning domain. On the other hand, we stress the importance of domain-agnostic pedagogical strategies set up by didactic experts. We build up a knowledge base that serves as a playground for tutors to evaluate if the didactic recommendations meet their desires and guide the learner correctly.

Di erent logical frameworks for ontologies have been used for e-Learning, with varying expressiveness and complexity, ranging from simple taxonomies [ 13 ], to more complex representations with axioms that constrain the interpretation of the model. The most common knowledge representation formalisms adopted for e-Learning are F-logic [ 15, 14, 28, 9 ], OWL-DL [ 12 ], the Semantic Web Rule Language (SWRL) [ 24 ], and OWL-S [ 5 ]. In our approach, we use OWL 2 DL, a W3C standard that builds on Description Logics and extends the earlier OWL standard by several language features while still preserving decidability.

A fundamental aspect of e-Learning are so-called \learning pathways" that aim to optimise the individual learning experience. The most common formal approach to learning paths is based on the IEEE SCORM and IMS simple sequencing speci cation6. However, various di erent implementations are used.

In ontology-based frameworks, often an additional domain ontology layer is integrated [ 27, 11 ] where a \hasResource(C; LO )" relation is used to link a Learning Object to a domain concept. In order to reach a learning goal, the order of concepts is constrained by a partial ordering relation \isRequiredBy (C1; C2)". A similar modelling approach is o ered by Yu et al. [ 30 ] based on a binary relation \hasPrerequisite(C1; C2)", which describes content dependency information at the course level to generate a learning pathway.

We o er an extension of the ontology-based approach that takes into account the fundamental characteristics of a learning pathway, such as modular composition, nested composition, optional parts, and sequencing. Thus, our approach is more expressive, since we seek to model entire structured sequences, following the learning pathway speci cation as suggested by Janssen et al. [ 16 ]. Finite state (FS) frameworks also have been used for learning pathway modelling and almost exclusively rely on Directed Acyclic Graphs (DAGs) [ 2, 17 ]. The main focus of this work, however, has been on e ciency rather than expressiveness with the aim of shortest path analysis, using Weighted FS networks. 6 http://scorm.com/scorm-explained/technical-scorm/sequencing/

The main approach to ranking considers a hybrid approach (i. e., semantic and arithmetic). Shen et al. [ 27, 11 ] use competency gap analysis to this aim, favouring learning content that might help the user to progress towards his/her learning goal. Yu et al. [ 30 ] give preference to Learning Objects that are in a close taxonomic relationship of the respective domain to the learning goal of the user, speci ed as dc:subject metadata entry.

Both approaches thus require a separate domain model. Opposed to this, we rank the Learning Objects according to the degree of relevance to the learner and select the highest scoring Learning Objects , considering that di erent Didactic Factors have a di erent impact on the overall recommendation and to what degree learning objects ful l the recommendation constraints. 2.1

Strengths of Ontology-based Frameworks

Information sharing, integration and reuse. Ontologies support the use

of well-established standards for de ning and sharing Learning Objects within di erent e-Learning platforms. International metadata standards exist that offer a set of metadata descriptors such as LOM (Learning Object Metadata) and SCORM (Shareable Content Object Reference Model), cf. Aroyo et al. [ 4 ], Dolog et al. [ 9 ]. Speci cally, in di erent educational environments, this increases interoperability and enables reusability of learning material.

Semantic Search and Reasoning. Ontology-based approaches have become increasingly popular, since they o er additional reasoning capabilities and thus support semantic search of LO s [ 8 ]. Unlike the search paradigm on the Web, the focus is on searching for structured data, where LO s are semantically annotated on the metadata level. Consequently, a more precise information need can be expressed by means of a complex constraint query. Moreover, LO s are generally related to each other via structural relationships re ecting, i. e., the learning pathways, and this semantic graph can also be exploited for search. 2.2

Weaknesses of Ontology-based Frameworks

E ciency. Reasoning in expressive Description Logics has exponential runtime complexity in the worst case. However, implementations of state-of-the-art OWL reasoners are typically optimised to show acceptable runtime behaviour in many real-world scenarios. Modelling in the tractable OWL 2 Pro les, however, comes with considerable compromises regarding expressiveness.

