Proactive Mobile Learning on the Semantic Web

Rachid Benlamri

, Jawad Berri

and Xiaoyun Zhang

Department of Software Engineering, Lakehead University

955 Oliver Rd, Thunder Bay, Ontario, P7B 5E1, Canada

Department of Computer Engineering, Etisalat University College

PO Box 573 Sharjah, United Arab Emirates

Abstract. Flexible and personalized instruction is one of the most important re-

quirements to next generation intelligent educational systems. The intelligence

of any e-learning system is thus measured by its ability to sense, aggregate and

use, the various contextual elements to characterize the learner, and to react ac-

cordingly by providing a set of customized learning services. In this paper we

propose a proactive context aware mobile learning system on the Semantic

Web. The contribution of this work is a combined model using both a probabil-

istic learning technique and an ontology-based approach to enable intelligent

context processing and management. The system uses a Naïve Bayesian classi-

fier to recognize high level contexts in terms of their constituent atomic context

elements. Recognized contexts are then interpreted as triggers of actions yield-

ing a Web service composition. This is achieved by reasoning on the ontologi-

cal description of atomic context elements participating in the high level con-

text.

1 Introduction

Research work in the field of mobile learning [1][2][3][4][5] has shown that the edu-

cational potential of mobile technologies is driven by the continuing expansion of

broadband wireless networks and the capacity of the new generation of cellular

phones. However, the utilization of these technologies for educational purposes has

been sparsely explored and many problems related to: context acquisition and man-

agement, conceptual knowledge modeling for personalized instruction, and adaptive

information discovery remain unresolved. This paper contributes towards this direc-

tion, aiming at using the evolving semantic web and mobile computing to enable

context-aware learning which delivers adaptive instructional resources on a learner’s

schedule. Context-aware learning is a critical support mechanism for educational

institutions and organizations to compete in the new economy. Today’s global market

requires adaptive, fast, just-in time, and relevant learning processes that can be initi-

ated by user profiles and business demands [6].

In this paper we propose an integrated approach to context modeling and reasoning

based on Naïve Bayesian classifiers and ontological structures. First, higher-level

contexts are recognized using a Naïve Bayesian classifier. Then, ontology-based

Benlamri R., Berri J. and Zhang X. (2007).

Proactive Mobile Learning on the Semantic Web.

In Proceedings of the 1st International Joint Workshop on Wireless Ubiquitous Computing, pages 63-73

DOI: 10.5220/0002434400630073

 SciTePress

reasoning with the recognized contexts triggers actions yielding Web service compo-

sition that are customized to learner’s context, needs, and preferences. The contextual

information used in the personalization process encompasses all elements that charac-

terize the learner’s interaction, task at hand, the resources on which the Web services

are to be performed, and surrounding environment.

The remaining of the paper is organized as follows. Section 2 describes back-

ground knowledge and related work. Section 3 describes context representation and

modeling schemes. In section 4, we describe the higher-level context recognition

process. Section 5 presents the framework for ontology reasoning and Web services

composition to generate adaptive learning services. Finally, conclusions are drawn

and further research work is suggested.

2 Background and Related Work

A considerable amount of research in knowledge-based and intelligent e-learning

systems is now moving towards ontology-based context acquisition and management

for personalized learning [7][8][9][10]. The main issues and challenges are however

related to the ability of such systems to model and consistently reason with high level

contexts at the semantic level. Although, some research attempts were made to solve

some of these problems [9][10][11][12], the shortcoming of most of these efforts is

their limitations to specific context elements and specific learning scenarios. General-

purpose modeling and reasoning with context is a complex problem, and much re-

search work is needed before achieving any real progress in this field. Most devel-

oped learning systems restrict the use of ontology relations and rules to describing

and adapting content and sequencing of learning material according to some sensed

context. However, little contextual semantics has been embedded in the ontology

itself.

Other approaches to context modeling have also been considered. McCalla [13]

has introduced an approach to learning design where learners’ models are attached to

Learning Objects (LOs) they interact with, and useful learning patterns are then de-

rived by mining those models. The problem with this approach is its limitation to

context that can be inferred from the learner’s profile only, ignoring other type of

context. Stojanovic et al. [6] however, have extended ontology usage to describe

content, context, and sequence of learning material. Content-ontology was used for

checking consistency as well as searching and navigating repositories of LOs. Con-

text-ontology was used to present learning material in various learning contexts.

