PERSO: towards an adaptive e-learning system.

In today's information technology society, members are increasingly required to be up to date on new technologies, particularly for computers, regardless of their background social situation. In this context, our aim is to design and develop an adaptive hypermedia e-learning system, called PERSO (PERSOnalizing e-learning system), where learners with different learning goals and different learning aptitudes are treated differently, by building a model of knowledge and preferences for each of them. This model is used to propose to the learner a personalized course fitting his/her needs. We focus on both the way information about the learner is collected and the way a hypermedia course composed of several learning object data are updated. We consider two types of adaptation: adaptive content and adaptive presentation.

**********

With the spread of New Technologies of Information and Communication (NTIC), we are living a multi-domain revolution particularly in education and training thanks to the development of a new mode of education: e-learning (Arroyo, Conejo, Guzmand, & Woolf 2001; Brusilovsky, Eklund, & Schwartz, 1998).

E-learning, the electronic delivery of information, communication, education, and training, provides a new set of tools that can add value to traditional learning modes: classroom experiments, textbook study, CD-ROMs, and traditional Computer-Based Training.

E-learning is a revolutionary way to empower a workforce with the skills and knowledge necessary to bring about positive points. It will not replace the classroom environment, but it can considerably enhance it, by taking advantage of new content and delivery technologies.

However, the rise of this new way of education needs new didactic models. Both pedagogical and technological aspects must be carefully reviewed. Consequently, the demand for intelligent systems that make teaching in distance learning environments more efficient and adaptive is increasing.

With this premise in mind, we are developing an hypermedia adaptive system for supporting e-learning, called PERSO (PERSOnalizing e-learning system) (Chorfi & Jemni, 2003).

In this article, we start by providing an overview and a diagnosis of an e-learning experiment we recently conducted at the University of Tunis. Second, we explain the role and the fundamental structure of PERSO. Next, we detail the system functioning: we focus on the way we collect information about the learner and the way we adapt the personalized course. Finally, we present the findings of our research.

EXPERIMENTAL DIAGNOSIS

To explore the different aspects of this new teaching mode and to analyze how it can be performed effectively, the e-learning team of the Ecole Superieure des Sciences et Techniques de Tunis (ESSTT) conducted a pilot experiment on e-learning (from November 2001 to February 2002). Two courses on MS-Word and MS-Excel were developed and taught to a group of 130 first-year Computer Science students (divided into 8 groups). The courses were developed by the teachers themselves with the help of two multimedia specialists to process images, audio and video sequences, and to prepare flash animation and Java applets.

For these courses, we used a locally web-based learning platform that provides a learning management system and a range of content creation and publication tools. (For more details see Chorfi & Jemni, 2002).

At the end of the experiment, we performed two types of analysis:

* Analysis of users appreciation: questionnaires were distributed to both students and teachers to obtain their appreciation about different aspects. For students, the questionnaires focused, particularly, on the facility of using the platform, course's structure, and communication. Whereas, for teachers, the questionnaires concerned two points of view: the pedagogical strategies and quantification of efforts required for this new learning and teaching mode, such as time of content creation, time spent to communicate with students (respond to e-mails, Chat, forum animation), and time needed to correct assessments.

* Statistical analysis of data delivered by the platform: our aim here is to know the rate of the use of tools and resources and to eventually constitute relations between the relevant criteria and aspects of e-learning, such as the use of a particular tool and time of connection or number of visits. Based on the questionnaires and the teachers' remarks, the results were promising. We highlighted the students for the following positive points: elimination of the psychological students-teacher barrier, discrete communication with the teacher or with colleagues (via e-mail or private chat), possibility of feedback, and respect of the student's individual learning rate. Similarly, positive features for the teacher included: improvement of pedagogical methods, permanent courses availability, and teaching schedule flexibility.

However, two important questions emerged:

First, why do we provide the same content and the same amount of details to all students when they do not all have the same background?

Secondly, why do we present a course in the same manner to the entire group when it is heterogeneous and all students do not have the same preferences?

Our suggestion is to propose courses adapted to users having very different backgrounds and prior knowledge of the subject. For example, students with some knowledge of the subject would not be taught the known material again, and less prepared students would get more details. To guaranty more efficiency in practice, courses should be automatically adapted to the student by means of a system, that analyses his own background and preferences, and determines the parts of course needed with the desired form of presentation.

The system presented in this article, namely PERSO, addresses this specific problem and tries to match the student's particular needs by recommending to each one an appropriate training content. It offers an adaptive e-learning environment, where learners with different learning goals and different learning aptitudes are treated differently, by building a model of knowledge and preferences about each user.