Support of complex conjunctive queries. Especially in cases where no resource completely satis es all conjuncts of a given complex conjunctive class expression, it would be of interest to determine the winner among the competing candidates, i. e., the learning resources which ful l most of the constraints. A naive approach to instance retrieval inference may often return an empty result set, since some constraints might not be satis ed. We thus need to nd an optimal solution that satis es a maximal subset of the constraints.

Sequences. In e-Learning, Learning Objects are organised into sequences that describe an optimal navigational path towards a learning goal. At present, there is no support for de ning sequences in OWL as would be needed for reasoning over learning pathways. However, by use of Ontology Design Patterns (ODP), best practice solutions can be adopted that allow for regexp-like queries [ 10 ]. 2.3

Limits of Ontology-based Frameworks

Ranked Retrieval. In order to retrieve a ranked list of suitable Learning Objects with a recommendation factor, standard Boolean Retrieval, as facilitated in OWL, is not enough, since it does not support the scoring of objects but delivers an unordered result set. Instead, the results should be ranked re ecting the degree of relevance to the learner. A semantic search algorithm that integrates ranking based on RDF graphs and similar to the PageRank scoring algorithm has been investigated by Kasneci et al. [ 18 ], while an account based on exploiting semantic relationships between entities has also been proposed [ 1, 3 ]. SPARQL allows to rank results by means of the \ORDER BY" predicate [ 22 ] but the data used for computing the order has to be available in the RDF graph explicitly.

Support of Soft Constraints. Preferences behave like soft selection constraints. In this sense, no exact match is required and therefore soft constraints should be satis ed if possible, but may be violated if necessary. A metric generally aims to provide an idea for how close a result is to the learners' needs, i. e., to which degree a Learning Object is relevant. The challenge related to the ways in which such vague knowledge in terms of OWL and its reasoners should be incorporated is well recognised, and some fuzzy and rough extensions have been proposed (cf. [ 19, 25, 21 ]), however, a standard semantic web compliant solution regarding vagueness is not available yet. 3

Enhanced Framework for Personalised e-Learning

This section presents an approach for recommending learning material. It is based on a modular ontology design, and an extension for ranking the results. 3.1

Modular Ontology Design

We propose a modular ontology design in order to cover static pedagogical background knowledge as well as course and learner speci c, dynamic knowledge.

In the pedagogical ontology [ 29 ] learning material is organized into Courses (KD s), Lessons (CC s), and Knowledge Objects (KO s)7, forming a hierarchical graph structure. It provides concepts for Knowledge Type (KT ), e. g., orientation, example, assignment, etc., and Media Type (MT ), e. g., text, video, audio, etc. and metadata vocabulary for describing KO s, such as, hasDi cultyLevel, hasEstimatedLearningTime, hasLanguage, hasRecommendedAge, isSuitableFor

Blind, isSuitebleForDeaf, or isSuitableForMute. Moreover, classes and properties

7 We use the term Knowledge Object instead of the frequently used term Learning Object in order to di erentiate between concrete learning material and the more abstract course topics. for specifying learning pathways (LP s) are formalized, distinguishing between macro- and micro-level learning pathways on the CC and KO level, respectively.

The cognitive map and content map are instantiations of the pedagogical ontology.

The learner model ontology with associated instance data, i.e. the learner state ontology, de nes classes and properties for describing the current learner state, a snapshot characterized by Didactic Factors ( DFs) that currently hold, including the completion state of KO s and CC s, and current learning pathways. 3.2

Speci cation of Learning Pathways

We follow a exible approach to LP modelling that uses auxiliary individuals for connecting KO s (here: CKO (i;j)).

MicroLP v LP MyMicroLP v MicroLP

MyMicroLP (CKO (1;2)) hasPredLP (CKO (1;2); KO 1) hasSuccLP (CKO (1;2); KO 2)

A learning pathway is described as a subclass of MicroLP , which contains only the connector individuals, in this example MyMicroLP . Using the same principle, it is possible to specify the currently selected learning pathway: CurrentLP v LP

CurrentLP v 9isCurrentLP :Self

MyMicroLP v CurrentLP 3.3

Hard and Soft Criteria Based on Didactic Factors

Hard and soft criteria are de ned in the learner model ontology and describe classes of KO s that ful ll the respective constraints.

