Third Generation Adaptive Hypermedia Systems Curtis A. Carver Jr, John M.D. Hill, and Udo W. Pooch Department of Computer Science, Texas A&M University, USA Email: carverc/hillj/
[email protected]
Abstract: This paper examines the development of adaptive hypermedia systems and proposes adaptive characteristics for third generation adaptive hypermedia systems. First generation adaptive hypermedia systems predated the World-Wide Web (WWW) and were primarily single user adaptive hypermedia systems. Second generation systems have exploited the distributed nature and ease of authoring of the WWW to develop more robust and mature adaptive hypermedia systems. While these systems are a dramatic improvement over first generation systems, they have several limitations that limited the effectiveness of adaptation. These limitations include limited adaptation through one-dimensional, stereotypical user models, coarse granularity of adaptive support, closed adaptive hyperspaces with sharp boundaries, limited authoring support, and limited and nonconstructive communications between the user and the adaptive model. Third generation adaptive hypermedia systems must address these shortcomings to fully exploit the potential of adaptive hypermedia. Future systems must provide explicit, fine-grained adaptation support that the user can easily tailor and refine to provide highly relevant multidimensional adaptation. Future adaptive hypermedia systems must be open systems with soft boundaries that are expandable and incorporate resources from non-adaptive hypermedia with gradual degradation of support. Finally, adaptive systems must be relatively easy to build and maintain. This paper examines these characteristics of future adaptive hypermedia systems and proposes a framework for development.
1. Background 1.1 First Generation Adaptive Hypermedia Systems Numerous adaptive hypermedia systems have been implemented over the last fifteen years. These systems can be characterized as first generation or second-generation adaptive hypermedia systems based on when they were developed and what delivery mechanism was used for deployment of the systems. First generation systems span the period from 1985-1993 and are principally adaptive systems that were not distributed in nature. These systems were generally PC or Macintosh-based provided limited adaptability through stereotype-based user models and limited functionality adaptation techniques such as conditional text filters, direct guidance, stretchtext, hiding, and primitive link annotation. EPIAIM [Rosis et al. 1994], Hypadapter [Bocker et al. 1990], ITEM/IP [Brusilovsky 1992], ISISTutor [Brusilovsky and Pesin 1994; Brusilovsky and Pesin 1994], MetaDoc [Boyle and Encarnacion 1994], and MetaDoc V [Boyle and Teh 1993] are examples of first generation adaptive hypermedia systems. ISIS-Tutor is a good example a first generation adaptive system. It is an adaptive learning environment for the information retrieval system CDS/ISIS/M (ISIS) [Brusilovsky and Pesin 1994; Brusilovsky and Pesin 1994]. ISISTutor provided adaptive support through frame-based presentation, link annotation, and information hiding. It presented an adaptive sequence of tasks and concepts based on the user model and maintained a four state, user knowledge model (not-ready-to-be-learned, ready-to-be-learned, in-work, and learned) on each concept within ISIS. As the user progressed through the system, ISIS-Tutor annotated each concept with color and hiding could be enabled to remove concepts that the user was not ready for. ISIS-Tutor was one of the first adaptive models to incorporate more than a bipolar knowledge model and one of the first adaptive systems to have empirical support for the effectiveness of adaptive interfaces [Brusilovsky and Pesin 1994]. The advent of the World-Wide Web, however, provided new opportunities for the development of adaptive hypermedia systems and led to new systems and more advanced adaptive techniques.
