A knowledge-based approach to describe and adapt learning objects.

Our claim is that semantic metadata are required to allow a real reusing and assembling of learning objects. Our system is based on three models used to describe the domain, learners, and learning objects. The learning object model is inspired from knowledge representation proposals. A learning object can be reused directly or can be combined with other learning objects using composition operators with well-defined semantic. Using these models we are able to define powerful search tools and an adaptive environment taking as input the learner model to construct the learning object to deliver. We are currently implementing this system using Sesame, an RDF repository.

**********

The development of technology-enhanced learning has been very high these last years. There are now numerous pedagogical materials available on the Web (the so-called learning objects--LO). An important problem is to offer tools allowing users (learners and authors) to search for existing learning objects. Authors are interested in existing objects to reuse them directly or to combine them with other objects. Learners of course want to use these objects to improve their knowledge. To facilitate the search and reuse, some standards for metadata have been proposed such as LOM (IEEE, 2002) and SCORM (Advanced Distributed Learning Initiative, 2001). Learning objects are stored into repositories such as Ariadne knowledge pool (Duval et al., 2001) or Educanext (Law, Maillet, Quemada, & Simon, 2003) which implement LOM like metadata. In our opinion these proposals are not so powerful because the description of learning objects does not include their semantic. In the SIMBAD project (SIMBAD, 2004) we are investigating semantic extensions to existing metadata standard such as LOM. This will allow us to define powerful search mechanisms, to propose formal composition operators to create new objects, to offer different learning strategies to learners and to adapt learning objects content to learners (one to one delivering).

This article is structured in the following way. In the first section "The Three-Levels Model," we present our approach based on a three-levels model (domain model, learner model, learning object model). In the next section "Learning Strategies and the Adaptive Process," we present the different learning strategies proposed to learners and the associated adaptive process. Finally we conclude and present some discussions about the manipulation of learning objects distributed across several repositories.

THE THREE-LEVELS MODEL

Our claim is that semantic metadata are required to allow a real reusing and assembling of learning objects. Our system is based on three models presented in detail in (Bouzeghoub, Carpentier, Defude and Duitama, 2003): (a) the domain model, which represents the concepts covered by the LO, (b) the learner model, which keeps the profile of learners, and (c) the learning object model, which describes LO content related to the domain model. Using this knowledge we can propose sophisticated tools for searching and browsing into the LO repository. Authors can reuse and compose existing LO using operators (such as sequence, alternative, parallel) to produce new LOs. A LO may be automatically adapted to a specific learner.

In the following we present successively the domain model, the learner model, the learning objects model, and the associated properties.

Domain Model

Our approach uses ontologies to describe the domain model. The goal of this model is to define a normalized and common referential among all learners of the system (administrator, authors, and learners). The precision level of the model defines the precision of the system; that is, if we choose a very precise domain model, the system will be able to provide a more sophisticated inference task.

We define a terminological ontology for every specific knowledge domain; it is intended to describe its most relevant concepts. This ontology defines concepts and relationships among them. We use two types of relationships: a narrower/broader relationship to support hierarchical links between concepts and a set of rhetorical relationships such as contrast or extend.

Learner Model

An adaptive e-learning system may adapt contents depending on learners' background, preferences, and goals. Our approach considers the three aspects. It uses an overlay model to maintain an evaluation of learners, and allows the learner to select preferences such as language, format, and maximal learning time. Finally, a learner selects goals from concepts of the domain model.

We describe a learner under two facets. The first one, called his/her preferences, describes facts (name, e-mail, language, ethnicity ...) and is modelled with a set of couples (attribute-value). The second one called knowledge, describes concepts known by the learner qualified with one or several roles (e.g., introduce, define, resume) and a weight (learner level for this concept-role). A learner knowledge grows dynamically when he/she acquires new concepts.

Learning Object Model

To be found and reused, a LO must be described by a set of metadata. In the LO model, we distinguish two types of metadata: the first one describes LO general characteristics (e.g., author, title, language, media) using LOM standard and the second one describes the semantic of the LO. This semantic is structured in three parts and described in the same way as software components: prerequisites are the LO inputs (what is required by the LO) whereas content and acquisition functions are its exits (what is provided by the LO).