However, learning style ontology was used to describe the way knowledge can be

dynamically connected to adapt to learners’ cognitive needs and preferences. Sets of

relations, rules and axioms have been separately defined for each type of adaptation.

The shortcoming of this approach is that efficient modeling of mobile learning sce-

narios would require the definition of atomic context elements at the semantic level

and the use of the various ontologies in an orthogonal way. This is due to the fact that

context, content, and learning styles are semantically inter-related aspects of cognitive

learning [14]. This paper explores such a new dimension. The emphasis is on context

discovery and its semantic modeling and management. Mobile users equipped with

wireless devices go through several contextual changes as they move around in physi-

cal and social surroundings. These contextual changes could be used to drive ontol-

ogy navigation and reasoning for better modeling of mobile learning scenarios.

Another challenging aspect addressed in this paper is automation of metadata gen-

eration for mobile learning. Metadata provides a common set of tags for describing,

indexing, searching, and reusing learning materials on the Web in an interoperable

way [14]. However, it is really difficult to create and maintain metadata rich enough

to meet the diverse and ever changing needs of potential mobile learners. Mobile

learning requires additional metadata to capture context. In this study, an attempt is

made to solve this problem by defining contextual information at three hierarchical

levels – atomic context – composite context – and higher-level context. Atomic con-

text elements are sensed from the learner’s interaction, task at hand, the used mobile-

device, and the surrounding environment. These are then grouped into four composite

context classes – learner context – activity context – device context and environment

context. Composite contexts are further aggregated to build meaningful time-stamped

higher-level contexts which are matched against context classes describing typical

learning scenarios. Context classes are simply built from previously sensed similar

higher-level contexts that have exhibited high degree of confidence. Matching higher-

level contexts against these context classes is performed using a Naïve Bayesian clas-

sifier. The Naïve Bayesian classifier technique is used to cope with the uncertainty

embedded in most sensed atomic contextual elements. Recognized contexts are then

interpreted as triggers of actions that are translated into Web service compositions.

This is achieved through ontological descriptions and reasoning with higher-level

context.

Fig. 1 describes the overall system architecture which consists of four main com-

ponents – context acquisition and aggregation – context recognizer – ontology rea-

soning engine – and Web-service composer. The context acquisition and aggregation

component controls the user’s interaction with the system and senses atomic context

information from different sources. These are then aggregated into domain related

contexts. Mobile learners go through continuous contextual changes as they move in

their environments. It is the context acquisition and aggregator’s job to communicate

and update such changes yielding new contexts. The context recognizer identifies the

aggregated contexts by matching them against well defined context classes stored in a

context repository. The recognition process is performed using a Naïve Bayesian

classifier. The context recognizer also allows for newly formed context classes to be

added to the context repository.

The third component of the system is an ontology reasoning engine which uses the

recognized higher-level contexts to customize learning services. Two ontologies are

used to perform such a task – device/environment ontology – and domain ontology.

The former is used to generate metadata that is used to discover Web-services that

can run in the learner’s device/network environment. However, the later is used to

customize the learning content and the learning sequence according to the learner’s

current activity, background and preferences. This requires an ontological description

and interpretation of higher-level contexts in terms of their constituent atomic context

elements. Finally, the Web-service composer uses the generated device/environment

metadata and the inferred learning concepts’ sequence to compose Web-services

accordingly.

User

Context

Device

Context

Environment

Context

Activity

Context

Context Aggregator

Device/

Environment

Ontology

…

Domain

Ontology

…

Web Services

Agents

Context Recognizer

Context

Repository

Device/

Environment

Repository

Web-Services

Composer

Repositories

Learner Profile

Repository

Ontology

Reasoning

Fig. 1. System Architecture.

3 Context Acquisition and Aggregation

Contextual information used in this study is defined at three hierarchical levels –

atomic context – composite context – and higher-level context. At the lower level,

atomic contextual elements consist of the basic information describing the learner’s

profile, the current learner’s activity, the used mobile device, and the surrounding

environment. These can be either direct or indirect atomic contextual elements. Direct

atomic contextual elements are those that can be directly sensed from the user interac-

tion with the system and may originate from different sources such as the used device

(i.e. device type, communication protocol), the task at hand (i.e. current learner’s

activity), and the surrounding environment (i.e. location, time, wireless network,

network security). Indirect atomic contextual elements are however those elements

that can be indirectly inferred from the direct atomic context elements. Inference of

indirect atomic context elements is performed by the context aggregator relying on

the device/environment repository and the learner profile repository. For instance,

information such as device’s operating system, device memory, and screen resolution

of a specific mobile device which is previously stored in a device repository can be

inferred using the atomic context element device-type. Similarly, other information

related to the learner’s pre-requisite knowledge, previously accessed services, and

learner’s preferences can be inferred from the learner profile repository. The use of

indirect contextual elements aims at reducing the amount of contextual information

that has to be sensed from the learner’s interaction, device, and surrounding environ-

ment, which significantly speedup the context recognition process.