1.2 Second Generation Adaptive Hypermedia Systems Second generation systems span the period from 1993 to present and predominantly use the WWW as their delivery and presentation means. These systems have worldwide availability and are generally platform independent. Second
generation systems introduced new capabilities such as adaptive multimedia presentation, map adaptation, and link sorting. They also refined existing adaptation techniques to provide greater functionality. User models became more robust and incorporated more user characteristics. AHA [Bra 1996; Bra and Calvi 1998], ARNIE-HyperMan [Mathe and Chen 1994; Rabinowitz et al. 1995; Mathe and Chen 1996], AVANTI [Fink et al. 1997; Rizzo et al. 1997; Nill 1998], Basar [Thomas 1995], CS383 [Carver et al. 1996; Carver et al. 1996], Cameleon [Laroussi and Benahmed 1999], ELM-ART [Brusilovsky et al. 1996], I-Doc [Erdem and Johnson 1998; Erdem et al. 1998], InterBook [Brusilovsky and Eklund 1998] , KN-AHS [Kobsa et al. 1994], MediaDoc [Erdem et al. 1998], RATH [Hockemeyer et al. 1998] and WebWatcher [Armstrong et al. 1995] are examples of second generation adaptive hypermedia systems. ELM-ART (The Episodic Learning Model: The Adaptive Remote Tutor) is good example of a second-generation adaptive hypermedia system. ELM-ART is a distributed intelligent tutoring system on LISP that provides course adaptation through a combination of adaptive annotation and link sorting [Brusilovsky et al. 1996]. Links are colorcoded according to user preparation for the information in the node. Red annotations indicate nodes that the user has to meet the prerequisites for, amber nodes represent information the student is ready for but not recommended, and green nodes represent nodes that the user is ready for and are recommended. ELM-ART adaptively sorts links as well so that the links that are most similar to the node that the user is currently on are presented first. ELM-ART features extensive user feedback and is highly interactive. As the user completes exercises and reads nodes, status bars change to reflect the user’s progress through the course. Link annotations change color and the navigational view changes to reflect the user’s newly gained knowledge. Users can directly edit the user model and override the navigational choices present to meet their educational goals. An extensive number of exercises engage the user and proved constant feedback both to the user and to the user model on the user’s evaluated knowledge. While second generation adaptive hypermedia systems such as ELM-ART have dramatically improved upon the functionality of first generation systems, there are fundamental flaws associated with these systems. These limitations include limited adaptation through one-dimensional, stereotypical user models, coarse granularity of adaptive support, closed adaptive hyperspaces with sharp boundaries, limited authoring support, and limited and non-constructive communications between the user and the adaptive model. The remainder of this paper will discuss these limitations with second-generation systems and propose framework for solutions.
2. Third Generation Adaptive Hypermedia Systems 2.1 Multidimensional User Models Third generation adaptive hypermedia systems should support multidimensional user models. Current user models measure limited user characteristics to normally a single dimension such as declared or demonstrated knowledge, or hypermedia nodes visited. For example, AHA [Bra and Calvi 1998] uses the number of nodes visited while CS383 [Carver et al. 1996] consider learning styles as the basis for the user model. Actual users in a learning environment are much more complex and are both multi-dimensional and multi-faceted. Future user models must incorporate multiple dimensions of the user including expertise, user goals, interests, and preferred learning style by subject matter. These dimensions may be declared by the user, measured by the adaptive system, or combination of both approaches. Not only must the user model incorporate multiple dimensions, the importance of an individual user model dimension may vary over time. As a user progresses through hyperspace, their goals and interests may change as they learn new concepts. The user model must quickly adapt to these changes in the user model so as to present relevant information to the user. Discrepancies between declared and demonstrated user characteristics must be resolved and presentation of material adapted. The users of adaptive hypermedia systems are not one-dimensional but instead are multidimensional. Future user models in adaptive hypermedia systems should be multidimensional and adaptive as well. Providing adaptive, multidimensional user models raises a number of open research issues. Most current adaptive systems allow their users to explicitly manipulate the user model. This is most commonly done through a long list of checkboxes. Different presentation techniques will be required for users to effectively manipulate multiple user dimensions. The effective manipulation of a multidimensional user model clearly presents significant user interface
issues. Additionally, it remains an open research issue as to what is the proper type and number of dimensions to measure. Adding additional dimensions will not always increase the accuracy of the user model but will always increase the complexity of the user model and the requirements to collect additional user information. There is a balance between the number of dimensions, model complexity, and the accuracy of the model. Finally, techniques for modifying the weights associated with different dimensions dynamically to better represent the user are open research issues.