The LO prerequisites are a set of triples (concept, role, level); the content is described with a set of couples (concept, role); the acquisition function indicates which triple (concept, role, level) will be added to the learner model if a condition of validation is satisfied.

A LO can be a set of web pages, a file, or a program (a simulator for example). We just suppose that it is a unit accessible by the way of an URL. This unit can be used independently or for composition by third parties. We have defined composition operators (SEQ, ALT, and PAR) to compose recursively LOs. A composed LO is an acyclic oriented graph where nodes are learning objects, or operators. Failure nodes can be added to the composition graph to define an alternative path if a LO is not successfully visited.

Intensional LOs have been introduced in order to support more generic and flexible LOs for authors and to increase flexibility in the adaptation process. It allows authors to define a virtual LO, which can be considered as views on actual LO.

An intensional learning object (ILO) is a composed LO whose composition graph has at least one node defined by a query instead of a specific object. In other words, an ILO can have three kinds of nodes: an operator-node (SEQ, ALT, or PAR), a LO-node and a query-node. A query-node is defined by an intensional query (IQ) specifying the condition to be satisfied by candidate LOs. An IQ is defined by:

IQ = [Q.sub.content] ^ ([Q.sub.prerequisite] v [Q.sub.educational])

[Q.sub.content]|[Q.sub.prerequisite]=([C.sub.1,1], [r.sub.1,1] ^ ... ^ [C.sub.1,k], [r.sub.1,k]) v ... v ([C.sub.n,1], [r.sub.n,1] ^ ... ^ [C.sub.n,m], [r.sub.n,m]),

where c is a concept, r a role, k, n and m [greater than or equal to] 1.

[Q.sub.educational] is a logical combination of attribute--value comparisons.

[Q.sub.content] cannot be empty but [Q.sub.prerequisite] and [Q.sub.educational] can.

The semantic of a query-node is partly defined at authoring time: it has always a content (its [Q.sub.content]) but the other parts of its description may be undefined. At delivering time, all query-nodes of an ILO will be processed. If (at least) one query-node is empty (its corresponding query returns an empty set), the ILO is undefined and cannot be delivered. If all query-nodes return some LOs, these LOs will be composed by an ALT operator (the ILO is fully instantiated). A fully instantiated ILO can be adapted and delivered as a classical LO.

Learning Objects and Learners Properties

Our models of learners and LOs allow us to define several properties (in the following L[O.sub.1], L[O.sub.2] are two learning objects and L is a learner):

* satisfaction: L satisfies L[O.sub.1] when his model includes prerequisites of L[O.sub.1]. This property is mainly used during the adaptive process (see next section);

* master: L masters L[O.sub.1] when his model includes L[O.sub.1] content;

* substitution: L[O.sub.1] may be substituted to L[O.sub.2] when L[O.sub.1] prerequisites are equals to L[O.sub.2] prerequisites;

* equivalent: L[O.sub.1] is equivalent to L[O.sub.2] when L[O.sub.2] can be substituted by L[O.sub.1] and L[O.sub.1] content is equals to L[O.sub.2] content;

* weak precedence: L[O.sub.1] weakly precedes L[O.sub.2] if L[O.sub.1] content is included inside L[O.sub.2] prerequisites; and

* strong precedence: L[O.sub.1] strongly precedes L[O.sub.2] if L[O.sub.1] content is equals to L[O.sub.2] prerequisites.

The four last properties (from substitution to strong precedence) are used to automatically classify the set of learning objects. This is the basis of our browsing tool.

LEARNING STRATEGIES AND THE ADAPTIVE PROCESS

Our approach supports two learning strategies: concept-based and goals-based learning. The latter allows learners to define their goals from the domain model, whereas the former provides guidance and helps to meet course objectives. (Duitama, Defude, Bouzeghoub and Lecocq 2005), describes in detail our vision of LO adaptation and learning strategies. This section describes the learning object model and introduces scenarios where adaptation is required. The adaptive system is materialized by combining the three levels of modeling previously described, which are the domain model, the learner model, and the learning object model.