An atomic context element c

is defined by:

),(

ipivi

ccc

(1)

where

c is the context value, and

c is the probability of context

c of value

c being part of a higher-level context. The context value

c , as shown below, can

be either a specific value (i.e. device type, learner identifier), a binary value (i.e.

whether the used device is browser-enabled or not, secured/non-secured wireless

network), or a value within a predefined range (i.e. network bandwidth, screen resolu-

tion).

(2)

Composite contextual elements are aggregates of atomic context elements describ-

ing a specific context type. There are four context types –

learner context –

de-

vice context –

c environment context – and

c activity context. Each of which is

defined by:

(3)

Finally, higher-level contexts consist of four-tuples

tAEDLt

ccccC ),,,(

which

are built out of configurations of composite context elements sensed at time t and

which characterize typical learning scenarios in a specific domain. Classes of higher-

level contexts are defined at the ontological level in that they can be interpreted di-

rectly as triggers of learning actions implemented as Web service compositions.

4 Context Recognition

While ontologies have the ability to communicate context information by naming

different concepts in machine readable fashion and allowing for the use of everyday

words and concepts when interacting with the technology, they are unable to effi-

ciently recognize learners’ context. This is because the mapping between the defined

concepts and the sensed real world atomic context elements is not so straightforward

due to the uncertainty embedded in some atomic context elements. The mapping fails

because ontologies do not handle uncertainty. They rather rely on well defined logic

which assumes all information required to make a logical decision is available and

produces either true, false or undeterminable statements. Uncertainty on the other

hand produces similar statements but with degrees of truth or falseness [15]. To cope

with uncertainty, higher-level contexts are recognized using a Naïve Bayesian Classi-

fier. Bayesian Classification is a probabilistic learning technique where prior knowl-

⎪

⎩

⎪

⎨

⎧

∈

−

]2..1[ vvvalue

valuebinary

valuespecific

⎭

⎬

⎫

⎩

⎨

⎧

∑∑

−

indirectj

directicomposite

ccc

edge can be combined with observed data. The aim is to recognize a currently ob-

served context state against a set of learned context classes. The input to the classifier

is thus a set of sensed/observed atomic context elements which describe the user’s

context at a given instant of time, while the output is a learned context class. The

classification process is thus performed with no user intervention or understanding

required. The Bayesian classification also makes the implicit assumption that the data

being handled is noisy and can tolerate any missing pieces of information. One differ-

ence between the Bayesian classification and the ontology approach is that once the

ontology is defined then it can be available immediately whereas in the Bayesian

classification approach each context has to be experienced at least once before being

recognized again [15].

Let X be a current context whose class label is unknown, and let H be a hypothesis

that X belongs to context class C, the classification problem consists of determining

P(H/X) that is the probability that the hypothesis holds given the observed context X.

This is defined by:

(4)

where :

• P(H) is the prior probability of hypothesis H (i.e. the initial probability be-

fore we sense the current context and reflects the background knowledge).

• P(X) is the probability associated to the current context.

• P(X|H) is the probability of observing the context X, given that the hypothe-

sis holds.

The above Bayesian model assumes that the observed context elements are re-

lated and depend on each other, and therefore, requires initial knowledge of many

probabilities, as well as, significant computational cost. However, since most sensed

atomic context elements are independent, the above model can be further simplified

by applying the Naïve Bayesian classifier which is defined by:

(5)

Where: C

is a context class, and the set of x

s are the atomic context elements

forming the higher-level context X as defined in section 3.1.

The Naïve Bayesian classifier greatly reduces the complexity of the model, as well

as its computational requirements. The context recognition problem is solved by as-

signing the current context X to the class C

that satisfies the following condition:

(6)

where m is the number of recognized context classes.