2.2 Finely Grained, Multimedia Adaptation Third generation adaptive hypermedia systems should provide a fine degree of adaptation granularity and adapt more than just hypertext. All second-generation adaptive hypermedia systems provide text-based adaptation based on a user model with limited levels of user differentiation. In addition to using multidimensional user models, third generation systems must incorporate multiple levels in each user model dimension so that the adaptation provided is truly tailored to the user. Users are not just novice, intermediate, or expert users but range a scale of many intermediate values. Users are not simply sequential or global learners but instead are some combination of both characteristics. Third generation adaptive hypermedia systems should not only model multiple dimensions of the user, but each dimension should have as much delineation as necessary to truly model the user. In the last five years, there has been significant growth in the granularity of user models in adaptive systems such as ELM-ART [Brusilovsky et al. 1996], InterBook [Brusilovsky and Eklund 1998], AHA [Bra and Calvi 1998], and RATH [Hockemeyer et al. 1998]. These systems incorporate numerous exercises to capture more accurately user knowledge of hypermedia material covered. Third generation adaptive hypermedia must continue to build upon the successes of these earlier systems. Third generation adaptive hypermedia systems should adapt more than hypertext. Adaptive hypermedia systems have traditionally focused on text presentation or navigation support. Only one system, CS383 [Carver et al. 1996] implemented adaptation of other media types such as sound, graphics, or video but its implementation was limited to adaptation of media type (not individual media elements) by learning style. Third generation systems must expand adaptation so that appropriate pictures, movies, slideshows, or sound files play to different users. Introducing fine-grained multimedia adaptation raises a number of open research issues. Like adding additional dimensions to the user model, adding greater granularity to each dimension requires a much greater implementation effort and increases the complexity of the user model. Assessing the proper balance between granularity and adaptive system performance remains on open research issue. Multimedia adaptation adds additional complexity and requires a greater implementation effort. Media elements, other that text, are more difficult to generate and are not as malleable to automatic recombination. It is extremely difficult to automatically adapt video segments on the fly and present the results to users for example. Techniques for adaptation and presentation of media elements other than text are open research issues.
2.3 Open Adaptive Hypermedia Systems Third generation adaptive hypermedia systems must support open adaptive systems. All known adaptive hypermedia systems, with the exception of Personal WebWatcher [Joachims et al. 1997], are closed navigation systems. Adaptation is provided while the user is in the restricted hyperspace of the adaptive system and there are explicit, well-defined navigational boundaries on the system. The adaptive system is self-contained. Because the system is self-contained, the author of the adaptive system must ensure that all information the user might require is inside the system. If users are not satisfied with the information resources within the system, they can either remain dissatisfied and use the adaptive system, or leave the system. Third generation adaptive hypermedia systems should allow users to stay within the adaptive system but travel to informational resources outside of the navigational boundaries of the adaptive hyperspace. This tunneling through adaptive navigational boundaries provides the user the ability to gather additional informational resources and return to the adaptive system when ready. These informational resources may be links provided by the adaptive system to
non-adaptive information resources outside the system that are constrained in a separate frame or in a new window, or may be a general-purpose search capability. This ability to integrate external informational sources within the adaptive system provides a number of advantages that second generation adaptive hypermedia systems cannot. The informational resources of the adaptive system are significantly increased without a large amount of work by the adaptive system author. Users have greater flexibility to more deeply explore information within the conceptual model of the adaptive system and pursue individual goals and interests. But more importantly, based on the number of tunneling attempts outside the adaptive hyperspace by concept, the adaptive system gains a quantitative measure of informational regions within the navigational space that are not meeting user needs. Armed with critical feedback, the adaptive system author can then add new information to the system, remove useless information, provide additional external links to known information resources, or provide a general-purpose search engine. Future adaptive hypermedia systems should be open systems with soft boundaries as opposed to closed systems with hard limitations. There are significant implications in the construction of open adaptive hypermedia systems. Adaptive hypermedia is no longer self-contained but instead is part of larger body of information. The synergistic integration of search engines and glossaries becomes important. Some second-generation hypermedia systems, such as InterBook [Brusilovsky and Eklund 1998], have already begun to include adaptive glossaries and several systems use extensive assessment to provide more accurate adaptation. As glossaries, digital libraries, search engines, assessment engines, and resources external to the adaptive system are added, third generation adaptive systems will be more of a solution and less of a research prototype. Opening adaptive hypermedia systems to include informational resources outside of the navigational space of the adaptive system raises a number of interesting research issues. Capturing the tunneling activity and correctly categorizing it becomes an important research area. Are multiple, successive tunneling attempts indicative of a confused, frustrated user, a novice user exploring every link, or a happy, enthusiastic user that wants to leave everything they can about the subject matter at hand? Controlling the effects of context shift is another open research area. Even if the material is appropriate, it may be in a presentation style that causes the student to lose focus or become confused. Ensuring that users do not become lost when leaving the adaptive navigational space is likewise an open research issue. Finally, how to capture and react to user tunneling efforts is an open research issue.