In course-based learning strategy, a learner selects a learning object L[O.sub.j] from the learning objects repository. At authoring time, authors may have specified a LO as a composition graph (CG) of learning objects. When a LO is chosen by the learner, its composition graph is transformed to obtain a set [S.sub.1] of delivering graphs (a delivering graph is a graph without the ALT operator). This set of delivering graphs will be filtered at delivering time in order to select the "best" composition according to the learner model. This filtering process is called "the adaptive process," and is divided into several steps shown in Figure 1. First, the system builds [S.sub.2], the set of delevering graphs meeting prerequisites satisfied by the learner model. Second, the learner preferences (e.g., the type of media, the language) are applied to construct [S.sub.3]. If there are several graphs satisfying this step, the system (or the learner in an interactive process) will choose only one. If the resultant set is empty, it implies that the current learner cannot access this course because he/she has not sufficient knowledge (the system can state the missing knowledge). Finally, the selected graph is simplified, that is all the nodes having their content already known by current learner are annotated (see adaptive navigation and presentation in (Brusilovsky, 1996). Of course, if the same conceptual graph is delivered to another learner, the selected delivering graph can be different.

[FIGURE 1 OMITTED]

In goals-based learning strategy, a learner formulates a query over concepts of the domain model. The general form of this query is the following:

Q = ([C.sub.1,1], [r.sub.1,1] v ... v [C.sub.1,k], [r.sub.1,k]) ^ ... ^ ([C.sub.n,1], [r.sub.n,1] v ... v [C.sub.n,l], [r.sub.n,l]),

where c denotes a concept, r a role, k, n and l [greater than or equal to] 1.

This query is a conjunction of disjunctions of concepts and roles; where negation is not allowed. Goals-based learning process is separated in two distinct processes depending on the number of conjunctions included in the query, either single or multiple. In single goal mode, the process is the same than course-based except that the query may return an empty set. In this case, the system has to rewrite the query (using adaptation rules) to obtain a non empty set (if possible). In multiple-goals mode, the process is more complex because in some cases the query cannot be satisfied by any existing LO but may be satisfied by a composition of existing LOs. In this case the system has to dynamically construct a new LO using composition operators.

CONCLUSION AND ONGOING WORK

Our claim is that semantic metadata are required to allow a real reusing and assembling of learning objects. This semantic allows describing domain model, learner model, and LO model and provides authors and learners with powerful mechanisms to manage learning objects, concepts, and learners (e.g., browsing, querying, composing, classifying, etc.). An adaptive process has been defined allowing to adapt a specific LO to a learner considering his/her preferences and knowledge. A similar approach, integrating also pedagogical models, is proposed in (O'Keeffe, Brady, Conlan, and Wade 2006).

We are currently implementing a prototype using Sesame (Broekstra, Kampman, & Van Harmelen, 2001) that will allow us to validate our approach. Sesame offers a storage layer for RDF statements and RDFS. SeRQL query language is the only reasoning layer of Sesame. This language has interesting capabilities but lacks in expressivity to handle all type of queries we are interested in. RDF allows us to easily support our three models see (Bouzeghoub, Ammour, Defude, Duitama, and Lecocq, 2004), for a detailed presentation of our RDF mappings. DAML+OIL will be a better candidate to support our approach but existing tools are not so mature.