Hard criteria de ne requirements a KO must meet in order to be included in the recommendation. An example are the disabilities DFs.

Soft criteria de ne requirements a KO should meet in order to be included in the recommendation. For instance, the learner might prefer reading text rather than seeing a video in the user settings. In order to specify the importance of soft criteria, a weight can be assigned to each soft criterion (see Sect. 3.5). 3.4

Speci cation of Recommendation Axioms

Knowledge Object Restriction Type 1 (learning pathway successors).

This restriction type describes KO s that are successors of the current or previous KO w.r.t. the current learning pathways8. Only KO s are considered, that have 8 Both macro- and micro-level learning pathways are considered. (1) (2) (3) (4) (5) (6) (7) (8) not yet been completed. There are three sets of KO s determined. For all three queries, let the current macro- and micro-level learning pathways be de ned as in axioms (2), (3), (4), and (5), and furthermore in axiom (8).

1. Direct successors of the current KO . A property connecting any KO

with its direct successors w.r.t. the current learning pathway can be inferred via the property chain hasPredLP isCurrentLP

hasSuccLP v hasDirectKOSuccessor (9) In case the current KO is the last KO within a CC , the current macro-level LP has to be considered. This requires two additional constructs: (a) Marking the rst and last KO w.r.t. a micro-level LP within a CC by asserting the connector individuals as described in axiom (3) to an additional class FirstLPElement , or LastLPElement , resp.:

FirstLPElement (CKO (i;j))

LastLPElement (CKO (k;l)) hasPredKT hasSuccKT hasKT hasKT v hasPredLP v hasSuccLP 9 We only discuss KT pathways here. MT pathways are handled analogously.

v nextKOsInNextCCs (b) An auxiliary property connecting all KO s of a CC with all KO s of the successor CC according to the current macro-level LP : isContainedByCC

hasDirectCCSuccessor isContainedByCC .

The direct successors of the current KO can now be determined by retrieving all individuals for the following class expression: 9hasDirectKOSuccessor :CurrentKO t (9nextKOsInNextCCs :(CurrentKO u

9hasSuccLP :(CurrentLP u LastLPElement )) u (9hasPredLP :(CurrentLP u FirstLPElement ))) Micro-level learning pathways for KO s do not necessarily have to be explicitly de ned, but can also be inferred based on general didactical knowledge in terms of Knowledge Type or Media Type pathways9 represented in the pedagogical ontology [ 29 ]. KO successors based on KT pathways can be retrieved as in class expression (13), however, two additional subproperty axioms are necessary. Let KT s be put in a pathway relation using auxiliary connector individuals as in axioms (3), (4), and (5), with according connector properties hasPredKT and hasSuccKT . Property assertions for hasPredLP and hasSuccLP , i. e., on the KO level, can be inferred as follows: (10) (11) (12) (13) (14) (15) where hasKT is the property assigning KT s to KO s. In order to infer KO successors based on a KT pathway now simply requires an axiom making the class describing the KT pathway subclass of CurrentLP . 2. All (direct and indirect) successors of the current KO . A property connecting any KO with all its successors (direct and indirect) w.r.t. the current learning path can be inferred via a transitive superproperty of hasDirectKOSuccessor : hasDirectKOSuccessor v hasKOSuccessor trans(hasKOSuccessor ) (16) (17)

3. Direct successors of the previous KO . Direct successors of the previ

ous KO , i. e., the KO visited before skipping, can be retrieved analogously, replacing CurrentKO with PreviousKO in class expression (13). Knowledge Object Restriction Type 2 (un nished predecessors). This restriction type describes KO s that are predecessors of the current KO the learner has not yet completed. According to the completion state described by the classes CompletedKO , PartiallyCompletedKO , and UnseenKO (as disjoint union of the class KO ), the un nished predecessors can be retrieved as follows: :CompletedKO u 9hasKOSuccessor :CurrentKO (18)