2.4 Constructive Hypermedia Third generation adaptive hypermedia systems should be constructive in nature. The most accurate user model is the one that users construct themselves. Users should have the capability to easily add material to the adaptive system with a clear delineation of material from the original author and from users. User should be able to view the original material only, all user comments, or their comments only. They should be able to add comments to pages as in the ELM-ART system [Brusilovsky et al. 1996] or annotate links, a technique which has not yet been implemented in an adaptive hypermedia system. Constructive hypermedia provides the most precise and fine-grained form of adaptation. As the complexity or size of adaptive hypermedia systems grows, so too will the need for constructive hypermedia. There are a number of open research issues associated with constructive, adaptive hypermedia. User construction may be indicative that the underlying adaptive system is not meeting the information needs of the user. Mechanisms for evaluating user constructions and modifying the user model as necessary need to be explored and developed. Providing access to all user comments may not be productive and may confuse the user more than help them. Methods must be developed for assessing the impact and utility of user construction on the user and adapting the system appropriately. Further assessment of this approach is necessary before implementation.
2.5 Other Third Generation Attributes 2.5.1 Adaptation Feedback Third generation adaptive hypermedia systems should provide better feedback to the user and better incorporation of user feedback into the user model. Users should easily be able to determine where they are, where they have been,
what they have left to do, and what the user model is doing to them. In terms of adaptation, the user should be able to quickly understand the adaptation model and modify it to provide better adaptation. ELM-ART [Brusilovsky et al. 1996] is an example of a second-generation adaptive hypermedia system that provides excellent adaptation feedback. Additional work is necessary particularly in light of the multi-dimensional nature of the user model, to provide the user with clear feedback as to how user activities are influencing the adaptation.
2.5.2 Ease of Authoring Third generation adaptive hypermedia systems must be easy to build and maintain. Most second-generation adaptive hypermedia systems are handcrafted and not easily expandable or adaptable to other projects. As adaptive hypermedia systems grow in size, dimensions, or the granularity of those dimensions, the ease of construction and maintenance of the systems will increase in importance.
2.5.3 Assessment of Effectiveness Additional, statistically valid assessments of the effectiveness of adaptive hypermedia are necessary to validate the approaches that are effective and those that are not. Most assessments of second-generation systems were conducted with very small sample populations. Assessments of adaptive techniques, adaptive systems, and user models have begun but remain open research areas.
3. Conclusion This paper has examined the development of adaptive hypermedia systems and proposed several attributes for third generation adaptive hypermedia systems. Third generation adaptive hypermedia systems must incorporate fine grained, multidimensional user models that support open, constructive, adaptive systems that are easy to build. Numerous research issues require resolution to provide these capabilities. Adaptive hypermedia systems have made great strides in the last fifteen years, but new capabilities and a better understanding of users and adaptive systems opens new avenues of research. Research into these issues raised by the advent of third generation adaptive hypermedia systems will bring us closer to realizing the full potential of adaptive hypermedia.
4. References Armstrong, R., D. Freitag, T. Joachims and T. Mitchell (1995). WebWatcher: A Learning Apprentice for the World Wide Web. 1995 AAAI Spring Symposium on Information Gathering from Heterogeneous, Distributed Environments, Stanford, CA. Available at http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-6/web-agent/www/project-home.html. Bocker, H. D., H. Hohl and T. Schwab (1990). Hypadapter - Individualizing Hypertext. INTERACT 90, Amsterdam, Holland. 931-936. Boyle, C. and A. O. Encarnacion (1994). MetaDoc: an Adaptive Hypertext Reading System. User Modeling and User Adaptation 4(1): 1-19. Boyle, C. and S. H. Teh (1993). Multimedia Intelligent Documentation: Metadoc V. 11th Annual International Conference on Systems Documentation, Kitchener, Ontario, CA. 21-27. Bra, P. D. (1996). Teaching Hypertext and Hypermedia through the Web. WEBNET 96 - World Conference of the Web Society, San Francisco, CA. 110-115. Bra, P. D. and L. Calvi (1998). AHA: a Generic Adaptive Hypermedia System. 2nd Workshop on Adaptive Hypertext and Hypermedia, Pittsburgh, PA. 5-12. Brusilovsky, P. (1992). Intelligent Tutor, Environment and Manual for Introductory Programming. Educational and Training Technology International 29(1): 26-34.