Fig. 2 describes the context acquisition and recognition cycle. First, direct-atomic

context elements are sensed, these are then used to infer related indirect-context ele-

ments. Next, the Naïve Bayesian classifier is applied to recognize the associated

higher-level context-class, and finally, changes to the learner’s context are sensed and

a new context recognition cycle is performed. It should be noted here that the context-

)

()|(

CXP

∏

)}()({)(

..1

CPCXPMaxCXP

⋅=

∈

)(

)()|(

)|(

XHP =

change detection process significantly speedup the recognition time of successive

high level contexts. This is because we just infer the indirect-atomic contexts of those

context elements that have undergone some changes. The subset of newly observed

context elements designated by C

changes

is defined by:

(7)

where “\” means set subtraction.

direct-atomic

context

acquisition

indirect-

atomic

context

inference

high-level

context

recognition

Context-

change

detection

Surrounding

Environment

device/env.

repository

learning-style

repository

learner

repository

high-level

context

repository

nomadic

learners

Fig. 2. Context Acquisition and Recognition Cycle.

5 Ontology Reasoning and Web Service Composition

Recognized higher-level contexts are fed to the ontology reasoning engine in order to

customize learning services based on the learner’s context, preferences and back-

ground. Reasoning with recognized higher-level contexts is performed using the two

ontologies – device/environment ontology – and domain ontology. A set of ontologi-

cal rules are applied to the device/environment ontology to infer the computing re-

sources and the operational environment features compatible with the used mobile

device and its surrounding environment. We call this process, context-driven re-

sources adaptation. The output of this reasoning process is a set of metadata that will

help discovering the Web services that can run into such an operational environment.

The inference rules that are built around the domain ontology however are used to

provide the learner with a learning sequence and content tailored to his/her current

activity, previous background and preferences.

The two ontologies are coded in the Web Language Ontology – OWL; and the in-

ference engine is implemented in Rule Markup Language – RuleML. Metadata de-

rived from the ontology reasoning process is compliant with the IEEE-LTSC Learn-

ing Object Metadata (LOM) specification which is coded in XML. In particular, the

XML description of both the inferred learning concepts and the device-related opera-

tional environment are used for Web services discovery. However, the inferred learn-

ing sequence which we call in this paper domain-context (i.e. the order of learning

concepts inferred using the properties and relations between the domain-ontology’s

classes [16]) is used for Web service composition. This is described in OWL-S. A

tAEDLtAEDLchanges

ccccccccC ),,,(\),,,(

domain context represents a control structure that makes it possible to adapt the do-

main knowledge to a particular higher-level context. This adaptation is facilitated by

the ontology

for a given domain

is defined by

),(

MMM

RCoO

, where

},...,{

1 nM

cocoCo =

is a set of concepts and

{ ,..., }

is an ordered set of rules

defined as follows:

)...()...(

1 lkrj

cocoqcocop

→

, where

and

are predicates

reflecting respectively the factual information and the resulting one based on the

inferential rule

The semantic of the ontological links is obtained by the rules in R

. These rules are

prioritized to reflect their importance or abstraction levels in a given knowledge tax-

onomy. For example, if the sensed higher-level context reflects a time-constrained

learning scenario, one would like to focus only on say “the necessary-part-of” rules of

the ontology to get a quick abstraction on the general structure of the requested

knowledge. In a less time-stringent learning scenario however, this abstraction could

further include the “part-of”, and/or “case-study” rules, etc. These knowledge-

supporting rules generate additional concepts of the ontology in multi-level clusters

which are used to infer a progressive knowledge based on the learners’ context de-

noted by C

and the activity context denoted by C

as described in section 3.

A software agent as shown in Fig. 1 is spawned at the server side to supervise a

learning session for each learner. The agent typically represents the learner on the

Semantic Web. The agent successively invokes the inference engine to get the current

learner’s focus, then discovers, composes, and invokes the chosen Web services ac-

cordingly.

To illustrate the main functions provided by our framework, we provide the fol-

lowing example ontologies describing a C++ programming course as a domain ontol-

ogy, and a device/environment ontology. These are shown in Fig. 3 and Fig. 4 respec-

tively.