Knowledge Object Restriction Type 3 (soft criteria). This type of KO

restriction describes sets of KO s that ful ll soft criteria. The class expressions speci ed in the learner model ontology associated to these soft criteria are evaluated independently from each other and deliver di erent sets of candidate KO s that are used for the ranking algorithm in the subsequent post-processing. 3.5

Extension: Ranking

We provide an extension of the framework, integrating ranking into the learning object recommendation in a post-processing step based on the reasoning results. Our approach is most similar to Kerkiril [ 20 ] and Alian et al. [ 2 ] based on Simple Additive Weighting, a widely used multi-attribute decision technique.

In our work, DF features have a weight that indicate their relative importance. The highest scoring \most recommended" KO is calculated based on this model, combining all feature weights to an overall score.

We use the RecScore in (19) to calculate the suitability of a learning object for a certain learner in a certain context. Linear ranking functions de ne the aggregated score of ranking predicates as a weighted sum. In this case, the weightings need to be de ned a-priori by the tutor. The basis for ranking is provided by 1. Degree of Match. Parameter d is used to de ne when constraints given as a key value pairs match. For instance, a Portuguese KO can be recommended to a native speaker of Spanish because it tends to be comprehensible, even though it does not match the learner pro le perfectly.

2. Weights describing the importance of a feature. Di erent weights can

be assigned (by the tutor) to individual features, re ecting their importance with respect to all other feature constraints. For instance, the DF "gender" in most pedagogical frameworks seems to be of minor importance for recommendations.

The recommendation score of a learning object and formula used for ranking is thus:

RecScore(KOi) = (19) n X w(k)d(i; k) k=1 where { w(k) is the weight of feature k, and thus its contribution to the nal result. { d(i; k) is the matching degree of the feature k, represented by a oating-point value ranging from 0 to 1.

{ n is the number of DF features.

Accordingly, the results that best match the axiom ( guring the learners' needs) are ranked higher in the list. The ranking algorithm takes into account the weight values for each feature and looks for the highest overall score, so that the number of features that are satis ed can still be low, if they are assigned a high overall weight. Since the weight values are subjective, we allow to manually edit them. In our framework, all DF features are independent of each other. 4

Implementation and Validation

In this section, we present implementation details of our hybrid recommendation framework. A message-oriented architecture, illustrated in Fig. 1. is used to implement an asynchronous publication process over a set of loosely coupled software components which get active in reaction to user actions. The di erent ontology modules make up a central part of the application logic.

At rst, the learner status is analyzed and the learner state ontology is instantiated accordingly. Information about the course becomes available via the Cognitive Map and the Content Map10

Then, the reasoning module is invoked to o er enhanced learner adaptation based on the \recommendation axioms ontology" which comprises all class expressions and axioms from Sect. 3.4. Since this can partially be done in parallel, e. g., when evaluating soft criteria, we utilise a reasoning broker framework [ 6 ] 10 For the design of the curricula a speci c authoring tool has been developed. Thus, course annotation is supported by a user-friendly visual editor, that strictly adheres to our OWL modelling approach, used as an intermediate exchange format. which comprises several standard OWL reasoners (FaCT++, Pellet or HermiT), and using the latest standards for OWL access protocols (OWLAPI, OWLLink).

Finally, in the ranking stage, the various result sets from the reasoning stage are taken to compute a recommendation score as presented in Sect. 3.5.

A curriculum of 125 hours from the domain of Philosophy of Didactics, comprising 103 CC s, 1133 KO s, and 4 macro- and 2 micro LPs, has been modelled in adherence to our semantic model to test the functionality of our approach. It can be adopted to di erent pedagogical strategies and is highly adjustable, e.g. allowing to con gure individual DF weights and recommendation axioms. This is important because a didactical theory seeks to investigate how learning works, and is never xed from the start.

In the near future we plan to conduct an integrative testing in real settings to assess a) the contribution to the learning experience, b) the learner's satisfaction levels with the system, and c) the usefulness from the learner point of view. 5

Conclusion

We have presented a novel approach to personalized learning based on Semantic Web technologies, aiming to optimise the individual learning experience.