PERSO DESCRIPTION

Role

PERSO is production-centered rather than learning-centered. It is mainly based, on the elaboration of dynamic questionnaire generator to model user. User modeling concerns the student background about the subject to be taught and his/her preferences in terms of type of media presentation. The user background modeling is based on an open approach where the student expresses his/her answer to the system question on free verbal statements. (Figure 1). (we will detail later the technique used to analyze student answers).

PERSO is also based on a course generator. The user model built by the questionnaire is used to automatically generate, for every student, appropriate training content. To build a new course, PERSO tries to exploit previous experimentations and solutions when creating a new personalized course.

PERSO Architecture

PERSO is comprised of five components: (a) the curriculum, (b) the student model, (c) the analyzer, (d) the CBR system, and (e) the generator. Figure 2 illustrates these components and the interactions between them.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

The functioning and the interactions of the different modules of PERSO are described in the next sections.

PERSO FUNCTIONING

Curriculum Representation

The knowledge representation used, for the curriculum, is a form of semantic network, a graph in which nodes represent pieces of knowledge and edges represent the relations between these nodes. The nodes in the network of concepts are linked by the following relations: composition, equivalence, prerequisite, and analogy.

The following figure shows a part of the network of concepts. Note that all the examples in this article are about MS-Word courses.

In the network of concepts, we define three entry points according to each student category (beginner, intermediate, and advanced). An entry point corresponds to the most advanced concept that the student has to master in a given category. The semantic network is in turn divided into three sub-networks corresponding to each category (Figure 4).

[FIGURE 3 OMITTED]

[FIGURE 4 OMITTED]

Learner Modelling

The first time, the student connects to the system, he/she is asked to fill-out a questionnaire to determine his/her profile. The questionnaire approach used in PERSO is categorization because it is often possible to observe patterns among students and to group students with similar features within categories, sometimes called stereotypes (Kay, 2000).

Two types of information are modeled: the student knowledge about the course to be taught and the student media preferences.

* Media preferences: we note that, whenever possible, a notion is stored in the curriculum with several types of media presentation. We propose four types of media: text/image, video, audio, and simulation. To notify the system of his presentation preferences, the student is asked to sort the different types.

The student media preference is represented as a succession of mediavalue pairs.

Example: {(video, 1), (sound, 2), (simulation, 3), (text/image, 4)}

* Student knowledge about the course to be taught (cognitive profile): to determine the student background about the course, the system proceeds in two steps. First, it determines the category to which the student belongs. Second, it asks the student on its category's concepts. For the category determination, the system asks one or two initial questions. These questions are about the entry points of the semantic network.

The system gives the student two possibilities for answering the question. If the student does not know the answer to the question, he/she can give the pre-defined response "I do not know." If he/she knows it, they can formulate it as a free verbal statement. In that case, the system performs an analysis of the student's answer. The analysis consists of calculating the semantic closeness between the student's answer and the correct one. Correct answers are already stored in the system by the professor. The semantic closeness is calculated by use of a robust technique named Latent Semantic Analysis (LSA) (Landauer, Foltz, & Laham, 1998).

LSA represents documents and paragraphs in a large two-dimensional semantic space matrix (Gounon & Lemaire, 2002). An algebraic technique known as Singular Value Decomposition is applied to this matrix resulting in a set of weightings (the singular values) and a set of K-long vectors (one for each term or text). The normalized sum of the vectors of the terms in any text produces the vector for this text. The distance between any two vectors, calculated by their cosine, is interpreted as the semantic closeness between terms or texts. An interesting feature of this method is that the semantic information is only derived from the cooccurrence of words in a large corpus of texts. LSA is a fully automatic program designed by Bellcore Labs.

The cognitive profile is represented as a succession of concept-value pairs: the concept is that on which the student is questioned. The value represents the semantic closeness between the student's response and the correct one (stored previously in the system).

[GRAPHIC OMITTED]

[GRAPHIC OMITTED]

Course Adaptation

The course generation is based on the Case-Based Reasoning (CBR) approach (Kolodner, 1993). The main hypothesis behind CBR is simply that "similar problems have similar solutions" or that "you can reuse the solution of a similar problem in order to solve your actual problem" (Wilke & Bergmann, 1998).

A case is the most basic element representing an experienced situation. The techniques that make up CBR are: case representation, indexing, retrieval techniques, and adaptation.

The following sections give details on PERSO's CBR process.