Brusilovsky, P. and J. Eklund (1998). A Study of User Model Based Link Annotation in Educational Hypermedia. Journal of Universal Computer Science 4(4): 429-448. Brusilovsky, P. and L. Pesin (1994). ISIS-Tutor: An Adaptive Hypertext Learning Environment. Japanese-CIS Symposium in Knowledge-based Software, Tokyo. 83-87. Brusilovsky, P. and L. Pesin (1994). ISIS-Tutor: An Intelligent Learning Environment for CDS/ISIS Users. CLCE. Available at http://cs.joensuu.fi/~mtuki/www_clce.270296/Contents.html. Brusilovsky, P., E. Schwarz and G. Weber (1996). ELM-ART: An Intelligent Tutoring system on the World Wide Web. Intelligent Tutoring Systems. C. Frasson, G. Gauthier and A. Lesgold. Berlin, Springer Verlag. 1086: 261-269. Carver, C. A., Jr., R. A. Howard and W. D. Lane (1996). A Methodology for Active, Student-Controlled Leaning: Motivating Our Weakest Students. 27th ACM Technical Symposium on Computer Science Education, Philadelphia, PA. 195-199. Carver, C. A., Jr., R. A. Howard and E. Lavelle (1996). Enhancing Student Learning by Incorporating Learning Styles into Adaptive Hypermedia. EDMEDIA 96 - World Conference on Educational Multimedia and Hypermedia, Boston, MA. 118-123. Erdem, A. and L. Johnson (1998). Task Oriented Software Understanding. Automated Software Engineering Conference. Available at http://www.isi.edu/isd/I-DOC/i-doc.html. Erdem, A., W. L. Johnson and S. Marsella (1998). User and Task Tailored Software Explanations. Automated Software Engineering Journal. Available at http://www.isi.edu/isd/erdem.html. Fink, J., A. Kobsa and J. Schreck (1997). Personalized Hypermedia Information Provision through Adaptive and Adaptable System features: User Modeling, Privacy and Security Issues. Fourth International Conference on Intelligence in Services and Networks, Como. Italy. Available at http://zeus.gmd.de/hci/projects/avanti/publications/ISandN97/ISandN.html. Hockemeyer, C., T. Held and D. Albert (1998). RATH: A Relational Adaptive Tutoring Hypertext WWW - Environment Based on Knowledge Space Theory. Proceedings of the Fourth International Conference CALISCE 1998, Goteborg, Sweden. 417-423. Joachims, T., D. Freitag and T. Mitchell (1997). WebWatcher: A Tour Guide for the World Wide Web. International Joint Conference on Artificial Intelligence. Available at http://lia.univ-savoie.fr/~bleyer/agent/references.html. Kobsa, A., D. Müller and A. Nill (1994). KN-AHS: An Adaptive Hypertext Client of the User Modeling System BGP-MS. Fourth International Conference on User Modeling, Hyannis, MA. 99-105. Laroussi, M. and M. Benahmed (1999). CAMELEON: Computer Aided MEdium for LEarning On Networks. Available at http://lithwww.epfl.ch/~laroussi/cameleon.html. Accessed on February 20, 1999. Mathe, N. and J. R. Chen (1994). A User-Centered Approach to Adaptive Hypertext based on Information Relevance Model. 4th International Conference on User Modeling, Hyannis, MA. 107-114. Mathe, N. and J. R. Chen (1996). User-Centered Indexing for Adaptive Information Access. International Journal of User Modeling and User Adapted Interaction 62(2/3): 225-261. Nill, A. (1998). AVANTI Home Page. Available at http://zeus.gmd.de/hci/projects/avanti/cd-rom/contents/contents/index.htm. Accessed on February 25, 1999. Rabinowitz, J. L., N. Mathe and J. R. Chen (1995). Adaptive HyperMan: A Customizable Hypertext System for Reference Manuals. AAAI Fall Symposium on Artificial Intelligence Applications in Knowledge Navigation and Retrieval, Cambridge, MA. Available at http://ic-www.arc.nasa.gov/ic/projects/aim/cid/cid.html. Rizzo, A., E. Marchigiani and A. Andreais (1997). The AVANTI project: prototyping and evaluation with a cognitive walkthrough based on the Norman’s model of action. ACM Conference on Designing Interactive Systems. 305-309. Rosis, F. D., B. D. Carolis and S. Pizzutilo (1994). User-Tailored Hypermedia Explanations. Workshop Adaptive Hypertext and Hypermedia at 4th International Conference on User Modeling, Hyannis, MA. Available at http://www.education.uts.edu.au/ projects/ah/deRosis.html. Thomas, C. G. (1995). Basar: A framework for integrating agents in the World Wide Web. IEEE Computer 28(5): 84-86.