We show how OWL can be extended to cope with some inherent limits. In particular, we o er a formalization of a learning pathway as a structured sequence which can be used for dynamic couseware generation based on more abstract classes, e. g., KT s and MT s. Moreover, we present a solution to deal with complex conjunctive queries and show how to incorporate soft constraints: Di erent result sets delivered by the OWL reasoner are combined and ranked according to a speci c weighting scheme in a post-processing step.

We plan to extend our framework by a dialogue module that provides metacognitive feedback in terms of motivational messages and explanations why a speci c recommendation was given. Since the feedback messages are generated together with the reasoning/inferencing, the same justi cations for the conclusions can be drawn. 6

Acknowledgements

The research leading to these results has received funding from the EC's 7th Framework Programme (FP7/2007-2013) under grant agreement N 318496.

References

1. Aleman-Meza , B. , Halaschek-Weiner , C. , Arpinar , I.B. , Ramakrishnan , C. , Sheth , A.P. : Ranking complex relationships on the semantic web . Internet Computing, IEEE 9 ( 3 ), 37 { 44 ( 2005 )

2. Alian , M. , Jabri , R.: A Shortest Adaptive Learning Path in eLearning Systems: Mathematical View. Journal of American Science 5 ( 6 ), 32 { 42 ( 2009 )

3. Anyanwu , K. , Maduko , A. , Sheth , A. : Semrank: ranking complex relationship search results on the semantic web . In: Proc. of the 14th Intl. Conf. on World Wide Web . pp. 117 { 127 . ACM ( 2005 )

4. Aroyo , L. , Dolog , P. , Houben , G.J. , Kravcik , M. , Naeve , A. , Nilsson , M. , Wild , F. : Interoperability in personalized adaptive learning . Educational Technology & Society 9 ( 2 ), 4 { 18 ( 2006 )

5. Barros , H. , Silva , A. , Costa , E. , Bittencourt , I.I. , Holanda , O. , Sales , L. : Steps, techniques, and technologies for the development of intelligent applications based on semantic web services: A case study in e-learning systems . Engineering Applications of Arti cial Intelligence 24 ( 8 ), 1355 { 1367 ( 2011 )

6. Bock , J. , Tserendorj , T. , Xu , Y. , Wissmann , J. , Grimm , S.: A Reasoning Broker Framework for OWL . In: Proc. of the 6th Intl. Workshop on OWL: Experiences and Directions (OWLED) . vol. 529 . CEUR Workshop Proceedings ( 2009 )

7. Brusilovsky , P. , Millan , E. : User models for adaptive hypermedia and adaptive educational systems . In: The adaptive web . pp. 3 { 53 . Springer ( 2007 )

8. Devedzic , V.: Education and the Semantic Web . Intl. Journal of Arti cial Intelligence in Education 14 ( 2 ), 165 { 191 ( 2004 )

9. Dolog , P. , Henze , N. , Nejdl , W. , Sintek , M.: Personalization in distributed elearning environments . In: Proc. of the 13th international World Wide Web conference on Alternate track papers & posters . pp. 170 { 179 . ACM ( 2004 )

10. Drummond , N. , Rector , A.L. , Stevens , R. , Moulton , G. , Horridge , M. , Wang , H. , Seidenberg , J.: Putting OWL in Order: Patterns for Sequences in OWL . In: Proc. of the Workshop on OWL: Experiences and Directions (OWLED 2006 ) ( 2006 )

11. Gaeta , M. , Orciuoli , F. , Paolozzi , S. , Salerno , S. : Ontology Extraction for Knowledge Reuse: The e-Learning Perspective . IEEE Transactions on Systems Man and Cybernetics|Part A-Systems and Humans 41 ( 4 ), 798 { 809 ( 2011 )

12. Gaeta , M. , Orciuoli , F. , Ritrovato , P. : Advanced ontology management system for personalised e-Learning . Knowledge-Based Systems 22(4) , 292 { 301 ( 2009 )

13. Gueye , B. , Rigaux , P. , Spyratos , N.: Taxonomy-based annotation of xml documents: Application to elearning resources . In: Methods and Applications of Arti - cial Intelligence , pp. 33 { 42 . Springer ( 2004 )

