At the outset, we identified several inadequacies of earlier systems. Let us now revisit each and discuss how they have been approached in the PUSH project.
Active but Not Magical
Active systems, such as agents, intelligent filters, or the POP system, run the risk of alienating users. It is important to gain users’ trust in these systems. Adaptive systems should be active but not magical, and in general, usability principles (or variants of usability principles) should be used when designing adaptive systems.
The POP system will hide some information from users’ immediate view, and unless users can stay in control of the adaptivity it may become very frustrating. Almost all adaptive systems will from time to time make mistakes, our system is no exception. It is therefore crucial that users are allowed to ”correct” the system, or override the adaptations made. POP offers both: users can correct the system through setting the task inferred to one that fits better with their needs, and they can also override the adaptations through stretching the text.
Our last study, in section , confirmed that users felt in control of the adaptivity. In the study, subjects would frequently use the stretchtext possibility and occasionally, although not very often, set the task themselves thereby adapting the adaptivity.
Continuous Improvisation
Some adaptive systems assume that users will be goal-oriented, rational, and will stick to their goal throughout whole sessions. Systems adapting to users’ knowledge frequently assume that it will be possible to classify their knowledge into categories such as novice, expert, or intermediate, and that their knowledge will be coherent theories of the subject matter. The situated cognition challenge both these views, and suggests that adaptive system designers should investigate and meet the ”continuous improvisation” behaviour that people exhibit.
In POP the system adapts continuously throughout the session. Our studies confirm that users’ goals change rapidly. Even when explicitly asked to search for some specific information in our last study, subjects would still wander off checking out other aspects of the information space.
The risk of making faulty adaptations is perhaps just as big in POP as in other adaptive systems, but as the system keeps adapting and the assumptions made have fading importance, any erroneous assumptions will have less impact on the overall performance of the system.
Scale up
A general problem to the field of artificial intelligence and also to adaptive systems is in how much they cover of the problem space. Adaptive systems should ”scale up” both in the sense of covering more than a few examples, but also in terms of being able to tackle real-world problems. Research efforts should also be directed at providing efficient methods for how to develop adaptive systems.
In the PUSH project we did tackle a real-world industrial domain. We furthermore took care not to build the design on a few example problems collected from some experts or novices in the field. Instead we attempted to get a grip of the whole problem situation.
Even if we did not build any proper authoring tool we had the authors needs in mind when deciding on the database organisation and content of the information entities. The information organisation and representation must be such that authors would easily be able to input new information entities or alter old ones.
The methods used in PUSH for task and user analysis, knowledge acquisition, design, etc., are maybe not efficient enough to be considered as industrially applicable methods but we believe that there are some lessons to be learnt from the experiences made in PUSH. First, it is important to study users’ real needs and problems and attempt to cater for those with a combination of good system design and adaptive behaviour. Second, it is important to study the users from various perspectives; their knowledge, their roles, their cognitive abilities, and their goals and tasks. Finally, bootstrapping the adaptivity in several rounds of studies is crucial since we do not (yet) know enough about exactly how adaptations should work in order to be of real use.
Imitating Human-Human Communication
Finally, a problem with some of the existing adaptive system is that they attempt to imitate human-human communication, sometimes without paying attention to what would make the system usable. In PUSH the philosophy was to integrate the adaptive behaviour with a (to the targeted users) familiar direct-manipulation environment. By extending the possibilities offered in WWW using Java and manipulation of CGI-scripts, etc., we could integrate some fairly advanced features including hotlists, stretchtext and interactive maps of the information space.
Taken together, the interaction possibilities in POP offers a rich environment for a dialogue between the user and the system, but it is not like a human-human dialogue, instead we ”use computer for what computers are good at”.
Future Work
Let us divide our discussion on future work into what needs to be done on POP and what could be interesting to study within the general field of adaptive hypermedia.
Adaptivitet av navigeringen?
Adaptiv till projekt status.
Större mer aktiv påverkan på projektgenomförandet på Ellemtel.
In order to really study how well our adapted explanations meet users’ needs we need to make a longitudinal study where we can see how users’ information seeking needs naturally arise and whether those are met by our system. An interesting extension is to utilise some form of machine learning to adjust the coupling between task and information entity over time (Rudström, 1996). Unfortunately, the SDP project at former Ellemtel was cancelled in December 1995. This means that we should switch domain in order to be able to further test the system.
Switching domain relates to
another aspect which is not dealt with in this thesis, but that lies in the future, namely to utilise our prototype as a shell system. By changing the database we can make new applications. In order to make it into a useful shell system, we would have discuss how to build the corresponding authoring tool. Such a authoring tool also needs to be tested with users. As far as we know, there are no studies on how difficult it is to write stretchtext. Boyle and Encarnacion (1994) discussed the authoring problem when presenting their hypertext system MetaDoc, but no studies were made with authors.
If we decide to switch domain and rebuild POP into a more general shell-system, we would probably re-evaluated the system architecture design. The issue lies in how much should be executed at the server site versus the client site. It is not possible to send the whole database to the client since it is too large, but the query answering and adaptations could probably be sent to the user. The query answering machinery could use database-keys to look up information needed from the server-site database. See Figure V for an abstracted architecture of this kind.
Figure V. A new architecture for POP.Future Work on Adaptive Hypermedia
In Brusilovsky’s (1996) survey of adaptive hypermedia systems he ends by outlining three research challenges for research in the area:
-
first, there are too few studies on the efficiency of adaptive hypermedia systems (with an exception of the studies by Boyle and Encarnacion (1994), Kaplan et al. (1993), and Brusilovsky and Pesin (1995)).
-
second, the usability of adaptive systems has to be further considered.
-
third, a promising direction for adaptive hypermedia is on the WWW since WWW is widespread, has potentially an unlimited number of nodes and a variety of users.
All three issues have been discussed in this thesis although our usage of the WWW is more of an Intranet-application rather than adding adaptivity to the usage of WWW. Still, clearly more research is needed into all three issues.
We would like to point at another problem to be tackled by adaptive hypermedia researchers, namely to provide a better basis for the design process of hypermedia systems and any decision to make the system adaptive or not. Such a basis requires more success stories where adaptive hypermedia systems convincingly show that certain adaptation techniques are useful. It also requires a method for their development.
There are many aspects that affect which adaptations are useful, and what user characteristics to adapt to. When designing any hypermedia space, there is a choice between whether to put more information into one node or to spread the information in several nodes. Depending on characteristics of the domain either can be preferred. If the information space is large the system designer can choose either to adapt the contents of the node or the navigation between nodes or both. Different user characteristics may be connected to these two main categories of adaptation. Also depending on the goal of the users, it may be more or less important to gain an understanding of the whole hypermedia information space.
Yet another challenge lies in tackling problems for domains where the hypermedia space is not static but rapidly changing, e.g. the WWW. Knowledge extraction in order to characterise the contents of the nodes and the links between nodes is a difficult research problem.
In POP we did not adapt the navigation in the hyperspace, and something which I find particularly intriguing is how to aid users with low spatial ability to navigate. The study on the relation between spatial ability and navigation in hypermedia presented in this thesis provides some basis for assuming that adapting to spatial ability should be a promising endeavour, but further studies are needed to determine how to help users’ with low spatial ability. It may be that we are barking up the wrong tree when we try to force users’ with low ability to navigate in a manner that requires spatial reasoning skills? Maybe it would be better to search for the strategies that these users employ in their daily life, and then make it possible for them to navigate and search for information in the same manner in the hypermedia system?
|