School of Information Sciences, University of Pittsburgh,
135 North Bellefield Ave., Pittsburgh, PA 15260, USA
Adaptive help systems (also called intelligent help systems ) is a specific kind of help systems and a recognized research direction on the crossroads of Artificial Intelligence (AI) and Human-Computer Interaction (HCI). The goal of adaptive help systems (AHS) is to provide personalized help to users working with complex interfaces. Unlike traditional "static" help systems that serve by request the same information to different users, AHS attempt to adapt to knowledge and goals of individual users offering information about most relevant aspects in a most relevant way.
The first wave of research on adaptive help emerged in early 1980 when UNIX systems with their complicated command line interfaces reached many computer-naive users. Early works on adaptive and intelligent help systems were focused almost exclusively on UNIX and its utilities such as mail and text editors. Between 1980 and 1995 this research direction has involved more than a hundred of researchers working on at least two dozens of projects. Most representative projects of this generation are UNIX Consultant (Wilensky, et al., 1988) and EUROHELP (Winkels, 1992). A good collection of work on intelligent UNIX help is (Hegner, Mc Kevitt, Norvig, & Wilensky, 2001). The wide-spread of user-friendly Graphical User Interfaces (GUI) in early 1990 caused a pause in IHS research: it has become clear that a number of problems addressed by the early generation of adaptive help systems can be resolved by GUI. However in just a few years GUI had reached the level of complexity where adaptive help becomes important again. It has initiated the second wave of research on AHS that focused on providing adaptive help in the context of modern GUI. Lumiere (Horvitz, Breese, Heckerman, Hovel, & Rommelse, 1998), the most well known project of this wave introduced the idea of intelligent help to millions of Microsoft applications users.
Following Gerhard Fischer (Fischer, 2001), AHS are traditionally divided into two classes: active and passive AHS. In a passive AHS, it is the user who initiates the help session by asking for help. An active help system initiates the help session itself.
Both kinds of AHS have to solve three challenging problems in order to provide adaptive help: (1) to build a model of user goals and knowledge (2) to decide what to present in the next help message (3) to decide how to present it. In addition to that, active AHS systems also need (4) to decide when to intervene with adaptive help.
To be useful, an AHS help message has to present information that is new to the user and relevant to the user`s current goal. To determine what is new and relevant, AHS track the user`s goals and the user`s knowledge about the interface and maintain a user model . Two major approaches to user modeling in AHS are "ask the user" and "observe the user". Most of passive AHS have exploited the first of these approaches. UNIX Consultant (Wilensky, et al., 1988) demonstrates that a passive AHS can be fairly advanced being able to involve user in a NL dialog to elicit her goals and knowledge and to provide the most relevant information. In contrast, active AHS introduced by Fischer in 1985 strive to deduce user's goals by observing the user interface actions "over the shoulder" and to identify the lack of knowledge by detecting errors and suboptimal behavior. In practical AHS the two approaches often co-exist: the user model is initialized through a short interview with a user and then kept updated through observation.
Many AHS use two classic approaches to model the user: (1) tracking the user`s actions to understand which commands and concepts are known to the user and which are not, and (2) using task models to deduce user current goal and missing knowledge. The first technology is reasonably simple--the system just records all used commands and parameters, assuming that "used means known". The second is based on plan recognition (Goodman, & Litman, 1992) and advanced domain knowledge representation in a form "goal-plan-action" tree or some other form of task model. To identify the current goal and missing pieces of knowledge, the system (1) infers the user`s goal from an observed sequence of commands, (2) tries to find a more efficient (or simply correct) sequence of commands to achieve this goal, (3) identifies the aspects of the interface that the user need to know to build this sequence. These aspects are suspected to be unknown and become the candidates to be presented in the following help messages.
Deciding what should be the focus of the next help message is the most challenging job of an AHS. A number of passive AHS simply avoid this problem allowing the users to determine what they need and focussing on adaptive presentation only. Classic AHS based on plan recognition can determine quite precisely what the user needs, but this functionality requires elaborated knowledge representation. To bypass the knowledge representation barrier, modern practical AHS use a range of alternative (though less precise) technologies that are either statistically (Encarnação, & Stoev, 1999; Horvitz, et al., 1998) or socially (Linton, & Schaefer, 2000) based.
As soon as the focus of the next help message is determined, the AHS has to decide how to present the target content. While some AHS ignore this part focusing solely on the selection part, it has been shown that adaptive presentation of help information can provide a significant difference by increasing the comprehension speed and decreasing errors. Most often the content presentation is adapted to the user knowledge: for example, expert users receive more specific details while novice users receive more explanations. To present the adaptive content, classic AHS relied mostly on NL generation while modern AHS use almost exclusively Adaptive Hypermedia (Brusilovsky, Kobsa, & Vassileva, 1998).
AHS as a research direction has contributed to the progress in a number of sub-fields within AI and HCI and helped to establish such active streams of research as Intelligent Interfaces and User Modeling. A treasury of knowledge accumulated by various AHS projects over the last 30 years is being used now to develop practical adaptive help and adaptive performance support systems.
Brusilovsky, P., Kobsa, A., & Vassileva, J. (Eds.). (1998). Adaptive Hypertext and Hypermedia . Dordrecht: Kluwer Academic Publishers.
Encarnação, L. M., & Stoev, S. L. (1999). An application-independent intelligent user support system exploiting action-sequence based user modeling. In J. Kay (Eds.), Proceedings of 7th International Conference on User Modeling, UM99, June 20-24, 1999. Wien, SpringerWienNewYork. - pp. 245-254.
Fischer, G. (2001). User Modeling in Human-Computer Interaction. User Modeling and User-Adapted Interaction , 11 (1-2), 65-86.
Goodman, B. A., & Litman, D. J. (1992). On the interaction between plan recognition and intelligent interfaces. User Modeling and User-Adapted Interaction , 2 (1), 83-115.
Hegner, S. J., Mc Kevitt, P., Norvig, P., & Wilensky, R. L. (Eds.). (2001). Intelligent Help Systems for UNIX . Dordrecht: Kluwer Academic Publishers.
Horvitz, E., Breese, J., Heckerman, D., Hovel, D., & Rommelse, K. (1998). The Lumière project: Bayesian user modeling for inferring the goals and needs of software users. Proceedings of Fourteenth Conference on Uncertainty in Artificial Intelligence, Madison, WI, - pp. 256-265.
Linton, F., & Schaefer, H.-P. (2000). Recommender systems for learning: Building user and expert models through long-term observation of application use. User Modeling and User-Adapted Interaction , 10 (2-3), 181-208.
Oppermann, R. (Ed.). (1994). Adaptive user support - ergonomic design of manually and automatically adaptable software . Hillsdale, NJ: Lawrence Erlbaum Associates.
Wilensky, R., Chin, D., Luria, M., Martin, J., Mayfield, J., & Wu, D. (1988). The Berkeley UNIX Consultant project. Computational Linguistics , 14 (4), 35-84.
Winkels, R. (1992). Explorations in Intelligent Tutoring Systems and Help . Amsterdam: IOS Press.