This paper describes an Intelligent Tutoring System that provides adaptable facilities to students over the World Wide Web. The system is able to adapt to a large corpus of students, in terms of allowing commonly used paths to emerge, and adapts to individual students by employing a novel student model.

The lack of intelligent tutoring systems outside of academic research has largely been attributed to the fact that they are domain specific. Whilst very powerful for the tutoring of a specific subject, the inability of a knowledge-based approach to artificial intelligence to generalise that has rendered it less useful for generic systems. Neural networks, or connectionist models, are the antithesis of knowledge-based approaches in that they are extremely adept at generalising which gives them the ability to work with very noisy data.

The research project described in the paper employs both knowledge-based representations and neural networks to model students using non-domain specific parameters, such as browse strategies and ability to answer questions. The domain is structured in a hypermedia network using semantic linking that enables the system to automatically produce and weight new links. The weighting system is tailored according to a student's requirements and the student's ability level and is continuously updated.

This novel paraigm is of great potential in a tele-education environment, since the system s generic and is therefore useful to a multitude of authors/domains and the system is able t adapt to a large number of students, such as may be found on the World Wide Web.

Keywords: Hypermedia, neural networks, domain independence, browsing strategy, student modelling

Hypermedia systems when used for learning generally offer no constraint on the user. Indeed this according to some is the beauty of such systems (Jonassen, Grabinger 1991). However as has often been pointed out, hypermedia does suffer from some problems, especially when used for education, most notably "getting lost in (hyper) space" (Conklin 1987), where a user becomes so bemused by the wealth of choice on offer that they become lost in a maze of information. Research has suggested that this is caused by "cognitive overload"; i.e. the brain can only cope with a limited number of tasks (Kibby, Mayes 1990). In the early stages of using an unfamiliar system, much load occurs in the use of the unfamiliar features. It is therefore perhaps better to reduce the plethora of complexities found in a hypermedia system until the user has reached a level such that the complexities will not induce so much load (Dillon 1990).

Traditional computer based learning/tutoring systems (CBL/CBT) are generally the antithesis of hypermedia learning systems, in that they constrain the student and force them to learn a predetermined method (Ridgeway 1989). Intelligent tutoring systems (ITS) use a model of the student’s knowledge so that they are presented with new information only when they require it, to reinforce a point or to progress in the learning and to identify misconceptions and mal-rules (Sleeman, Brown 1982). Such systems have been criticised for constraining the way students solve a given problem (Ridgeway 1989). In most complex problem domains, there can be many methods to achieve a correct solution and some learners may find one particular method suits their way of thinking better than others. It has been argued that students should be able to experiment with their own ideas and find the method that suites them individually (Ridgeway 1989).

Elsom-Cook (1989) reviews some computer based training packages and grades them between two extremes, total constraint, (such as a typical intelligent tutoring system) and totally unconstrained (like a typical hypermedia system). Most systems tended towards total constraint. He argues that the perfect tutoring system should be able to "slide" between these two extremes according to the student’s needs and state of knowledge, appearing as a traditional ITS to a novice student or a discovery learning, hypermedia system to an advanced student. Further research has shown that learning is improved when a student is allowed to follow pathways of their own choice, at their own pace and is able to monitor progress by instant feedback questions (Kibby, Mayes 1990).

A further concern with computer based learning systems is that they are generally applicable only to the domain for which they were specifically produced (Bergeron 1991). A reason for this is that most are of the constrained type, which lend themselves less readily to their implementation in a number of domains. The outcome is that every time a tutoring package is required for a new domain a complete new system must be produced. This is a major reason why intelligent tutoring systems are still predominantly in the domain of academic research (Elsom-Cook 1989, Kinshuk and Patel 1997).

This research therefore highlights the need for both a generic intelligent tutoring systems to provide teachers with a cheap and effective teaching medium and a method for providing a large corpus of students with an adaptable environment. The system is based on a semantically structured hypermedia domain with a novel student model. The World Wide Web may be used to allow the systems to be used with a large corpus of students.

The above discussion suggests the need for a hybrid system that can both constrain the student to ensure that the basics are learned, and then allow them to explore the domain in a less constrained manner once they have achieved a sufficient level of experience. There is also a need for more generic systems, enabling the advantages of computer aided learning to be employed in many domains. This research is concerned with tackling these issues.