Case Representation

Once the category and the student profile has been determined, the system is faced with a new case. In PERSO, a case is composed of the case number, the category, the student identification, and the student profile (the cognitive profile and the media preference). The solution consists of a structured number of concepts to be recommended to the student. Figure 5 shows how the case is built.

The case base (as the semantic network) is composed of three sub-bases corresponding to the three categories (Figure 6).

[FIGURE 5 OMITTED]

Each category in the case base is initialized with typical cases and is eventually updated with new cases. When a new case arrives, the system tries to find similar cases in the case base. The new case is stored if there is no similar case and the system considers it relevant.

In the following examples (Figure 7), we present an initial case for each category. Values 0 and 1 represent the degree of mastering the notion (0: the lowest degree, 1: the highest degree).

Indexation

A case is indexed according to the case number and the category. When a new student logs in, the system asks the student about his/her background on the concepts of the course to construct a new case. The case is constructed in two steps. First, the system uses the answer to one or two questions to determine the student category. Then it asks the student questions on the concepts of that category.

[FIGURE 6 OMITTED]

Similarity

When a new case is constructed, we calculate its similarity (Richter, 1995) with the cases of the same category stored in the case base.

The similarity function we use is the Euclidean distance defined as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

In PERSO, this function is applied with the values of the cognitive attributes. We consider that two cases are similar if their similarity is lower than a certain threshold. This similarity is calculated to permit the system to minimize time of constructing a new solution for the current case by using or adapting the solution of an old and similar case stored in the case base.

Adaptation

The result of the similarity measure allows the system to decide on the adaptation to perform. Different forms of adaptation exist, such as null adaptation, transformational adaptation (including substitutional and structural adaptation), and generative adaptation. In PERSO, we use structural adaptation (Smyth & Keane, 1996). It supports the reorganization of solution elements and permits the addition and deletion of such elements under certain conditions. Also, structural adaptation systems uses a fixed set of adaptation operators and/or transformation rules that modify the structure of the solution depending on the relation between the description of the query and the similar case (Wilke & Bergmann, 1998).

PERSO develops a solution to a novel case by choosing the most similar case and applying a set of rules.

Example of rules:

1) If similar.attribute(i) = 0 and new.attribute(i) = 1 then delete attribute(i) to similar.solution

2) If similar.attribute(i) = 1 and new.attribute(i) = 0 then add attribute(i) from similar.solution

Note that similar.attribute(i) is an attribute of the most similar case and new.attribute(i) is the corresponding attribute of the new case for which a solution is being adapted.

General Algorithm

We propose, now the general algorithm and the two procedures: Calculate-similarity and Adapt-solution.

General Algorithm           Calculate-similarity Procedure
BEGIN                       BEGIN
  Enter new-identification    ts := max-threshold
  Determine-category           For i := 1 to N do
  Construct-new-case            If f-similarity (new, case (i))<ts
  Calculate-similarity           Then ts := f-similarity (new, case (i))
  Adapt-solution                       Similar := case (i)
  Recommend-course               Endif
END                              Endfor
                            END

Adapt-solution Procedure
BEGIN
  If not exist (similar)
   Then Construct-new-solution
  Else If f-similarity (new, similar)=0
    Then retain similar.solution
    Else apply-rules-adaptation
   Endif
  Endif
END

In the algorithm, new is the case for which we are calculating similarity; case(i) is the stored case number i; f-similarity is the function we apply to calculate the similarity (i.e., the Euclidian distance); n is the number of cases in the sub-category to which the student belongs; ts is the threshold under which two cases are considered similar; similar is the similar case to retain for adaptation; and similar.solution is the solution obtained from adapting the similar solution.

To generate a course, we have two possibilities:

1. We have a similar case, so the system uses or adapts the solution of this case.

2. We don't have a similar case, so the system uses two processes. The first process extracts the notions to include in the course and structures them into chapters. The system considers three cases, if the semantic closeness for the concept on which the student is questioned:

* [member of] [-1; 0.5[, the notion is poorly covered by the student and is therefore included in the course to generate for the student.

* [member of] [0.5; 0.7], the notion is well covered. In this case, the choice of whether to include the notion in the course is left to the student.

[FIGURE 8 OMITTED]

* [member of] [0.7; 1], the notion is mastered and is not included in the course.

Note that interval limits can be adjusted and modified by the teacher depending on the system's results (currently, we believe these values deliver good results).

The second process addresses the type of presentation. We note that a concept is stored in the curriculum with different presentations when possible.

For each concept to include in the course, the system tries to satisfy the first choice of the student. If this choice is not available in the database then the system proposes the second choice and so on.

Example of System Functioning

We recapitulate the general functioning of the system by illustrating it by an example (Figure 8).