C++ Programming 1

Prog. & PS 2

Program

Computer

Part-of Relation

Necessary Part-of Relation

Concept

Prerequisite Relation

i: Concept Id

Is-a Relation

Prog. Language 10

PS Techniques 11

Means-Ends Analysis 13

Solve by Analogy 12

Divide & Conquer 14

Building Blocks App. 15

Merging Solutions 16

Algorithmic PS 17

Program Development Process 3

Program I/O 4

Interactive I/O

File I/O

Input Failure

Output Formatting 25

Selection 5

Loop Design 37

Control Flow Design 38

Loop Exection 36

Event-Controlled Loops 39

Looping Subtasks 40

Looping 6

C++ Loops 41

Nested Loops 45

While 42

Do-While 43

For 44

User-Defined Fcts 47

Function Call 49

Functional Decomposition 46

Declarations and Defs. 50

Local Variables 51

Functions 7

Void Functions 48

Parameters 54

Designing Functions 55

The Return Statement 52

Header Files 53

Scope of Identifiers 56

Value-Returning Fcts. 57

Conditions 27

If Statement 33

Control Flow 26

Nested If Stat. 34

Switch Stat. 35

Logical Expr. 28

Boolean Data 29

Relational Op. 30

Logical Op. 31

Ops. Precedence 32

C++ Programs

Program Construction

Program Execution

Testing & Debugging

Fig. 3. Ontology for C++ Programming Course.

Mobile Devices

Hardware

Screen

Keyboard

Navigation device

Memory

Software

Customization

Normal Screen

Touch Screen

Size

Colors

Virtual

Physical

Scroll Wheel

Thumb Wheel

Stylus

Size

Speed

Operating System

Palm OS

Windows Mobile

RIM

GPE

OPIE

Windows

Speakers

Connectivity

Networks

GSM 900

GSM 1800

GSM 1900

EGSM

Applications

Bluetooth

Wimax

WAP

Personal Inf. Manager

Multimedia

Wi Fi

Video Player

Audio player

SW Applications

Devices

Memory Slot

USB

Infrared

WLAN

Device-Type

PDA

Laptop

Pocket PC

Cell Phone

Programming

J2ME

Symbian

BREW

Televison

E-Book

Symbian OS

Linux

GPRS

3 4

Part-of Relation

Necessary Part-of Relation

Concept

Prerequisite Relation

i: Concept Id

Is-a Relation

Fig. 4. Device/Environment Ontology.

A fragment of the ontology shown in Fig. 3, describing concept 3 “Program De-

velopment Process”, is described in OWL in Fig. 5. The OWL definition of the se-

mantics of the different relationships used in the C++ programming ontology is also

given in Fig. 5.

Details about the rules used by the ontology reasoning engine to customize the

learning sequence can be found in our previous work [17].

Fig. 5. Fragments of OWL description of the C++ Programming ontology.

6 Conclusions

In this paper, we proposed a proactive mobile-learning system on the Semantic Web.

We argued that a probabilistic learning model is more suitable that an ontology-based

approach for context recognition. This is mainly due to uncertainty embedded in

some atomic contextual information. Higher-level recognized contexts are however

described at the semantic level using ontology rules and axioms. The ontology rea-

soning process allows the system to react to any observed contextual changes by

interpreting the newly sensed contexts as triggers of actions yielding a Web service

composition. We are currently implementing a prototype of our framework as part of

our personalized-learning provision project.