To address the problems outlined the research is based on a hypermedia architecture of nodes and links with an intelligent module to aid navigation through the nodes and to model the student’s current level of experience in the domain. To a novice student the system would offer a limited number of links from the total set of links available from that node, based upon information relating to their previous movements through the hypermedia. As the student’s knowledge increases then control is increasingly relinquished by the system, until the student is in total control of the system. Thus, the novice student is freed from much of the complexity associated with an unfamiliar system teaching an unfamiliar subject, while the more advanced student is freed from unhelpful constraints. Offering fewer links that are strongly semantically related is helpful for ensuring that the novice student learns the structure of that particular part of the domain, thus helping them to map the knowledge structures stored in the author’s brain onto their own knowledge structures stored in their own brains (Jonassen, Wang 1993). Once such a transfer of structure is complete then the advanced student is freed from having to navigate the structure on the computer and can focus their cognitive powers more fully on the actual domain content.

The hypermedia architecture uses typed links such as those found in semantic networks (Rich 1991). Types include, is-a, which represents class definition; a-kind-of, which represents membership of and inheritance from super classes; has-a, relating an object to an attribute or property and part-of which is used to show how an object is composed from smaller parts (Frost 1982, Beynon-Davies et al 1994). A hypermedia system built using such link types provides a simple knowledge base about the domain and such a knowledge base may be used to reason new links automatically by utilising automatic reasoning algorithms.

Using a Solar System tutor as an example (figure 1), the nodes Mars and The Earth are linked by an is-a link to the class node Planet, this gives a simple knowledge base, i.e. the knowledge that both Mars and The Earth are planets. Thus, the system can automatically link Mars to The Earth. Further direct links are made from Phobos to Mars and The Moon and The Earth. These two new nodes are then linked to the class node Satellites, which are linked as a child of the planet class node (since a satellite is a sub-type of planet). This allows the system to link Phobos to the Moon, by virtue of them both being Satellites. It also allows the system to link Phobos to the Earth and The Moon to Mars, via the longer path and hence weaker semantic relationship via the class nodes Satellite and Planet. Links can thus be given a weight according to the length of the path that indirectly connects them. The weight allows the system to reduce the number of links it offers to novice students without randomly pruning them. Further more the semantic structure coupled with a student model allows the system to offer links dynamically tailored to an individual student.

The nodes representing classes provide information for both the students and authors. For the student they provide general information and a starting point from where they may seek more detailed information should they require it. For the author they provide information about how to add new nodes to the domain. For example, the ‘Satellite’ class node, if properly defined by the author who created it, should explicitly states that a satellite ‘has-a’ host body, i.e. a host planet.

This method differs from other intelligent hypermedia systems in that the "intelligent" links are produced when the node is added to the system, rather than at run time where it would require a massive processor overhead (Kibby, Mayes 1989). The links are then further tailored according to an individual student’s needs, in accordance with the student model. The former process is processor intensive and the time taken to accomplish it rises rapidly with each additional node. It is therefore better to accommodate the bulk of the dynamic linking process before a student uses the system.

Implementing various domains in the proposed system would require some effort from the domain author in that it is necessary to think about the domain carefully in order to produce the appropriate class nodes and link types. However, there would be no need to produce new software as the intelligent module bases its evaluation of the student on generic measures such as their ability to answer questions and their browsing patterns, each of which are independent of the domain content. It is noted however that some domains may be more difficult to produce than others, for example less structured or human centred domains. This problem is the subject of on going research.

The intelligent module determines the amount of control the student has over the system. It is comprised of a student model and a dynamic link manager.