CONCLUSION

PERSO, the system we presented in this article is production-centred rather than learning-centred. It can be integrated in an e-learning environment to give it more efficiency. It recommends to each student a personalized course considering his/her background and his/her preferences (in terms of course appearance).

It is based on a questionnaire in which responses can be free text statements. A critical point for this kind of system is a natural language processing mechanism that can robustly understand student input. LSA provides such a mechanism. On the other hand, PERSO uses CBR approach to determine the appropriate course to propose for the student. The advantage of using CBR techniques is twofold: on one hand, it allows us to minimise the number of questions to ask to the student. On the other hand, it minimises the time for finding a new solution (personalised course) by adapting previous ones.

A prototype of PERSO has recently been developed. It is based on open source software that is, My Sql database, Tomcat web server, JSP code, and Linux operating system. We expect in the near future to experiment with and evaluate the system before making it available for free use on the Internet.

                 Case number=1                 Example 1 of case
                 Category=beginner
                 Level=1
Student          *
Identification   *
Media            *
preference       *
Cognitive        Start word = 0
profile          Create a new document = 0
                 Select a block = 0
                 Police format = 0
                 Paragraph alignment = 0
                 Print a document = 0
                 Insert a table = 0
                 Case number=10                Example 2 of case
                 Category=intermediate
                 Level=1
                 ****
                 Print options = 1
                 Insert page number = 0
                 Sort a table = 0
                 Convert a table = 0
                 Draw a table = 0
                 Insert image = 0
                 Draw simple form = 0
                 Case number=20                Example 3 of case
                 Category=advanced
                 Level=1
                 ***
                 Multi-columns paragraphs = 0
                 Insert equations = 1
                 Documents fusion = 0
                 Draw complex form = 0
                 Protect a document = 0
                 Fusion letter/database = 0
                 Create table of contents = 0

Figure 7. Examples of cases

Acknowledgment 

We would like to thank Dr. Zaher Mahjoub, University of Tunis El Manar, for his remarks and Dr. Henda Ben Ghezala, University of Manouba, for her help.

References

Arroyo, I., Conejo, R., Guzmand, E., & Woolf, B.P. (2001). An adaptative web-based component for cognitive ability estimation. Proceedings of AIED'01, (pp. 456-466).

Brusilovsky, P., Eklund, J., & Schwartz, E. (1998). Web-based education for all: A tool for developing adaptive courseware. Computer Networks and ISDN Systems, 30(1-7), 291-300.

Chorfi, H., & Jemni, M. (2002, July/August). Evaluation and perspectives of an innovative Tunisian e-learning experimentation. Proceedings of the International Conference on Advances in Infrastructure for e-Business, e-Education, e-Science and e-Medecine on the Internet, l'Aquila, Italy.

Chorfi, H., & Jemni, M. (2003, May). PERSO: A system to customize e-training. Proceedings of the 5th International Conference on New Educational Environments, Lucerne, Switzerland.

Gounon, P., & Lemaire, B. (2002). Semantic comparison of texts for learning environments. In F.J. Garijo, J.C. Riquelme Santos, & M. Toro (Eds.), Advances in artificial intelligence--IBERAMIA, 2002, (pp. 724-733). Berli Springer Verlag LNCS 2527.

Landauer, T.K., Foltz, P., & Laham, D. (1998). Introduction to latent semantic analysis. Discourse Processes, 25, 259-284.

Kay, J. (2000). Stereotypes, students models and scrutability. In G. Gauthier, C. Frasson, & K. VanLehn, (Eds.), Intelligent Tutoring Systems ITS 2000, (pp. 19-30), Montreal, Canada. Berlin: Springer-Verlag.

Kolodner, J.L. (1993). Case-based reasoning. San Mateo, CA: Morgan Kaufmann.

Richter, M.M., (1995). The knowledge contained in similarity measures. Invited talk on the ICCBR-95. [Online]. Available: http://www.agr.informatik.uni_kl.de/~lsa/CBR/

Smyth, B., & Keane, M.T. (1996). Using adaptation knowledge to retrieve and adapt design cases. Journal of Knowledge Based Systems, 9(2), 127-135.

Wilke, W., & Bergmann, R. (1998). Techniques and knowledge used for adaptation during case-based problem solving. Proceedings of IEA-98-AIE. Berlin, Heidelberg, New York: Springer-Verlag.

HENDA CHORFI AND MOHAMED JEMNI

Ecole Superieure des Sciences et Techniques de Tunis, Tunis, Tunisia

[e-mail]

[e-mail]