<owl:ObjectProperty rdf:ID="NecessaryPartOf">

<rdf:type rdf:resource="&owl;TransitiveProperty"/>

<owl:inverseOf rdf:resource="#hasNecessaryPart"/>

</owl:ObjectProperty>

<owl:ObjectProperty rdf:ID="isPartOf">

<rdf:type rdf:resource="&owl;TransitiveProperty"/>

<owl:inverseOf rdf:resource="#hasPart"/>

</owl:ObjectProperty>

<owl:ObjectProperty rdf:ID="is-a">

<rdf:type rdf:resource="&owl;TransitiveProperty"/>

<owl:inverseOf rdf:resource="#has"/>

</owl:ObjectProperty>

<owl:ObjectProperty rdf:ID="isPrerequisiteOf">

<rdf:type rdf:resource="&owl;TransitiveProperty"/>

<owl:inverseOf rdf:resource="#hasPrerequisite"/>

<owl:Class rdf:ID="Program Development Precess_3">

<rdfs:subClassOf rdf:resource="#C++ Programming_1"/>

<owl:disjointWith rdf:resource="#Prog and PS_2"/>

<owl:disjointWith rdf:resource="#Program I/O_4"/>

<owl:disjointWith rdf:resource="#Selection_5"/>

<owl:disjointWith rdf:resource="#Looping_6"/>

<owl:disjointWith rdf:resource="#Functions_7"/>

</owl:Class>

<owl:Class rdf:ID="C++ Programs_18"/>

<rdfs:subClassOf rdf:resource="#Program Development Process_3"/>

<rdfs:subClassOf>

<owl:Restriction>

<owl:onProperty rdf:resource="#isNecessaryPartOf"/>

<owl:allValuesFrom rdf:resource="#program Development Process_3"/>

</owl:Restriction>

</rdfs:subClassOf>

<rdfs:subClassOf>

<owl:Restriction>

<owl:onProperty rdf:resource="#isPrerequisiteOf"/>

<owl:allValuesFrom rdf:resource="#Testing and Debugging_21"/>

</owl:Restriction>

</rdfs:subClassOf>

<owl:disjointWith rdf:resource="#Program Construction_19"/>

<owl:disjointWith rdf:resource="#Program Execution_20"/>

<owl:disjointWith rdf:resource="#Testing and Debugging_21"/>

</owl:Class>

Acknowledgements

This work is funded by Lakehead University and NSERC under a startup research

grant.

References

1. L. Low and M. O’Connell, “Learner-centric design of digital mobile learning”, Proc. of the

On-line Learning and Teaching Conference, Brisbane, Australia, September, (2006) 71-82.

2. M. Milrad, “Mobile Learning: Challenges, Perspectives and Reality”, Book chapter in

Kristóf Nyíri (ed.), Mobile Learning: Essays on Philosophy, Psychology and Education,

Vienna, (2003) 25-38.

3. B. Alexander, “Going Nomadic: Mobile Learning in Higher Education”, EDUCAUSE

Review, 39(5), (2004) 28–35.

4. L. Chittaro, “Visualizing Information on Mobile Devices”, IEEE Computer, March (2006)

40-45.

5. G. Vavoula and M. Sharples, KLeOS: A personal Mobile Knowledge and Learning Or-

ganisation System”, Proc. 2nd IEEE Int. Workshop on Wireless and Mobile Technologies

in Education, Sweden, (2002) 152-157.

6. L. Stojanovic, S. Staab and R. Studer, Elearning based on the semantic web, Proc. of the

World Conf. on the WWW and Internet (WebNet2001), Orlando, Florida, USA, (2001)

1174-1183.

7. R. Mizoguchi and J. Bourdeau, Using Ontological Engineering to Overcome Common AI-

ED Problems”, Int. Journal of Artificial Intelligence in Education, 11 (2), (2000) 107-121.

8. L. Aroyo and D. Dicheva, Courseware Authoring Tasks Ontology”, Proc. Int. Conf. on

Computers in Education, Los Alamitos, CA, (2002) 1319-1321.

9. J. Kay and A. Lum, Ontology based User Modeling for the Semantic Web, Proc. of the Int.

Workshop on Personalisation on the Semantic Web, Edinburgh, UK, (2005) 14-23.

10. J. Wang et al., “Personalized Knowledge Service Framework for Mobile Learning”, Pro-

ceedings of the Second Int. Conf. on Semantics, Knowledge, and Grid (SKG'06), Novem-

ber, (2006) 102-103.

11. I. Tirrellil et al., “Fractal Adaptive Web Service for Mobile Learning”, Int. Journal, of

Emerging Technologies in Learning, 1(2), (2006) 25-35.

12. P. Veith, “Conception and Development of Reusable and Modular Mobile Content”, Proc.

of Int. Conf. on Mobile Learning mLearn’2005, Cape Town, 25-28 October, (2005), Re-

trieved from <http://www.mlearn.org.za/CD/ > Dec. 2006.

13. G. McCalla, The Ecological Approach to the Design of E-learning, Environments: Pur-

pose-based Capture and Use of Information about Learners, Journal of Interactive Media in

Education, 7, (2004) 1-23.

14. L. Aroyo and D. Dicheva, The new challenges for e-Learning: The educational Sementic

Web, Educational Technology and Society, 7(4), (2004) 59-69.

15. J. Flanagan, Unsupervised clustering of symbol strings, in Intl’ Joint Conference on Neural

Networks,IJCNN’03, Portland Oregon, (2003) 3250–3255.

16. R. Benlamri, J. Berri and Y. Atif, A Framework for Ontology-aware Instructional Design

and Planning, Journal of E-Learning and Knowledge Society, 2(1), (2006) 83-96.

17 J. Berri, R. Benlamri and Y. Atif, Ontology-Based Framework for Context-aware Mobile

Learning, Proc. Int. Conf. on Wireless Communication and Mobile Computing, Vancouver,

Canada, (2006) 1307-1310.