In our research a neural network (Browsing Pattern Recogniser) is employed to recognise movement strategies through the hypermedia network, to assess the student’s abilities and confidence. For more information of browsing strategies, the reader is referred to (McAleese 1989, Canter 1985). A neural network (Tutorial Supervisor) is also used to grade the student into levels according to how they use the hypermedia and how well they respond to tutorial nodes (nodes that require the student complete some task). Grading a student according to how they respond to tutorials is deemed an accurate metric (since it is directly measuring the student’s ability), however it is measured at some cost, since the student must be directly engaged. Examining the student’s browsing behaviour and linking it to ability is deemed a less accurate, but cheap metric, since it can be measured at any time, but is not directly measuring the student’s ability. For this reason, a novel structure has been devised, based on connectionist and fuzzy logic technologies, to utilise information from the Tutorial Supervisor to train the output of the Browsing Pattern Recogniser, thus providing a system that can lean to evaluate the student based upon their movements throuh the hypermedia only. For further information concerning the use of neural networks in modelling human behaviour, consult Gallant (1988), Tsutsui (1991), Bergeron et al (1991), Battiti and Serra (1991) and Jennings (1993).

The semantic hypermedia coupled with the student model allows a simple level of dynamic linking, in that a high level student who revisits a node after a promotion will see new links at that node. These links were always present, they were generated by the system after the domain was added to it, they were simply not strong enough to appear on the student’s list of available links. Further dynamic linking capabilities can be added that are unique to each individual student and not just to each individual level of student. The following dynamic capabilities do not permanently effect the link weights stored in the system. Rather they are updated as they are retrieved and before they are presented to the current student.

Paragraph inserted from bigAt run time the system builds up a model of the student’s current state of experience by storing data relating to the nodes visited, any question and answer sessions, tutorials and browsing strategy. It then uses this knowledge to a) decide upon the student’s current level and b) suggest one or more links to other related nodes. A student’s inability to rise in levels may alert the system, which may then take a student back over the links they have traversed, or offer different links. Thus, the student may be shown a node again to reinforce a point, or a different node if the current path is not producing a rise in levels. Traces of the student’s pathway through the hypermedia may be of use to the human teacher in determining if the student is utilising the hypermedia well and in determining if the domain is deficient in some way.

Further, this student experience information can be utilised to modify the present link weights to provide links that are related to those previously visited by the student. For example, a student who has visited the nodes ‘Phobos’ and ‘Titan’ has shown an interest in ‘Satellites’, a slight increase in the weights of all nodes connected to the class ‘Satellites’ will ensure that all satellites have a slightly stronger link weight and thus will rise in the student’s list of links. It is a matter of pedagogical discussion whether this situation is desirable or not, the system is able to cater for both views. For example, it may be desirable to prevent the student from becoming attracted to a single area of the domain.

For a large domain, it may be desirable for the human teacher to direct the student into a particular area, whilst still enabling them, to varying degrees depending on the student’s experience, to bring in the potentially useful information stored elsewhere within the domain. This can be accomplished by supplying a modifier for areas of the semantic network that the teacher wants the student to study. The system can then add the modifier to the present link weights before presenting them to the student. This has the effect of ‘pulling’ the student towards certain areas of the domain. The ‘pull’ becomes progressively weaker as the student rises in levels, since a greater number of links will be offered, thus diluting the effect.

Every possible useful path (a trail of several links) in the hypermedia may not be readily apparent at the time of authoring, although the author attempts to provide such paths by defining the semantic network for the domain. Such paths may however arise naturally as a student follows a connected train of thought and may be determined by examining the student’s rise in levels. Once such paths have emerged, a weight increase in the links between its nodes will make the path more likely to be followed by another student. Thus, some of the burden is removed from the domain author who is never likely to be able to produce every useful link. The system is therefore potentially capable of becoming better at directing students through the hypermedia with use.

Such an emergence of structure has been termed the "bottom-up" approach to authoring, since the students are imposing structure upon the domain. It is the subject of on-going research to test these facilities with the requisite large number of students. This is achieved by making the system available via an Internet server. Links are offered to the student in ordered groups, for example a group of links representing the local semantic structure and a group of links representing the weighted, automatically generated links. The lists of links may be formatted by a Java script and presented on a World Wide Web page. Selection of a link results in the server presenting the relevant information or graphic on the web page. The system is then able to record the links that each student visits and is able to slightly modify the link weight, so as to make the selection of the link more likely for the next student. In this fashion useful paths tend to emerge. The presentation of the links is therefore controlled by the server, upon which the student model resides. It is possible for the student model to store unique information about each student by providing a free registration and log in service. The utilisation of the World Wide Web is therefore useful for the research project in terms of it providing the requisite number of students to evaluate the intelligent facilities. It is also a paramount concern of the project that the system is itself useable over the Internet, since this imbues the system with the greatest potential as a mass educational aid.