For the moment our proposal is based on a centralized architecture as we suppose that the domain model, the learners, and all LOs are defined and stored on the same system. Of course this is a very restrictive view. We are currently investigating different distributed architectures for our system. One possible solution is to use SQI (Simon, Massart, & Duval, 2004) to allow access in both directions between our repository and other ones. Prototypes have already been developed to federate existing learning repositories (Massart, 2006; Ternier & Duval, 2005). The problem is that SQI is just a language neutral API, that is, it does not resolve problems of heterogeneity between repositories (e.g., at the metadata level). Another solution consists in adapting a mediation architecture. Mediation is widely used in distributed database systems to allow access to distributed and heterogeneous data sources. In this architecture, the mediator implements a generic view of the system it exposes to users. Users send queries to the mediator using a generic query language. The mediator uses information about the different data sources to optimize and split the query into subqueries. These subqueries are then sent to the data sources. An adapter is used at each data source to transform subqueries into queries processable by the source. It is also used to transform results into the mediator model. We propose to use our approach to construct such a mediator. A new model is introduced to describe repositories capabilities in terms of query language, metadata model, and so on. We suppose that all repositories use LOM as their metadata model (or some LOM extensions). Adapters will define the mappings between our model and a repository model (the problem is simpler because we know that both models are LOM extensions). The problem is much more complex if the mediator implements the adaptive process. In this case, we suppose that all repositories describe semantic information about learners and LOs and that the mediator is able to construct mappings between these models.

References

Advanced Distributed Learning Initiative (2001). Sharable Content Object Reference Model. The SCORM Content Aggregation Model. Version 1.2 Retrieved October 3, 2005 from http://www.adlnet.org/

Bouzeghoub, A., Ammour, S., Defude, B., Duitama, JF., & Lecocq, C. (2004). An RDF description model for manipulating learning objects. Proceedings of the International Conference on Advanced Learning Technologies, Joensuu, Finland.

Bouzeghoub, A., Carpentier, C., Defude, B., & Duitama, JF. (2003). A model of reusable educational components for the generation of adaptive courses. Proceedings of the First International Workshop on Semantic Web for Web-Based Learning in conjunction with CAISE'03 Conference, Klagenfurt, Austria.

Broekstra, J., Kampman, A., & Van Harmelen, F. (2001). Sesame: An architecture for storing and querying RDF data and schema information. In D. Fensel, J. Hendler, H. Lieberman, & W. Wahlster (Eds.), Semantics for the WWW.Cambridge, MA: MIT Press.

Brusilovsky, P. (1996). Methods and techniques of adaptive hypermedia. User Modeling and User Adapted Interaction, 6(2-3), 87-129.

Duitama, J.F., Defude, B., Bouzeghoub, A., & Lecocq, C. (2005). A framework for the generation of adaptive courses based on semantic metadata. Multimedia Tools and Applications, 25(3), 377-390.

Duval, E., Warkentyne, K., Haenni, F., Forte, E., Cardinaels, K., Verhoeven, B., et al. (2001). The ARIADNE knowledge pool system. Communication of the ACM, 44(5), 72-78.

IEEE (2002). Draft standards for learning object metadata (IEEE P1484. 12.1.) Retrieved October 3, 2005, from http://ieeeltsc.org/

Law, E., Maillet, K., Quemada, J., & Simon, B. (2003). EducaNext: A service for knowledge sharing. Proceedings of the 3rd Annual ARIADNE Conference. Leuven, Belgium: ARIADNE Foundation.

Massart, D. (2006). Accessing learning contents using a "simple query interface" adapter. International Journal on E-Learning, 5(1).

O'Keeffe, I., Brady, A., Conlan, O., & Wade, V. (2006) Just-in-time generation of pedagogically sound, context sensitive personalized learning experiences. International Journal on E-Learning, 5(1).

SIMBAD (2004). Semantic Interoperability for coMponent-Based and ADaptive resources. Retrieved October 3, 2005, from http://www-inf.int-evry.fr/SIMBAD/English

Simon, B., Massart, D., & Duval, E. (2004). Simple query interface (SQI) for learning repositories. Retrieved October 3, 2005, from http://nm.wu-wien.ac.at/e-learning/interoperability

Ternier, S., & Duval E. (2005). Interoperability of repositories: The simple query interface in ARIADNE. ProLearn-iClass Thematic Workshop, Leuven

AMEL BOUZEGHOUB, BRUNO DEFUDE, JOHN FREDDY DUITAMA, AND CLAIRE LECOCQ

Groupe des Ecoles des Telecommunications/Institut National des Telecommunications, France

[e-mail]

[e-mail]

[e-mail]

[e-mail]