14. Henze , N. , Dolog , P. , Nejdl , W. : Reasoning and Ontologies for Personalized ELearning . Educational Technology & Society 7 ( 4 ), 82 { 97 ( 2004 )

15. Henze , N. , Nejdl , W. : Logically Characterizing Adaptive Educational Hypermedia Systems . In: Proc. of the Workshop on Adaptive Hypermedia and Adaptive WebBased Systems (AH2003) ( 2003 )

16. Janssen , J. , Berlanga , A. , Vogten , H. , Koper , R.: Towards a learning path speci - cation . Intl. Journal of Continuing Engineering Education and Life-Long Learning 18(1) , 77 { 97 ( 2008 )

17. Karampiperis , P. , Sampson , D.G. : Adaptive instructional planning using ontologies . In: ICALT . vol. 4 , pp. 126 { 1 ( 2004 )

18. Kasneci , G. , Suchanek , F.M. , Ifrim , G. , Ramanath , M. , Weikum , G.: Naga: Searching and ranking knowledge . In: 24th Intl. Conf. on Data Engineering , ICDE 2008 . pp. 953 { 962 . IEEE ( 2008 )

19. Keet , C.M.: Ontology engineering with rough concepts and instances . In: Knowledge Engineering and Management by the Masses , pp. 503 { 513 . Springer ( 2010 )

20. Kerkiri , T. , Manitsaris , A. , Mavridou , A. : Reputation Metadata for Recommending Personalized e-Learning Resources . In: Proc. of the 2nd Intl. Workshop on Semantic Media Adaptation and Personalization ( 2007 )

21. Lukasiewicz , T. , Straccia , U. : Top-k retrieval in description logic programs under vagueness for the semantic web . In: Scalable Uncertainty Management , pp. 16 { 30 . Springer ( 2007 )

22. Magliacane , S. , Bozzon , A. , Della Valle , E.: E cient execution of top-k sparql queries . In: The Semantic Web{ISWC 2012 , pp. 344 { 360 . Springer ( 2012 )

23. Markellou , P. , Mousourouli , I. , Spiros , S. , Tsakalidis , A. : Using semantic web mining technologies for personalized e-learning experiences . Proc. of the web-based education pp. 461 { 826 ( 2005 )

24. Min , W.X. , Wei , C. , Lei , C. : Research of Ontology-based Adaptive Learning System . In: Proc. of the Intl. Symposium on Computational Intelligence and Design (ISCID '08) . vol. 2 , pp. 366 { 370 ( 2008 )

25. Schockaert , S. , De Cock , M. , Kerre , E.E. : Reasoning about fuzzy temporal information from the web: towards retrieval of historical events . Soft Computing 14 ( 8 ), 869 { 886 ( 2010 )

26. Shaw , C.J. : System design and architecture of an online, adaptive, and personalized learning platform . Ph.D. thesis , Massachusetts Institute of Technology ( 2013 )

27. Shen , L.P. , Shen , R.M. : Ontology-based learning content recommendation . Intl. Journal of Continuing Engineering Education and Life Long Learning 15 ( 3 {6), 308 { 317 ( 2005 )

28. Stojanovic , L. , Staab , S. , Studer , R.: eLearning based on the Semantic Web . In: Proc. of the World Conference on the WWW and Internet (WebNet2001) ( 2001 )

29. Swertz , C. , Schmolz, A. , Forstner , A. , Heberle , F. , Henning , P. , Streicher , A. , Bargel , B.A. , Bock , J. , Zander , S.: A Pedagogical Ontology as a Playground in Adaptive Elearning Environments . In: Proc. of the 7th Intl. Workshop on Applications of Semantic Technologies (AST) ( 2013 )

30. Yu , Z. , Nakamura , Y. , Jang , S. , Kajita , S. , Mase , K. : Ontology-Based Semantic Recommendation for Context-Aware E-Learning . In: Proc. of the 4th Intl. Conf. on Ubiquitous Intelligence and Computing. LNCS , vol. 4611 . Springer ( 2007 )