For a more detailed discussion of this research project the reader is referred to Mullier (1996), Moore et al (1997) and http://www.lmu.ac.uk/ies/comp/staff/dhobbs/aims.htm.

Battiti, R. and Serra, R. (1991) Neural Networks for Intelligent Tutoring Systems. In Artificial Neural Networks: Proceedings of ICANN-91; Kohonen T (ed).

Beer, M. and Diaper, D. (1991) Reading and Writing Documents using header Record Expertext. In Proceedings of the fourth Annual conference: Computers and the Writing Process 1991; Computers and Writing Association.

Beynon-Davies, P., Tudhope, D. and Jones, C. (1994) A Semantic Database Approach to Knowledge-based Hypermedia Systems; In Information and Software Technology Vol.36 No.6. 1994

Bergeron, B., Morse, A. And Greenes, R. (1989) A Generic Neural Network Based Tutorial Supervisor for C.A.I.; In 14th Annual Symposium on Computer Applications in Medical Care; IEEE Publishing

Chiu, C., Norcio, A. F. and Petrucci, K. E. (1991) Using Neural Networks and Expert Systems to Model Usesr in an OO Enviroment. In 1991 IEEE International Conference on Systems, Man and Cybernetics. Vol. III; IEEE Publishing

Conklin, J. (1987) Hypertext: An Introduction and Survey. In IEEE Computer September 87.

Dillon, A. (1990) Designing the human computer interface to hypermedia applications. In Hypermedia for Learning; D. Jonassen (ed). Springer-Verlag.

Elsom-Cook, M. (1989) Guided discovery tutoring and bounded user modelling. In Artificial Intelligence and Human Learning; J. Self (ed). Chapman and Hall.

Frost, R. A. (1982) Binary Relational Storage Structures. In Computer Journal Vol .25 No. 3. 1982

Gallant, S. I. (1988) Connectionist Expert Systems. In Communications of the ACM, Volume 31 Number 2.

Jennings, A. And Higueti, H. (1993) A User Neural Network Model for Personal News Service. In Australian telecommunications Research; Volume 3.

Jonassen, D. And Grabinger, R. S. (1990) Problems and issues in designing hypertext/hypermedia for learning. In Hypermedia for Learning; D. Jonassen (ed). Springer-Verlag.

Kibby, M. and Mayes, T. (1989) Towards Intelligent Hypertext. In Hypertext theory into practice. Blackwell Scientific.

Kibby, M. and Mayes, T. (1990) Learning about Learning from Hypertext. In Hypermedia for Learning; D.Jonassen (ed). Springer-Verlag.

Kinshuk, Patel A, 97; A Conceptual Framework for Internet based Intelligent Tutoring Systems; In Knowledge Transfer (Volume II); Behrooz A (ed); Pace, London.

Marren, A. J. (1991) The Handbook of Neural Computing Applications.

McAleese, J. (1989) Navigation and Browsing in Hypertext. In Hypertext Theory into Practice; R.McAleese (ed). Blackwell Scientific.

Rich, E. and Knight, K. (1991) Artificial Intelligence, second edition. McGraw-Hill.

Moore D J, Hobbs D J, Mullier D, Bell C (1997) Interaction Paradigms with Educational Hypermedia; EuroMicro 97, pp 65-71, Budapest, September 1-4, 1997

Mullier D. J. (1996 ) Integrating Neural Network Technology with Hypermedia; in Henno J (ed.) Proceedings of Hypermedia at Tallinn, Estonia

Ridgeway, J. (1989) Of course ICAI is impossible...worse though, it might be seditious. In Artificial Intelligence and Human Learning; J.Self (ed). Chapman and Hall.

Sleeman, D. H. and Brown, J. S. (eds) (1982) Intelligent Tutoring Systems. Academic Press.

Tsutsui, S. (1991) Knowledge Based versus Neurocomputing. In 1991 IEEE International Conference on Systems, Man and Cybernetics, Vol III. IEEE Publishing